diff --git a/.ci/build.sh b/.ci/build.sh
index 728c09113..cf420d4dd 100755
--- a/.ci/build.sh
+++ b/.ci/build.sh
@@ -24,8 +24,8 @@
 # - For Windows (MSYS MinGW) builds:
 #   - Packaging requires 7-Zip on Program Files
 #   - Packaging the Ghostscript DLL requires 32-bit and/or 64-bit Ghostscript on Program Files
-#   - Packaging the FluidSynth DLL requires it to be at /home/86Box/dll32/libfluidsynth.dll
-#     and/or /home/86Box/dll64/libfluidsynth64.dll (for 32-bit and 64-bit builds respectively)
+#   - Packaging the XAudio2 DLL for FAudio requires it to be at /home/86Box/dll32/xaudio2*.dll
+#     and/or /home/86Box/dll64/xaudio2*.dll (for 32-bit and 64-bit builds respectively)
 # - For Linux builds:
 #   - Only Debian and derivatives are supported
 #   - dpkg and apt-get are called through sudo to manage dependencies; make sure those
@@ -288,7 +288,6 @@ then
 	echo [-] Using MSYSTEM [$MSYSTEM]
 
 	# Install dependencies only if we're in a new build and/or architecture.
-	freetype_dll="$cache_dir/freetype.$MSYSTEM.dll"
 	if check_buildtag "$MSYSTEM"
 	then
 		# Update databases and keyring only if we're in a new build.
@@ -333,9 +332,6 @@ then
 		# Clean pacman cache when running under Jenkins to save disk space.
 		[ "$CI" = "true" ] && rm -rf /var/cache/pacman/pkg
 
-		# Generate a new freetype DLL for this architecture.
-		rm -f "$freetype_dll"
-
 		# Save build tag to skip this later. Doing it here (once everything is
 		# in place) is important to avoid potential issues with retried builds.
 		save_buildtag "$MSYSTEM"
@@ -595,7 +591,7 @@ else
 	# ...and the ones we do want listed. Non-dev packages fill missing spots on the list.
 	libpkgs=""
 	longest_libpkg=0
-	for pkg in libc6-dev libstdc++6 libopenal-dev libfreetype6-dev libx11-dev libsdl2-dev libpng-dev librtmidi-dev qtdeclarative5-dev libwayland-dev libevdev-dev libxkbcommon-x11-dev libglib2.0-dev libslirp-dev libfaudio-dev libaudio-dev libjack-jackd2-dev libpipewire-0.3-dev libsamplerate0-dev libsndio-dev libvdeplug-dev
+	for pkg in libc6-dev libstdc++6 libopenal-dev libfreetype6-dev libx11-dev libsdl2-dev libpng-dev librtmidi-dev qtdeclarative5-dev libwayland-dev libevdev-dev libxkbcommon-x11-dev libglib2.0-dev libslirp-dev libfaudio-dev libaudio-dev libjack-jackd2-dev libpipewire-0.3-dev libsamplerate0-dev libsndio-dev libvdeplug-dev libfluidsynth-dev
 	do
 		libpkgs="$libpkgs $pkg:$arch_deb"
 		length=$(echo -n $pkg | sed 's/-dev$//' | sed "s/qtdeclarative/qt/" | wc -c)
@@ -796,10 +792,6 @@ then
 	sevenzip="$pf/7-Zip/7z.exe"
 	[ "$arch" = "32" -a -d "/c/Program Files (x86)" ] && pf="/c/Program Files (x86)"
 
-	# Archive freetype from cache or generate it from local MSYS installation.
-	[ ! -e "$freetype_dll" ] && .ci/static2dll.sh -p freetype2 /$MSYSTEM/lib/libfreetype.a "$freetype_dll"
-	cp -p "$freetype_dll" archive_tmp/freetype.dll
-
 	# Archive Ghostscript DLL from local official distribution installation.
 	for gs in "$pf"/gs/gs*.*.*
 	do
@@ -810,8 +802,8 @@ then
 	"$sevenzip" e -y -o"archive_tmp" "$discord_zip" "lib/$arch_discord/discord_game_sdk.dll"
 	[ ! -e "archive_tmp/discord_game_sdk.dll" ] && echo [!] No Discord Game SDK for architecture [$arch_discord]
 
-	# Archive other DLLs from local directory.
-	cp -p "/home/$project/dll$arch/"* archive_tmp/
+	# Archive XAudio2 DLL if required.
+	grep -q "OPENAL:BOOL=ON" build/CMakeCache.txt || cp -p "/home/$project/dll$arch/xaudio2"* archive_tmp/
 
 	# Archive executable, while also stripping it if requested.
 	if [ $strip -ne 0 ]
@@ -1003,6 +995,11 @@ else
 	ln -s "$relroot/usr/lib/libvulkan.so.1" "archive_tmp/usr/lib/libvulkan.so"
 	ln -s "$relroot/usr/lib/$libdir/libvulkan.so.1" "archive_tmp/usr/lib/$libdir/libvulkan.so"
 
+	# The FluidSynth packaged by Debian bullseye is ABI incompatible with
+	# the newer version we compile, despite sharing a major version. Since we
+	# don't run into the one breaking ABI change they made, just symlink it.
+	ln -s "$(readlink "archive_tmp/usr/lib/libfluidsynth.so.3")" "archive_tmp/usr/lib/libfluidsynth.so.2"
+
 	# Archive Discord Game SDK library.
 	7z e -y -o"archive_tmp/usr/lib" "$discord_zip" "lib/$arch_discord/discord_game_sdk.so"
 	[ ! -e "archive_tmp/usr/lib/discord_game_sdk.so" ] && echo [!] No Discord Game SDK for architecture [$arch_discord]
diff --git a/.ci/dependencies_macports.txt b/.ci/dependencies_macports.txt
index e78d4a6da..5ec71d07c 100644
--- a/.ci/dependencies_macports.txt
+++ b/.ci/dependencies_macports.txt
@@ -13,4 +13,5 @@ qt5
 wget
 fluidsynth
 ghostscript
+libslirp
 vde2
diff --git a/.ci/dependencies_msys.txt b/.ci/dependencies_msys.txt
index df4932352..1b9ccdcc8 100644
--- a/.ci/dependencies_msys.txt
+++ b/.ci/dependencies_msys.txt
@@ -8,5 +8,6 @@ SDL2
 zlib
 libpng
 rtmidi
+libslirp
 qt5-static
 qt5-translations
diff --git a/.ci/static2dll.sh b/.ci/static2dll.sh
deleted file mode 100755
index 030898752..000000000
--- a/.ci/static2dll.sh
+++ /dev/null
@@ -1,160 +0,0 @@
-#!/bin/sh
-#
-# 86Box    A hypervisor and IBM PC system emulator that specializes in
-#          running old operating systems and software designed for IBM
-#          PC systems and compatibles from 1981 through fairly recent
-#          system designs based on the PCI bus.
-#
-#          This file is part of the 86Box distribution.
-#
-#          Script for converting MinGW static libraries into a DLL.
-#
-#
-# Authors: RichardG, <richardg867@gmail.com>
-#
-#          Copyright 2021 RichardG.
-#
-
-def_file="static2dll.def"
-seen_file="static2dll.seen"
-libs_file="static2dll.libs"
-
-find_lib() {
-	# Try to find a static library's file.
-	local msystem_lib="/$(echo $MSYSTEM | tr '[:upper:]' '[:lower:]')/lib/lib"
-	if [ -e "$msystem_lib$1.a" ]
-	then
-		echo "$msystem_lib$1.a"
-	elif [ -e "$msystem_lib$1.dll.a" ]
-	then
-		echo "$msystem_lib$1.dll.a"
-	else
-		# Return dynamic reference to the library.
-		echo "-l$1"
-		return 1
-	fi
-}
-
-add_lib() {
-	# Always make sure this lib is listed after the last lib that depends on it.
-	old_libs=$(cat "$libs_file")
-	rm -f "$libs_file"
-	for lib in $old_libs
-	do
-		[ "$lib" != "$*" ] && echo "$lib" >> "$libs_file"
-	done
-	echo "$*" >> "$libs_file"
-
-	# Add libstdc++ in the end if required.
-	if echo "$*" | grep -q "/"
-	then
-		grep -Eq -- "__cxa_|__gxx_" "$1" 2> /dev/null && add_lib -static -lstdc++
-	fi
-
-	# Add libiconv for libintl.
-	if echo "$*" | grep -q "libintl"
-	then
-		add_lib $(find_lib iconv)
-	fi
-
-	# Add libuuid for glib.
-	if echo "$*" | grep -q "libglib"
-	then
-		add_lib $(find_lib uuid)
-	fi
-}
-
-run_pkgconfig() {
-	local cache_file="static2dll.$1.cache"
-	if [ -e "$cache_file" ]
-	then
-		cat "$cache_file"
-	else
-		pkg-config --static --libs "$1" 2> /dev/null | tee "$cache_file"
-	fi
-}
-
-parse_pkgconfig() {
-	# Parse arguments.
-	local layers=$1
-	shift
-	local input_lib_name=$1
-	shift
-
-	# Don't process the same file again.
-	grep -q '^'$input_lib_name'$' "$seen_file" && return
-	echo $input_lib_name >> "$seen_file"
-
-	echo "$layers" parse_pkgconfig $input_lib_name
-
-	# Parse pkg-config arguments.
-	for arg in $*
-	do
-		local arg_base="$(echo $arg | cut -c1-2)"
-		if [ "x$arg_base" = "x-l" ]
-		then
-			# Don't process the same lib again.
-			local lib_name="$(echo $arg | cut -c3-)"
-			[ "x$lib_name" == "x$input_lib_name" ] && continue
-
-			# Add lib path.
-			add_lib "$(find_lib $lib_name)"
-
-			# Get this lib's dependencies through pkg-config.
-			local pkgconfig="$(run_pkgconfig "$lib_name")"
-			[ $? -eq 0 ] && parse_pkgconfig "$layers"'>' "$lib_name" $pkgconfig || echo $lib_name >> "$seen_file"
-		elif [ "x$(echo $arg_base | cut -c1)" = "x-" ]
-		then
-			# Ignore other arguments.
-			continue
-		else
-			# Add lib path.
-			add_lib "$arg"
-		fi
-	done
-}
-
-# Parse arguments.
-case $1 in
-	-p) # -p pkg_config_name static_lib_path out_dll
-		shift
-		base_pkgconfig=$(run_pkgconfig "$1")
-		base_path="$2"
-		base_name="$1"
-		;;
-
-	*) # pc_path static_lib_path out_dll
-		base_pkgconfig="$(grep ^Libs.private: $1 | cut -d: -f2-)"
-		base_path="$2"
-		base_name="$2"
-		;;
-esac
-
-# Check arguments.
-if [ -z "$base_pkgconfig" -o -z "$base_path" -o -z "$base_name" ]
-then
-	echo Usage:
-	echo static2dll.sh -p {pkgconfig_package_name} {static_lib_path} {out_dll_name}
-	echo static2dll.sh {pc_file_path} {static_lib_path} {out_dll_name}
-	exit 1
-fi
-
-# Produce .def file.
-echo LIBRARY $(basename "$3") > "$def_file"
-echo EXPORTS >> "$def_file"
-nm "$base_path" | grep " [TC] " | sed "/ _/s// /" | awk '{ print $3 }' >> "$def_file"
-
-# Parse dependencies recursively.
-rm -f "$seen_file" "$libs_file" "$libs_file.tmp"
-touch "$seen_file" "$libs_file"
-parse_pkgconfig '>' $base_name $base_pkgconfig
-
-# Produce final DLL.
-dllwrap --def "$def_file" -o "$3" -Wl,--allow-multiple-definition "$base_path" $(cat "$libs_file")
-status=$?
-[ $status -eq 0 ] && rm -f "$def_file" "$seen_file" "$libs_file" "static2dll.*.cache"
-
-# Update final DLL timestamp.
-touch -r "$base_path" "$3"
-
-exit $status
diff --git a/.github/workflows/c-cpp.yml b/.github/workflows/c-cpp.yml
index 16b56cc10..7b8b6eeb4 100644
--- a/.github/workflows/c-cpp.yml
+++ b/.github/workflows/c-cpp.yml
@@ -90,6 +90,8 @@ jobs:
             libpng:p
             openal:p
             rtmidi:p
+            libslirp:p
+            fluidsynth:p
             libvncserver:p
 
       - name: Checkout repository
diff --git a/.github/workflows/cmake.yml b/.github/workflows/cmake.yml
index 49ea8c890..5f703a65f 100644
--- a/.github/workflows/cmake.yml
+++ b/.github/workflows/cmake.yml
@@ -104,6 +104,8 @@ jobs:
             libpng:p
             openal:p
             rtmidi:p
+            libslirp:p
+            fluidsynth:p
             libvncserver:p
             ${{ matrix.ui.packages }}
 
@@ -328,6 +330,8 @@ jobs:
           libc6-dev
           librtmidi-dev
           libopenal-dev
+          libslirp-dev
+          libfluidsynth-dev
           libvncserver-dev
           ${{ matrix.ui.packages }}
 
@@ -413,6 +417,7 @@ jobs:
           libpng
           rtmidi
           openal-soft
+          fluidsynth
           libvncserver
           ${{ matrix.ui.packages }}
 
diff --git a/.github/workflows/codeql.yml b/.github/workflows/codeql.yml
index 107755af6..fcfb16341 100644
--- a/.github/workflows/codeql.yml
+++ b/.github/workflows/codeql.yml
@@ -107,6 +107,8 @@ jobs:
             libpng:p
             openal:p
             rtmidi:p
+            libslirp:p
+            fluidsynth:p
             libvncserver:p
             ${{ matrix.ui.packages }}
 
@@ -191,6 +193,8 @@ jobs:
           libc6-dev
           librtmidi-dev
           libopenal-dev
+          libslirp-dev
+          libfluidsynth-dev
           libvncserver-dev
           ${{ matrix.ui.packages }}
 
@@ -266,6 +270,7 @@ jobs:
           libpng
           rtmidi
           openal-soft
+          fluidsynth
           libvncserver
           ${{ matrix.ui.packages }}
 
diff --git a/src/86box.c b/src/86box.c
index a9bfb2f72..8db2fc75a 100644
--- a/src/86box.c
+++ b/src/86box.c
@@ -1275,7 +1275,9 @@ pc_run(void)
 #ifdef USE_GDBSTUB /* avoid a KBC FIFO overflow when CPU emulation is stalled */
     // if (gdbstub_step == GDBSTUB_EXEC)
 #endif
-        // mouse_process();
+#if 0
+        mouse_process();
+#endif
     joystick_process();
     endblit();
 
@@ -1416,6 +1418,9 @@ set_screen_size_monitor(int x, int y, int monitor_index)
             monitors[monitor_index].mon_scrnsz_x = (monitors[monitor_index].mon_unscaled_size_x << 3);
             monitors[monitor_index].mon_scrnsz_y = (monitors[monitor_index].mon_unscaled_size_y << 3);
             break;
+
+        default:
+            break;
     }
 
     plat_resize_request(monitors[monitor_index].mon_scrnsz_x, monitors[monitor_index].mon_scrnsz_y, monitor_index);
diff --git a/src/acpi.c b/src/acpi.c
index 539da644b..495419f4e 100644
--- a/src/acpi.c
+++ b/src/acpi.c
@@ -169,9 +169,9 @@ acpi_raise_smi(void *priv, int do_smi)
 }
 
 static uint32_t
-acpi_reg_read_common_regs(int size, uint16_t addr, void *p)
+acpi_reg_read_common_regs(UNUSED(int size), uint16_t addr, void *priv)
 {
-    acpi_t  *dev = (acpi_t *) p;
+    acpi_t  *dev = (acpi_t *) priv;
     uint32_t ret = 0x00000000;
     int      shift16;
     int      shift32;
@@ -211,6 +211,9 @@ acpi_reg_read_common_regs(int size, uint16_t addr, void *p)
                 update_tsc();
 #endif
             break;
+
+        default:
+            break;
     }
 
 #ifdef ENABLE_ACPI_LOG
@@ -221,9 +224,9 @@ acpi_reg_read_common_regs(int size, uint16_t addr, void *p)
 }
 
 static uint32_t
-acpi_reg_read_ali(int size, uint16_t addr, void *p)
+acpi_reg_read_ali(int size, uint16_t addr, void *priv)
 {
-    acpi_t  *dev = (acpi_t *) p;
+    acpi_t  *dev = (acpi_t *) priv;
     uint32_t ret = 0x00000000;
     int      shift16;
     int      shift32;
@@ -277,7 +280,7 @@ acpi_reg_read_ali(int size, uint16_t addr, void *p)
             ret = dev->regs.pmcntrl;
             break;
         default:
-            ret = acpi_reg_read_common_regs(size, addr, p);
+            ret = acpi_reg_read_common_regs(size, addr, priv);
             break;
     }
 
@@ -289,9 +292,9 @@ acpi_reg_read_ali(int size, uint16_t addr, void *p)
 }
 
 static uint32_t
-acpi_reg_read_intel(int size, uint16_t addr, void *p)
+acpi_reg_read_intel(int size, uint16_t addr, void *priv)
 {
-    acpi_t  *dev = (acpi_t *) p;
+    acpi_t  *dev = (acpi_t *) priv;
     uint32_t ret = 0x00000000;
     int      shift16;
     int      shift32;
@@ -374,7 +377,7 @@ acpi_reg_read_intel(int size, uint16_t addr, void *p)
                 ret = dev->regs.gporeg[addr & 3];
             break;
         default:
-            ret = acpi_reg_read_common_regs(size, addr, p);
+            ret = acpi_reg_read_common_regs(size, addr, priv);
             break;
     }
 
@@ -386,9 +389,9 @@ acpi_reg_read_intel(int size, uint16_t addr, void *p)
 }
 
 static uint32_t
-acpi_reg_read_via_common(int size, uint16_t addr, void *p)
+acpi_reg_read_via_common(int size, uint16_t addr, void *priv)
 {
-    acpi_t  *dev = (acpi_t *) p;
+    acpi_t  *dev = (acpi_t *) priv;
     uint32_t ret = 0x00000000;
     int      shift16;
     int      shift32;
@@ -470,7 +473,7 @@ acpi_reg_read_via_common(int size, uint16_t addr, void *p)
             ret = (dev->regs.gptren >> shift32) & 0xff;
             break;
         default:
-            ret = acpi_reg_read_common_regs(size, addr, p);
+            ret = acpi_reg_read_common_regs(size, addr, priv);
             break;
     }
 
@@ -482,9 +485,9 @@ acpi_reg_read_via_common(int size, uint16_t addr, void *p)
 }
 
 static uint32_t
-acpi_reg_read_via(int size, uint16_t addr, void *p)
+acpi_reg_read_via(int size, uint16_t addr, void *priv)
 {
-    acpi_t  *dev = (acpi_t *) p;
+    acpi_t  *dev = (acpi_t *) priv;
     uint32_t ret = 0x00000000;
     int      shift16;
 
@@ -527,7 +530,7 @@ acpi_reg_read_via(int size, uint16_t addr, void *p)
             ret = (dev->regs.gpi_val >> shift16) & 0xff;
             break;
         default:
-            ret = acpi_reg_read_via_common(size, addr, p);
+            ret = acpi_reg_read_via_common(size, addr, priv);
             break;
     }
 
@@ -539,9 +542,9 @@ acpi_reg_read_via(int size, uint16_t addr, void *p)
 }
 
 static uint32_t
-acpi_reg_read_via_596b(int size, uint16_t addr, void *p)
+acpi_reg_read_via_596b(int size, uint16_t addr, void *priv)
 {
-    acpi_t  *dev = (acpi_t *) p;
+    acpi_t  *dev = (acpi_t *) priv;
     uint32_t ret = 0x00000000;
     int      shift16;
     int      shift32;
@@ -577,7 +580,7 @@ acpi_reg_read_via_596b(int size, uint16_t addr, void *p)
             ret = (dev->regs.gpo_val >> shift32) & 0xff;
             break;
         default:
-            ret = acpi_reg_read_via_common(size, addr, p);
+            ret = acpi_reg_read_via_common(size, addr, priv);
             break;
     }
 
@@ -589,13 +592,13 @@ acpi_reg_read_via_596b(int size, uint16_t addr, void *p)
 }
 
 static uint32_t
-acpi_reg_read_smc(int size, uint16_t addr, void *p)
+acpi_reg_read_smc(int size, uint16_t addr, void *priv)
 {
     uint32_t ret = 0x00000000;
 
     addr &= 0x0f;
 
-    ret = acpi_reg_read_common_regs(size, addr, p);
+    ret = acpi_reg_read_common_regs(size, addr, priv);
 
 #ifdef ENABLE_ACPI_LOG
     if (size != 1)
@@ -605,9 +608,9 @@ acpi_reg_read_smc(int size, uint16_t addr, void *p)
 }
 
 static uint32_t
-acpi_aux_reg_read_smc(int size, uint16_t addr, void *p)
+acpi_aux_reg_read_smc(UNUSED(int size), uint16_t addr, void *priv)
 {
-    acpi_t  *dev = (acpi_t *) p;
+    acpi_t  *dev = (acpi_t *) priv;
     uint32_t ret = 0x00000000;
     int      shift16;
 
@@ -638,6 +641,9 @@ acpi_aux_reg_read_smc(int size, uint16_t addr, void *p)
             /* Miscellaneous Control Register */
             ret = dev->regs.glbctl & 0xff;
             break;
+
+        default:
+            break;
     }
 
     acpi_log("(%i) ACPI Read  (%i) %02X: %02X\n", in_smm, size, addr, ret);
@@ -645,9 +651,9 @@ acpi_aux_reg_read_smc(int size, uint16_t addr, void *p)
 }
 
 static void
-acpi_reg_write_common_regs(int size, uint16_t addr, uint8_t val, void *p)
+acpi_reg_write_common_regs(UNUSED(int size), uint16_t addr, uint8_t val, void *priv)
 {
-    acpi_t *dev = (acpi_t *) p;
+    acpi_t *dev = (acpi_t *) priv;
     int     shift16;
     int     sus_typ;
 
@@ -719,13 +725,16 @@ acpi_reg_write_common_regs(int size, uint16_t addr, uint8_t val, void *p)
             }
             dev->regs.pmcntrl = ((dev->regs.pmcntrl & ~(0xff << shift16)) | (val << shift16)) & 0x3f07 /* 0x3c07 */;
             break;
+
+        default:
+            break;
     }
 }
 
 static void
-acpi_reg_write_ali(int size, uint16_t addr, uint8_t val, void *p)
+acpi_reg_write_ali(int size, uint16_t addr, uint8_t val, void *priv)
 {
-    acpi_t *dev = (acpi_t *) p;
+    acpi_t *dev = (acpi_t *) priv;
     int     shift16;
     int     shift32;
 
@@ -782,7 +791,7 @@ acpi_reg_write_ali(int size, uint16_t addr, uint8_t val, void *p)
             dev->regs.pmcntrl = val & 1;
             break;
         default:
-            acpi_reg_write_common_regs(size, addr, val, p);
+            acpi_reg_write_common_regs(size, addr, val, priv);
             /* Setting GBL_RLS also sets BIOS_STS and generates SMI. */
             if ((addr == 0x00) && !(dev->regs.pmsts & 0x20))
                 dev->regs.gpcntrl &= ~0x0002;
@@ -795,9 +804,9 @@ acpi_reg_write_ali(int size, uint16_t addr, uint8_t val, void *p)
 }
 
 static void
-acpi_reg_write_intel(int size, uint16_t addr, uint8_t val, void *p)
+acpi_reg_write_intel(int size, uint16_t addr, uint8_t val, void *priv)
 {
-    acpi_t *dev = (acpi_t *) p;
+    acpi_t *dev = (acpi_t *) priv;
     int     shift16;
     int     shift32;
 
@@ -878,7 +887,7 @@ acpi_reg_write_intel(int size, uint16_t addr, uint8_t val, void *p)
                 dev->regs.gporeg[addr & 3] = val;
             break;
         default:
-            acpi_reg_write_common_regs(size, addr, val, p);
+            acpi_reg_write_common_regs(size, addr, val, priv);
             /* Setting GBL_RLS also sets BIOS_STS and generates SMI. */
             if ((addr == 0x00) && !(dev->regs.pmsts & 0x20))
                 dev->regs.glbctl &= ~0x0002;
@@ -892,9 +901,9 @@ acpi_reg_write_intel(int size, uint16_t addr, uint8_t val, void *p)
 }
 
 static void
-acpi_reg_write_via_common(int size, uint16_t addr, uint8_t val, void *p)
+acpi_reg_write_via_common(int size, uint16_t addr, uint8_t val, void *priv)
 {
-    acpi_t *dev = (acpi_t *) p;
+    acpi_t *dev = (acpi_t *) priv;
     int     shift16;
     int     shift32;
 
@@ -965,7 +974,7 @@ acpi_reg_write_via_common(int size, uint16_t addr, uint8_t val, void *p)
             dev->regs.gptren = ((dev->regs.gptren & ~(0xff << shift32)) | (val << shift32)) & 0x000000d9;
             break;
         default:
-            acpi_reg_write_common_regs(size, addr, val, p);
+            acpi_reg_write_common_regs(size, addr, val, priv);
             /* Setting GBL_RLS also sets BIOS_STS and generates SMI. */
             if ((addr == 0x00) && !(dev->regs.pmsts & 0x20))
                 dev->regs.glbctl &= ~0x0002;
@@ -986,9 +995,9 @@ acpi_i2c_set(acpi_t *dev)
 }
 
 static void
-acpi_reg_write_via(int size, uint16_t addr, uint8_t val, void *p)
+acpi_reg_write_via(int size, uint16_t addr, uint8_t val, void *priv)
 {
-    acpi_t *dev = (acpi_t *) p;
+    acpi_t *dev = (acpi_t *) priv;
     int     shift16;
     int     shift32;
 
@@ -1044,15 +1053,15 @@ acpi_reg_write_via(int size, uint16_t addr, uint8_t val, void *p)
             dev->regs.gpo_val = ((dev->regs.gpo_val & ~(0xff << shift16)) | (val << shift16)) & 0xffff;
             break;
         default:
-            acpi_reg_write_via_common(size, addr, val, p);
+            acpi_reg_write_via_common(size, addr, val, priv);
             break;
     }
 }
 
 static void
-acpi_reg_write_via_596b(int size, uint16_t addr, uint8_t val, void *p)
+acpi_reg_write_via_596b(int size, uint16_t addr, uint8_t val, void *priv)
 {
-    acpi_t *dev = (acpi_t *) p;
+    acpi_t *dev = (acpi_t *) priv;
     int     shift16;
     int     shift32;
 
@@ -1102,20 +1111,20 @@ acpi_reg_write_via_596b(int size, uint16_t addr, uint8_t val, void *p)
             dev->regs.gpo_val = ((dev->regs.gpo_val & ~(0xff << shift32)) | (val << shift32)) & 0x7fffffff;
             break;
         default:
-            acpi_reg_write_via_common(size, addr, val, p);
+            acpi_reg_write_via_common(size, addr, val, priv);
             break;
     }
 }
 
 static void
-acpi_reg_write_smc(int size, uint16_t addr, uint8_t val, void *p)
+acpi_reg_write_smc(int size, uint16_t addr, uint8_t val, void *priv)
 {
-    acpi_t *dev = (acpi_t *) p;
+    acpi_t *dev = (acpi_t *) priv;
 
     addr &= 0x0f;
     acpi_log("(%i) ACPI Write (%i) %02X: %02X\n", in_smm, size, addr, val);
 
-    acpi_reg_write_common_regs(size, addr, val, p);
+    acpi_reg_write_common_regs(size, addr, val, priv);
     /* Setting GBL_RLS also sets BIOS_STS and generates SMI. */
     if ((addr == 0x00) && !(dev->regs.pmsts & 0x20))
         dev->regs.glbctl &= ~0x0001;
@@ -1127,9 +1136,9 @@ acpi_reg_write_smc(int size, uint16_t addr, uint8_t val, void *p)
 }
 
 static void
-acpi_aux_reg_write_smc(int size, uint16_t addr, uint8_t val, void *p)
+acpi_aux_reg_write_smc(UNUSED(int size), uint16_t addr, uint8_t val, void *priv)
 {
-    acpi_t *dev = (acpi_t *) p;
+    acpi_t *dev = (acpi_t *) priv;
     int     shift16;
 
     addr &= 0x07;
@@ -1171,76 +1180,79 @@ acpi_aux_reg_write_smc(int size, uint16_t addr, uint8_t val, void *p)
                     acpi_update_irq(dev);
             }
             break;
+
+        default:
+            break;
     }
 }
 
 static uint32_t
-acpi_reg_read_common(int size, uint16_t addr, void *p)
+acpi_reg_read_common(int size, uint16_t addr, void *priv)
 {
-    acpi_t *dev = (acpi_t *) p;
+    acpi_t *dev = (acpi_t *) priv;
     uint8_t ret = 0xff;
 
     if (dev->vendor == VEN_ALI)
-        ret = acpi_reg_read_ali(size, addr, p);
+        ret = acpi_reg_read_ali(size, addr, priv);
     else if (dev->vendor == VEN_VIA)
-        ret = acpi_reg_read_via(size, addr, p);
+        ret = acpi_reg_read_via(size, addr, priv);
     else if (dev->vendor == VEN_VIA_596B)
-        ret = acpi_reg_read_via_596b(size, addr, p);
+        ret = acpi_reg_read_via_596b(size, addr, priv);
     else if (dev->vendor == VEN_INTEL)
-        ret = acpi_reg_read_intel(size, addr, p);
+        ret = acpi_reg_read_intel(size, addr, priv);
     else if (dev->vendor == VEN_SMC)
-        ret = acpi_reg_read_smc(size, addr, p);
+        ret = acpi_reg_read_smc(size, addr, priv);
 
     return ret;
 }
 
 static void
-acpi_reg_write_common(int size, uint16_t addr, uint8_t val, void *p)
+acpi_reg_write_common(int size, uint16_t addr, uint8_t val, void *priv)
 {
-    acpi_t *dev = (acpi_t *) p;
+    acpi_t *dev = (acpi_t *) priv;
 
     if (dev->vendor == VEN_ALI)
-        acpi_reg_write_ali(size, addr, val, p);
+        acpi_reg_write_ali(size, addr, val, priv);
     else if (dev->vendor == VEN_VIA)
-        acpi_reg_write_via(size, addr, val, p);
+        acpi_reg_write_via(size, addr, val, priv);
     else if (dev->vendor == VEN_VIA_596B)
-        acpi_reg_write_via_596b(size, addr, val, p);
+        acpi_reg_write_via_596b(size, addr, val, priv);
     else if (dev->vendor == VEN_INTEL)
-        acpi_reg_write_intel(size, addr, val, p);
+        acpi_reg_write_intel(size, addr, val, priv);
     else if (dev->vendor == VEN_SMC)
-        acpi_reg_write_smc(size, addr, val, p);
+        acpi_reg_write_smc(size, addr, val, priv);
 }
 
 static uint32_t
-acpi_aux_reg_read_common(int size, uint16_t addr, void *p)
+acpi_aux_reg_read_common(int size, uint16_t addr, void *priv)
 {
-    acpi_t *dev = (acpi_t *) p;
+    acpi_t *dev = (acpi_t *) priv;
     uint8_t ret = 0xff;
 
     if (dev->vendor == VEN_SMC)
-        ret = acpi_aux_reg_read_smc(size, addr, p);
+        ret = acpi_aux_reg_read_smc(size, addr, priv);
 
     return ret;
 }
 
 static void
-acpi_aux_reg_write_common(int size, uint16_t addr, uint8_t val, void *p)
+acpi_aux_reg_write_common(int size, uint16_t addr, uint8_t val, void *priv)
 {
-    acpi_t *dev = (acpi_t *) p;
+    acpi_t *dev = (acpi_t *) priv;
 
     if (dev->vendor == VEN_SMC)
-        acpi_aux_reg_write_smc(size, addr, val, p);
+        acpi_aux_reg_write_smc(size, addr, val, priv);
 }
 
 static uint32_t
-acpi_reg_readl(uint16_t addr, void *p)
+acpi_reg_readl(uint16_t addr, void *priv)
 {
     uint32_t ret = 0x00000000;
 
-    ret = acpi_reg_read_common(4, addr, p);
-    ret |= (acpi_reg_read_common(4, addr + 1, p) << 8);
-    ret |= (acpi_reg_read_common(4, addr + 2, p) << 16);
-    ret |= (acpi_reg_read_common(4, addr + 3, p) << 24);
+    ret = acpi_reg_read_common(4, addr, priv);
+    ret |= (acpi_reg_read_common(4, addr + 1, priv) << 8);
+    ret |= (acpi_reg_read_common(4, addr + 2, priv) << 16);
+    ret |= (acpi_reg_read_common(4, addr + 3, priv) << 24);
 
     acpi_log("ACPI: Read L %08X from %04X\n", ret, addr);
 
@@ -1248,12 +1260,12 @@ acpi_reg_readl(uint16_t addr, void *p)
 }
 
 static uint16_t
-acpi_reg_readw(uint16_t addr, void *p)
+acpi_reg_readw(uint16_t addr, void *priv)
 {
     uint16_t ret = 0x0000;
 
-    ret = acpi_reg_read_common(2, addr, p);
-    ret |= (acpi_reg_read_common(2, addr + 1, p) << 8);
+    ret = acpi_reg_read_common(2, addr, priv);
+    ret |= (acpi_reg_read_common(2, addr + 1, priv) << 8);
 
     acpi_log("ACPI: Read W %08X from %04X\n", ret, addr);
 
@@ -1261,11 +1273,11 @@ acpi_reg_readw(uint16_t addr, void *p)
 }
 
 static uint8_t
-acpi_reg_read(uint16_t addr, void *p)
+acpi_reg_read(uint16_t addr, void *priv)
 {
     uint8_t ret = 0x00;
 
-    ret = acpi_reg_read_common(1, addr, p);
+    ret = acpi_reg_read_common(1, addr, priv);
 
     acpi_log("ACPI: Read B %02X from %04X\n", ret, addr);
 
@@ -1273,14 +1285,14 @@ acpi_reg_read(uint16_t addr, void *p)
 }
 
 static uint32_t
-acpi_aux_reg_readl(uint16_t addr, void *p)
+acpi_aux_reg_readl(uint16_t addr, void *priv)
 {
     uint32_t ret = 0x00000000;
 
-    ret = acpi_aux_reg_read_common(4, addr, p);
-    ret |= (acpi_aux_reg_read_common(4, addr + 1, p) << 8);
-    ret |= (acpi_aux_reg_read_common(4, addr + 2, p) << 16);
-    ret |= (acpi_aux_reg_read_common(4, addr + 3, p) << 24);
+    ret = acpi_aux_reg_read_common(4, addr, priv);
+    ret |= (acpi_aux_reg_read_common(4, addr + 1, priv) << 8);
+    ret |= (acpi_aux_reg_read_common(4, addr + 2, priv) << 16);
+    ret |= (acpi_aux_reg_read_common(4, addr + 3, priv) << 24);
 
     acpi_log("ACPI: Read Aux L %08X from %04X\n", ret, addr);
 
@@ -1288,12 +1300,12 @@ acpi_aux_reg_readl(uint16_t addr, void *p)
 }
 
 static uint16_t
-acpi_aux_reg_readw(uint16_t addr, void *p)
+acpi_aux_reg_readw(uint16_t addr, void *priv)
 {
     uint16_t ret = 0x0000;
 
-    ret = acpi_aux_reg_read_common(2, addr, p);
-    ret |= (acpi_aux_reg_read_common(2, addr + 1, p) << 8);
+    ret = acpi_aux_reg_read_common(2, addr, priv);
+    ret |= (acpi_aux_reg_read_common(2, addr + 1, priv) << 8);
 
     acpi_log("ACPI: Read Aux W %04X from %04X\n", ret, addr);
 
@@ -1301,11 +1313,11 @@ acpi_aux_reg_readw(uint16_t addr, void *p)
 }
 
 static uint8_t
-acpi_aux_reg_read(uint16_t addr, void *p)
+acpi_aux_reg_read(uint16_t addr, void *priv)
 {
     uint8_t ret = 0x00;
 
-    ret = acpi_aux_reg_read_common(1, addr, p);
+    ret = acpi_aux_reg_read_common(1, addr, priv);
 
     acpi_log("ACPI: Read Aux B %02X from %04X\n", ret, addr);
 
@@ -1313,59 +1325,59 @@ acpi_aux_reg_read(uint16_t addr, void *p)
 }
 
 static void
-acpi_reg_writel(uint16_t addr, uint32_t val, void *p)
+acpi_reg_writel(uint16_t addr, uint32_t val, void *priv)
 {
     acpi_log("ACPI: Write L %08X to %04X\n", val, addr);
 
-    acpi_reg_write_common(4, addr, val & 0xff, p);
-    acpi_reg_write_common(4, addr + 1, (val >> 8) & 0xff, p);
-    acpi_reg_write_common(4, addr + 2, (val >> 16) & 0xff, p);
-    acpi_reg_write_common(4, addr + 3, (val >> 24) & 0xff, p);
+    acpi_reg_write_common(4, addr, val & 0xff, priv);
+    acpi_reg_write_common(4, addr + 1, (val >> 8) & 0xff, priv);
+    acpi_reg_write_common(4, addr + 2, (val >> 16) & 0xff, priv);
+    acpi_reg_write_common(4, addr + 3, (val >> 24) & 0xff, priv);
 }
 
 static void
-acpi_reg_writew(uint16_t addr, uint16_t val, void *p)
+acpi_reg_writew(uint16_t addr, uint16_t val, void *priv)
 {
     acpi_log("ACPI: Write W %04X to %04X\n", val, addr);
 
-    acpi_reg_write_common(2, addr, val & 0xff, p);
-    acpi_reg_write_common(2, addr + 1, (val >> 8) & 0xff, p);
+    acpi_reg_write_common(2, addr, val & 0xff, priv);
+    acpi_reg_write_common(2, addr + 1, (val >> 8) & 0xff, priv);
 }
 
 static void
-acpi_reg_write(uint16_t addr, uint8_t val, void *p)
+acpi_reg_write(uint16_t addr, uint8_t val, void *priv)
 {
     acpi_log("ACPI: Write B %02X to %04X\n", val, addr);
 
-    acpi_reg_write_common(1, addr, val, p);
+    acpi_reg_write_common(1, addr, val, priv);
 }
 
 static void
-acpi_aux_reg_writel(uint16_t addr, uint32_t val, void *p)
+acpi_aux_reg_writel(uint16_t addr, uint32_t val, void *priv)
 {
     acpi_log("ACPI: Write Aux L %08X to %04X\n", val, addr);
 
-    acpi_aux_reg_write_common(4, addr, val & 0xff, p);
-    acpi_aux_reg_write_common(4, addr + 1, (val >> 8) & 0xff, p);
-    acpi_aux_reg_write_common(4, addr + 2, (val >> 16) & 0xff, p);
-    acpi_aux_reg_write_common(4, addr + 3, (val >> 24) & 0xff, p);
+    acpi_aux_reg_write_common(4, addr, val & 0xff, priv);
+    acpi_aux_reg_write_common(4, addr + 1, (val >> 8) & 0xff, priv);
+    acpi_aux_reg_write_common(4, addr + 2, (val >> 16) & 0xff, priv);
+    acpi_aux_reg_write_common(4, addr + 3, (val >> 24) & 0xff, priv);
 }
 
 static void
-acpi_aux_reg_writew(uint16_t addr, uint16_t val, void *p)
+acpi_aux_reg_writew(uint16_t addr, uint16_t val, void *priv)
 {
     acpi_log("ACPI: Write Aux W %04X to %04X\n", val, addr);
 
-    acpi_aux_reg_write_common(2, addr, val & 0xff, p);
-    acpi_aux_reg_write_common(2, addr + 1, (val >> 8) & 0xff, p);
+    acpi_aux_reg_write_common(2, addr, val & 0xff, priv);
+    acpi_aux_reg_write_common(2, addr + 1, (val >> 8) & 0xff, priv);
 }
 
 static void
-acpi_aux_reg_write(uint16_t addr, uint8_t val, void *p)
+acpi_aux_reg_write(uint16_t addr, uint8_t val, void *priv)
 {
     acpi_log("ACPI: Write Aux B %02X to %04X\n", val, addr);
 
-    acpi_aux_reg_write_common(1, addr, val, p);
+    acpi_aux_reg_write_common(1, addr, val, priv);
 }
 
 void
@@ -1374,9 +1386,9 @@ acpi_update_io_mapping(acpi_t *dev, uint32_t base, int chipset_en)
     int size;
 
     switch (dev->vendor) {
+        default:
         case VEN_ALI:
         case VEN_INTEL:
-        default:
             size = 0x040;
             break;
         case VEN_SMC:
@@ -1545,9 +1557,9 @@ acpi_pwrbtn_timer(void *priv)
 }
 
 static void
-acpi_apm_out(uint16_t port, uint8_t val, void *p)
+acpi_apm_out(uint16_t port, uint8_t val, void *priv)
 {
-    acpi_t *dev = (acpi_t *) p;
+    acpi_t *dev = (acpi_t *) priv;
 
     acpi_log("[%04X:%08X] APM write: %04X = %02X (AX = %04X, BX = %04X, CX = %04X)\n", CS, cpu_state.pc, port, val, AX, BX, CX);
 
@@ -1557,7 +1569,9 @@ acpi_apm_out(uint16_t port, uint8_t val, void *p)
         if (port == 0x0001) {
             acpi_log("ALi SOFT SMI# status set (%i)\n", dev->apm->do_smi);
             dev->apm->cmd = val;
-            // acpi_raise_smi(dev, dev->apm->do_smi);
+#if 0
+            acpi_raise_smi(dev, dev->apm->do_smi);
+#endif
             if (dev->apm->do_smi)
                 smi_raise();
             dev->regs.ali_soft_smi = 1;
@@ -1575,9 +1589,9 @@ acpi_apm_out(uint16_t port, uint8_t val, void *p)
 }
 
 static uint8_t
-acpi_apm_in(uint16_t port, void *p)
+acpi_apm_in(uint16_t port, void *priv)
 {
-    acpi_t *dev = (acpi_t *) p;
+    acpi_t *dev = (acpi_t *) priv;
     uint8_t ret = 0xff;
 
     port &= 0x0001;
@@ -1731,6 +1745,9 @@ acpi_init(const device_t *info)
             dev->suspend_types[3] = SUS_SUSPEND | SUS_RESET_CACHE;
             dev->suspend_types[4] = SUS_SUSPEND;
             break;
+
+        default:
+            break;
     }
 
     timer_add(&dev->timer, acpi_timer_overflow, dev, 0);
diff --git a/src/apm.c b/src/apm.c
index 76fc24257..8ea3f435d 100644
--- a/src/apm.c
+++ b/src/apm.c
@@ -52,9 +52,9 @@ apm_set_do_smi(apm_t *dev, uint8_t do_smi)
 }
 
 static void
-apm_out(uint16_t port, uint8_t val, void *p)
+apm_out(uint16_t port, uint8_t val, void *priv)
 {
-    apm_t *dev = (apm_t *) p;
+    apm_t *dev = (apm_t *) priv;
 
     apm_log("[%04X:%08X] APM write: %04X = %02X (BX = %04X, CX = %04X)\n", CS, cpu_state.pc, port, val, BX, CX);
 
@@ -69,9 +69,9 @@ apm_out(uint16_t port, uint8_t val, void *p)
 }
 
 static uint8_t
-apm_in(uint16_t port, void *p)
+apm_in(uint16_t port, void *priv)
 {
-    apm_t  *dev = (apm_t *) p;
+    apm_t  *dev = (apm_t *) priv;
     uint8_t ret = 0xff;
 
     port &= 0x0001;
@@ -87,17 +87,17 @@ apm_in(uint16_t port, void *p)
 }
 
 static void
-apm_reset(void *p)
+apm_reset(void *priv)
 {
-    apm_t *dev = (apm_t *) p;
+    apm_t *dev = (apm_t *) priv;
 
     dev->cmd = dev->stat = 0x00;
 }
 
 static void
-apm_close(void *p)
+apm_close(void *priv)
 {
-    apm_t *dev = (apm_t *) p;
+    apm_t *dev = (apm_t *) priv;
 
     free(dev);
 }
diff --git a/src/chipset/intel_420ex.c b/src/chipset/intel_420ex.c
index 4f7951047..9ab21e449 100644
--- a/src/chipset/intel_420ex.c
+++ b/src/chipset/intel_420ex.c
@@ -41,6 +41,9 @@
 #include <86box/machine.h>
 #include <86box/chipset.h>
 #include <86box/spd.h>
+#ifndef USE_DRB_HACK
+#include <86box/row.h>
+#endif
 
 #define MEM_STATE_SHADOW_R 0x01
 #define MEM_STATE_SHADOW_W 0x02
@@ -158,6 +161,26 @@ i420ex_smram_handler_phase1(i420ex_t *dev)
                  (regs[0x70] & 0x70) == 0x40, !(regs[0x70] & 0x20));
 }
 
+#ifndef USE_DRB_HACK
+static void
+i420ex_drb_recalc(i420ex_t *dev)
+{
+    int i;
+    uint32_t boundary;
+
+    for (i = 4; i >= 0; i--)
+        row_disable(i);
+
+    for (i = 0; i <= 4; i++) {
+	boundary = ((uint32_t) dev->regs[0x60 + i]) & 0xff;
+	row_set_boundary(i, boundary);
+    }
+
+    flushmmucache();
+}
+#endif
+
+
 static void
 i420ex_write(int func, int addr, uint8_t val, void *priv)
 {
@@ -289,7 +312,12 @@ i420ex_write(int func, int addr, uint8_t val, void *priv)
         case 0x62:
         case 0x63:
         case 0x64:
+#ifdef USE_DRB_HACK
             spd_write_drbs(dev->regs, 0x60, 0x64, 1);
+#else
+            dev->regs[addr] = val;
+            i420ex_drb_recalc(dev);
+#endif
             break;
         case 0x66:
         case 0x67:
@@ -452,7 +480,7 @@ i420ex_reset(void *priv)
         i420ex_write(0, 0x59 + i, 0x00, priv);
 
     for (uint8_t i = 0; i <= 4; i++)
-        i420ex_write(0, 0x60 + i, 0x01, priv);
+        dev->regs[0x60 + i] = 0x01;
 
     dev->regs[0x70] &= 0xef; /* Forcibly unlock the SMRAM register. */
     dev->smram_locked = 0;
@@ -530,6 +558,11 @@ i420ex_init(const device_t *info)
 
     device_add(&ide_pci_2ch_device);
 
+#ifndef USE_DRB_HACK
+    row_device.local = 4 | (1 << 8) | (0x01 << 16) | (8 << 24);
+    device_add((const device_t *) &row_device);
+#endif
+
     i420ex_reset_hard(dev);
 
     return dev;
diff --git a/src/chipset/intel_4x0.c b/src/chipset/intel_4x0.c
index ab34ddd02..0d07914b2 100644
--- a/src/chipset/intel_4x0.c
+++ b/src/chipset/intel_4x0.c
@@ -1537,7 +1537,12 @@ i4x0_reset(void *priv)
         i4x0_write(0, 0x5a + i, 0x00, priv);
 
     for (uint8_t i = 0; i <= dev->max_drb; i++)
-        i4x0_write(0, 0x60 + i, dev->drb_default, priv);
+        dev->regs[0x60 + i] = dev->drb_default;
+
+    if (dev->type >= INTEL_430NX) {
+        for (uint8_t i = 0; i < 4; i++)
+            dev->regs[0x68 + i] = 0x00;
+    }
 
     if (dev->type >= INTEL_430FX) {
         dev->regs[0x72] &= 0xef; /* Forcibly unlock the SMRAM register. */
@@ -1621,7 +1626,7 @@ i4x0_init(const device_t *info)
             regs[0x59] = 0x0f;
             regs[0x60] = regs[0x61] = regs[0x62] = regs[0x63] = 0x02;
             dev->max_drb                                      = 3;
-            dev->drb_unit                                     = 4;
+            dev->drb_unit                                     = 1;
             dev->drb_default                                  = 0x02;
             break;
         case INTEL_430LX:
diff --git a/src/chipset/intel_i450kx.c b/src/chipset/intel_i450kx.c
index d2c82d4bd..eb126c80a 100644
--- a/src/chipset/intel_i450kx.c
+++ b/src/chipset/intel_i450kx.c
@@ -326,13 +326,6 @@ pb_write(int func, int addr, uint8_t val, void *priv)
                 dev->pb_pci_conf[addr] = val & /*0x1a*/ 0x1f;
                 break;
 
-            case 0xb4:
-                dev->pb_pci_conf[addr] = val & 0xe0;
-                break;
-            case 0xb5:
-                dev->pb_pci_conf[addr] = val & 0x1f;
-                break;
-
             case 0xb8:
             case 0xb9:
                 dev->pb_pci_conf[addr] = val;
@@ -689,7 +682,7 @@ i450kx_reset(void *priv)
     dev->pb_pci_conf[0xb0] = 0x00;
     dev->pb_pci_conf[0xb1] = 0x00;
 #endif
-    dev->pb_pci_conf[0xb4] = 0x00;
+    dev->pb_pci_conf[0xb4] = 0xff;
     dev->pb_pci_conf[0xb5] = 0x00;
     dev->pb_pci_conf[0xb8] = 0x05;
     dev->pb_pci_conf[0xb9] = 0x00;
diff --git a/src/chipset/sis_85c496.c b/src/chipset/sis_85c496.c
index 6c0c1d0d3..106d5f793 100644
--- a/src/chipset/sis_85c496.c
+++ b/src/chipset/sis_85c496.c
@@ -38,10 +38,17 @@
 #include <86box/machine.h>
 #include <86box/chipset.h>
 #include <86box/spd.h>
+#ifndef USE_DRB_HACK
+#include <86box/row.h>
+#endif
 
 typedef struct sis_85c496_t {
     uint8_t    cur_reg;
     uint8_t    rmsmiblk_count;
+#ifndef USE_DRB_HACK
+    uint8_t    drb_default;
+    uint8_t    drb_bits;
+#endif
     uint8_t    regs[127];
     uint8_t    pci_conf[256];
     smram_t   *smram;
@@ -184,6 +191,26 @@ sis_85c496_ide_handler(sis_85c496_t *dev)
     }
 }
 
+#ifndef USE_DRB_HACK
+static void
+sis_85c496_drb_recalc(sis_85c496_t *dev)
+{
+    int i;
+    uint32_t boundary;
+
+    for (i = 7; i >= 0; i--)
+        row_disable(i);
+
+    for (i = 0; i <= 7; i++) {
+        boundary = ((uint32_t) dev->pci_conf[0x48 + i]);
+        row_set_boundary(i, boundary);
+    }
+
+    flushmmucache();
+}
+#endif
+
+
 /* 00 - 3F = PCI Configuration, 40 - 7F = 85C496, 80 - FF = 85C497 */
 static void
 sis_85c49x_pci_write(UNUSED(int func), int addr, uint8_t val, void *priv)
@@ -259,10 +286,12 @@ sis_85c49x_pci_write(UNUSED(int func), int addr, uint8_t val, void *priv)
         case 0x4d:
         case 0x4e:
         case 0x4f:
-#if 0
-            dev->pci_conf[addr] = val;
-#endif
+#ifdef USE_DRB_HACK
             spd_write_drbs(dev->pci_conf, 0x48, 0x4f, 1);
+#else
+            dev->pci_conf[addr] = val;
+            sis_85c496_drb_recalc(dev);
+#endif
             break;
         case 0x50:
         case 0x51: /* Exclusive Area 0 Setup */
@@ -552,7 +581,7 @@ sis_85c496_reset(void *priv)
     // sis_85c49x_pci_write(0, 0x5a, 0x06, dev);
 
     for (uint8_t i = 0; i < 8; i++)
-        sis_85c49x_pci_write(0, 0x48 + i, 0x00, dev);
+        dev->pci_conf[0x48 + i] = 0x02;
 
     sis_85c49x_pci_write(0, 0x80, 0x00, dev);
     sis_85c49x_pci_write(0, 0x81, 0x00, dev);
@@ -643,6 +672,11 @@ static void
 
     timer_add(&dev->rmsmiblk_timer, sis_85c496_rmsmiblk_count, dev, 0);
 
+#ifndef USE_DRB_HACK
+    row_device.local = 7 | (1 << 8) | (0x02 << 16) | (7 << 24);
+    device_add((const device_t *) &row_device);
+#endif
+
     sis_85c496_reset(dev);
 
     return dev;
diff --git a/src/chipset/vl82c480.c b/src/chipset/vl82c480.c
index c1716469e..055d91183 100644
--- a/src/chipset/vl82c480.c
+++ b/src/chipset/vl82c480.c
@@ -121,10 +121,13 @@ vl82c480_write(uint16_t addr, uint8_t val, void *priv)
             }
             break;
 
+/* TODO: This is actually Fast A20 disable. */
+#if 0
         case 0xee:
             if (mem_a20_alt)
                 outb(0x92, inb(0x92) & ~2);
             break;
+#endif
 
         default:
             break;
@@ -146,10 +149,13 @@ vl82c480_read(uint16_t addr, void *priv)
             ret = dev->regs[dev->idx];
             break;
 
+/* TODO: This is actually Fast A20 enable. */
+#if 0
         case 0xee:
             if (!mem_a20_alt)
                 outb(0x92, inb(0x92) | 2);
             break;
+#endif
 
         case 0xef:
             softresetx86();
diff --git a/src/codegen/codegen_ops.c b/src/codegen/codegen_ops.c
index 46a49f118..894ebb100 100644
--- a/src/codegen/codegen_ops.c
+++ b/src/codegen/codegen_ops.c
@@ -128,6 +128,54 @@ RecompOpFn recomp_opcodes_0f[512] = {
     // clang-format on
 };
 
+RecompOpFn recomp_opcodes_0f_no_mmx[512] = {
+    // clang-format off
+        /*16-bit data*/
+/*      00              01              02              03              04              05              06              07              08              09              0a              0b              0c              0d              0e              0f*/
+/*00*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*10*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*20*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*30*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+
+/*40*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*50*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*60*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*70*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+
+/*80*/  ropJO_w,        ropJNO_w,       ropJB_w,        ropJNB_w,       ropJE_w,        ropJNE_w,       ropJBE_w,       ropJNBE_w,      ropJS_w,        ropJNS_w,       ropJP_w,        ropJNP_w,       ropJL_w,        ropJNL_w,       ropJLE_w,       ropJNLE_w,
+/*90*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*a0*/  ropPUSH_FS_16,  ropPOP_FS_16,   NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           ropPUSH_GS_16,  ropPOP_GS_16,   NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*b0*/  NULL,           NULL,           ropLSS,         NULL,           ropLFS,         ropLGS,         ropMOVZX_w_b,   NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           ropMOVSX_w_b,   NULL,
+
+/*c0*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*d0*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*e0*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*f0*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+
+        /*32-bit data*/
+/*      00              01              02              03              04              05              06              07              08              09              0a              0b              0c              0d              0e              0f*/
+/*00*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*10*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*20*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*30*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+
+/*40*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*50*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*60*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*70*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+
+/*80*/  ropJO_l,        ropJNO_l,       ropJB_l,        ropJNB_l,       ropJE_l,        ropJNE_l,       ropJBE_l,       ropJNBE_l,      ropJS_l,        ropJNS_l,       ropJP_l,        ropJNP_l,       ropJL_l,        ropJNL_l,       ropJLE_l,       ropJNLE_l,
+/*90*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*a0*/  ropPUSH_FS_32,  ropPOP_FS_32,   NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           ropPUSH_GS_32,  ropPOP_GS_32,   NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*b0*/  NULL,           NULL,           ropLSS,         NULL,           ropLFS,         ropLGS,         ropMOVZX_l_b,   ropMOVZX_l_w,   NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           ropMOVSX_l_b,   ropMOVSX_l_w,
+
+/*c0*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*d0*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*e0*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*f0*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+    // clang-format on
+};
+
 RecompOpFn recomp_opcodes_d8[512] = {
     // clang-format off
         /*16-bit data*/
diff --git a/src/codegen/codegen_ops.h b/src/codegen/codegen_ops.h
index f92ba4f6d..5c19fb666 100644
--- a/src/codegen/codegen_ops.h
+++ b/src/codegen/codegen_ops.h
@@ -7,6 +7,7 @@ typedef uint32_t (*RecompOpFn)(uint8_t opcode, uint32_t fetchdat, uint32_t op_32
 
 extern RecompOpFn recomp_opcodes[512];
 extern RecompOpFn recomp_opcodes_0f[512];
+extern RecompOpFn recomp_opcodes_0f_no_mmx[512];
 extern RecompOpFn recomp_opcodes_d8[512];
 extern RecompOpFn recomp_opcodes_d9[512];
 extern RecompOpFn recomp_opcodes_da[512];
diff --git a/src/codegen/codegen_x86-64.c b/src/codegen/codegen_x86-64.c
index c02e8a7c2..3934b4ac5 100644
--- a/src/codegen/codegen_x86-64.c
+++ b/src/codegen/codegen_x86-64.c
@@ -845,7 +845,7 @@ codegen_generate_call(uint8_t opcode, OpFn op, uint32_t fetchdat, uint32_t new_p
         switch (opcode) {
             case 0x0f:
                 op_table        = x86_dynarec_opcodes_0f;
-                recomp_op_table = recomp_opcodes_0f;
+                recomp_op_table = fpu_softfloat ? recomp_opcodes_0f_no_mmx : recomp_opcodes_0f;
                 over            = 1;
                 break;
 
diff --git a/src/codegen/codegen_x86.c b/src/codegen/codegen_x86.c
index c4b32c8a2..712fbe087 100644
--- a/src/codegen/codegen_x86.c
+++ b/src/codegen/codegen_x86.c
@@ -1884,7 +1884,7 @@ codegen_generate_call(uint8_t opcode, OpFn op, uint32_t fetchdat, uint32_t new_p
         switch (opcode) {
             case 0x0f:
                 op_table        = x86_dynarec_opcodes_0f;
-                recomp_op_table = recomp_opcodes_0f;
+                recomp_op_table = fpu_softfloat ? recomp_opcodes_0f_no_mmx : recomp_opcodes_0f;
                 over            = 1;
                 break;
 
diff --git a/src/codegen_new/codegen.c b/src/codegen_new/codegen.c
index d49fcecbf..b0250fb7d 100644
--- a/src/codegen_new/codegen.c
+++ b/src/codegen_new/codegen.c
@@ -399,7 +399,7 @@ codegen_generate_call(uint8_t opcode, OpFn op, uint32_t fetchdat, uint32_t new_p
                 last_prefix = 0x0f;
 #endif
                 op_table        = x86_dynarec_opcodes_0f;
-                recomp_op_table = recomp_opcodes_0f;
+                recomp_op_table = fpu_softfloat ? recomp_opcodes_0f_no_mmx : recomp_opcodes_0f;
                 over            = 1;
                 break;
 
@@ -634,11 +634,11 @@ generate_call:
         }
 
         opcode_3dnow = fastreadb(cs + opcode_pc);
-        if (recomp_opcodes_3DNOW[opcode_3dnow]) {
+        if (!fpu_softfloat && recomp_opcodes_3DNOW[opcode_3dnow]) {
             next_pc = opcode_pc + 1;
 
             op_table           = (OpFn *) x86_dynarec_opcodes_3DNOW;
-            recomp_op_table    = recomp_opcodes_3DNOW;
+            recomp_op_table    = fpu_softfloat ? NULL : recomp_opcodes_3DNOW;
             opcode             = opcode_3dnow;
             recomp_opcode_mask = 0xff;
             opcode_mask        = 0xff;
diff --git a/src/codegen_new/codegen_ops.c b/src/codegen_new/codegen_ops.c
index 698a7899b..ae93aa80f 100644
--- a/src/codegen_new/codegen_ops.c
+++ b/src/codegen_new/codegen_ops.c
@@ -144,6 +144,54 @@ RecompOpFn recomp_opcodes_0f[512] = {
     // clang-format on
 };
 
+RecompOpFn recomp_opcodes_0f_no_mmx[512] = {
+    // clang-format off
+        /*16-bit data*/
+/*      00              01              02              03              04              05              06              07              08              09              0a              0b              0c              0d              0e              0f*/
+/*00*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*10*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*20*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*30*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+
+/*40*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*50*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*60*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*70*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+
+/*80*/  ropJO_16,       ropJNO_16,      ropJB_16,       ropJNB_16,      ropJE_16,       ropJNE_16,      ropJBE_16,      ropJNBE_16,     ropJS_16,       ropJNS_16,      ropJP_16,       ropJNP_16,      ropJL_16,       ropJNL_16,      ropJLE_16,      ropJNLE_16,
+/*90*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*a0*/  ropPUSH_FS_16,  ropPOP_FS_16,   NULL,           NULL,           ropSHLD_16_imm, NULL,           NULL,           NULL,           ropPUSH_GS_16,  ropPOP_GS_16,   NULL,           NULL,           ropSHRD_16_imm, NULL,           NULL,           NULL,
+/*b0*/  NULL,           NULL,           ropLSS_16,      NULL,           ropLFS_16,      ropLGS_16,      ropMOVZX_16_8,  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           ropMOVSX_16_8,  NULL,
+
+/*c0*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*d0*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*e0*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*f0*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+
+        /*32-bit data*/
+/*      00              01              02              03              04              05              06              07              08              09              0a              0b              0c              0d              0e              0f*/
+/*00*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*10*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*20*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*30*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+
+/*40*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*50*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*60*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*70*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+
+/*80*/  ropJO_32,       ropJNO_32,      ropJB_32,       ropJNB_32,      ropJE_32,       ropJNE_32,      ropJBE_32,      ropJNBE_32,     ropJS_32,       ropJNS_32,      ropJP_32,       ropJNP_32,      ropJL_32,       ropJNL_32,      ropJLE_32,      ropJNLE_32,
+/*90*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*a0*/  ropPUSH_FS_32,  ropPOP_FS_32,   NULL,           NULL,           ropSHLD_32_imm, NULL,           NULL,           NULL,           ropPUSH_GS_32,  ropPOP_GS_32,   NULL,           NULL,           ropSHRD_32_imm, NULL,           NULL,           NULL,
+/*b0*/  NULL,           NULL,           ropLSS_32,      NULL,           ropLFS_32,      ropLGS_32,      ropMOVZX_32_8,  ropMOVZX_32_16, NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           ropMOVSX_32_8,  ropMOVSX_32_16,
+
+/*c0*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*d0*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*e0*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*f0*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+    // clang-format on
+};
+
 RecompOpFn recomp_opcodes_3DNOW[256] = {
 // clang-format off
 #if defined __ARM_EABI__ || defined _ARM_ || defined _M_ARM || defined __aarch64__ || defined _M_ARM64
diff --git a/src/codegen_new/codegen_ops.h b/src/codegen_new/codegen_ops.h
index 9cef07e15..352d95f13 100644
--- a/src/codegen_new/codegen_ops.h
+++ b/src/codegen_new/codegen_ops.h
@@ -9,6 +9,7 @@ typedef uint32_t (*RecompOpFn)(codeblock_t *block, struct ir_data_t *ir, uint8_t
 
 extern RecompOpFn recomp_opcodes[512];
 extern RecompOpFn recomp_opcodes_0f[512];
+extern RecompOpFn recomp_opcodes_0f_no_mmx[512];
 extern RecompOpFn recomp_opcodes_3DNOW[256];
 extern RecompOpFn recomp_opcodes_d8[512];
 extern RecompOpFn recomp_opcodes_d9[512];
diff --git a/src/config.c b/src/config.c
index bb200dcbd..32824c2d4 100644
--- a/src/config.c
+++ b/src/config.c
@@ -515,13 +515,8 @@ load_machine(void)
 
     cpu_use_dynarec = !!ini_section_get_int(cat, "cpu_use_dynarec", 0);
     fpu_softfloat = !!ini_section_get_int(cat, "fpu_softfloat", 0);
-    /*The IBM PS/2 model 70 type 4 BIOS does heavy tests to the FPU in 80-bit precision mode, requiring softfloat 
-      otherwise it would always throw error 12903 on POST, so always disable dynarec and enable softfloat for this
-      machine only.*/
-    if (!strcmp(machines[machine].internal_name, "ibmps2_m70_type4")) {
-        cpu_use_dynarec = 0;
+    if (machine_has_flags(machine, MACHINE_SOFTFLOAT_ONLY))
         fpu_softfloat = 1;
-    }
 
     p = ini_section_get_string(cat, "time_sync", NULL);
     if (p != NULL) {
@@ -898,11 +893,11 @@ load_ports(void)
         sprintf(temp, "serial%d_enabled", c + 1);
         com_ports[c].enabled = !!ini_section_get_int(cat, temp, (c >= 2) ? 0 : 1);
 
-        /*
+#if 0
                 sprintf(temp, "serial%d_device", c + 1);
                 p = (char *) ini_section_get_string(cat, temp, "none");
                 com_ports[c].device = com_device_get_from_internal_name(p);
-        */
+#endif
 
         sprintf(temp, "serial%d_passthrough_enabled", c + 1);
         serial_passthrough_enabled[c] = !!ini_section_get_int(cat, temp, 0);
@@ -1080,8 +1075,8 @@ load_hard_disks(void)
 
         hdd[c].bus = hdd_string_to_bus(s, 0);
         switch (hdd[c].bus) {
-            case HDD_BUS_DISABLED:
             default:
+            case HDD_BUS_DISABLED:
                 max_spt = max_hpc = max_tracks = 0;
                 break;
 
@@ -1311,8 +1306,10 @@ load_floppy_drives(void)
         else
             strncpy(floppyfns[c], p, 511);
 
-        /* if (*wp != L'\0')
-            config_log("Floppy%d: %ls\n", c, floppyfns[c]); */
+#if 0
+        if (*wp != L'\0')
+            config_log("Floppy%d: %ls\n", c, floppyfns[c]);
+#endif
         sprintf(temp, "fdd_%02i_writeprot", c + 1);
         ui_writeprot[c] = !!ini_section_get_int(cat, temp, 0);
         ini_section_delete_var(cat, temp);
@@ -1377,8 +1374,10 @@ load_floppy_and_cdrom_drives(void)
         else
             strncpy(floppyfns[c], p, 511);
 
-        /* if (*wp != L'\0')
-            config_log("Floppy%d: %ls\n", c, floppyfns[c]); */
+#if 0
+        if (*wp != L'\0')
+            config_log("Floppy%d: %ls\n", c, floppyfns[c]);
+#endif
         sprintf(temp, "fdd_%02i_writeprot", c + 1);
         ui_writeprot[c] = !!ini_section_get_int(cat, temp, 0);
         sprintf(temp, "fdd_%02i_turbo", c + 1);
@@ -2533,6 +2532,9 @@ save_network(void)
             case NET_TYPE_VDE:
                 ini_section_set_string(cat, temp, "vde");
                 break;
+
+            default:
+                break;
         }
 
         sprintf(temp, "net_%02i_host_device", c + 1);
@@ -2542,7 +2544,9 @@ save_network(void)
             else
                 ini_section_set_string(cat, temp, net_cards_conf[c].host_dev_name);
         } else {
-            /* ini_section_set_string(cat, temp, "none"); */
+#if 0
+            ini_section_set_string(cat, temp, "none");
+#endif
             ini_section_delete_var(cat, temp);
         }
 
@@ -2573,7 +2577,7 @@ save_ports(void)
         else
             ini_section_set_int(cat, temp, com_ports[c].enabled);
 
-        /*
+#if 0
                         sprintf(temp, "serial%d_type", c + 1);
                         if (!com_ports[c].enabled))
                             ini_section_delete_var(cat, temp);
@@ -2586,7 +2590,7 @@ save_ports(void)
                         else
                             ini_section_set_string(cat, temp,
                               (char *) com_device_get_internal_name(com_ports[c].device));
-                */
+#endif
 
         sprintf(temp, "serial%d_passthrough_enabled", c + 1);
         if (serial_passthrough_enabled[c]) {
diff --git a/src/cpu/808x.c b/src/cpu/808x.c
index 925594a11..e0419a9f7 100644
--- a/src/cpu/808x.c
+++ b/src/cpu/808x.c
@@ -70,6 +70,38 @@ static int       in_rep = 0, repeating = 0, rep_c_flag = 0;
 static int       oldc, clear_lock = 0;
 static int       refresh = 0, cycdiff;
 
+static int access_code = 0;
+static int hlda = 0;
+static int not_ready = 0;
+static int bus_request_type = 0;
+static int pic_data = -1;
+static int last_was_code = 0;
+static uint16_t mem_data = 0;
+static uint32_t mem_seg = 0;
+static uint16_t mem_addr = 0;
+static int schedule_fetch = 1;
+static int pasv = 0;
+
+#define BUS_OUT     1
+#define BUS_HIGH    2
+#define BUS_WIDE    4
+#define BUS_CODE    8
+#define BUS_IO      16
+#define BUS_MEM     32
+#define BUS_PIC     64
+#define BUS_ACCESS_TYPE (BUS_CODE | BUS_IO | BUS_MEM | BUS_PIC)
+
+#define BUS_CYCLE                   (biu_cycles & 3)
+#define BUS_CYCLE_T1                biu_cycles = 0
+#define BUS_CYCLE_NEXT biu_cycles = (biu_cycles + 1) & 3
+
+enum {
+    BUS_T1 = 0,
+    BUS_T2,
+    BUS_T3,
+    BUS_T4
+};
+
 /* Various things needed for 8087. */
 #define OP_TABLE(name) ops_##name
 
@@ -151,7 +183,7 @@ x808x_log(const char *fmt, ...)
 #    define x808x_log(fmt, ...)
 #endif
 
-static void pfq_add(int c, int add);
+static void pfq_add(void);
 static void set_pzs(int bits);
 
 uint16_t
@@ -172,62 +204,344 @@ clock_end(void)
     int diff = cycdiff - cycles;
 
     /* On 808x systems, clock speed is usually crystal frequency divided by an integer. */
-    tsc += (uint64_t) diff * ((uint64_t) xt_cpu_multi >> 32ULL); /* Shift xt_cpu_multi by 32 bits to the right and then multiply. */
+    tsc += ((uint64_t) diff * ((uint64_t) xt_cpu_multi >> 32ULL)); /* Shift xt_cpu_multi by 32 bits to the right and then multiply. */
     if (TIMER_VAL_LESS_THAN_VAL(timer_target, (uint32_t) tsc))
         timer_process();
 }
 
 static void
-fetch_and_bus(int c, int bus)
-{
-    if (refresh > 0) {
-        /* Finish the current fetch, if any. */
-        cycles -= ((4 - (biu_cycles & 3)) & 3);
-        pfq_add((4 - (biu_cycles & 3)) & 3, 1);
-        /* Add 4 memory access cycles. */
-        cycles -= 4;
-        pfq_add(4, 0);
+process_timers(void)
+{  
+    clock_end();
+    clock_start();
+}
 
-        refresh--;
+static void
+cycles_forward(int c)
+{
+    cycles -= c;
+
+    if (!is286)
+        process_timers();
+}
+
+static void
+bus_outb(uint16_t port, uint8_t val)
+{
+    int old_cycles = cycles;
+
+    cycles--;
+    outb(port, val);
+    resub_cycles(old_cycles);
+}
+
+static void
+bus_outw(uint16_t port, uint16_t val)
+{
+    int old_cycles = cycles;
+
+    cycles--;
+    outw(port, val);
+    resub_cycles(old_cycles);
+}
+
+static uint8_t
+bus_inb(uint16_t port)
+{
+    int old_cycles = cycles;
+    uint8_t ret;
+
+    cycles--;
+    ret = inb(port);
+    resub_cycles(old_cycles);
+
+    return ret;
+}
+
+static uint16_t
+bus_inw(uint16_t port)
+{
+    int old_cycles = cycles;
+    uint16_t ret;
+
+    cycles--;
+    ret = inw(port);
+    resub_cycles(old_cycles);
+
+    return ret;
+}
+
+static void
+bus_do_io(int io_type)
+{
+    last_was_code = 0;
+
+    x808x_log("(%02X) bus_do_io(%02X): %04X\n", opcode, io_type, cpu_state.eaaddr);
+
+    if (io_type & BUS_OUT) {
+        if (io_type & BUS_WIDE)
+            bus_outw((uint16_t) cpu_state.eaaddr, AX);
+        else if (io_type & BUS_HIGH)
+            bus_outb(((uint16_t) cpu_state.eaaddr + 1) & 0xffff, AH);
+        else
+            bus_outb((uint16_t) cpu_state.eaaddr, AL);
+    } else {
+        if (io_type & BUS_WIDE)
+            AX = bus_inw((uint16_t) cpu_state.eaaddr);
+        else if (io_type & BUS_HIGH)
+            AH = bus_inb(((uint16_t) cpu_state.eaaddr + 1) & 0xffff);
+        else
+            AL = bus_inb((uint16_t) cpu_state.eaaddr);
     }
 
-    pfq_add(c, !bus);
-    if (bus < 2) {
-        clock_end();
-        clock_start();
+    process_timers();
+}
+
+static void
+bus_writeb(uint32_t seg, uint32_t addr, uint8_t val)
+{
+    write_mem_b(seg + addr, val);
+}
+
+static void
+bus_writew(uint32_t seg, uint32_t addr, uint16_t val)
+{
+    write_mem_w(seg + addr, val);
+}
+
+static uint8_t
+bus_readb(uint32_t seg, uint32_t addr)
+{
+    uint8_t ret = read_mem_b(seg + addr);
+
+    return ret;
+}
+
+static uint16_t
+bus_readw(uint32_t seg, uint32_t addr)
+{
+    uint16_t ret = read_mem_w(seg + addr);
+
+    return ret;
+}
+
+static void
+bus_do_mem(int io_type)
+{
+    last_was_code = 0;
+
+    if (io_type & BUS_OUT) {
+        if (io_type & BUS_WIDE)
+            bus_writew(mem_seg, (uint32_t) mem_addr, mem_data);
+        else if (io_type & BUS_HIGH) {
+            if (is186 && !is_nec)
+                bus_writeb(mem_seg, ((uint32_t) mem_addr) + 1, mem_data >> 8);
+            else
+                bus_writeb(mem_seg, (uint32_t) ((mem_addr + 1) & 0xffff), mem_data >> 8);
+        } else
+            bus_writeb(mem_seg, (uint32_t) mem_addr, mem_data & 0xff);
+    } else {
+        if (io_type & BUS_WIDE)
+            mem_data = bus_readw(mem_seg, (uint32_t) mem_addr);
+        else if (io_type & BUS_HIGH) {
+            if (is186 && !is_nec)
+                mem_data = (mem_data & 0x00ff) | (((uint16_t) bus_readb(mem_seg, ((uint32_t) mem_addr) + 1)) << 8);
+            else
+                mem_data = (mem_data & 0x00ff) | (((uint16_t) bus_readb(mem_seg, (uint32_t) ((mem_addr + 1) & 0xffff))) << 8);
+        } else
+            mem_data = (mem_data & 0xff00) | ((uint16_t) bus_readb(mem_seg, (uint32_t) mem_addr));
     }
 }
 
 static void
-wait(int c, int bus)
+run_bus_cycle(int io_type)
 {
-    cycles -= c;
-    fetch_and_bus(c, bus);
+    int do_bus_access = (io_type != 0) && (!(io_type & BUS_CODE) || schedule_fetch);
+
+    x808x_log("[%04X:%04X] %02X bus access %02X (%i)\n", CS, cpu_state.pc, opcode, io_type, do_bus_access);
+
+    if (do_bus_access) {
+        if (not_ready > 0) {
+            x808x_log("[%04X:%04X] %02X TW x%i\n", CS, cpu_state.pc, opcode, not_ready);
+            cycles_forward(not_ready);
+            not_ready = 0;
+        }
+
+        switch(BUS_CYCLE) {
+            case BUS_T1:
+                access_code = !!(io_type & BUS_CODE);
+                break;
+            case BUS_T3:
+                switch (io_type & BUS_ACCESS_TYPE) {
+                    case BUS_CODE:
+                        pfq_add();
+                        last_was_code = 1;
+                        break;
+                    case BUS_IO:
+                        bus_do_io(io_type);
+                        break;
+                    case BUS_MEM:
+                        bus_do_mem(io_type);
+                        break;
+                    case BUS_PIC:
+                        pic_data = pic_irq_ack();
+                        last_was_code = 0;
+                        break;
+                    default:
+                        break;
+                }
+                break;
+            default:
+                break;
+        }
+    }
 }
 
-/* This is for external subtraction of cycles. */
+static void
+run_dma_cycle(int idle)
+{
+    if (not_ready > 0) {
+        /* Subtract one not ready cycle. */
+        not_ready--;
+    } else if (hlda > 0) {
+        hlda--;
+        /* DMAWAIT is two cycles in, the actual wait states
+           are inserted with one cycle of delay. */
+        if (hlda == 0) {
+            /* Deassert READY. */
+            not_ready = 6;
+        }
+    } else if ((refresh > 0) && (in_lock == 0) && (idle || (BUS_CYCLE >= BUS_T3))) {
+        /* Refresh pending and it's either non-bus cycle or T3-T4,
+           raise HLDA. */
+        hlda = 2;
+        /* Decrease the refresh count. */
+        refresh--;
+    }
+}
+
+static void
+cycles_idle(int c)
+{
+    int d;
+
+    for (d = 0; d < c; d++) {
+        x808x_log("[%04X:%04X] %02X TI\n", CS, cpu_state.pc, opcode);
+
+        cycles_forward(1);
+        run_dma_cycle(1);
+    }
+}
+
+static void
+cycles_biu(int bus, int init)
+{
+    /* T1, T2 = Nothing, T3 = Start and schedule, T4 = Nothing */
+    pasv = (bus || ((BUS_CYCLE == BUS_T1) && schedule_fetch)) ? 0 : 1;
+
+    x808x_log("cycles_biu(%i, %i): %i, %i, %i, %i\n", bus, init, prefetching, pfq_pos, pfq_size, BUS_CYCLE);
+    if (bus) {
+        /* CPU wants non-code bus access. */
+        if (init) {
+            if (schedule_fetch) {
+                switch (BUS_CYCLE) {
+                    case BUS_T1:
+                    case BUS_T2:
+                        BUS_CYCLE_T1;                  /* Simply abort the prefetch before actual scheduling, no penalty. */
+                        break;
+                    case BUS_T3:
+                    case BUS_T4:
+                        cycles_idle(5 - BUS_CYCLE);    /* Leftover BIU cycles + 2 idle cycles. */
+                        BUS_CYCLE_T1;                  /* Abort the prefetch. */
+                        break;
+                }
+
+                schedule_fetch = 0;
+                access_code = 0;
+            }
+        }
+
+        run_bus_cycle(bus_request_type);
+    } else {
+        /* CPU wants idle or code bus access. */
+        if (schedule_fetch)
+            run_bus_cycle(BUS_CODE);
+    }
+
+    if (BUS_CYCLE == BUS_T3)
+        schedule_fetch = prefetching && (pfq_pos < pfq_size);
+
+    run_dma_cycle(pasv);
+
+    BUS_CYCLE_NEXT;
+}
+
+#ifdef REENIGNE_MODELING
+static void
+bus_init(void)
+{
+    /* Replacement for the old access() stuff. */
+    if ((BUS_CYCLE == BUS_T4) && last_was_code && (opcode != 0x8f) && (opcode != 0xc7) && (opcode != 0xcc) && (opcode != 0xcd) && (opcode != 0xce) && ((opcode & 0xf0) != 0xa0))
+        cycles_idle(1);
+
+    cycles_idle(2);
+
+    while ((BUS_CYCLE == BUS_T2) || (BUS_CYCLE == BUS_T3))
+        cycles_idle(1);
+}
+#endif
+
+/* Bus:
+   0    CPU cycles without bus access.
+   1    CPU cycle T1-T4, bus access.
+   2    CPU cycle Tw (wait state).
+   3    CPU cycle Ti (idle).
+ */
+static void
+wait(int c, int bus)
+{
+    int d;
+
+    if (c < 0)
+        pclog("Negative cycles: %i!\n", c);
+
+    x808x_log("[%04X:%04X] %02X %i cycles (%i)\n", CS, cpu_state.pc, opcode, c, bus);
+
+    for (d = 0; d < c; d++) {
+        x808x_log("[%04X:%04X] %02X cycle %i BIU\n", CS, cpu_state.pc, opcode, d);
+        cycles_biu(bus, !d);
+        x808x_log("[%04X:%04X] %02X cycle %i EU\n", CS, cpu_state.pc, opcode, d);
+        cycles_forward(1);
+    }
+}
+
+/* This is for external subtraction of cycles, ie. wait states. */
 void
 sub_cycles(int c)
 {
-    if (c <= 0)
-        return;
-
-    cycles -= c;
-
-    if (!is286)
-        fetch_and_bus(c, 2);
+    if (is286)
+        cycles -= c;
+    else {
+        if (c > 0)
+            cycles_idle(c);
+    }
 }
 
 void
 resub_cycles(int old_cycles)
 {
-    int cyc_diff = 0;
+    int i, cyc_diff = 0;
 
     if (old_cycles > cycles) {
         cyc_diff = old_cycles - cycles;
-        cycles   = old_cycles;
-        sub_cycles(cyc_diff);
+
+        for (i = 0; i < cyc_diff; i++) {
+            if (not_ready > 0)
+                not_ready--;
+        }
     }
+
+    process_timers();
 }
 
 #undef readmemb
@@ -238,53 +552,63 @@ resub_cycles(int old_cycles)
 static void
 cpu_io(int bits, int out, uint16_t port)
 {
-    int old_cycles = cycles;
+#ifdef REENIGNE_MODELING
+    bus_init();
+#endif
 
     if (out) {
-        wait(4, 1);
         if (bits == 16) {
             if (is8086 && !(port & 1)) {
-                old_cycles = cycles;
-                outw(port, AX);
-            } else {
+                bus_request_type = BUS_IO | BUS_OUT | BUS_WIDE;
+                wait(4, 1);
+            } else {
+                bus_request_type = BUS_IO | BUS_OUT;
+                wait(4, 1);
+                schedule_fetch = 0;
+                bus_request_type = BUS_IO | BUS_OUT | BUS_HIGH;
                 wait(4, 1);
-                old_cycles = cycles;
-                outb(port++, AL);
-                outb(port, AH);
             }
         } else {
-            old_cycles = cycles;
-            outb(port, AL);
+            bus_request_type = BUS_IO | BUS_OUT;
+            wait(4, 1);
         }
     } else {
-        wait(4, 1);
         if (bits == 16) {
             if (is8086 && !(port & 1)) {
-                old_cycles = cycles;
-                AX         = inw(port);
-            } else {
+                bus_request_type = BUS_IO | BUS_WIDE;
+                wait(4, 1);
+            } else {
+                bus_request_type = BUS_IO;
+                wait(4, 1);
+                schedule_fetch = 0;
+                bus_request_type = BUS_IO | BUS_HIGH;
                 wait(4, 1);
-                old_cycles = cycles;
-                AL         = inb(port++);
-                AH         = inb(port);
             }
         } else {
-            old_cycles = cycles;
-            AL         = inb(port);
+            bus_request_type = BUS_IO;
+            wait(4, 1);
         }
     }
 
-    resub_cycles(old_cycles);
+    bus_request_type = 0;
 }
 
 /* Reads a byte from the memory and advances the BIU. */
 static uint8_t
-readmemb(uint32_t a)
+readmemb(uint32_t s, uint16_t a)
 {
     uint8_t ret;
 
+#ifdef REENIGNE_MODELING
+    bus_init();
+#endif
+
+    mem_seg = s;
+    mem_addr = a;
+    bus_request_type = BUS_MEM;
     wait(4, 1);
-    ret = read_mem_b(a);
+    ret = mem_data & 0xff;
+    bus_request_type = 0;
 
     return ret;
 }
@@ -298,6 +622,8 @@ readmembf(uint32_t a)
     a   = cs + (a & 0xffff);
     ret = read_mem_b(a);
 
+    last_was_code = 1;
+
     return ret;
 }
 
@@ -307,14 +633,24 @@ readmemw(uint32_t s, uint16_t a)
 {
     uint16_t ret;
 
-    wait(4, 1);
-    if (is8086 && !(a & 1))
-        ret = read_mem_w(s + a);
-    else {
+#ifdef REENIGNE_MODELING
+    bus_init();
+#endif
+
+    mem_seg = s;
+    mem_addr = a;
+    if (is8086 && !(a & 1)) {
+        bus_request_type = BUS_MEM | BUS_WIDE;
+        wait(4, 1);
+    } else {
+        bus_request_type = BUS_MEM | BUS_HIGH;
+        wait(4, 1);
+        schedule_fetch = 0;
+        bus_request_type = BUS_MEM;
         wait(4, 1);
-        ret = read_mem_b(s + a);
-        ret |= read_mem_b(s + ((is186 && !is_nec) ? (a + 1) : (a + 1) & 0xffff)) << 8;
     }
+    ret = mem_data;
+    bus_request_type = 0;
 
     return ret;
 }
@@ -326,6 +662,8 @@ readmemwf(uint16_t a)
 
     ret = read_mem_w(cs + (a & 0xffff));
 
+    last_was_code = 1;
+
     return ret;
 }
 
@@ -335,7 +673,7 @@ readmem(uint32_t s)
     if (opcode & 1)
         return readmemw(s, cpu_state.eaaddr);
     else
-        return (uint16_t) readmemb(s + cpu_state.eaaddr);
+        return (uint16_t) readmemb(s, cpu_state.eaaddr);
 }
 
 static uint32_t
@@ -357,6 +695,8 @@ readmemq(uint32_t s, uint16_t a)
     temp = (uint64_t) (readmeml(s, a + 4)) << 32;
     temp |= readmeml(s, a);
 
+    last_was_code = 0;
+
     return temp;
 }
 
@@ -366,8 +706,16 @@ writememb(uint32_t s, uint32_t a, uint8_t v)
 {
     uint32_t addr = s + a;
 
+#ifdef REENIGNE_MODELING
+    bus_init();
+#endif
+
+    mem_seg = s;
+    mem_addr = a;
+    mem_data = v;
+    bus_request_type = BUS_MEM | BUS_OUT;
     wait(4, 1);
-    write_mem_b(addr, v);
+    bus_request_type = 0;
 
     if ((addr >= 0xf0000) && (addr <= 0xfffff))
         last_addr = addr & 0xffff;
@@ -379,15 +727,24 @@ writememw(uint32_t s, uint32_t a, uint16_t v)
 {
     uint32_t addr = s + a;
 
-    wait(4, 1);
-    if (is8086 && !(a & 1))
-        write_mem_w(addr, v);
-    else {
-        write_mem_b(addr, v & 0xff);
+#ifdef REENIGNE_MODELING
+    bus_init();
+#endif
+
+    mem_seg = s;
+    mem_addr = a;
+    mem_data = v;
+    if (is8086 && !(a & 1)) {
+        bus_request_type = BUS_MEM | BUS_OUT | BUS_WIDE;
+        wait(4, 1);
+    } else {
+        bus_request_type = BUS_MEM | BUS_OUT | BUS_HIGH;
+        wait(4, 1);
+        schedule_fetch = 0;
+        bus_request_type = BUS_MEM | BUS_OUT;
         wait(4, 1);
-        addr = s + ((is186 && !is_nec) ? (a + 1) : ((a + 1) & 0xffff));
-        write_mem_b(addr, v >> 8);
     }
+    bus_request_type = 0;
 
     if ((addr >= 0xf0000) && (addr <= 0xfffff))
         last_addr = addr & 0xffff;
@@ -420,21 +777,26 @@ static void
 pfq_write(void)
 {
     uint16_t tempw;
+    /* Byte fetch on odd addres on 8086 to simulate the HL toggle. */
+    int fetch_word = is8086 && !(pfq_ip & 1);
 
-    if (is8086 && (pfq_pos < (pfq_size - 1))) {
+    if (fetch_word && (pfq_pos < (pfq_size - 1))) {
         /* The 8086 fetches 2 bytes at a time, and only if there's at least 2 bytes
            free in the queue. */
         tempw                         = readmemwf(pfq_ip);
         *(uint16_t *) &(pfq[pfq_pos]) = tempw;
         pfq_ip = (pfq_ip + 2) & 0xffff;
         pfq_pos += 2;
-    } else if (!is8086 && (pfq_pos < pfq_size)) {
+    } else if (!fetch_word && (pfq_pos < pfq_size)) {
         /* The 8088 fetches 1 byte at a time, and only if there's at least 1 byte
            free in the queue. */
         pfq[pfq_pos] = readmembf(pfq_ip);
         pfq_ip = (pfq_ip + 1) & 0xffff;
         pfq_pos++;
     }
+
+    if (pfq_pos >= pfq_size)
+        pfq_pos = pfq_size;
 }
 
 static uint8_t
@@ -446,12 +808,18 @@ pfq_read(void)
     for (int i = 0; i < (pfq_size - 1); i++)
         pfq[i] = pfq[i + 1];
     pfq_pos--;
+    if (pfq_pos < 0)
+        pfq_pos = 0;
     cpu_state.pc = (cpu_state.pc + 1) & 0xffff;
     return temp;
 }
 
 /* Fetches a byte from the prefetch queue, or from memory if the queue has
-   been drained. */
+   been drained.
+
+   Cycles: 1                         If fetching from the queue;
+           (4 - (biu_cycles & 3))    If fetching from the bus - fetch into the queue;
+           1                         If fetching from the bus - delay. */
 static uint8_t
 pfq_fetchb_common(void)
 {
@@ -461,7 +829,8 @@ pfq_fetchb_common(void)
         /* Reset prefetch queue internal position. */
         pfq_ip = cpu_state.pc;
         /* Fill the queue. */
-        wait(4 - (biu_cycles & 3), 0);
+        while (pfq_pos == 0)
+            wait(1, 0);
     }
 
     /* Fetch. */
@@ -469,6 +838,7 @@ pfq_fetchb_common(void)
     return temp;
 }
 
+/* The timings are above. */
 static uint8_t
 pfq_fetchb(void)
 {
@@ -504,18 +874,10 @@ pfq_fetch(void)
 
 /* Adds bytes to the prefetch queue based on the instruction's cycle count. */
 static void
-pfq_add(int c, int add)
+pfq_add(void)
 {
-    int d;
-
-    if ((c <= 0) || (pfq_pos >= pfq_size))
-        return;
-
-    for (d = 0; d < c; d++) {
-        biu_cycles = (biu_cycles + 1) & 0x03;
-        if (prefetching && add && (biu_cycles == 0x00))
-            pfq_write();
-    }
+    if (prefetching && (pfq_pos < pfq_size))
+        pfq_write();
 }
 
 /* Clear the prefetch queue - called on reset and on anything that affects either CS or IP. */
@@ -524,6 +886,16 @@ pfq_clear(void)
 {
     pfq_pos     = 0;
     prefetching = 0;
+    schedule_fetch = 0;
+
+    BUS_CYCLE_T1;
+}
+
+static void
+pfq_suspend(void)
+{
+   pfq_clear();
+   cycles_idle(3);
 }
 
 static void
@@ -543,7 +915,7 @@ load_seg(uint16_t seg, x86seg *s)
 void
 reset_808x(int hard)
 {
-    biu_cycles = 0;
+    BUS_CYCLE_T1;
     in_rep     = 0;
     in_lock    = 0;
     completed  = 1;
@@ -573,10 +945,24 @@ reset_808x(int hard)
     rammask = 0xfffff;
 
     prefetching = 1;
+
+    schedule_fetch = 1;
+    pasv           = 0;
+
     cpu_alu_op  = 0;
 
     use_custom_nmi_vector = 0x00;
     custom_nmi_vector     = 0x00000000;
+
+    access_code           = 0;
+    hlda                  = 0;
+    not_ready             = 0;
+    bus_request_type      = 0;
+    pic_data              = -1;
+    last_was_code         = 0;
+    mem_data              = 0;
+    mem_seg               = 0;
+    mem_addr              = 0;
 }
 
 static void
@@ -584,6 +970,7 @@ set_ip(uint16_t new_ip)
 {
     pfq_ip = cpu_state.pc = new_ip;
     prefetching           = 1;
+    schedule_fetch = prefetching && (pfq_pos < pfq_size);
 }
 
 /* Memory refresh read - called by reads and writes on DMA channel 0. */
@@ -626,34 +1013,34 @@ do_mod_rm(void)
     if (cpu_mod == 3)
         return;
 
-    wait(1, 0);
+    wait(2, 0);
     if ((rmdat & 0xc7) == 0x06) {
-        wait(1, 0);
         cpu_state.eaaddr = pfq_fetchw();
         easeg            = ovr_seg ? *ovr_seg : ds;
-        wait(1, 0);
+        wait(2, 0);
         return;
-    } else
-        switch (cpu_rm) {
-            case 0:
-            case 3:
-                wait(2, 0);
-                break;
-            case 1:
-            case 2:
-                wait(3, 0);
-                break;
-        }
+    } else  switch (cpu_rm) {
+        case 0:
+        case 3:
+            wait(2, 0);
+            break;
+        case 1:
+        case 2:
+            wait(3, 0);
+            break;
+    }
     cpu_state.eaaddr = (*mod1add[0][cpu_rm]) + (*mod1add[1][cpu_rm]);
     easeg            = ovr_seg ? *ovr_seg : *mod1seg[cpu_rm];
     switch (rmdat & 0xc0) {
         case 0x40:
-            wait(3, 0);
+            wait(2, 0);
             cpu_state.eaaddr += sign_extend(pfq_fetchb());
+            wait(1, 0);
             break;
         case 0x80:
-            wait(3, 0);
+            wait(2, 0);
             cpu_state.eaaddr += pfq_fetchw();
+            wait(1, 0);
             break;
     }
     cpu_state.eaaddr &= 0xffff;
@@ -677,7 +1064,7 @@ geteab(void)
     if (cpu_mod == 3)
         return (getr8(cpu_rm));
 
-    return readmemb(easeg + cpu_state.eaaddr);
+    return readmemb(easeg, cpu_state.eaaddr);
 }
 
 /* Reads a word from the effective address. */
@@ -721,7 +1108,7 @@ read_ea(int memory_only, int bits)
         if (bits == 16)
             cpu_data = readmemw(easeg, cpu_state.eaaddr);
         else
-            cpu_data = readmemb(easeg + cpu_state.eaaddr);
+            cpu_data = readmemb(easeg, cpu_state.eaaddr);
         return;
     }
     if (!memory_only) {
@@ -739,7 +1126,7 @@ read_ea2(int bits)
     if (bits == 16)
         cpu_data = readmemw(easeg, cpu_state.eaaddr);
     else
-        cpu_data = readmemb(easeg + cpu_state.eaaddr);
+        cpu_data = readmemb(easeg, cpu_state.eaaddr);
 }
 
 /* Writes a byte to the effective address. */
@@ -814,135 +1201,6 @@ pop(void)
     return readmemw(ss, cpu_state.eaaddr);
 }
 
-static void
-access(int num, int bits)
-{
-    switch (num) {
-        case 0:
-        case 61:
-        case 63:
-        case 64:
-        case 67:
-        case 69:
-        case 71:
-        case 72:
-        default:
-            break;
-        case 1:
-        case 6:
-        case 7:
-        case 8:
-        case 9:
-        case 17:
-        case 20:
-        case 21:
-        case 24:
-        case 28:
-        case 47:
-        case 48:
-        case 49:
-        case 50:
-        case 51:
-        case 55:
-        case 56:
-        case 62:
-        case 66:
-        case 68:
-            wait(1, 0);
-            break;
-        case 3:
-        case 11:
-        case 15:
-        case 22:
-        case 23:
-        case 25:
-        case 26:
-        case 35:
-        case 44:
-        case 45:
-        case 46:
-        case 52:
-        case 53:
-        case 54:
-            wait(2, 0);
-            break;
-        case 16:
-        case 18:
-        case 19:
-        case 27:
-        case 32:
-        case 37:
-        case 42:
-            wait(3, 0);
-            break;
-        case 10:
-        case 12:
-        case 13:
-        case 14:
-        case 29:
-        case 30:
-        case 33:
-        case 34:
-        case 39:
-        case 41:
-        case 60:
-            wait(4, 0);
-            break;
-        case 4:
-        case 70:
-            wait(5, 0);
-            break;
-        case 31:
-        case 38:
-        case 40:
-            wait(6, 0);
-            break;
-        case 5:
-            if (opcode == 0xcc)
-                wait(7, 0);
-            else
-                wait(4, 0);
-            break;
-        case 36:
-            wait(1, 0);
-            pfq_clear();
-            wait(1, 0);
-            if (cpu_mod != 3)
-                wait(1, 0);
-            wait(3, 0);
-            break;
-        case 43:
-            wait(2, 0);
-            pfq_clear();
-            wait(1, 0);
-            break;
-        case 57:
-            if (cpu_mod != 3)
-                wait(2, 0);
-            wait(4, 0);
-            break;
-        case 58:
-            if (cpu_mod != 3)
-                wait(1, 0);
-            wait(4, 0);
-            break;
-        case 59:
-            wait(2, 0);
-            pfq_clear();
-            if (cpu_mod != 3)
-                wait(1, 0);
-            wait(3, 0);
-            break;
-        case 65:
-            wait(1, 0);
-            pfq_clear();
-            wait(2, 0);
-            if (cpu_mod != 3)
-                wait(1, 0);
-            break;
-    }
-}
-
 /* Calls an interrupt. */
 static void
 interrupt(uint16_t addr)
@@ -954,26 +1212,24 @@ interrupt(uint16_t addr)
     addr <<= 2;
     cpu_state.eaaddr = addr;
     old_cs           = CS;
-    access(5, 16);
     new_ip = readmemw(0, cpu_state.eaaddr);
     wait(1, 0);
     cpu_state.eaaddr = (cpu_state.eaaddr + 2) & 0xffff;
-    access(6, 16);
     new_cs      = readmemw(0, cpu_state.eaaddr);
     prefetching = 0;
     pfq_clear();
     ovr_seg = NULL;
-    access(39, 16);
+    wait(2, 0);
     tempf = cpu_state.flags & (is_nec ? 0x8fd7 : 0x0fd7);
     push(&tempf);
     cpu_state.flags &= ~(I_FLAG | T_FLAG);
-    access(40, 16);
+    wait(5, 0);
     push(&old_cs);
     old_ip = cpu_state.pc;
     load_cs(new_cs);
-    access(68, 16);
+    pfq_suspend();
     set_ip(new_ip);
-    access(41, 16);
+    wait(2, 0);
     push(&old_ip);
 }
 
@@ -992,28 +1248,26 @@ custom_nmi(void)
 
     cpu_state.eaaddr = 0x0002;
     old_cs           = CS;
-    access(5, 16);
     (void) readmemw(0, cpu_state.eaaddr);
     new_ip = custom_nmi_vector & 0xffff;
     wait(1, 0);
     cpu_state.eaaddr = (cpu_state.eaaddr + 2) & 0xffff;
-    access(6, 16);
     (void) readmemw(0, cpu_state.eaaddr);
     new_cs      = custom_nmi_vector >> 16;
     prefetching = 0;
     pfq_clear();
     ovr_seg = NULL;
-    access(39, 16);
+    wait(2, 0);
     tempf = cpu_state.flags & (is_nec ? 0x8fd7 : 0x0fd7);
     push(&tempf);
     cpu_state.flags &= ~(I_FLAG | T_FLAG);
-    access(40, 16);
+    wait(5, 0);
     push(&old_cs);
     old_ip = cpu_state.pc;
     load_cs(new_cs);
-    access(68, 16);
+    pfq_suspend();
     set_ip(new_ip);
-    access(41, 16);
+    wait(2, 0);
     push(&old_ip);
 }
 
@@ -1027,6 +1281,18 @@ irq_pending(void)
     return temp;
 }
 
+static int
+bus_pic_ack(void)
+{
+    int old_in_lock = in_lock;
+
+    in_lock = 1;
+    bus_request_type = BUS_PIC;
+    wait(4, 1);
+    in_lock = old_in_lock;
+    return pic_data;
+}
+
 static void
 check_interrupts(void)
 {
@@ -1034,11 +1300,13 @@ check_interrupts(void)
 
     if (irq_pending()) {
         if ((cpu_state.flags & T_FLAG) && !noint) {
+            wait(2, 0);
             interrupt(1);
             return;
         }
         if (nmi && nmi_enable && nmi_mask) {
             nmi_enable = 0;
+            wait(2, 0);
             if (use_custom_nmi_vector)
                 custom_nmi();
             else
@@ -1052,26 +1320,61 @@ check_interrupts(void)
             repeating = 0;
             completed = 1;
             ovr_seg   = NULL;
-            wait(3, 0);
+            wait(4, 0);
             /* ACK to PIC */
-            temp = pic_irq_ack();
-            wait(4, 1);
+            temp = bus_pic_ack();
             wait(1, 0);
             /* ACK to PIC */
-            temp = pic_irq_ack();
-            wait(4, 1);
+            temp = bus_pic_ack();
             wait(1, 0);
             in_lock    = 0;
             clear_lock = 0;
-            wait(1, 0);
+            if (BUS_CYCLE != BUS_T3)
+                wait(1, 0);
+            wait(5, 0);
             /* Here is where temp should be filled, but we cheat. */
-            wait(3, 0);
             opcode = 0x00;
             interrupt(temp);
         }
     }
 }
 
+static uint16_t tmpc;
+
+static int
+rep_setup(void)
+{
+    if (repeating)
+        return 0;
+    wait(2, 0);
+    if (in_rep == 0)
+        return 0;
+    wait(4, 0);
+    tmpc = CX;
+    if (tmpc == 0)
+        return 1;
+    wait(3, 0);
+    return 0;
+}
+
+static int
+rep_interrupt(void)
+{
+    if (!irq_pending()) {
+        repeating = 1;
+        completed = 0;
+        return 0;
+    }
+    completed = 1;
+    CX = tmpc;
+    pfq_clear();
+    if (is_nec && (ovr_seg != NULL))
+        set_ip(cpu_state.pc - 3);
+    else
+        set_ip(cpu_state.pc - 2);
+    return 1;
+}
+
 static int
 rep_action(int bits)
 {
@@ -1082,7 +1385,6 @@ rep_action(int bits)
     wait(2, 0);
     t = CX;
     if (irq_pending() && (repeating != 0)) {
-        access(71, bits);
         pfq_clear();
         if (is_nec && (ovr_seg != NULL))
             set_ip(cpu_state.pc - 3);
@@ -1108,9 +1410,9 @@ static uint16_t
 jump(uint16_t delta)
 {
     uint16_t old_ip;
-    access(67, 8);
-    pfq_clear();
-    wait(5, 0);
+    wait(1, 0);
+    pfq_suspend();
+    cycles_idle(1);
     old_ip = cpu_state.pc;
     set_ip((cpu_state.pc + delta) & 0xffff);
     return old_ip;
@@ -1132,8 +1434,6 @@ jump_near(void)
 static void
 jcc(uint8_t opcode, int cond)
 {
-    /* int8_t offset; */
-
     wait(1, 0);
     cpu_data = pfq_fetchb();
     wait(1, 0);
@@ -1536,7 +1836,7 @@ lods(int bits)
     if (bits == 16)
         cpu_data = readmemw((ovr_seg ? *ovr_seg : ds), cpu_state.eaaddr);
     else
-        cpu_data = readmemb((ovr_seg ? *ovr_seg : ds) + cpu_state.eaaddr);
+        cpu_data = readmemb((ovr_seg ? *ovr_seg : ds), cpu_state.eaaddr);
     SI = string_increment(bits);
 }
 
@@ -1658,6 +1958,7 @@ execx86(int cycs)
     uint16_t tempw_int, size, tempbp, lowbound;
     uint16_t highbound, regval, orig_sp, wordtopush;
     uint16_t immediate, old_flags;
+        uint16_t tmpa;
     int      bits;
     uint32_t dest_seg, i, carry, nibble;
     uint32_t srcseg, byteaddr;
@@ -1669,7 +1970,8 @@ execx86(int cycs)
 
         if (!repeating) {
             cpu_state.oldpc = cpu_state.pc;
-            opcode          = pfq_fetchb();
+            // opcode          = pfq_fetchb();
+            opcode          = pfq_fetchb_common();
             handled         = 0;
             oldc            = cpu_state.flags & C_FLAG;
             if (clear_lock) {
@@ -1680,7 +1982,7 @@ execx86(int cycs)
         }
 
         completed = 1;
-        // pclog("[%04X:%04X] Opcode: %02X\n", CS, cpu_state.pc, opcode);
+        x808x_log("[%04X:%04X] Opcode: %02X\n", CS, cpu_state.pc, opcode);
         if (is186) {
             switch (opcode) {
                 case 0x60: /*PUSHA/PUSH R*/
@@ -1820,7 +2122,7 @@ execx86(int cycs)
                         SI += (cpu_state.flags & D_FLAG) ? -2 : 2;
                     } else {
                         wait(4, 0);
-                        outb(DX, readmemb(dest_seg + SI));
+                        outb(DX, readmemb(dest_seg, SI));
                         SI += (cpu_state.flags & D_FLAG) ? -1 : 1;
                     }
                     if (in_rep == 0)
@@ -1858,7 +2160,6 @@ execx86(int cycs)
                     do_mod_rm();
                     if (cpu_mod == 3)
                         wait(1, 0);
-                    access(53, bits);
                     cpu_data = get_ea();
                     cpu_src  = pfq_fetchb();
 
@@ -1937,7 +2238,6 @@ execx86(int cycs)
                             wait(4, 0);
                         --cpu_src;
                     }
-                    access(17, bits);
                     set_ea(cpu_data);
                     handled = 1;
                     break;
@@ -1955,7 +2255,7 @@ execx86(int cycs)
                 case 0x0E:
                 case 0x16:
                 case 0x1E: /* PUSH seg */
-                    access(29, 16);
+                    wait(3, 0);
                     push(&(_opseg[(opcode >> 3) & 0x03]->seg));
                     break;
                 case 0x07:
@@ -2229,7 +2529,7 @@ execx86(int cycs)
                                 AX = 0;
                                 for (i = 0; i < bit_length; i++) {
                                     byteaddr = (ds) + SI;
-                                    AX |= (!!(readmemb(byteaddr) & (1 << bit_offset))) << i;
+                                    AX |= (!!(readmemb((ds), SI) & (1 << bit_offset))) << i;
                                     bit_offset++;
                                     if (bit_offset == 8) {
                                         SI++;
@@ -2251,19 +2551,16 @@ execx86(int cycs)
                                 cpu_state.pc = (cpu_state.pc - 1) & 0xffff;
                                 break;
                         }
-                    } else
-                        handled = 0;
-                    if (handled)
-                        break;
-                    access(22, 16);
-                    if (opcode == 0x0F) {
-                        load_cs(pop());
-                        pfq_pos = 0;
-                    } else
-                        load_seg(pop(), _opseg[(opcode >> 3) & 0x03]);
-                    wait(1, 0);
-                    /* All POP segment instructions suppress interrupts for one instruction. */
-                    noint = 1;
+                    } else {
+                        wait(1, 0);
+                        if (opcode == 0x0F) {
+                            load_cs(pop());
+                            pfq_pos = 0;
+                        } else
+                            load_seg(pop(), _opseg[(opcode >> 3) & 0x03]);
+                        /* All POP segment instructions suppress interrupts for one instruction. */
+                        noint = 1;
+                    }
                     break;
 
                 case 0x26: /*ES:*/
@@ -2310,7 +2607,6 @@ execx86(int cycs)
                     /* alu rm, r / r, rm */
                     bits = 8 << (opcode & 1);
                     do_mod_rm();
-                    access(46, bits);
                     tempw      = get_ea();
                     cpu_alu_op = (opcode >> 3) & 7;
                     if ((opcode & 2) == 0) {
@@ -2320,22 +2616,19 @@ execx86(int cycs)
                         cpu_dest = get_reg(cpu_reg);
                         cpu_src  = tempw;
                     }
-                    if (cpu_mod != 3)
-                        wait(2, 0);
                     wait(1, 0);
+                    if (cpu_mod != 3)
+                        wait(1, 0);
                     alu_op(bits);
+                    wait(1, 0);
                     if (cpu_alu_op != 7) {
                         if ((opcode & 2) == 0) {
-                            access(10, bits);
-                            set_ea(cpu_data);
                             if (cpu_mod == 3)
-                                wait(1, 0);
-                        } else {
+                                wait(2, 0);
+                            set_ea(cpu_data);
+                        } else
                             set_reg(cpu_reg, cpu_data);
-                            wait(1, 0);
-                        }
-                    } else
-                        wait(1, 0);
+                    }
                     break;
 
                 case 0x04:
@@ -2485,7 +2778,7 @@ execx86(int cycs)
                 case 0x55:
                 case 0x56:
                 case 0x57:
-                    access(30, 16);
+                    wait(3, 0);
                     push(&(cpu_state.regs[opcode & 0x07].w));
                     break;
                 case 0x58:
@@ -2496,9 +2789,8 @@ execx86(int cycs)
                 case 0x5D:
                 case 0x5E:
                 case 0x5F:
-                    access(23, 16);
-                    cpu_state.regs[opcode & 0x07].w = pop();
                     wait(1, 0);
+                    cpu_state.regs[opcode & 0x07].w = pop();
                     break;
 
                 case 0x60: /*JO alias*/
@@ -2561,18 +2853,14 @@ execx86(int cycs)
                     /* alu rm, imm */
                     bits = 8 << (opcode & 1);
                     do_mod_rm();
-                    access(47, bits);
                     cpu_data = get_ea();
                     cpu_dest = cpu_data;
                     if (cpu_mod != 3)
-                        wait(3, 0);
-                    if (opcode == 0x81) {
-                        if (cpu_mod == 3)
-                            wait(1, 0);
+                        wait(1, 0);
+                    wait(1, 0);
+                    if (opcode == 0x81)
                         cpu_src = pfq_fetchw();
-                    } else {
-                        if (cpu_mod == 3)
-                            wait(1, 0);
+                    else {
                         if (opcode == 0x83)
                             cpu_src = sign_extend(pfq_fetchb());
                         else
@@ -2582,7 +2870,8 @@ execx86(int cycs)
                     cpu_alu_op = (rmdat & 0x38) >> 3;
                     alu_op(bits);
                     if (cpu_alu_op != 7) {
-                        access(11, bits);
+                        if (cpu_mod != 3)
+                            wait(1, 0);
                         set_ea(cpu_data);
                     } else {
                         if (cpu_mod != 3)
@@ -2595,11 +2884,10 @@ execx86(int cycs)
                     /* TEST rm, reg */
                     bits = 8 << (opcode & 1);
                     do_mod_rm();
-                    access(48, bits);
                     cpu_data = get_ea();
                     test(bits, cpu_data, get_reg(cpu_reg));
-                    if (cpu_mod == 3)
-                        wait(2, 0);
+                    if (cpu_mod != 3)
+                        wait(1, 0);
                     wait(2, 0);
                     break;
                 case 0x86:
@@ -2607,12 +2895,12 @@ execx86(int cycs)
                     /* XCHG rm, reg */
                     bits = 8 << (opcode & 1);
                     do_mod_rm();
-                    access(49, bits);
                     cpu_data = get_ea();
                     cpu_src  = get_reg(cpu_reg);
                     set_reg(cpu_reg, cpu_data);
                     wait(3, 0);
-                    access(12, bits);
+                    if (cpu_mod != 3)
+                        wait(3, 0);
                     set_ea(cpu_src);
                     break;
 
@@ -2622,7 +2910,8 @@ execx86(int cycs)
                     bits = 8 << (opcode & 1);
                     do_mod_rm();
                     wait(1, 0);
-                    access(13, bits);
+                    if (cpu_mod != 3)
+                        wait(3, 0);
                     set_ea(get_reg(cpu_reg));
                     break;
                 case 0x8A:
@@ -2630,18 +2919,17 @@ execx86(int cycs)
                     /* MOV reg, rm */
                     bits = 8 << (opcode & 1);
                     do_mod_rm();
-                    access(50, bits);
                     set_reg(cpu_reg, get_ea());
                     wait(1, 0);
                     if (cpu_mod != 3)
-                        wait(2, 0);
+                        wait(1, 0);
                     break;
 
                 case 0x8C: /*MOV w,sreg*/
                     do_mod_rm();
-                    if (cpu_mod == 3)
-                        wait(1, 0);
-                    access(14, 16);
+                    wait(1, 0);
+                    if (cpu_mod != 3)
+                        wait(2, 0);
                     seteaw(_opseg[(rmdat & 0x18) >> 3]->seg);
                     break;
 
@@ -2650,12 +2938,11 @@ execx86(int cycs)
                     cpu_state.regs[cpu_reg].w = cpu_state.eaaddr;
                     wait(1, 0);
                     if (cpu_mod != 3)
-                        wait(2, 0);
+                        wait(1, 0);
                     break;
 
                 case 0x8E: /*MOV sreg,w*/
                     do_mod_rm();
-                    access(51, 16);
                     tempw = geteaw();
                     if ((rmdat & 0x18) == 0x08) {
                         load_cs(tempw);
@@ -2664,22 +2951,22 @@ execx86(int cycs)
                         load_seg(tempw, _opseg[(rmdat & 0x18) >> 3]);
                     wait(1, 0);
                     if (cpu_mod != 3)
-                        wait(2, 0);
+                        wait(1, 0);
                     if (((rmdat & 0x18) >> 3) == 2)
                         noint = 1;
                     break;
 
                 case 0x8F: /*POPW*/
                     do_mod_rm();
-                    wait(1, 0);
+                    wait(2, 0);
                     cpu_src = cpu_state.eaaddr;
-                    access(24, 16);
+                    if (cpu_mod != 3)
+                        wait(1, 0);
+                    wait(1, 0);
                     if (cpu_mod != 3)
                         wait(2, 0);
                     cpu_data         = pop();
                     cpu_state.eaaddr = cpu_src;
-                    wait(2, 0);
-                    access(15, 16);
                     seteaw(cpu_data);
                     break;
 
@@ -2717,14 +3004,14 @@ execx86(int cycs)
                     new_ip = pfq_fetchw();
                     wait(1, 0);
                     new_cs = pfq_fetchw();
-                    pfq_clear();
-                    access(31, 16);
+                    wait(1, 0);
+                    pfq_suspend();
                     push(&(CS));
-                    access(60, 16);
+                    wait(4, 0);
                     cpu_state.oldpc = cpu_state.pc;
                     load_cs(new_cs);
                     set_ip(new_ip);
-                    access(32, 16);
+                    wait(1, 0);
                     push((uint16_t *) &(cpu_state.oldpc));
                     break;
                 case 0x9B: /*WAIT*/
@@ -2746,7 +3033,7 @@ execx86(int cycs)
 #endif
                     break;
                 case 0x9C: /*PUSHF*/
-                    access(33, 16);
+                    wait(4, 0);
                     if (is_nec)
                         tempw = (cpu_state.flags & 0x8fd7) | 0x7000;
                     else
@@ -2754,12 +3041,11 @@ execx86(int cycs)
                     push(&tempw);
                     break;
                 case 0x9D: /*POPF*/
-                    access(25, 16);
+                    wait(1, 0);
                     if (is_nec)
                         cpu_state.flags = pop() | 0x8002;
                     else
                         cpu_state.flags = pop() | 0x0002;
-                    wait(1, 0);
                     break;
                 case 0x9E: /*SAHF*/
                     wait(1, 0);
@@ -2775,20 +3061,18 @@ execx86(int cycs)
                 case 0xA1:
                     /* MOV A, [iw] */
                     bits = 8 << (opcode & 1);
-                    wait(1, 0);
+                    wait(2, 0);
                     cpu_state.eaaddr = pfq_fetchw();
-                    access(1, bits);
                     set_accum(bits, readmem((ovr_seg ? *ovr_seg : ds)));
-                    wait(1, 0);
                     break;
                 case 0xA2:
                 case 0xA3:
                     /* MOV [iw], A */
                     bits = 8 << (opcode & 1);
-                    wait(1, 0);
+                    wait(2, 0);
                     cpu_state.eaaddr = pfq_fetchw();
-                    access(7, bits);
                     writemem((ovr_seg ? *ovr_seg : ds), get_accum(bits));
+                    wait(2, 0);
                     break;
 
                 case 0xA4:
@@ -2796,37 +3080,33 @@ execx86(int cycs)
                 case 0xAC:
                 case 0xAD: /* LODS */
                     bits = 8 << (opcode & 1);
-                    if (!repeating) {
-                        wait(1, 0);
-                        if ((opcode & 8) == 0 && in_rep != 0)
-                            wait(1, 0);
-                    }
-                    if (rep_action(bits)) {
-                        wait(1, 0);
-                        if ((opcode & 8) != 0)
-                            wait(1, 0);
+                    if (rep_setup())
                         break;
-                    }
-                    if (in_rep != 0 && (opcode & 8) != 0)
+                    if (in_rep != 0 && (BUS_CYCLE == BUS_T4))
                         wait(1, 0);
-                    access(20, bits);
                     lods(bits);
                     if ((opcode & 8) == 0) {
-                        access(27, bits);
+                        wait(1, 0);
                         stos(bits);
-                    } else {
+                    } else
                         set_accum(bits, cpu_data);
-                        if (in_rep != 0)
+                    wait(3, 0);
+                    if (in_rep == 0)
+                        break;
+                    --tmpc;
+                    if (rep_interrupt())
+                        break;
+                    CX = tmpc;
+                    if (tmpc == 0) {
+                        completed = 1;
+                        wait(1, 0);
+                        if ((opcode & 8) != 0)
+                            wait(2, 0);
+                    } else {
+                        wait(2, 0);
+                        if ((opcode & 8) != 0)
                             wait(2, 0);
                     }
-                    if (in_rep == 0) {
-                        wait(3, 0);
-                        if ((opcode & 8) != 0)
-                            wait(1, 0);
-                        break;
-                    }
-                    repeating = 1;
-                    clock_end();
                     break;
 
                 case 0xA6:
@@ -2834,40 +3114,39 @@ execx86(int cycs)
                 case 0xAE:
                 case 0xAF: /* SCAS */
                     bits = 8 << (opcode & 1);
-                    if (!repeating)
-                        wait(1, 0);
-                    if (rep_action(bits)) {
-                        wait(2, 0);
+                    if (rep_setup())
                         break;
-                    }
-                    if (in_rep != 0)
-                        wait(1, 0);
-                    wait(1, 0);
-                    cpu_dest = get_accum(bits);
+                    tmpa = AX;
                     if ((opcode & 8) == 0) {
-                        access(21, bits);
-                        lods(bits);
                         wait(1, 0);
-                        cpu_dest = cpu_data;
+                        lods(bits);
+                        tmpa = cpu_data;
                     }
-                    access(2, bits);
+                    wait(2, 0);
                     cpu_state.eaaddr = DI;
                     cpu_data         = readmem(es);
-                    DI               = string_increment(bits);
+                    DI = string_increment(bits);
                     cpu_src          = cpu_data;
+                    cpu_dest         = tmpa;
                     sub(bits);
                     wait(2, 0);
                     if (in_rep == 0) {
+                        wait(2, 0);
+                        break;
+                    }
+                    --tmpc;
+                    CX = tmpc;
+                    if ((!!(cpu_state.flags & (rep_c_flag ? C_FLAG : Z_FLAG))) == (in_rep == 1)) {
                         wait(3, 0);
                         break;
                     }
-                    if ((!!(cpu_state.flags & (rep_c_flag ? C_FLAG : Z_FLAG))) == (in_rep == 1)) {
-                        completed = 1;
-                        wait(4, 0);
+                    if (rep_interrupt())
                         break;
-                    }
-                    repeating = 1;
-                    clock_end();
+                    wait(4, 0);
+                    if (tmpc == 0)
+                        completed = 1;
+                    else
+                        wait(1, 0);
                     break;
 
                 case 0xA8:
@@ -2883,24 +3162,24 @@ execx86(int cycs)
                 case 0xAA:
                 case 0xAB: /* STOS */
                     bits = 8 << (opcode & 1);
-                    if (!repeating) {
-                        wait(1, 0);
-                        if (in_rep != 0)
-                            wait(1, 0);
-                    }
-                    if (rep_action(bits)) {
-                        wait(1, 0);
+                    if (rep_setup())
                         break;
-                    }
                     cpu_data = AX;
-                    access(28, bits);
+                    if (in_rep == 0 && (BUS_CYCLE == BUS_T4))
+                        wait(1, 0);
                     stos(bits);
-                    if (in_rep == 0) {
-                        wait(3, 0);
+                    wait(3, 0);
+                    if (in_rep == 0)
                         break;
-                    }
-                    repeating = 1;
-                    clock_end();
+                    --tmpc;
+                    if (rep_interrupt())
+                        break;
+                    CX = tmpc;
+                    if (tmpc == 0) {
+                        completed = 1;
+                        wait(1, 0);
+                    } else
+                        wait(2, 0);
                     break;
 
                 case 0xB0:
@@ -2942,32 +3221,27 @@ execx86(int cycs)
                 case 0xCB:
                     /* RET */
                     bits = 8 + (opcode & 0x08);
-                    if ((opcode & 9) != 1)
-                        wait(1, 0);
+                    wait(1, 0);
                     if (!(opcode & 1)) {
                         cpu_src = pfq_fetchw();
-                        wait(1, 0);
+                        wait(2, 0);
                     }
                     if ((opcode & 9) == 9)
-                        wait(1, 0);
+                        wait(2, 0);
                     pfq_clear();
-                    access(26, bits);
                     new_ip = pop();
-                    wait(2, 0);
-                    if ((opcode & 8) == 0)
+                    wait(1, 0);
+                    if ((opcode & 8) == 0) {
                         new_cs = CS;
-                    else {
-                        access(42, bits);
-                        new_cs = pop();
                         if (opcode & 1)
                             wait(1, 0);
+                    } else {
+                        wait(2, 0);
+                        new_cs = pop();
                     }
-                    if (!(opcode & 1)) {
+                    if (!(opcode & 1))
                         SP += cpu_src;
-                        wait(1, 0);
-                    }
                     load_cs(new_cs);
-                    access(72, bits);
                     set_ip(new_ip);
                     break;
 
@@ -2976,13 +3250,12 @@ execx86(int cycs)
                     /* LsS rw, rmd */
                     do_mod_rm();
                     bits = 16;
-                    access(52, bits);
                     read_ea(1, bits);
                     cpu_state.regs[cpu_reg].w = cpu_data;
-                    access(57, bits);
+                    if (cpu_mod != 3)
+                        wait(2, 0);
                     read_ea2(bits);
                     load_seg(cpu_data, (opcode & 0x01) ? &cpu_state.seg_ds : &cpu_state.seg_es);
-                    wait(1, 0);
                     break;
 
                 case 0xC6:
@@ -2991,40 +3264,40 @@ execx86(int cycs)
                     bits = 8 << (opcode & 1);
                     do_mod_rm();
                     wait(1, 0);
-                    if (cpu_mod != 3)
-                        wait(2, 0);
                     cpu_data = pfq_fetch();
-                    if (cpu_mod == 3)
-                        wait(1, 0);
-                    access(16, bits);
+                    wait(2, 0);
                     set_ea(cpu_data);
                     break;
 
                 case 0xCC: /*INT 3*/
+                    wait(7, 0);
                     interrupt(3);
                     break;
                 case 0xCD: /*INT*/
                     wait(1, 0);
-                    interrupt(pfq_fetchb());
+                    temp = pfq_fetchb();
+                    wait(1, 0);
+                    if (BUS_CYCLE != BUS_T4)
+                        wait(1, 0);
+                    wait(1, 0);
+
+                    interrupt(temp);
                     break;
                 case 0xCE: /*INTO*/
                     wait(3, 0);
                     if (cpu_state.flags & V_FLAG) {
-                        wait(2, 0);
+                        wait(5, 0);
                         interrupt(4);
                     }
                     break;
 
                 case 0xCF: /*IRET*/
-                    access(43, 8);
-                    new_ip = pop();
                     wait(3, 0);
-                    access(44, 8);
+                    pfq_clear();
+                    new_ip = pop();
                     new_cs = pop();
                     load_cs(new_cs);
-                    access(62, 8);
                     set_ip(new_ip);
-                    access(45, 8);
                     if (is_nec)
                         cpu_state.flags = pop() | 0x8002;
                     else
@@ -3041,10 +3314,9 @@ execx86(int cycs)
                     /* rot rm */
                     bits = 8 << (opcode & 1);
                     do_mod_rm();
+                    cpu_data = get_ea();
                     if (cpu_mod == 3)
                         wait(1, 0);
-                    access(53, bits);
-                    cpu_data = get_ea();
                     if ((opcode & 2) == 0) {
                         cpu_src = 1;
                         wait((cpu_mod != 3) ? 4 : 0, 0);
@@ -3125,7 +3397,6 @@ execx86(int cycs)
                             wait(4, 0);
                         --cpu_src;
                     }
-                    access(17, bits);
                     set_ea(cpu_data);
                     break;
 
@@ -3159,9 +3430,8 @@ execx86(int cycs)
                     break;
                 case 0xD7: /*XLATB*/
                     cpu_state.eaaddr = (BX + AL) & 0xffff;
-                    access(4, 8);
-                    AL = readmemb((ovr_seg ? *ovr_seg : ds) + cpu_state.eaaddr);
-                    wait(1, 0);
+                    wait(4, 0);
+                    AL = readmemb((ovr_seg ? *ovr_seg : ds), cpu_state.eaaddr);
                     break;
 
                 case 0xD8:
@@ -3174,11 +3444,12 @@ execx86(int cycs)
                 case 0xDF:
                     /* esc i, r, rm */
                     do_mod_rm();
-                    access(54, 16);
                     tempw = cpu_state.pc;
-                    if (!hasfpu)
-                        geteaw();
-                    else {
+                    geteaw();
+                    wait(1, 0);
+                    if (cpu_mod != 3)
+                        wait(1, 0);
+                    if (hasfpu) {
                         if (fpu_softfloat) {
                             switch (opcode) {
                                 case 0xD8:
@@ -3237,9 +3508,6 @@ execx86(int cycs)
                     }
                     cpu_state.pc = tempw; /* Do this as the x87 code advances it, which is needed on
                                              the 286+ core, but not here. */
-                    wait(1, 0);
-                    if (cpu_mod != 3)
-                        wait(2, 0);
                     break;
 
                 case 0xE0:
@@ -3272,43 +3540,46 @@ execx86(int cycs)
 
                 case 0xE4:
                 case 0xE5:
+                    bits = 8 << (opcode & 1);
+                    wait(1, 0);
+                    cpu_data = pfq_fetchb();
+                    cpu_state.eaaddr = cpu_data;
+                    wait(1, 0);
+                    cpu_io(bits, 0, cpu_state.eaaddr);
+                    break;
                 case 0xE6:
                 case 0xE7:
+                    bits = 8 << (opcode & 1);
+                    wait(1, 0);
+                    cpu_data = pfq_fetchb();
+                    cpu_state.eaaddr = cpu_data;
+                    cpu_data = (bits == 16) ? AX : AL;
+                    wait(2, 0);
+                    cpu_io(bits, 1, cpu_state.eaaddr);
+                    break;
                 case 0xEC:
                 case 0xED:
+                    bits = 8 << (opcode & 1);
+                    cpu_data = DX;
+                    cpu_state.eaaddr = cpu_data;
+                    wait(1, 0);
+                    cpu_io(bits, 0, cpu_state.eaaddr);
+                    break;
                 case 0xEE:
                 case 0xEF:
                     bits = 8 << (opcode & 1);
-                    if ((opcode & 0x0e) != 0x0c)
-                        wait(1, 0);
-                    if ((opcode & 8) == 0)
-                        cpu_data = pfq_fetchb();
-                    else
-                        cpu_data = DX;
+                    wait(2, 0);
+                    cpu_data = DX;
                     cpu_state.eaaddr = cpu_data;
-                    if ((opcode & 2) == 0) {
-                        access(3, bits);
-                        if (opcode & 1)
-                            cpu_io(16, 0, cpu_data);
-                        else
-                            cpu_io(8, 0, cpu_data);
-                        wait(1, 0);
-                    } else {
-                        if ((opcode & 8) == 0)
-                            access(8, bits);
-                        else
-                            access(9, bits);
-                        if (opcode & 1)
-                            cpu_io(16, 1, cpu_data);
-                        else
-                            cpu_io(8, 1, cpu_data);
-                    }
+                    cpu_data = (bits == 16) ? AX : AL;
+                    cpu_io(bits, 1, cpu_state.eaaddr);
+                    wait(1, 0);
                     break;
 
                 case 0xE8: /*CALL rel 16*/
                     wait(1, 0);
                     cpu_state.oldpc = jump_near();
-                    access(34, 8);
+                    wait(2, 0);
                     push((uint16_t *) &(cpu_state.oldpc));
                     break;
                 case 0xE9: /*JMP rel 16*/
@@ -3321,8 +3592,8 @@ execx86(int cycs)
                     wait(1, 0);
                     tempw = pfq_fetchw();
                     load_cs(tempw);
-                    access(70, 8);
                     pfq_clear();
+                    wait(4, 0);
                     set_ip(addr);
                     break;
                 case 0xEB: /*JMP rel*/
@@ -3349,17 +3620,22 @@ execx86(int cycs)
 
                 case 0xF4: /*HLT*/
                     if (!repeating) {
-                        wait(1, 0);
+                        if ((BUS_CYCLE == BUS_T4) || !last_was_code)
+                            cpu_data = 1;
+                        else
+                            cpu_data = 2;
+                        wait(2, 0);
                         pfq_clear();
                     }
                     wait(1, 0);
                     if (irq_pending()) {
-                        wait(cycles & 1, 0);
+                        wait(1, 0);
+                        if (cpu_data == 2)
+                            wait(1, 0);
                         check_interrupts();
                     } else {
                         repeating = 1;
                         completed = 0;
-                        clock_end();
                     }
                     break;
                 case 0xF5: /*CMC*/
@@ -3371,15 +3647,12 @@ execx86(int cycs)
                 case 0xF7:
                     bits = 8 << (opcode & 1);
                     do_mod_rm();
-                    access(55, bits);
                     cpu_data = get_ea();
                     switch (rmdat & 0x38) {
                         case 0x00:
                         case 0x08:
                             /* TEST */
                             wait(2, 0);
-                            if (cpu_mod != 3)
-                                wait(1, 0);
                             cpu_src = pfq_fetch();
                             wait(1, 0);
                             test(bits, cpu_data, cpu_src);
@@ -3396,7 +3669,8 @@ execx86(int cycs)
                                 cpu_dest = 0;
                                 sub(bits);
                             }
-                            access(18, bits);
+                            if (cpu_mod != 3)
+                                wait(2, 0);
                             set_ea(cpu_data);
                             break;
                         case 0x20: /* MUL */
@@ -3418,8 +3692,6 @@ execx86(int cycs)
                             }
                             set_sf(bits);
                             set_pf();
-                            if (cpu_mod != 3)
-                                wait(1, 0);
                             /* NOTE: When implementing the V20, care should be taken to not change
                                      the zero flag. */
                             if (is_nec)
@@ -3427,8 +3699,6 @@ execx86(int cycs)
                             break;
                         case 0x30: /* DIV */
                         case 0x38: /* IDIV */
-                            if (cpu_mod != 3)
-                                wait(1, 0);
                             cpu_src = cpu_data;
                             if (x86_div(AL, AH))
                                 wait(1, 0);
@@ -3460,7 +3730,6 @@ execx86(int cycs)
                     /* misc */
                     bits = 8 << (opcode & 1);
                     do_mod_rm();
-                    access(56, bits);
                     read_ea(((rmdat & 0x38) == 0x18) || ((rmdat & 0x38) == 0x28), bits);
                     switch (rmdat & 0x38) {
                         case 0x00: /* INC rm */
@@ -3477,62 +3746,60 @@ execx86(int cycs)
                             do_af();
                             set_pzs(bits);
                             wait(2, 0);
-                            access(19, bits);
                             set_ea(cpu_data);
                             break;
                         case 0x10: /* CALL rm */
                             cpu_data_opff_rm();
-                            access(63, bits);
-                            wait(1, 0);
+                            wait(2, 0);
                             pfq_clear();
-                            wait(4, 0);
-                            if (cpu_mod != 3)
-                                wait(1, 0);
-                            wait(1, 0); /* Wait. */
+                            wait(5, 0);
                             cpu_state.oldpc = cpu_state.pc;
                             set_ip(cpu_data);
                             wait(2, 0);
-                            access(35, bits);
                             push((uint16_t *) &(cpu_state.oldpc));
                             break;
                         case 0x18: /* CALL rmd */
                             new_ip = cpu_data;
-                            access(58, bits);
+                            wait(3, 0);
                             read_ea2(bits);
                             if (!(opcode & 1))
                                 cpu_data |= 0xff00;
                             new_cs = cpu_data;
-                            access(36, bits);
+                            wait(1, 0);
+                            pfq_clear();
                             push(&(CS));
-                            access(64, bits);
                             wait(4, 0);
                             cpu_state.oldpc = cpu_state.pc;
                             load_cs(new_cs);
                             set_ip(new_ip);
-                            access(37, bits);
+                            wait(1, 0);
                             push((uint16_t *) &(cpu_state.oldpc));
                             break;
                         case 0x20: /* JMP rm */
                             cpu_data_opff_rm();
-                            access(65, bits);
+                            wait(2, 0);
+                            pfq_clear();
+                            if (BUS_CYCLE != BUS_T4)
+                                wait(1, 0);
                             set_ip(cpu_data);
                             break;
                         case 0x28: /* JMP rmd */
                             new_ip = cpu_data;
-                            access(59, bits);
+                            wait(3, 0);
+                            pfq_clear();
+                            wait(1, 0);
                             read_ea2(bits);
                             if (!(opcode & 1))
                                 cpu_data |= 0xff00;
                             new_cs = cpu_data;
                             load_cs(new_cs);
-                            access(66, bits);
                             set_ip(new_ip);
                             break;
                         case 0x30: /* PUSH rm */
                         case 0x38:
                             if (cpu_mod != 3)
                                 wait(1, 0);
-                            access(38, bits);
+                            wait(4, 0);
                             push((uint16_t *) &(cpu_data));
                             break;
                     }
diff --git a/src/cpu/808x/CMakeLists.txt b/src/cpu/808x/CMakeLists.txt
new file mode 100644
index 000000000..d29bdf4e0
--- /dev/null
+++ b/src/cpu/808x/CMakeLists.txt
@@ -0,0 +1,16 @@
+#
+# 86Box    A hypervisor and IBM PC system emulator that specializes in
+#          running old operating systems and software designed for IBM
+#          PC systems and compatibles from 1981 through fairly recent
+#          system designs based on the PCI bus.
+#
+#          This file is part of the 86Box distribution.
+#
+#          CMake build script.
+#
+# Authors: David Hrdlička, <hrdlickadavid@outlook.com>
+#
+#          Copyright 2020-2021 David Hrdlička.
+#
+
+add_library(808x OBJECT queue.c)
diff --git a/src/cpu/808x/queue.c b/src/cpu/808x/queue.c
new file mode 100644
index 000000000..2eebde0ce
--- /dev/null
+++ b/src/cpu/808x/queue.c
@@ -0,0 +1,190 @@
+/*
+ * 86Box    A hypervisor and IBM PC system emulator that specializes in
+ *          running old operating systems and software designed for IBM
+ *          PC systems and compatibles from 1981 through fairly recent
+ *          system designs based on the PCI bus.
+ *
+ *          This file is part of the 86Box distribution.
+ *
+ *          808x CPU emulation, mostly ported from reenigne's XTCE, which
+ *          is cycle-accurate.
+ *
+ * Authors: gloriouscow, <https://github.com/dbalsom>
+ *          Miran Grca, <mgrca8@gmail.com>
+ *
+ *          Copyright 2023 gloriouscow.
+ *          Copyright 2023 Miran Grca.
+ */
+#include <math.h>
+#include <stdarg.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <string.h>
+#include <wchar.h>
+
+#define HAVE_STDARG_H
+#include <86box/86box.h>
+#include "cpu.h"
+#include "x86.h"
+#include <86box/machine.h>
+#include <86box/io.h>
+#include <86box/mem.h>
+#include <86box/rom.h>
+#include <86box/nmi.h>
+#include <86box/pic.h>
+#include <86box/ppi.h>
+#include <86box/timer.h>
+#include <86box/gdbstub.h>
+// #include "808x.h"
+#include "queue.h"
+
+/* TODO: Move to cpu.h so this can eventually be reused for 286+ as well. */
+#define QUEUE_MAX 6
+
+typedef struct queue_t
+{
+    size_t        size;
+    size_t        len;
+    size_t        back;
+    size_t        front;
+    uint8_t       q[QUEUE_MAX];
+    uint16_t      preload;
+    queue_delay_t delay;
+} queue_t;
+
+static queue_t queue;
+
+#ifdef ENABLE_QUEUE_LOG
+int queue_do_log = ENABLE_QUEUE_LOG;
+
+static void
+queue_log(const char *fmt, ...)
+{
+    va_list ap;
+
+    if (queue_do_log) {
+        va_start(ap, fmt);
+        pclog_ex(fmt, ap);
+        va_end(ap);
+    }
+}
+#else
+#    define queue_log(fmt, ...)
+#endif
+
+void
+queue_set_size(size_t size)
+{
+    if (size > QUEUE_MAX)
+        fatal("Requested prefetch queue of %i bytes is too big\n", size);
+
+    queue.size = size;
+}
+
+size_t
+queue_get_len(void)
+{
+    return queue.len;
+}
+
+int
+queue_is_full(void)
+{
+    return (queue.len != queue.size);
+}
+
+uint16_t
+queue_get_preload(void)
+{
+    uint16_t ret = queue.preload;
+    queue.preload = 0x0000;
+
+    return ret;
+}
+
+int
+queue_has_preload(void)
+{
+    return (queue.preload & FLAG_PRELOADED) ? 1 : 0;
+}
+
+void
+queue_set_preload(void)
+{
+    uint8_t byte;
+
+    if (queue.len > 0) {
+        byte = queue_pop();
+        queue.preload = ((uint16_t) byte) | FLAG_PRELOADED;
+    } else
+        fatal("Tried to preload with empty queue\n");
+}
+
+void
+queue_push8(uint8_t byte)
+{
+    if (queue.len < queue.size) {
+        queue.q[queue.front] = byte;
+        queue.front = (queue.front + 1) % queue.size;
+        queue.len++;
+
+        if (queue.len == 3)
+            queue.delay = DELAY_WRITE;
+        else
+            queue.delay = DELAY_NONE;
+    } else
+        fatal("Queue overrun\n");
+}
+
+void
+queue_push16(uint16_t word)
+{
+    queue_push8((uint8_t) (word & 0xff));
+    queue_push8((uint8_t) ((word >> 8) & 0xff));
+}
+
+uint8_t
+queue_pop(void)
+{
+    uint8_t byte = 0xff;
+
+    if (queue.len > 0) {
+        byte = queue.q[queue.back];
+
+        queue.back = (queue.back + 1) % queue.size;
+        queue.len--;
+
+        if (queue.len >= 3)
+            queue.delay = DELAY_READ;
+        else
+            queue.delay = DELAY_NONE;
+    } else
+        fatal("Queue underrun\n");
+
+    return byte;
+}
+
+queue_delay_t
+queue_get_delay(void)
+{
+    return queue.delay;
+}
+
+void
+queue_flush(void)
+{
+    memset(&queue, 0x00, sizeof(queue_t));
+
+    queue.delay = DELAY_NONE;
+}
+
+void
+queue_init(void)
+{
+    queue_flush();
+
+    if (is8086)
+        queue_set_size(6);
+    else
+        queue_set_size(4);
+}
diff --git a/src/cpu/808x/queue.h b/src/cpu/808x/queue.h
new file mode 100644
index 000000000..544455784
--- /dev/null
+++ b/src/cpu/808x/queue.h
@@ -0,0 +1,43 @@
+/*
+ * 86Box    A hypervisor and IBM PC system emulator that specializes in
+ *          running old operating systems and software designed for IBM
+ *          PC systems and compatibles from 1981 through fairly recent
+ *          system designs based on the PCI bus.
+ *
+ *          This file is part of the 86Box distribution.
+ *
+ *          Prefetch queue implementation header.
+ *
+ * Authors: gloriouscow, <https://github.com/dbalsom>
+ *          Miran Grca, <mgrca8@gmail.com>
+ *
+ *          Copyright 2023 gloriouscow.
+ *          Copyright 2023 Miran Grca.
+ */
+#ifndef EMU_QUEUE_H
+#define EMU_QUEUE_H
+
+typedef enum queue_delay_t
+{
+    DELAY_READ,
+    DELAY_WRITE,
+    DELAY_NONE
+} queue_delay_t;
+
+#define FLAG_PRELOADED    0x8000
+
+extern void               queue_set_size(size_t size);
+extern size_t             queue_get_len(void);
+extern int                queue_is_full(void);
+extern uint16_t           queue_get_preload(void);
+extern int                queue_has_preload(void);
+extern void               queue_set_preload(void);
+extern void               queue_push8(uint8_t byte);
+extern void               queue_push16(uint16_t word);
+extern uint8_t            queue_pop(void);
+extern queue_delay_t      queue_get_delay(void);
+extern void               queue_flush(void);
+
+extern void               queue_init(void);
+
+#endif /*EMU_QUEUE_H*/
diff --git a/src/cpu/CMakeLists.txt b/src/cpu/CMakeLists.txt
index 18aa06023..e4d8e71b2 100644
--- a/src/cpu/CMakeLists.txt
+++ b/src/cpu/CMakeLists.txt
@@ -14,7 +14,7 @@
 #
 
 add_library(cpu OBJECT cpu.c cpu_table.c fpu.c x86.c 808x.c 386.c 386_common.c
-    386_dynarec.c x86seg.c x87.c x87_timings.c 8080.c)
+    386_dynarec.c x86_ops_mmx.c x86seg.c x87.c x87_timings.c 8080.c)
 
 if(AMD_K5)
     target_compile_definitions(cpu PRIVATE USE_AMD_K5)
@@ -35,3 +35,6 @@ endif()
 
 add_subdirectory(softfloat)
 target_link_libraries(86Box softfloat)
+
+add_subdirectory(808x)
+target_link_libraries(86Box 808x)
diff --git a/src/cpu/cpu.c b/src/cpu/cpu.c
index 257d23845..d633b9bb2 100644
--- a/src/cpu/cpu.c
+++ b/src/cpu/cpu.c
@@ -1646,6 +1646,7 @@ cpu_set(void)
         cpu_exec = exec386;
     else
         cpu_exec = execx86;
+    mmx_init();
     gdbstub_cpu_init();
 }
 
diff --git a/src/cpu/cpu.h b/src/cpu/cpu.h
index 692600005..3d6d0622a 100644
--- a/src/cpu/cpu.h
+++ b/src/cpu/cpu.h
@@ -850,4 +850,9 @@ extern void cpu_fast_off_reset(void);
 extern void smi_raise(void);
 extern void nmi_raise(void);
 
+extern MMX_REG *MMP[8];
+extern uint16_t *MMEP[8];
+
+extern void mmx_init(void);
+
 #endif /*EMU_CPU_H*/
diff --git a/src/cpu/x86_ops_3dnow.h b/src/cpu/x86_ops_3dnow.h
index eb7a35ace..ff657d708 100644
--- a/src/cpu/x86_ops_3dnow.h
+++ b/src/cpu/x86_ops_3dnow.h
@@ -36,17 +36,20 @@ static int
 opPAVGUSB(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst = MMX_GETREGP(cpu_reg);
 
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].b[0] = (cpu_state.MM[cpu_reg].b[0] + src.b[0] + 1) >> 1;
-    cpu_state.MM[cpu_reg].b[1] = (cpu_state.MM[cpu_reg].b[1] + src.b[1] + 1) >> 1;
-    cpu_state.MM[cpu_reg].b[2] = (cpu_state.MM[cpu_reg].b[2] + src.b[2] + 1) >> 1;
-    cpu_state.MM[cpu_reg].b[3] = (cpu_state.MM[cpu_reg].b[3] + src.b[3] + 1) >> 1;
-    cpu_state.MM[cpu_reg].b[4] = (cpu_state.MM[cpu_reg].b[4] + src.b[4] + 1) >> 1;
-    cpu_state.MM[cpu_reg].b[5] = (cpu_state.MM[cpu_reg].b[5] + src.b[5] + 1) >> 1;
-    cpu_state.MM[cpu_reg].b[6] = (cpu_state.MM[cpu_reg].b[6] + src.b[6] + 1) >> 1;
-    cpu_state.MM[cpu_reg].b[7] = (cpu_state.MM[cpu_reg].b[7] + src.b[7] + 1) >> 1;
+    dst->b[0] = (dst->b[0] + src.b[0] + 1) >> 1;
+    dst->b[1] = (dst->b[1] + src.b[1] + 1) >> 1;
+    dst->b[2] = (dst->b[2] + src.b[2] + 1) >> 1;
+    dst->b[3] = (dst->b[3] + src.b[3] + 1) >> 1;
+    dst->b[4] = (dst->b[4] + src.b[4] + 1) >> 1;
+    dst->b[5] = (dst->b[5] + src.b[5] + 1) >> 1;
+    dst->b[6] = (dst->b[6] + src.b[6] + 1) >> 1;
+    dst->b[7] = (dst->b[7] + src.b[7] + 1) >> 1;
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -54,11 +57,14 @@ static int
 opPF2ID(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst = MMX_GETREGP(cpu_reg);
 
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].sl[0] = (int32_t) src.f[0];
-    cpu_state.MM[cpu_reg].sl[1] = (int32_t) src.f[1];
+    dst->sl[0] = (int32_t) src.f[0];
+    dst->sl[1] = (int32_t) src.f[1];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -66,11 +72,14 @@ static int
 opPF2IW(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst = MMX_GETREGP(cpu_reg);
 
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].sw[0] = (int32_t) src.f[0];
-    cpu_state.MM[cpu_reg].sw[1] = (int32_t) src.f[1];
+    dst->sw[0] = (int32_t) src.f[0];
+    dst->sw[1] = (int32_t) src.f[1];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -78,13 +87,16 @@ static int
 opPFACC(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst = MMX_GETREGP(cpu_reg);
     float   tempf;
 
     MMX_GETSRC();
 
-    tempf                      = cpu_state.MM[cpu_reg].f[0] + cpu_state.MM[cpu_reg].f[1];
-    cpu_state.MM[cpu_reg].f[1] = src.f[0] + src.f[1];
-    cpu_state.MM[cpu_reg].f[0] = tempf;
+    tempf     = dst->f[0] + dst->f[1];
+    dst->f[1] = src.f[0] + src.f[1];
+    dst->f[0] = tempf;
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -92,13 +104,16 @@ static int
 opPFNACC(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst = MMX_GETREGP(cpu_reg);
     float   tempf;
 
     MMX_GETSRC();
 
-    tempf                      = cpu_state.MM[cpu_reg].f[0] - cpu_state.MM[cpu_reg].f[1];
-    cpu_state.MM[cpu_reg].f[1] = src.f[0] - src.f[1];
-    cpu_state.MM[cpu_reg].f[0] = tempf;
+    tempf     = dst->f[0] - dst->f[1];
+    dst->f[1] = src.f[0] - src.f[1];
+    dst->f[0] = tempf;
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -106,13 +121,16 @@ static int
 opPFPNACC(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst = MMX_GETREGP(cpu_reg);
     float   tempf;
 
     MMX_GETSRC();
 
-    tempf                      = cpu_state.MM[cpu_reg].f[0] - cpu_state.MM[cpu_reg].f[1];
-    cpu_state.MM[cpu_reg].f[1] = src.f[0] + src.f[1];
-    cpu_state.MM[cpu_reg].f[0] = tempf;
+    tempf     = dst->f[0] - dst->f[1];
+    dst->f[1] = src.f[0] + src.f[1];
+    dst->f[0] = tempf;
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -120,15 +138,18 @@ static int
 opPSWAPD(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst = MMX_GETREGP(cpu_reg);
     float   tempf, tempf2;
 
     MMX_GETSRC();
 
     /* We have to do this in case source and destination overlap. */
-    tempf                      = src.f[0];
-    tempf2                     = src.f[1];
-    cpu_state.MM[cpu_reg].f[1] = tempf;
-    cpu_state.MM[cpu_reg].f[0] = tempf2;
+    tempf     = src.f[0];
+    tempf2    = src.f[1];
+    dst->f[1] = tempf;
+    dst->f[0] = tempf2;
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -136,11 +157,14 @@ static int
 opPFADD(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst = MMX_GETREGP(cpu_reg);
 
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].f[0] += src.f[0];
-    cpu_state.MM[cpu_reg].f[1] += src.f[1];
+    dst->f[0] += src.f[0];
+    dst->f[1] += src.f[1];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -148,11 +172,14 @@ static int
 opPFCMPEQ(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst = MMX_GETREGP(cpu_reg);
 
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].l[0] = (cpu_state.MM[cpu_reg].f[0] == src.f[0]) ? 0xffffffff : 0;
-    cpu_state.MM[cpu_reg].l[1] = (cpu_state.MM[cpu_reg].f[1] == src.f[1]) ? 0xffffffff : 0;
+    dst->l[0] = (dst->f[0] == src.f[0]) ? 0xffffffff : 0;
+    dst->l[1] = (dst->f[1] == src.f[1]) ? 0xffffffff : 0;
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -160,11 +187,14 @@ static int
 opPFCMPGE(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst = MMX_GETREGP(cpu_reg);
 
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].l[0] = (cpu_state.MM[cpu_reg].f[0] >= src.f[0]) ? 0xffffffff : 0;
-    cpu_state.MM[cpu_reg].l[1] = (cpu_state.MM[cpu_reg].f[1] >= src.f[1]) ? 0xffffffff : 0;
+    dst->l[0] = (dst->f[0] >= src.f[0]) ? 0xffffffff : 0;
+    dst->l[1] = (dst->f[1] >= src.f[1]) ? 0xffffffff : 0;
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -172,11 +202,14 @@ static int
 opPFCMPGT(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst = MMX_GETREGP(cpu_reg);
 
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].l[0] = (cpu_state.MM[cpu_reg].f[0] > src.f[0]) ? 0xffffffff : 0;
-    cpu_state.MM[cpu_reg].l[1] = (cpu_state.MM[cpu_reg].f[1] > src.f[1]) ? 0xffffffff : 0;
+    dst->l[0] = (dst->f[0] > src.f[0]) ? 0xffffffff : 0;
+    dst->l[1] = (dst->f[1] > src.f[1]) ? 0xffffffff : 0;
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -184,13 +217,16 @@ static int
 opPFMAX(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst = MMX_GETREGP(cpu_reg);
 
     MMX_GETSRC();
 
-    if (src.f[0] > cpu_state.MM[cpu_reg].f[0])
-        cpu_state.MM[cpu_reg].f[0] = src.f[0];
-    if (src.f[1] > cpu_state.MM[cpu_reg].f[1])
-        cpu_state.MM[cpu_reg].f[1] = src.f[1];
+    if (src.f[0] > dst->f[0])
+        dst->f[0] = src.f[0];
+    if (src.f[1] > dst->f[1])
+        dst->f[1] = src.f[1];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -198,13 +234,16 @@ static int
 opPFMIN(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst = MMX_GETREGP(cpu_reg);
 
     MMX_GETSRC();
 
-    if (src.f[0] < cpu_state.MM[cpu_reg].f[0])
-        cpu_state.MM[cpu_reg].f[0] = src.f[0];
-    if (src.f[1] < cpu_state.MM[cpu_reg].f[1])
-        cpu_state.MM[cpu_reg].f[1] = src.f[1];
+    if (src.f[0] < dst->f[0])
+        dst->f[0] = src.f[0];
+    if (src.f[1] < dst->f[1])
+        dst->f[1] = src.f[1];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -212,24 +251,29 @@ static int
 opPFMUL(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst = MMX_GETREGP(cpu_reg);
 
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].f[0] *= src.f[0];
-    cpu_state.MM[cpu_reg].f[1] *= src.f[1];
+    dst->f[0] *= src.f[0];
+    dst->f[1] *= src.f[1];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
 static int
 opPFRCP(uint32_t fetchdat)
 {
+    MMX_REG *dst = MMX_GETREGP(cpu_reg);
+
     union {
         uint32_t i;
         float    f;
     } src;
 
     if (cpu_mod == 3) {
-        src.f = cpu_state.MM[cpu_rm].f[0];
+        src.f = (MMX_GETREG(cpu_rm)).f[0];
         CLOCK_CYCLES(1);
     } else {
         SEG_CHECK_READ(cpu_state.ea_seg);
@@ -239,8 +283,10 @@ opPFRCP(uint32_t fetchdat)
         CLOCK_CYCLES(2);
     }
 
-    cpu_state.MM[cpu_reg].f[0] = 1.0 / src.f;
-    cpu_state.MM[cpu_reg].f[1] = cpu_state.MM[cpu_reg].f[0];
+    dst->f[0] = 1.0 / src.f;
+    dst->f[1] = dst->f[0];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -249,11 +295,14 @@ static int
 opPFRCPIT1(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst = MMX_GETREGP(cpu_reg);
 
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].f[0] = src.f[0];
-    cpu_state.MM[cpu_reg].f[1] = src.f[1];
+    dst->f[0] = src.f[0];
+    dst->f[1] = src.f[1];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -261,24 +310,29 @@ static int
 opPFRCPIT2(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst = MMX_GETREGP(cpu_reg);
 
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].f[0] = src.f[0];
-    cpu_state.MM[cpu_reg].f[1] = src.f[1];
+    dst->f[0] = src.f[0];
+    dst->f[1] = src.f[1];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
 static int
 opPFRSQRT(uint32_t fetchdat)
 {
+    MMX_REG *dst = MMX_GETREGP(cpu_reg);
+
     union {
         uint32_t i;
         float    f;
     } src;
 
     if (cpu_mod == 3) {
-        src.f = cpu_state.MM[cpu_rm].f[0];
+        src.f = (MMX_GETREG(cpu_rm)).f[0];
         CLOCK_CYCLES(1);
     } else {
         SEG_CHECK_READ(cpu_state.ea_seg);
@@ -288,8 +342,10 @@ opPFRSQRT(uint32_t fetchdat)
         CLOCK_CYCLES(2);
     }
 
-    cpu_state.MM[cpu_reg].f[0] = 1.0 / sqrt(src.f);
-    cpu_state.MM[cpu_reg].f[1] = cpu_state.MM[cpu_reg].f[0];
+    dst->f[0] = 1.0 / sqrt(src.f);
+    dst->f[1] = dst->f[0];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -308,11 +364,14 @@ static int
 opPFSUB(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst = MMX_GETREGP(cpu_reg);
 
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].f[0] -= src.f[0];
-    cpu_state.MM[cpu_reg].f[1] -= src.f[1];
+    dst->f[0] -= src.f[0];
+    dst->f[1] -= src.f[1];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -320,11 +379,14 @@ static int
 opPFSUBR(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst = MMX_GETREGP(cpu_reg);
 
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].f[0] = src.f[0] - cpu_state.MM[cpu_reg].f[0];
-    cpu_state.MM[cpu_reg].f[1] = src.f[1] - cpu_state.MM[cpu_reg].f[1];
+    dst->f[0] = src.f[0] - dst->f[0];
+    dst->f[1] = src.f[1] - dst->f[1];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -332,11 +394,14 @@ static int
 opPI2FD(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst = MMX_GETREGP(cpu_reg);
 
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].f[0] = (float) src.sl[0];
-    cpu_state.MM[cpu_reg].f[1] = (float) src.sl[1];
+    dst->f[0] = (float) src.sl[0];
+    dst->f[1] = (float) src.sl[1];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -344,37 +409,46 @@ static int
 opPI2FW(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst = MMX_GETREGP(cpu_reg);
 
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].f[0] = (float) src.sw[0];
-    cpu_state.MM[cpu_reg].f[1] = (float) src.sw[1];
+    dst->f[0] = (float) src.sw[0];
+    dst->f[1] = (float) src.sw[1];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
 static int
 opPMULHRW(uint32_t fetchdat)
 {
+    MMX_REG src;
+    MMX_REG *dst = MMX_GETREGP(cpu_reg);
+
     if (cpu_mod == 3) {
-        cpu_state.MM[cpu_reg].w[0] = (((int32_t) cpu_state.MM[cpu_reg].sw[0] * (int32_t) cpu_state.MM[cpu_rm].sw[0]) + 0x8000) >> 16;
-        cpu_state.MM[cpu_reg].w[1] = (((int32_t) cpu_state.MM[cpu_reg].sw[1] * (int32_t) cpu_state.MM[cpu_rm].sw[1]) + 0x8000) >> 16;
-        cpu_state.MM[cpu_reg].w[2] = (((int32_t) cpu_state.MM[cpu_reg].sw[2] * (int32_t) cpu_state.MM[cpu_rm].sw[2]) + 0x8000) >> 16;
-        cpu_state.MM[cpu_reg].w[3] = (((int32_t) cpu_state.MM[cpu_reg].sw[3] * (int32_t) cpu_state.MM[cpu_rm].sw[3]) + 0x8000) >> 16;
+        src = MMX_GETREG(cpu_rm);
+
+        dst->w[0] = (((int32_t) dst->sw[0] * (int32_t) src.sw[0]) + 0x8000) >> 16;
+        dst->w[1] = (((int32_t) dst->sw[1] * (int32_t) src.sw[1]) + 0x8000) >> 16;
+        dst->w[2] = (((int32_t) dst->sw[2] * (int32_t) src.sw[2]) + 0x8000) >> 16;
+        dst->w[3] = (((int32_t) dst->sw[3] * (int32_t) src.sw[3]) + 0x8000) >> 16;
         CLOCK_CYCLES(1);
     } else {
-        MMX_REG src;
-
         SEG_CHECK_READ(cpu_state.ea_seg);
         src.l[0] = readmeml(easeg, cpu_state.eaaddr);
         src.l[1] = readmeml(easeg, cpu_state.eaaddr + 4);
         if (cpu_state.abrt)
             return 0;
-        cpu_state.MM[cpu_reg].w[0] = ((int32_t) (cpu_state.MM[cpu_reg].sw[0] * (int32_t) src.sw[0]) + 0x8000) >> 16;
-        cpu_state.MM[cpu_reg].w[1] = ((int32_t) (cpu_state.MM[cpu_reg].sw[1] * (int32_t) src.sw[1]) + 0x8000) >> 16;
-        cpu_state.MM[cpu_reg].w[2] = ((int32_t) (cpu_state.MM[cpu_reg].sw[2] * (int32_t) src.sw[2]) + 0x8000) >> 16;
-        cpu_state.MM[cpu_reg].w[3] = ((int32_t) (cpu_state.MM[cpu_reg].sw[3] * (int32_t) src.sw[3]) + 0x8000) >> 16;
+        dst->w[0] = ((int32_t) (dst->sw[0] * (int32_t) src.sw[0]) + 0x8000) >> 16;
+        dst->w[1] = ((int32_t) (dst->sw[1] * (int32_t) src.sw[1]) + 0x8000) >> 16;
+        dst->w[2] = ((int32_t) (dst->sw[2] * (int32_t) src.sw[2]) + 0x8000) >> 16;
+        dst->w[3] = ((int32_t) (dst->sw[3] * (int32_t) src.sw[3]) + 0x8000) >> 16;
         CLOCK_CYCLES(2);
     }
+
+    MMX_SETEXP(cpu_reg);
+
     return 0;
 }
 
diff --git a/src/cpu/x86_ops_i686.h b/src/cpu/x86_ops_i686.h
index 9e00249ca..f2b07a1c4 100644
--- a/src/cpu/x86_ops_i686.h
+++ b/src/cpu/x86_ops_i686.h
@@ -43,6 +43,132 @@ opSYSEXIT(uint32_t fetchdat)
     return ret;
 }
 
+static int
+sf_fx_save_stor_common(uint32_t fetchdat, int bits)
+{
+    uint8_t   fxinst     = 0;
+    uint32_t  tag_byte;
+    unsigned  index;
+    floatx80  reg;
+
+    if (CPUID < 0x650)
+        return ILLEGAL(fetchdat);
+
+    FP_ENTER();
+
+    if (bits == 32) {
+        fetch_ea_32(fetchdat);
+    } else {
+        fetch_ea_16(fetchdat);
+    }
+
+    if (cpu_state.eaaddr & 0xf) {
+        x386_dynarec_log("Effective address %08X not on 16-byte boundary\n", cpu_state.eaaddr);
+        x86gpf(NULL, 0);
+        return cpu_state.abrt;
+    }
+
+    fxinst = (rmdat >> 3) & 7;
+
+    if ((fxinst > 1) || (cpu_mod == 3)) {
+        x86illegal();
+        return cpu_state.abrt;
+    }
+
+    FP_ENTER();
+
+    if (fxinst == 1) {
+        /* FXRSTOR */
+        fpu_state.cwd = readmemw(easeg, cpu_state.eaaddr);
+        fpu_state.swd = readmemw(easeg, cpu_state.eaaddr + 2);
+        fpu_state.tos = (fpu_state.swd >> 11) & 7;
+
+        /* always set bit 6 as '1 */
+        fpu_state.cwd = (fpu_state.cwd & ~FPU_CW_Reserved_Bits) | 0x0040;
+
+        /* Restore x87 FPU Opcode */
+        /* The lower 11 bits contain the FPU opcode, upper 5 bits are reserved */
+        fpu_state.foo = readmemw(easeg, cpu_state.eaaddr + 6) & 0x7FF;
+
+        fpu_state.fip = readmeml(easeg, cpu_state.eaaddr + 8);
+        fpu_state.fcs = readmemw(easeg, cpu_state.eaaddr + 12);
+
+        tag_byte = readmemb(easeg, cpu_state.eaaddr + 4);
+
+        fpu_state.fdp = readmeml(easeg, cpu_state.eaaddr + 16);
+        fpu_state.fds = readmemw(easeg, cpu_state.eaaddr + 20);
+
+        /* load i387 register file */
+        for (index = 0; index < 8; index++) {
+            reg.fraction = readmemq(easeg, cpu_state.eaaddr + (index * 16) + 32);
+            reg.exp      = readmemw(easeg, cpu_state.eaaddr + (index * 16) + 40);
+
+            // update tag only if it is not empty
+            FPU_save_regi_tag(reg, IS_TAG_EMPTY(index) ? X87_TAG_EMPTY : FPU_tagof(reg), index);
+        }
+
+        fpu_state.tag = unpack_FPU_TW(tag_byte);
+
+        /* check for unmasked exceptions */
+        if (fpu_state.swd & ~fpu_state.cwd & FPU_CW_Exceptions_Mask) {
+            /* set the B and ES bits in the status-word */
+            fpu_state.swd |= (FPU_SW_Summary | FPU_SW_Backward);
+        } else {
+            /* clear the B and ES bits in the status-word */
+            fpu_state.swd &= ~(FPU_SW_Summary | FPU_SW_Backward);
+        }
+
+        CLOCK_CYCLES((cr0 & 1) ? 34 : 44);
+    } else {
+        /* FXSAVE */
+        writememw(easeg, cpu_state.eaaddr, i387_get_control_word());
+        writememw(easeg, cpu_state.eaaddr + 2, i387_get_status_word());
+        writememw(easeg, cpu_state.eaaddr + 4, pack_FPU_TW(fpu_state.tag));
+
+        /* x87 FPU Opcode (16 bits) */
+        /* The lower 11 bits contain the FPU opcode, upper 5 bits are reserved */
+        writememw(easeg, cpu_state.eaaddr + 6, fpu_state.foo);
+
+      /*
+       * x87 FPU IP Offset (32/64 bits)
+       * The contents of this field differ depending on the current
+       * addressing mode (16/32/64 bit) when the FXSAVE instruction was executed:
+       *   + 64-bit mode - 64-bit IP offset
+       *   + 32-bit mode - 32-bit IP offset
+       *   + 16-bit mode - low 16 bits are IP offset; high 16 bits are reserved.
+       * x87 CS FPU IP Selector
+       *   + 16 bit, in 16/32 bit mode only
+       */
+        writememl(easeg, cpu_state.eaaddr + 8, fpu_state.fip);
+        writememl(easeg, cpu_state.eaaddr + 12, fpu_state.fcs);
+
+      /*
+       * x87 FPU Instruction Operand (Data) Pointer Offset (32/64 bits)
+       * The contents of this field differ depending on the current
+       * addressing mode (16/32 bit) when the FXSAVE instruction was executed:
+       *   + 64-bit mode - 64-bit offset
+       *   + 32-bit mode - 32-bit offset
+       *   + 16-bit mode - low 16 bits are offset; high 16 bits are reserved.
+       * x87 DS FPU Instruction Operand (Data) Pointer Selector
+       *   + 16 bit, in 16/32 bit mode only
+       */
+        writememl(easeg, cpu_state.eaaddr + 16, fpu_state.fdp);
+        writememl(easeg, cpu_state.eaaddr + 20, fpu_state.fds);
+
+       /* store i387 register file */
+        for (index = 0; index < 8; index++) {
+            const floatx80 fp = FPU_read_regi(index);
+
+            writememq(easeg, cpu_state.eaaddr + (index * 16) + 32, fp.fraction);
+            writememw(easeg, cpu_state.eaaddr + (index * 16) + 40, fp.exp);
+        }
+
+        CLOCK_CYCLES((cr0 & 1) ? 56 : 67);
+    }
+
+    return cpu_state.abrt;
+}
+
 static int
 fx_save_stor_common(uint32_t fetchdat, int bits)
 {
@@ -253,12 +379,18 @@ fx_save_stor_common(uint32_t fetchdat, int bits)
 static int
 opFXSAVESTOR_a16(uint32_t fetchdat)
 {
+    if (fpu_softfloat) 
+        return sf_fx_save_stor_common(fetchdat, 16);
+
     return fx_save_stor_common(fetchdat, 16);
 }
 
 static int
 opFXSAVESTOR_a32(uint32_t fetchdat)
 {
+    if (fpu_softfloat)
+        return sf_fx_save_stor_common(fetchdat, 32);
+
     return fx_save_stor_common(fetchdat, 32);
 }
 
diff --git a/src/cpu/x86_ops_mmx.c b/src/cpu/x86_ops_mmx.c
new file mode 100644
index 000000000..1afc4fba2
--- /dev/null
+++ b/src/cpu/x86_ops_mmx.c
@@ -0,0 +1,50 @@
+#include <stdarg.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <string.h>
+#include <stdlib.h>
+#include <wchar.h>
+#include <math.h>
+#ifndef INFINITY
+#    define INFINITY (__builtin_inff())
+#endif
+#define HAVE_STDARG_H
+#include <86box/86box.h>
+#include "cpu.h"
+#include <86box/timer.h>
+#include "x86.h"
+#include "x87.h"
+#include <86box/nmi.h>
+#include <86box/mem.h>
+#include <86box/smram.h>
+#include <86box/pic.h>
+#include <86box/pit.h>
+#include <86box/fdd.h>
+#include <86box/fdc.h>
+#include <86box/keyboard.h>
+#include <86box/timer.h>
+#include "386_common.h"
+#include "x86_flags.h"
+#include "x86seg.h"
+
+MMX_REG *MMP[8];
+uint16_t *MMEP[8];
+
+static uint16_t MME[8];
+
+#define MMX_GETREGP(r) fpu_softfloat ? ((MMX_REG *) &fpu_state.st_space[r].fraction) : &(cpu_state.MM[r])
+void
+mmx_init(void)
+{
+    memset(MME, 0xff, sizeof(MME));
+
+    for (uint8_t i = 0; i < 8; i++) {
+        if (fpu_softfloat) {
+            MMP[i] = (MMX_REG *) &fpu_state.st_space[i].fraction;
+            MMEP[i] = (uint16_t *) &fpu_state.st_space[i].exp;
+        } else {
+            MMP[i] = &(cpu_state.MM[i]);
+            MMEP[i] = &(MME[i]);
+        }
+    }
+}
diff --git a/src/cpu/x86_ops_mmx.h b/src/cpu/x86_ops_mmx.h
index d270b728f..47751d059 100644
--- a/src/cpu/x86_ops_mmx.h
+++ b/src/cpu/x86_ops_mmx.h
@@ -3,9 +3,15 @@
 #define USATB(val) (((val) < 0) ? 0 : (((val) > 255) ? 255 : (val)))
 #define USATW(val) (((val) < 0) ? 0 : (((val) > 65535) ? 65535 : (val)))
 
+#define MMX_GETREGP(r) MMP[r]
+#define MMX_GETREG(r) *(MMP[r])
+
+#define MMX_SETEXP(r)  \
+    *(MMEP[r]) = 0xffff
+
 #define MMX_GETSRC()                               \
     if (cpu_mod == 3) {                            \
-        src = cpu_state.MM[cpu_rm];                \
+        src = MMX_GETREG(cpu_rm);                  \
         CLOCK_CYCLES(1);                           \
     } else {                                       \
         SEG_CHECK_READ(cpu_state.ea_seg);          \
diff --git a/src/cpu/x86_ops_mmx_arith.h b/src/cpu/x86_ops_mmx_arith.h
index e473f8ec5..642e99c8a 100644
--- a/src/cpu/x86_ops_mmx_arith.h
+++ b/src/cpu/x86_ops_mmx_arith.h
@@ -2,19 +2,25 @@ static int
 opPADDB_a16(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].b[0] += src.b[0];
-    cpu_state.MM[cpu_reg].b[1] += src.b[1];
-    cpu_state.MM[cpu_reg].b[2] += src.b[2];
-    cpu_state.MM[cpu_reg].b[3] += src.b[3];
-    cpu_state.MM[cpu_reg].b[4] += src.b[4];
-    cpu_state.MM[cpu_reg].b[5] += src.b[5];
-    cpu_state.MM[cpu_reg].b[6] += src.b[6];
-    cpu_state.MM[cpu_reg].b[7] += src.b[7];
+    dst->b[0] += src.b[0];
+    dst->b[1] += src.b[1];
+    dst->b[2] += src.b[2];
+    dst->b[3] += src.b[3];
+    dst->b[4] += src.b[4];
+    dst->b[5] += src.b[5];
+    dst->b[6] += src.b[6];
+    dst->b[7] += src.b[7];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -22,19 +28,25 @@ static int
 opPADDB_a32(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].b[0] += src.b[0];
-    cpu_state.MM[cpu_reg].b[1] += src.b[1];
-    cpu_state.MM[cpu_reg].b[2] += src.b[2];
-    cpu_state.MM[cpu_reg].b[3] += src.b[3];
-    cpu_state.MM[cpu_reg].b[4] += src.b[4];
-    cpu_state.MM[cpu_reg].b[5] += src.b[5];
-    cpu_state.MM[cpu_reg].b[6] += src.b[6];
-    cpu_state.MM[cpu_reg].b[7] += src.b[7];
+    dst->b[0] += src.b[0];
+    dst->b[1] += src.b[1];
+    dst->b[2] += src.b[2];
+    dst->b[3] += src.b[3];
+    dst->b[4] += src.b[4];
+    dst->b[5] += src.b[5];
+    dst->b[6] += src.b[6];
+    dst->b[7] += src.b[7];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -43,15 +55,21 @@ static int
 opPADDW_a16(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].w[0] += src.w[0];
-    cpu_state.MM[cpu_reg].w[1] += src.w[1];
-    cpu_state.MM[cpu_reg].w[2] += src.w[2];
-    cpu_state.MM[cpu_reg].w[3] += src.w[3];
+    dst->w[0] += src.w[0];
+    dst->w[1] += src.w[1];
+    dst->w[2] += src.w[2];
+    dst->w[3] += src.w[3];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -59,15 +77,21 @@ static int
 opPADDW_a32(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].w[0] += src.w[0];
-    cpu_state.MM[cpu_reg].w[1] += src.w[1];
-    cpu_state.MM[cpu_reg].w[2] += src.w[2];
-    cpu_state.MM[cpu_reg].w[3] += src.w[3];
+    dst->w[0] += src.w[0];
+    dst->w[1] += src.w[1];
+    dst->w[2] += src.w[2];
+    dst->w[3] += src.w[3];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -76,13 +100,19 @@ static int
 opPADDD_a16(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].l[0] += src.l[0];
-    cpu_state.MM[cpu_reg].l[1] += src.l[1];
+    dst->l[0] += src.l[0];
+    dst->l[1] += src.l[1];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -90,13 +120,19 @@ static int
 opPADDD_a32(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].l[0] += src.l[0];
-    cpu_state.MM[cpu_reg].l[1] += src.l[1];
+    dst->l[0] += src.l[0];
+    dst->l[1] += src.l[1];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -105,19 +141,25 @@ static int
 opPADDSB_a16(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].sb[0] = SSATB(cpu_state.MM[cpu_reg].sb[0] + src.sb[0]);
-    cpu_state.MM[cpu_reg].sb[1] = SSATB(cpu_state.MM[cpu_reg].sb[1] + src.sb[1]);
-    cpu_state.MM[cpu_reg].sb[2] = SSATB(cpu_state.MM[cpu_reg].sb[2] + src.sb[2]);
-    cpu_state.MM[cpu_reg].sb[3] = SSATB(cpu_state.MM[cpu_reg].sb[3] + src.sb[3]);
-    cpu_state.MM[cpu_reg].sb[4] = SSATB(cpu_state.MM[cpu_reg].sb[4] + src.sb[4]);
-    cpu_state.MM[cpu_reg].sb[5] = SSATB(cpu_state.MM[cpu_reg].sb[5] + src.sb[5]);
-    cpu_state.MM[cpu_reg].sb[6] = SSATB(cpu_state.MM[cpu_reg].sb[6] + src.sb[6]);
-    cpu_state.MM[cpu_reg].sb[7] = SSATB(cpu_state.MM[cpu_reg].sb[7] + src.sb[7]);
+    dst->sb[0] = SSATB(dst->sb[0] + src.sb[0]);
+    dst->sb[1] = SSATB(dst->sb[1] + src.sb[1]);
+    dst->sb[2] = SSATB(dst->sb[2] + src.sb[2]);
+    dst->sb[3] = SSATB(dst->sb[3] + src.sb[3]);
+    dst->sb[4] = SSATB(dst->sb[4] + src.sb[4]);
+    dst->sb[5] = SSATB(dst->sb[5] + src.sb[5]);
+    dst->sb[6] = SSATB(dst->sb[6] + src.sb[6]);
+    dst->sb[7] = SSATB(dst->sb[7] + src.sb[7]);
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -125,19 +167,25 @@ static int
 opPADDSB_a32(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].sb[0] = SSATB(cpu_state.MM[cpu_reg].sb[0] + src.sb[0]);
-    cpu_state.MM[cpu_reg].sb[1] = SSATB(cpu_state.MM[cpu_reg].sb[1] + src.sb[1]);
-    cpu_state.MM[cpu_reg].sb[2] = SSATB(cpu_state.MM[cpu_reg].sb[2] + src.sb[2]);
-    cpu_state.MM[cpu_reg].sb[3] = SSATB(cpu_state.MM[cpu_reg].sb[3] + src.sb[3]);
-    cpu_state.MM[cpu_reg].sb[4] = SSATB(cpu_state.MM[cpu_reg].sb[4] + src.sb[4]);
-    cpu_state.MM[cpu_reg].sb[5] = SSATB(cpu_state.MM[cpu_reg].sb[5] + src.sb[5]);
-    cpu_state.MM[cpu_reg].sb[6] = SSATB(cpu_state.MM[cpu_reg].sb[6] + src.sb[6]);
-    cpu_state.MM[cpu_reg].sb[7] = SSATB(cpu_state.MM[cpu_reg].sb[7] + src.sb[7]);
+    dst->sb[0] = SSATB(dst->sb[0] + src.sb[0]);
+    dst->sb[1] = SSATB(dst->sb[1] + src.sb[1]);
+    dst->sb[2] = SSATB(dst->sb[2] + src.sb[2]);
+    dst->sb[3] = SSATB(dst->sb[3] + src.sb[3]);
+    dst->sb[4] = SSATB(dst->sb[4] + src.sb[4]);
+    dst->sb[5] = SSATB(dst->sb[5] + src.sb[5]);
+    dst->sb[6] = SSATB(dst->sb[6] + src.sb[6]);
+    dst->sb[7] = SSATB(dst->sb[7] + src.sb[7]);
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -146,19 +194,25 @@ static int
 opPADDUSB_a16(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].b[0] = USATB(cpu_state.MM[cpu_reg].b[0] + src.b[0]);
-    cpu_state.MM[cpu_reg].b[1] = USATB(cpu_state.MM[cpu_reg].b[1] + src.b[1]);
-    cpu_state.MM[cpu_reg].b[2] = USATB(cpu_state.MM[cpu_reg].b[2] + src.b[2]);
-    cpu_state.MM[cpu_reg].b[3] = USATB(cpu_state.MM[cpu_reg].b[3] + src.b[3]);
-    cpu_state.MM[cpu_reg].b[4] = USATB(cpu_state.MM[cpu_reg].b[4] + src.b[4]);
-    cpu_state.MM[cpu_reg].b[5] = USATB(cpu_state.MM[cpu_reg].b[5] + src.b[5]);
-    cpu_state.MM[cpu_reg].b[6] = USATB(cpu_state.MM[cpu_reg].b[6] + src.b[6]);
-    cpu_state.MM[cpu_reg].b[7] = USATB(cpu_state.MM[cpu_reg].b[7] + src.b[7]);
+    dst->b[0] = USATB(dst->b[0] + src.b[0]);
+    dst->b[1] = USATB(dst->b[1] + src.b[1]);
+    dst->b[2] = USATB(dst->b[2] + src.b[2]);
+    dst->b[3] = USATB(dst->b[3] + src.b[3]);
+    dst->b[4] = USATB(dst->b[4] + src.b[4]);
+    dst->b[5] = USATB(dst->b[5] + src.b[5]);
+    dst->b[6] = USATB(dst->b[6] + src.b[6]);
+    dst->b[7] = USATB(dst->b[7] + src.b[7]);
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -166,19 +220,25 @@ static int
 opPADDUSB_a32(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].b[0] = USATB(cpu_state.MM[cpu_reg].b[0] + src.b[0]);
-    cpu_state.MM[cpu_reg].b[1] = USATB(cpu_state.MM[cpu_reg].b[1] + src.b[1]);
-    cpu_state.MM[cpu_reg].b[2] = USATB(cpu_state.MM[cpu_reg].b[2] + src.b[2]);
-    cpu_state.MM[cpu_reg].b[3] = USATB(cpu_state.MM[cpu_reg].b[3] + src.b[3]);
-    cpu_state.MM[cpu_reg].b[4] = USATB(cpu_state.MM[cpu_reg].b[4] + src.b[4]);
-    cpu_state.MM[cpu_reg].b[5] = USATB(cpu_state.MM[cpu_reg].b[5] + src.b[5]);
-    cpu_state.MM[cpu_reg].b[6] = USATB(cpu_state.MM[cpu_reg].b[6] + src.b[6]);
-    cpu_state.MM[cpu_reg].b[7] = USATB(cpu_state.MM[cpu_reg].b[7] + src.b[7]);
+    dst->b[0] = USATB(dst->b[0] + src.b[0]);
+    dst->b[1] = USATB(dst->b[1] + src.b[1]);
+    dst->b[2] = USATB(dst->b[2] + src.b[2]);
+    dst->b[3] = USATB(dst->b[3] + src.b[3]);
+    dst->b[4] = USATB(dst->b[4] + src.b[4]);
+    dst->b[5] = USATB(dst->b[5] + src.b[5]);
+    dst->b[6] = USATB(dst->b[6] + src.b[6]);
+    dst->b[7] = USATB(dst->b[7] + src.b[7]);
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -187,15 +247,21 @@ static int
 opPADDSW_a16(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].sw[0] = SSATW(cpu_state.MM[cpu_reg].sw[0] + src.sw[0]);
-    cpu_state.MM[cpu_reg].sw[1] = SSATW(cpu_state.MM[cpu_reg].sw[1] + src.sw[1]);
-    cpu_state.MM[cpu_reg].sw[2] = SSATW(cpu_state.MM[cpu_reg].sw[2] + src.sw[2]);
-    cpu_state.MM[cpu_reg].sw[3] = SSATW(cpu_state.MM[cpu_reg].sw[3] + src.sw[3]);
+    dst->sw[0] = SSATW(dst->sw[0] + src.sw[0]);
+    dst->sw[1] = SSATW(dst->sw[1] + src.sw[1]);
+    dst->sw[2] = SSATW(dst->sw[2] + src.sw[2]);
+    dst->sw[3] = SSATW(dst->sw[3] + src.sw[3]);
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -203,15 +269,21 @@ static int
 opPADDSW_a32(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].sw[0] = SSATW(cpu_state.MM[cpu_reg].sw[0] + src.sw[0]);
-    cpu_state.MM[cpu_reg].sw[1] = SSATW(cpu_state.MM[cpu_reg].sw[1] + src.sw[1]);
-    cpu_state.MM[cpu_reg].sw[2] = SSATW(cpu_state.MM[cpu_reg].sw[2] + src.sw[2]);
-    cpu_state.MM[cpu_reg].sw[3] = SSATW(cpu_state.MM[cpu_reg].sw[3] + src.sw[3]);
+    dst->sw[0] = SSATW(dst->sw[0] + src.sw[0]);
+    dst->sw[1] = SSATW(dst->sw[1] + src.sw[1]);
+    dst->sw[2] = SSATW(dst->sw[2] + src.sw[2]);
+    dst->sw[3] = SSATW(dst->sw[3] + src.sw[3]);
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -220,15 +292,21 @@ static int
 opPADDUSW_a16(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].w[0] = USATW(cpu_state.MM[cpu_reg].w[0] + src.w[0]);
-    cpu_state.MM[cpu_reg].w[1] = USATW(cpu_state.MM[cpu_reg].w[1] + src.w[1]);
-    cpu_state.MM[cpu_reg].w[2] = USATW(cpu_state.MM[cpu_reg].w[2] + src.w[2]);
-    cpu_state.MM[cpu_reg].w[3] = USATW(cpu_state.MM[cpu_reg].w[3] + src.w[3]);
+    dst->w[0] = USATW(dst->w[0] + src.w[0]);
+    dst->w[1] = USATW(dst->w[1] + src.w[1]);
+    dst->w[2] = USATW(dst->w[2] + src.w[2]);
+    dst->w[3] = USATW(dst->w[3] + src.w[3]);
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -236,15 +314,21 @@ static int
 opPADDUSW_a32(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].w[0] = USATW(cpu_state.MM[cpu_reg].w[0] + src.w[0]);
-    cpu_state.MM[cpu_reg].w[1] = USATW(cpu_state.MM[cpu_reg].w[1] + src.w[1]);
-    cpu_state.MM[cpu_reg].w[2] = USATW(cpu_state.MM[cpu_reg].w[2] + src.w[2]);
-    cpu_state.MM[cpu_reg].w[3] = USATW(cpu_state.MM[cpu_reg].w[3] + src.w[3]);
+    dst->w[0] = USATW(dst->w[0] + src.w[0]);
+    dst->w[1] = USATW(dst->w[1] + src.w[1]);
+    dst->w[2] = USATW(dst->w[2] + src.w[2]);
+    dst->w[3] = USATW(dst->w[3] + src.w[3]);
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -253,20 +337,26 @@ static int
 opPMADDWD_a16(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    if (cpu_state.MM[cpu_reg].l[0] == 0x80008000 && src.l[0] == 0x80008000)
-        cpu_state.MM[cpu_reg].l[0] = 0x80000000;
+    if (dst->l[0] == 0x80008000 && src.l[0] == 0x80008000)
+        dst->l[0] = 0x80000000;
     else
-        cpu_state.MM[cpu_reg].sl[0] = ((int32_t) cpu_state.MM[cpu_reg].sw[0] * (int32_t) src.sw[0]) + ((int32_t) cpu_state.MM[cpu_reg].sw[1] * (int32_t) src.sw[1]);
+        dst->sl[0] = ((int32_t) dst->sw[0] * (int32_t) src.sw[0]) + ((int32_t) dst->sw[1] * (int32_t) src.sw[1]);
 
-    if (cpu_state.MM[cpu_reg].l[1] == 0x80008000 && src.l[1] == 0x80008000)
-        cpu_state.MM[cpu_reg].l[1] = 0x80000000;
+    if (dst->l[1] == 0x80008000 && src.l[1] == 0x80008000)
+        dst->l[1] = 0x80000000;
     else
-        cpu_state.MM[cpu_reg].sl[1] = ((int32_t) cpu_state.MM[cpu_reg].sw[2] * (int32_t) src.sw[2]) + ((int32_t) cpu_state.MM[cpu_reg].sw[3] * (int32_t) src.sw[3]);
+        dst->sl[1] = ((int32_t) dst->sw[2] * (int32_t) src.sw[2]) + ((int32_t) dst->sw[3] * (int32_t) src.sw[3]);
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -274,20 +364,26 @@ static int
 opPMADDWD_a32(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    if (cpu_state.MM[cpu_reg].l[0] == 0x80008000 && src.l[0] == 0x80008000)
-        cpu_state.MM[cpu_reg].l[0] = 0x80000000;
+    if (dst->l[0] == 0x80008000 && src.l[0] == 0x80008000)
+        dst->l[0] = 0x80000000;
     else
-        cpu_state.MM[cpu_reg].sl[0] = ((int32_t) cpu_state.MM[cpu_reg].sw[0] * (int32_t) src.sw[0]) + ((int32_t) cpu_state.MM[cpu_reg].sw[1] * (int32_t) src.sw[1]);
+        dst->sl[0] = ((int32_t) dst->sw[0] * (int32_t) src.sw[0]) + ((int32_t) dst->sw[1] * (int32_t) src.sw[1]);
 
-    if (cpu_state.MM[cpu_reg].l[1] == 0x80008000 && src.l[1] == 0x80008000)
-        cpu_state.MM[cpu_reg].l[1] = 0x80000000;
+    if (dst->l[1] == 0x80008000 && src.l[1] == 0x80008000)
+        dst->l[1] = 0x80000000;
     else
-        cpu_state.MM[cpu_reg].sl[1] = ((int32_t) cpu_state.MM[cpu_reg].sw[2] * (int32_t) src.sw[2]) + ((int32_t) cpu_state.MM[cpu_reg].sw[3] * (int32_t) src.sw[3]);
+        dst->sl[1] = ((int32_t) dst->sw[2] * (int32_t) src.sw[2]) + ((int32_t) dst->sw[3] * (int32_t) src.sw[3]);
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -295,114 +391,126 @@ opPMADDWD_a32(uint32_t fetchdat)
 static int
 opPMULLW_a16(uint32_t fetchdat)
 {
+    MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
-    if (cpu_mod == 3) {
-        cpu_state.MM[cpu_reg].w[0] *= cpu_state.MM[cpu_rm].w[0];
-        cpu_state.MM[cpu_reg].w[1] *= cpu_state.MM[cpu_rm].w[1];
-        cpu_state.MM[cpu_reg].w[2] *= cpu_state.MM[cpu_rm].w[2];
-        cpu_state.MM[cpu_reg].w[3] *= cpu_state.MM[cpu_rm].w[3];
-        CLOCK_CYCLES(1);
-    } else {
-        MMX_REG src;
 
+    dst = MMX_GETREGP(cpu_reg);
+
+    if (cpu_mod == 3)
+        src = MMX_GETREG(cpu_rm);
+    else {
         SEG_CHECK_READ(cpu_state.ea_seg);
         src.l[0] = readmeml(easeg, cpu_state.eaaddr);
         src.l[1] = readmeml(easeg, cpu_state.eaaddr + 4);
         if (cpu_state.abrt)
             return 0;
-        cpu_state.MM[cpu_reg].w[0] *= src.w[0];
-        cpu_state.MM[cpu_reg].w[1] *= src.w[1];
-        cpu_state.MM[cpu_reg].w[2] *= src.w[2];
-        cpu_state.MM[cpu_reg].w[3] *= src.w[3];
-        CLOCK_CYCLES(2);
+        CLOCK_CYCLES(1);
     }
+    dst->w[0] *= src.w[0];
+    dst->w[1] *= src.w[1];
+    dst->w[2] *= src.w[2];
+    dst->w[3] *= src.w[3];
+    CLOCK_CYCLES(1);
+
+    MMX_SETEXP(cpu_reg);
+
     return 0;
 }
 static int
 opPMULLW_a32(uint32_t fetchdat)
 {
+    MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
-    if (cpu_mod == 3) {
-        cpu_state.MM[cpu_reg].w[0] *= cpu_state.MM[cpu_rm].w[0];
-        cpu_state.MM[cpu_reg].w[1] *= cpu_state.MM[cpu_rm].w[1];
-        cpu_state.MM[cpu_reg].w[2] *= cpu_state.MM[cpu_rm].w[2];
-        cpu_state.MM[cpu_reg].w[3] *= cpu_state.MM[cpu_rm].w[3];
-        CLOCK_CYCLES(1);
-    } else {
-        MMX_REG src;
 
+    dst = MMX_GETREGP(cpu_reg);
+
+    if (cpu_mod == 3)
+        src = MMX_GETREG(cpu_rm);
+    else {
         SEG_CHECK_READ(cpu_state.ea_seg);
         src.l[0] = readmeml(easeg, cpu_state.eaaddr);
         src.l[1] = readmeml(easeg, cpu_state.eaaddr + 4);
         if (cpu_state.abrt)
             return 0;
-        cpu_state.MM[cpu_reg].w[0] *= src.w[0];
-        cpu_state.MM[cpu_reg].w[1] *= src.w[1];
-        cpu_state.MM[cpu_reg].w[2] *= src.w[2];
-        cpu_state.MM[cpu_reg].w[3] *= src.w[3];
-        CLOCK_CYCLES(2);
+        CLOCK_CYCLES(1);
     }
+    dst->w[0] *= src.w[0];
+    dst->w[1] *= src.w[1];
+    dst->w[2] *= src.w[2];
+    dst->w[3] *= src.w[3];
+    CLOCK_CYCLES(1);
+
+    MMX_SETEXP(cpu_reg);
+
     return 0;
 }
 
 static int
 opPMULHW_a16(uint32_t fetchdat)
 {
+    MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
-    if (cpu_mod == 3) {
-        cpu_state.MM[cpu_reg].w[0] = ((int32_t) cpu_state.MM[cpu_reg].sw[0] * (int32_t) cpu_state.MM[cpu_rm].sw[0]) >> 16;
-        cpu_state.MM[cpu_reg].w[1] = ((int32_t) cpu_state.MM[cpu_reg].sw[1] * (int32_t) cpu_state.MM[cpu_rm].sw[1]) >> 16;
-        cpu_state.MM[cpu_reg].w[2] = ((int32_t) cpu_state.MM[cpu_reg].sw[2] * (int32_t) cpu_state.MM[cpu_rm].sw[2]) >> 16;
-        cpu_state.MM[cpu_reg].w[3] = ((int32_t) cpu_state.MM[cpu_reg].sw[3] * (int32_t) cpu_state.MM[cpu_rm].sw[3]) >> 16;
-        CLOCK_CYCLES(1);
-    } else {
-        MMX_REG src;
 
+    dst = MMX_GETREGP(cpu_reg);
+
+    if (cpu_mod == 3)
+        src = MMX_GETREG(cpu_rm);
+    else {
         SEG_CHECK_READ(cpu_state.ea_seg);
         src.l[0] = readmeml(easeg, cpu_state.eaaddr);
         src.l[1] = readmeml(easeg, cpu_state.eaaddr + 4);
         if (cpu_state.abrt)
             return 0;
-        cpu_state.MM[cpu_reg].w[0] = ((int32_t) cpu_state.MM[cpu_reg].sw[0] * (int32_t) src.sw[0]) >> 16;
-        cpu_state.MM[cpu_reg].w[1] = ((int32_t) cpu_state.MM[cpu_reg].sw[1] * (int32_t) src.sw[1]) >> 16;
-        cpu_state.MM[cpu_reg].w[2] = ((int32_t) cpu_state.MM[cpu_reg].sw[2] * (int32_t) src.sw[2]) >> 16;
-        cpu_state.MM[cpu_reg].w[3] = ((int32_t) cpu_state.MM[cpu_reg].sw[3] * (int32_t) src.sw[3]) >> 16;
-        CLOCK_CYCLES(2);
+        CLOCK_CYCLES(1);
     }
+    dst->w[0] = ((int32_t) dst->sw[0] * (int32_t) src.sw[0]) >> 16;
+    dst->w[1] = ((int32_t) dst->sw[1] * (int32_t) src.sw[1]) >> 16;
+    dst->w[2] = ((int32_t) dst->sw[2] * (int32_t) src.sw[2]) >> 16;
+    dst->w[3] = ((int32_t) dst->sw[3] * (int32_t) src.sw[3]) >> 16;
+    CLOCK_CYCLES(1);
+
+    MMX_SETEXP(cpu_reg);
+
     return 0;
 }
 static int
 opPMULHW_a32(uint32_t fetchdat)
 {
+    MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
-    if (cpu_mod == 3) {
-        cpu_state.MM[cpu_reg].w[0] = ((int32_t) cpu_state.MM[cpu_reg].sw[0] * (int32_t) cpu_state.MM[cpu_rm].sw[0]) >> 16;
-        cpu_state.MM[cpu_reg].w[1] = ((int32_t) cpu_state.MM[cpu_reg].sw[1] * (int32_t) cpu_state.MM[cpu_rm].sw[1]) >> 16;
-        cpu_state.MM[cpu_reg].w[2] = ((int32_t) cpu_state.MM[cpu_reg].sw[2] * (int32_t) cpu_state.MM[cpu_rm].sw[2]) >> 16;
-        cpu_state.MM[cpu_reg].w[3] = ((int32_t) cpu_state.MM[cpu_reg].sw[3] * (int32_t) cpu_state.MM[cpu_rm].sw[3]) >> 16;
-        CLOCK_CYCLES(1);
-    } else {
-        MMX_REG src;
 
+    dst = MMX_GETREGP(cpu_reg);
+
+    if (cpu_mod == 3)
+        src = MMX_GETREG(cpu_rm);
+    else {
         SEG_CHECK_READ(cpu_state.ea_seg);
         src.l[0] = readmeml(easeg, cpu_state.eaaddr);
         src.l[1] = readmeml(easeg, cpu_state.eaaddr + 4);
         if (cpu_state.abrt)
             return 0;
-        cpu_state.MM[cpu_reg].w[0] = ((int32_t) cpu_state.MM[cpu_reg].sw[0] * (int32_t) src.sw[0]) >> 16;
-        cpu_state.MM[cpu_reg].w[1] = ((int32_t) cpu_state.MM[cpu_reg].sw[1] * (int32_t) src.sw[1]) >> 16;
-        cpu_state.MM[cpu_reg].w[2] = ((int32_t) cpu_state.MM[cpu_reg].sw[2] * (int32_t) src.sw[2]) >> 16;
-        cpu_state.MM[cpu_reg].w[3] = ((int32_t) cpu_state.MM[cpu_reg].sw[3] * (int32_t) src.sw[3]) >> 16;
-        CLOCK_CYCLES(2);
+        CLOCK_CYCLES(1);
     }
+    dst->w[0] = ((int32_t) dst->sw[0] * (int32_t) src.sw[0]) >> 16;
+    dst->w[1] = ((int32_t) dst->sw[1] * (int32_t) src.sw[1]) >> 16;
+    dst->w[2] = ((int32_t) dst->sw[2] * (int32_t) src.sw[2]) >> 16;
+    dst->w[3] = ((int32_t) dst->sw[3] * (int32_t) src.sw[3]) >> 16;
+    CLOCK_CYCLES(1);
+
+    MMX_SETEXP(cpu_reg);
+
     return 0;
 }
 
@@ -410,19 +518,25 @@ static int
 opPSUBB_a16(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].b[0] -= src.b[0];
-    cpu_state.MM[cpu_reg].b[1] -= src.b[1];
-    cpu_state.MM[cpu_reg].b[2] -= src.b[2];
-    cpu_state.MM[cpu_reg].b[3] -= src.b[3];
-    cpu_state.MM[cpu_reg].b[4] -= src.b[4];
-    cpu_state.MM[cpu_reg].b[5] -= src.b[5];
-    cpu_state.MM[cpu_reg].b[6] -= src.b[6];
-    cpu_state.MM[cpu_reg].b[7] -= src.b[7];
+    dst->b[0] -= src.b[0];
+    dst->b[1] -= src.b[1];
+    dst->b[2] -= src.b[2];
+    dst->b[3] -= src.b[3];
+    dst->b[4] -= src.b[4];
+    dst->b[5] -= src.b[5];
+    dst->b[6] -= src.b[6];
+    dst->b[7] -= src.b[7];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -430,19 +544,25 @@ static int
 opPSUBB_a32(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].b[0] -= src.b[0];
-    cpu_state.MM[cpu_reg].b[1] -= src.b[1];
-    cpu_state.MM[cpu_reg].b[2] -= src.b[2];
-    cpu_state.MM[cpu_reg].b[3] -= src.b[3];
-    cpu_state.MM[cpu_reg].b[4] -= src.b[4];
-    cpu_state.MM[cpu_reg].b[5] -= src.b[5];
-    cpu_state.MM[cpu_reg].b[6] -= src.b[6];
-    cpu_state.MM[cpu_reg].b[7] -= src.b[7];
+    dst->b[0] -= src.b[0];
+    dst->b[1] -= src.b[1];
+    dst->b[2] -= src.b[2];
+    dst->b[3] -= src.b[3];
+    dst->b[4] -= src.b[4];
+    dst->b[5] -= src.b[5];
+    dst->b[6] -= src.b[6];
+    dst->b[7] -= src.b[7];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -451,15 +571,21 @@ static int
 opPSUBW_a16(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].w[0] -= src.w[0];
-    cpu_state.MM[cpu_reg].w[1] -= src.w[1];
-    cpu_state.MM[cpu_reg].w[2] -= src.w[2];
-    cpu_state.MM[cpu_reg].w[3] -= src.w[3];
+    dst->w[0] -= src.w[0];
+    dst->w[1] -= src.w[1];
+    dst->w[2] -= src.w[2];
+    dst->w[3] -= src.w[3];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -467,15 +593,21 @@ static int
 opPSUBW_a32(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].w[0] -= src.w[0];
-    cpu_state.MM[cpu_reg].w[1] -= src.w[1];
-    cpu_state.MM[cpu_reg].w[2] -= src.w[2];
-    cpu_state.MM[cpu_reg].w[3] -= src.w[3];
+    dst->w[0] -= src.w[0];
+    dst->w[1] -= src.w[1];
+    dst->w[2] -= src.w[2];
+    dst->w[3] -= src.w[3];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -484,13 +616,19 @@ static int
 opPSUBD_a16(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].l[0] -= src.l[0];
-    cpu_state.MM[cpu_reg].l[1] -= src.l[1];
+    dst->l[0] -= src.l[0];
+    dst->l[1] -= src.l[1];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -498,13 +636,19 @@ static int
 opPSUBD_a32(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].l[0] -= src.l[0];
-    cpu_state.MM[cpu_reg].l[1] -= src.l[1];
+    dst->l[0] -= src.l[0];
+    dst->l[1] -= src.l[1];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -513,20 +657,25 @@ static int
 opPSUBSB_a16(uint32_t fetchdat)
 {
     MMX_REG src;
-    pclog("opPSUBSB_a16(%08X)\n", fetchdat);
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].sb[0] = SSATB(cpu_state.MM[cpu_reg].sb[0] - src.sb[0]);
-    cpu_state.MM[cpu_reg].sb[1] = SSATB(cpu_state.MM[cpu_reg].sb[1] - src.sb[1]);
-    cpu_state.MM[cpu_reg].sb[2] = SSATB(cpu_state.MM[cpu_reg].sb[2] - src.sb[2]);
-    cpu_state.MM[cpu_reg].sb[3] = SSATB(cpu_state.MM[cpu_reg].sb[3] - src.sb[3]);
-    cpu_state.MM[cpu_reg].sb[4] = SSATB(cpu_state.MM[cpu_reg].sb[4] - src.sb[4]);
-    cpu_state.MM[cpu_reg].sb[5] = SSATB(cpu_state.MM[cpu_reg].sb[5] - src.sb[5]);
-    cpu_state.MM[cpu_reg].sb[6] = SSATB(cpu_state.MM[cpu_reg].sb[6] - src.sb[6]);
-    cpu_state.MM[cpu_reg].sb[7] = SSATB(cpu_state.MM[cpu_reg].sb[7] - src.sb[7]);
+    dst->sb[0] = SSATB(dst->sb[0] - src.sb[0]);
+    dst->sb[1] = SSATB(dst->sb[1] - src.sb[1]);
+    dst->sb[2] = SSATB(dst->sb[2] - src.sb[2]);
+    dst->sb[3] = SSATB(dst->sb[3] - src.sb[3]);
+    dst->sb[4] = SSATB(dst->sb[4] - src.sb[4]);
+    dst->sb[5] = SSATB(dst->sb[5] - src.sb[5]);
+    dst->sb[6] = SSATB(dst->sb[6] - src.sb[6]);
+    dst->sb[7] = SSATB(dst->sb[7] - src.sb[7]);
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -534,20 +683,25 @@ static int
 opPSUBSB_a32(uint32_t fetchdat)
 {
     MMX_REG src;
-    pclog("opPSUBSB_a32(%08X)\n", fetchdat);
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].sb[0] = SSATB(cpu_state.MM[cpu_reg].sb[0] - src.sb[0]);
-    cpu_state.MM[cpu_reg].sb[1] = SSATB(cpu_state.MM[cpu_reg].sb[1] - src.sb[1]);
-    cpu_state.MM[cpu_reg].sb[2] = SSATB(cpu_state.MM[cpu_reg].sb[2] - src.sb[2]);
-    cpu_state.MM[cpu_reg].sb[3] = SSATB(cpu_state.MM[cpu_reg].sb[3] - src.sb[3]);
-    cpu_state.MM[cpu_reg].sb[4] = SSATB(cpu_state.MM[cpu_reg].sb[4] - src.sb[4]);
-    cpu_state.MM[cpu_reg].sb[5] = SSATB(cpu_state.MM[cpu_reg].sb[5] - src.sb[5]);
-    cpu_state.MM[cpu_reg].sb[6] = SSATB(cpu_state.MM[cpu_reg].sb[6] - src.sb[6]);
-    cpu_state.MM[cpu_reg].sb[7] = SSATB(cpu_state.MM[cpu_reg].sb[7] - src.sb[7]);
+    dst->sb[0] = SSATB(dst->sb[0] - src.sb[0]);
+    dst->sb[1] = SSATB(dst->sb[1] - src.sb[1]);
+    dst->sb[2] = SSATB(dst->sb[2] - src.sb[2]);
+    dst->sb[3] = SSATB(dst->sb[3] - src.sb[3]);
+    dst->sb[4] = SSATB(dst->sb[4] - src.sb[4]);
+    dst->sb[5] = SSATB(dst->sb[5] - src.sb[5]);
+    dst->sb[6] = SSATB(dst->sb[6] - src.sb[6]);
+    dst->sb[7] = SSATB(dst->sb[7] - src.sb[7]);
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -556,19 +710,25 @@ static int
 opPSUBUSB_a16(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].b[0] = USATB(cpu_state.MM[cpu_reg].b[0] - src.b[0]);
-    cpu_state.MM[cpu_reg].b[1] = USATB(cpu_state.MM[cpu_reg].b[1] - src.b[1]);
-    cpu_state.MM[cpu_reg].b[2] = USATB(cpu_state.MM[cpu_reg].b[2] - src.b[2]);
-    cpu_state.MM[cpu_reg].b[3] = USATB(cpu_state.MM[cpu_reg].b[3] - src.b[3]);
-    cpu_state.MM[cpu_reg].b[4] = USATB(cpu_state.MM[cpu_reg].b[4] - src.b[4]);
-    cpu_state.MM[cpu_reg].b[5] = USATB(cpu_state.MM[cpu_reg].b[5] - src.b[5]);
-    cpu_state.MM[cpu_reg].b[6] = USATB(cpu_state.MM[cpu_reg].b[6] - src.b[6]);
-    cpu_state.MM[cpu_reg].b[7] = USATB(cpu_state.MM[cpu_reg].b[7] - src.b[7]);
+    dst->b[0] = USATB(dst->b[0] - src.b[0]);
+    dst->b[1] = USATB(dst->b[1] - src.b[1]);
+    dst->b[2] = USATB(dst->b[2] - src.b[2]);
+    dst->b[3] = USATB(dst->b[3] - src.b[3]);
+    dst->b[4] = USATB(dst->b[4] - src.b[4]);
+    dst->b[5] = USATB(dst->b[5] - src.b[5]);
+    dst->b[6] = USATB(dst->b[6] - src.b[6]);
+    dst->b[7] = USATB(dst->b[7] - src.b[7]);
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -576,19 +736,25 @@ static int
 opPSUBUSB_a32(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].b[0] = USATB(cpu_state.MM[cpu_reg].b[0] - src.b[0]);
-    cpu_state.MM[cpu_reg].b[1] = USATB(cpu_state.MM[cpu_reg].b[1] - src.b[1]);
-    cpu_state.MM[cpu_reg].b[2] = USATB(cpu_state.MM[cpu_reg].b[2] - src.b[2]);
-    cpu_state.MM[cpu_reg].b[3] = USATB(cpu_state.MM[cpu_reg].b[3] - src.b[3]);
-    cpu_state.MM[cpu_reg].b[4] = USATB(cpu_state.MM[cpu_reg].b[4] - src.b[4]);
-    cpu_state.MM[cpu_reg].b[5] = USATB(cpu_state.MM[cpu_reg].b[5] - src.b[5]);
-    cpu_state.MM[cpu_reg].b[6] = USATB(cpu_state.MM[cpu_reg].b[6] - src.b[6]);
-    cpu_state.MM[cpu_reg].b[7] = USATB(cpu_state.MM[cpu_reg].b[7] - src.b[7]);
+    dst->b[0] = USATB(dst->b[0] - src.b[0]);
+    dst->b[1] = USATB(dst->b[1] - src.b[1]);
+    dst->b[2] = USATB(dst->b[2] - src.b[2]);
+    dst->b[3] = USATB(dst->b[3] - src.b[3]);
+    dst->b[4] = USATB(dst->b[4] - src.b[4]);
+    dst->b[5] = USATB(dst->b[5] - src.b[5]);
+    dst->b[6] = USATB(dst->b[6] - src.b[6]);
+    dst->b[7] = USATB(dst->b[7] - src.b[7]);
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -597,15 +763,21 @@ static int
 opPSUBSW_a16(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].sw[0] = SSATW(cpu_state.MM[cpu_reg].sw[0] - src.sw[0]);
-    cpu_state.MM[cpu_reg].sw[1] = SSATW(cpu_state.MM[cpu_reg].sw[1] - src.sw[1]);
-    cpu_state.MM[cpu_reg].sw[2] = SSATW(cpu_state.MM[cpu_reg].sw[2] - src.sw[2]);
-    cpu_state.MM[cpu_reg].sw[3] = SSATW(cpu_state.MM[cpu_reg].sw[3] - src.sw[3]);
+    dst->sw[0] = SSATW(dst->sw[0] - src.sw[0]);
+    dst->sw[1] = SSATW(dst->sw[1] - src.sw[1]);
+    dst->sw[2] = SSATW(dst->sw[2] - src.sw[2]);
+    dst->sw[3] = SSATW(dst->sw[3] - src.sw[3]);
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -613,15 +785,21 @@ static int
 opPSUBSW_a32(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].sw[0] = SSATW(cpu_state.MM[cpu_reg].sw[0] - src.sw[0]);
-    cpu_state.MM[cpu_reg].sw[1] = SSATW(cpu_state.MM[cpu_reg].sw[1] - src.sw[1]);
-    cpu_state.MM[cpu_reg].sw[2] = SSATW(cpu_state.MM[cpu_reg].sw[2] - src.sw[2]);
-    cpu_state.MM[cpu_reg].sw[3] = SSATW(cpu_state.MM[cpu_reg].sw[3] - src.sw[3]);
+    dst->sw[0] = SSATW(dst->sw[0] - src.sw[0]);
+    dst->sw[1] = SSATW(dst->sw[1] - src.sw[1]);
+    dst->sw[2] = SSATW(dst->sw[2] - src.sw[2]);
+    dst->sw[3] = SSATW(dst->sw[3] - src.sw[3]);
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -630,15 +808,21 @@ static int
 opPSUBUSW_a16(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].w[0] = USATW(cpu_state.MM[cpu_reg].w[0] - src.w[0]);
-    cpu_state.MM[cpu_reg].w[1] = USATW(cpu_state.MM[cpu_reg].w[1] - src.w[1]);
-    cpu_state.MM[cpu_reg].w[2] = USATW(cpu_state.MM[cpu_reg].w[2] - src.w[2]);
-    cpu_state.MM[cpu_reg].w[3] = USATW(cpu_state.MM[cpu_reg].w[3] - src.w[3]);
+    dst->w[0] = USATW(dst->w[0] - src.w[0]);
+    dst->w[1] = USATW(dst->w[1] - src.w[1]);
+    dst->w[2] = USATW(dst->w[2] - src.w[2]);
+    dst->w[3] = USATW(dst->w[3] - src.w[3]);
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -646,15 +830,21 @@ static int
 opPSUBUSW_a32(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].w[0] = USATW(cpu_state.MM[cpu_reg].w[0] - src.w[0]);
-    cpu_state.MM[cpu_reg].w[1] = USATW(cpu_state.MM[cpu_reg].w[1] - src.w[1]);
-    cpu_state.MM[cpu_reg].w[2] = USATW(cpu_state.MM[cpu_reg].w[2] - src.w[2]);
-    cpu_state.MM[cpu_reg].w[3] = USATW(cpu_state.MM[cpu_reg].w[3] - src.w[3]);
+    dst->w[0] = USATW(dst->w[0] - src.w[0]);
+    dst->w[1] = USATW(dst->w[1] - src.w[1]);
+    dst->w[2] = USATW(dst->w[2] - src.w[2]);
+    dst->w[3] = USATW(dst->w[3] - src.w[3]);
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
diff --git a/src/cpu/x86_ops_mmx_cmp.h b/src/cpu/x86_ops_mmx_cmp.h
index 40ae66a9c..b3081b8e8 100644
--- a/src/cpu/x86_ops_mmx_cmp.h
+++ b/src/cpu/x86_ops_mmx_cmp.h
@@ -2,20 +2,25 @@ static int
 opPCMPEQB_a16(uint32_t fetchdat)
 {
     MMX_REG src;
-
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].b[0] = (cpu_state.MM[cpu_reg].b[0] == src.b[0]) ? 0xff : 0;
-    cpu_state.MM[cpu_reg].b[1] = (cpu_state.MM[cpu_reg].b[1] == src.b[1]) ? 0xff : 0;
-    cpu_state.MM[cpu_reg].b[2] = (cpu_state.MM[cpu_reg].b[2] == src.b[2]) ? 0xff : 0;
-    cpu_state.MM[cpu_reg].b[3] = (cpu_state.MM[cpu_reg].b[3] == src.b[3]) ? 0xff : 0;
-    cpu_state.MM[cpu_reg].b[4] = (cpu_state.MM[cpu_reg].b[4] == src.b[4]) ? 0xff : 0;
-    cpu_state.MM[cpu_reg].b[5] = (cpu_state.MM[cpu_reg].b[5] == src.b[5]) ? 0xff : 0;
-    cpu_state.MM[cpu_reg].b[6] = (cpu_state.MM[cpu_reg].b[6] == src.b[6]) ? 0xff : 0;
-    cpu_state.MM[cpu_reg].b[7] = (cpu_state.MM[cpu_reg].b[7] == src.b[7]) ? 0xff : 0;
+    dst->b[0] = (dst->b[0] == src.b[0]) ? 0xff : 0;
+    dst->b[1] = (dst->b[1] == src.b[1]) ? 0xff : 0;
+    dst->b[2] = (dst->b[2] == src.b[2]) ? 0xff : 0;
+    dst->b[3] = (dst->b[3] == src.b[3]) ? 0xff : 0;
+    dst->b[4] = (dst->b[4] == src.b[4]) ? 0xff : 0;
+    dst->b[5] = (dst->b[5] == src.b[5]) ? 0xff : 0;
+    dst->b[6] = (dst->b[6] == src.b[6]) ? 0xff : 0;
+    dst->b[7] = (dst->b[7] == src.b[7]) ? 0xff : 0;
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -23,20 +28,25 @@ static int
 opPCMPEQB_a32(uint32_t fetchdat)
 {
     MMX_REG src;
-
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].b[0] = (cpu_state.MM[cpu_reg].b[0] == src.b[0]) ? 0xff : 0;
-    cpu_state.MM[cpu_reg].b[1] = (cpu_state.MM[cpu_reg].b[1] == src.b[1]) ? 0xff : 0;
-    cpu_state.MM[cpu_reg].b[2] = (cpu_state.MM[cpu_reg].b[2] == src.b[2]) ? 0xff : 0;
-    cpu_state.MM[cpu_reg].b[3] = (cpu_state.MM[cpu_reg].b[3] == src.b[3]) ? 0xff : 0;
-    cpu_state.MM[cpu_reg].b[4] = (cpu_state.MM[cpu_reg].b[4] == src.b[4]) ? 0xff : 0;
-    cpu_state.MM[cpu_reg].b[5] = (cpu_state.MM[cpu_reg].b[5] == src.b[5]) ? 0xff : 0;
-    cpu_state.MM[cpu_reg].b[6] = (cpu_state.MM[cpu_reg].b[6] == src.b[6]) ? 0xff : 0;
-    cpu_state.MM[cpu_reg].b[7] = (cpu_state.MM[cpu_reg].b[7] == src.b[7]) ? 0xff : 0;
+    dst->b[0] = (dst->b[0] == src.b[0]) ? 0xff : 0;
+    dst->b[1] = (dst->b[1] == src.b[1]) ? 0xff : 0;
+    dst->b[2] = (dst->b[2] == src.b[2]) ? 0xff : 0;
+    dst->b[3] = (dst->b[3] == src.b[3]) ? 0xff : 0;
+    dst->b[4] = (dst->b[4] == src.b[4]) ? 0xff : 0;
+    dst->b[5] = (dst->b[5] == src.b[5]) ? 0xff : 0;
+    dst->b[6] = (dst->b[6] == src.b[6]) ? 0xff : 0;
+    dst->b[7] = (dst->b[7] == src.b[7]) ? 0xff : 0;
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -45,20 +55,25 @@ static int
 opPCMPGTB_a16(uint32_t fetchdat)
 {
     MMX_REG src;
-
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].b[0] = (cpu_state.MM[cpu_reg].sb[0] > src.sb[0]) ? 0xff : 0;
-    cpu_state.MM[cpu_reg].b[1] = (cpu_state.MM[cpu_reg].sb[1] > src.sb[1]) ? 0xff : 0;
-    cpu_state.MM[cpu_reg].b[2] = (cpu_state.MM[cpu_reg].sb[2] > src.sb[2]) ? 0xff : 0;
-    cpu_state.MM[cpu_reg].b[3] = (cpu_state.MM[cpu_reg].sb[3] > src.sb[3]) ? 0xff : 0;
-    cpu_state.MM[cpu_reg].b[4] = (cpu_state.MM[cpu_reg].sb[4] > src.sb[4]) ? 0xff : 0;
-    cpu_state.MM[cpu_reg].b[5] = (cpu_state.MM[cpu_reg].sb[5] > src.sb[5]) ? 0xff : 0;
-    cpu_state.MM[cpu_reg].b[6] = (cpu_state.MM[cpu_reg].sb[6] > src.sb[6]) ? 0xff : 0;
-    cpu_state.MM[cpu_reg].b[7] = (cpu_state.MM[cpu_reg].sb[7] > src.sb[7]) ? 0xff : 0;
+    dst->b[0] = (dst->sb[0] > src.sb[0]) ? 0xff : 0;
+    dst->b[1] = (dst->sb[1] > src.sb[1]) ? 0xff : 0;
+    dst->b[2] = (dst->sb[2] > src.sb[2]) ? 0xff : 0;
+    dst->b[3] = (dst->sb[3] > src.sb[3]) ? 0xff : 0;
+    dst->b[4] = (dst->sb[4] > src.sb[4]) ? 0xff : 0;
+    dst->b[5] = (dst->sb[5] > src.sb[5]) ? 0xff : 0;
+    dst->b[6] = (dst->sb[6] > src.sb[6]) ? 0xff : 0;
+    dst->b[7] = (dst->sb[7] > src.sb[7]) ? 0xff : 0;
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -66,20 +81,25 @@ static int
 opPCMPGTB_a32(uint32_t fetchdat)
 {
     MMX_REG src;
-
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].b[0] = (cpu_state.MM[cpu_reg].sb[0] > src.sb[0]) ? 0xff : 0;
-    cpu_state.MM[cpu_reg].b[1] = (cpu_state.MM[cpu_reg].sb[1] > src.sb[1]) ? 0xff : 0;
-    cpu_state.MM[cpu_reg].b[2] = (cpu_state.MM[cpu_reg].sb[2] > src.sb[2]) ? 0xff : 0;
-    cpu_state.MM[cpu_reg].b[3] = (cpu_state.MM[cpu_reg].sb[3] > src.sb[3]) ? 0xff : 0;
-    cpu_state.MM[cpu_reg].b[4] = (cpu_state.MM[cpu_reg].sb[4] > src.sb[4]) ? 0xff : 0;
-    cpu_state.MM[cpu_reg].b[5] = (cpu_state.MM[cpu_reg].sb[5] > src.sb[5]) ? 0xff : 0;
-    cpu_state.MM[cpu_reg].b[6] = (cpu_state.MM[cpu_reg].sb[6] > src.sb[6]) ? 0xff : 0;
-    cpu_state.MM[cpu_reg].b[7] = (cpu_state.MM[cpu_reg].sb[7] > src.sb[7]) ? 0xff : 0;
+    dst->b[0] = (dst->sb[0] > src.sb[0]) ? 0xff : 0;
+    dst->b[1] = (dst->sb[1] > src.sb[1]) ? 0xff : 0;
+    dst->b[2] = (dst->sb[2] > src.sb[2]) ? 0xff : 0;
+    dst->b[3] = (dst->sb[3] > src.sb[3]) ? 0xff : 0;
+    dst->b[4] = (dst->sb[4] > src.sb[4]) ? 0xff : 0;
+    dst->b[5] = (dst->sb[5] > src.sb[5]) ? 0xff : 0;
+    dst->b[6] = (dst->sb[6] > src.sb[6]) ? 0xff : 0;
+    dst->b[7] = (dst->sb[7] > src.sb[7]) ? 0xff : 0;
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -88,16 +108,21 @@ static int
 opPCMPEQW_a16(uint32_t fetchdat)
 {
     MMX_REG src;
-
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].w[0] = (cpu_state.MM[cpu_reg].w[0] == src.w[0]) ? 0xffff : 0;
-    cpu_state.MM[cpu_reg].w[1] = (cpu_state.MM[cpu_reg].w[1] == src.w[1]) ? 0xffff : 0;
-    cpu_state.MM[cpu_reg].w[2] = (cpu_state.MM[cpu_reg].w[2] == src.w[2]) ? 0xffff : 0;
-    cpu_state.MM[cpu_reg].w[3] = (cpu_state.MM[cpu_reg].w[3] == src.w[3]) ? 0xffff : 0;
+    dst->w[0] = (dst->w[0] == src.w[0]) ? 0xffff : 0;
+    dst->w[1] = (dst->w[1] == src.w[1]) ? 0xffff : 0;
+    dst->w[2] = (dst->w[2] == src.w[2]) ? 0xffff : 0;
+    dst->w[3] = (dst->w[3] == src.w[3]) ? 0xffff : 0;
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -105,16 +130,21 @@ static int
 opPCMPEQW_a32(uint32_t fetchdat)
 {
     MMX_REG src;
-
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].w[0] = (cpu_state.MM[cpu_reg].w[0] == src.w[0]) ? 0xffff : 0;
-    cpu_state.MM[cpu_reg].w[1] = (cpu_state.MM[cpu_reg].w[1] == src.w[1]) ? 0xffff : 0;
-    cpu_state.MM[cpu_reg].w[2] = (cpu_state.MM[cpu_reg].w[2] == src.w[2]) ? 0xffff : 0;
-    cpu_state.MM[cpu_reg].w[3] = (cpu_state.MM[cpu_reg].w[3] == src.w[3]) ? 0xffff : 0;
+    dst->w[0] = (dst->w[0] == src.w[0]) ? 0xffff : 0;
+    dst->w[1] = (dst->w[1] == src.w[1]) ? 0xffff : 0;
+    dst->w[2] = (dst->w[2] == src.w[2]) ? 0xffff : 0;
+    dst->w[3] = (dst->w[3] == src.w[3]) ? 0xffff : 0;
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -123,16 +153,21 @@ static int
 opPCMPGTW_a16(uint32_t fetchdat)
 {
     MMX_REG src;
-
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].w[0] = (cpu_state.MM[cpu_reg].sw[0] > src.sw[0]) ? 0xffff : 0;
-    cpu_state.MM[cpu_reg].w[1] = (cpu_state.MM[cpu_reg].sw[1] > src.sw[1]) ? 0xffff : 0;
-    cpu_state.MM[cpu_reg].w[2] = (cpu_state.MM[cpu_reg].sw[2] > src.sw[2]) ? 0xffff : 0;
-    cpu_state.MM[cpu_reg].w[3] = (cpu_state.MM[cpu_reg].sw[3] > src.sw[3]) ? 0xffff : 0;
+    dst->w[0] = (dst->sw[0] > src.sw[0]) ? 0xffff : 0;
+    dst->w[1] = (dst->sw[1] > src.sw[1]) ? 0xffff : 0;
+    dst->w[2] = (dst->sw[2] > src.sw[2]) ? 0xffff : 0;
+    dst->w[3] = (dst->sw[3] > src.sw[3]) ? 0xffff : 0;
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -140,16 +175,21 @@ static int
 opPCMPGTW_a32(uint32_t fetchdat)
 {
     MMX_REG src;
-
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].w[0] = (cpu_state.MM[cpu_reg].sw[0] > src.sw[0]) ? 0xffff : 0;
-    cpu_state.MM[cpu_reg].w[1] = (cpu_state.MM[cpu_reg].sw[1] > src.sw[1]) ? 0xffff : 0;
-    cpu_state.MM[cpu_reg].w[2] = (cpu_state.MM[cpu_reg].sw[2] > src.sw[2]) ? 0xffff : 0;
-    cpu_state.MM[cpu_reg].w[3] = (cpu_state.MM[cpu_reg].sw[3] > src.sw[3]) ? 0xffff : 0;
+    dst->w[0] = (dst->sw[0] > src.sw[0]) ? 0xffff : 0;
+    dst->w[1] = (dst->sw[1] > src.sw[1]) ? 0xffff : 0;
+    dst->w[2] = (dst->sw[2] > src.sw[2]) ? 0xffff : 0;
+    dst->w[3] = (dst->sw[3] > src.sw[3]) ? 0xffff : 0;
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -158,14 +198,19 @@ static int
 opPCMPEQD_a16(uint32_t fetchdat)
 {
     MMX_REG src;
-
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].l[0] = (cpu_state.MM[cpu_reg].l[0] == src.l[0]) ? 0xffffffff : 0;
-    cpu_state.MM[cpu_reg].l[1] = (cpu_state.MM[cpu_reg].l[1] == src.l[1]) ? 0xffffffff : 0;
+    dst->l[0] = (dst->l[0] == src.l[0]) ? 0xffffffff : 0;
+    dst->l[1] = (dst->l[1] == src.l[1]) ? 0xffffffff : 0;
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -173,14 +218,19 @@ static int
 opPCMPEQD_a32(uint32_t fetchdat)
 {
     MMX_REG src;
-
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].l[0] = (cpu_state.MM[cpu_reg].l[0] == src.l[0]) ? 0xffffffff : 0;
-    cpu_state.MM[cpu_reg].l[1] = (cpu_state.MM[cpu_reg].l[1] == src.l[1]) ? 0xffffffff : 0;
+    dst->l[0] = (dst->l[0] == src.l[0]) ? 0xffffffff : 0;
+    dst->l[1] = (dst->l[1] == src.l[1]) ? 0xffffffff : 0;
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -189,14 +239,19 @@ static int
 opPCMPGTD_a16(uint32_t fetchdat)
 {
     MMX_REG src;
-
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].l[0] = (cpu_state.MM[cpu_reg].sl[0] > src.sl[0]) ? 0xffffffff : 0;
-    cpu_state.MM[cpu_reg].l[1] = (cpu_state.MM[cpu_reg].sl[1] > src.sl[1]) ? 0xffffffff : 0;
+    dst->l[0] = (dst->sl[0] > src.sl[0]) ? 0xffffffff : 0;
+    dst->l[1] = (dst->sl[1] > src.sl[1]) ? 0xffffffff : 0;
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -204,14 +259,19 @@ static int
 opPCMPGTD_a32(uint32_t fetchdat)
 {
     MMX_REG src;
-
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].l[0] = (cpu_state.MM[cpu_reg].sl[0] > src.sl[0]) ? 0xffffffff : 0;
-    cpu_state.MM[cpu_reg].l[1] = (cpu_state.MM[cpu_reg].sl[1] > src.sl[1]) ? 0xffffffff : 0;
+    dst->l[0] = (dst->sl[0] > src.sl[0]) ? 0xffffffff : 0;
+    dst->l[1] = (dst->sl[1] > src.sl[1]) ? 0xffffffff : 0;
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
diff --git a/src/cpu/x86_ops_mmx_logic.h b/src/cpu/x86_ops_mmx_logic.h
index c22c820c1..26d7c1693 100644
--- a/src/cpu/x86_ops_mmx_logic.h
+++ b/src/cpu/x86_ops_mmx_logic.h
@@ -2,24 +2,38 @@ static int
 opPAND_a16(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].q &= src.q;
+    dst->q &= src.q;
+
+    MMX_SETEXP(cpu_reg);
+
     return 0;
 }
 static int
 opPAND_a32(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].q &= src.q;
+    dst->q &= src.q;
+
+    MMX_SETEXP(cpu_reg);
+
     return 0;
 }
 
@@ -27,24 +41,38 @@ static int
 opPANDN_a16(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].q = ~cpu_state.MM[cpu_reg].q & src.q;
+    dst->q = ~dst->q & src.q;
+
+    MMX_SETEXP(cpu_reg);
+
     return 0;
 }
 static int
 opPANDN_a32(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].q = ~cpu_state.MM[cpu_reg].q & src.q;
+    dst->q = ~dst->q & src.q;
+
+    MMX_SETEXP(cpu_reg);
+
     return 0;
 }
 
@@ -52,24 +80,38 @@ static int
 opPOR_a16(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].q |= src.q;
+    dst->q |= src.q;
+
+    MMX_SETEXP(cpu_reg);
+
     return 0;
 }
 static int
 opPOR_a32(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].q |= src.q;
+    dst->q |= src.q;
+
+    MMX_SETEXP(cpu_reg);
+
     return 0;
 }
 
@@ -77,23 +119,37 @@ static int
 opPXOR_a16(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].q ^= src.q;
+    dst->q ^= src.q;
+
+    MMX_SETEXP(cpu_reg);
+
     return 0;
 }
 static int
 opPXOR_a32(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].q ^= src.q;
+    dst->q ^= src.q;
+
+    MMX_SETEXP(cpu_reg);
+
     return 0;
 }
diff --git a/src/cpu/x86_ops_mmx_mov.h b/src/cpu/x86_ops_mmx_mov.h
index bb51573e6..65bbb0c01 100644
--- a/src/cpu/x86_ops_mmx_mov.h
+++ b/src/cpu/x86_ops_mmx_mov.h
@@ -1,88 +1,112 @@
 static int
 opMOVD_l_mm_a16(uint32_t fetchdat)
 {
+    uint32_t dst;
+    MMX_REG *op;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    op = MMX_GETREGP(cpu_reg);
+
     if (cpu_mod == 3) {
-        cpu_state.MM[cpu_reg].l[0] = cpu_state.regs[cpu_rm].l;
-        cpu_state.MM[cpu_reg].l[1] = 0;
+        op->l[0] = cpu_state.regs[cpu_rm].l;
+        op->l[1] = 0;
         CLOCK_CYCLES(1);
     } else {
-        uint32_t dst;
-
         SEG_CHECK_READ(cpu_state.ea_seg);
         dst = readmeml(easeg, cpu_state.eaaddr);
         if (cpu_state.abrt)
             return 1;
-        cpu_state.MM[cpu_reg].l[0] = dst;
-        cpu_state.MM[cpu_reg].l[1] = 0;
 
+        op->l[0] = dst;
+        op->l[1] = 0;
         CLOCK_CYCLES(2);
     }
+
+    MMX_SETEXP(cpu_reg);
+
     return 0;
 }
 static int
 opMOVD_l_mm_a32(uint32_t fetchdat)
 {
+    uint32_t dst;
+    MMX_REG *op;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    op = MMX_GETREGP(cpu_reg);
+
     if (cpu_mod == 3) {
-        cpu_state.MM[cpu_reg].l[0] = cpu_state.regs[cpu_rm].l;
-        cpu_state.MM[cpu_reg].l[1] = 0;
+        op->l[0] = cpu_state.regs[cpu_rm].l;
+        op->l[1] = 0;
         CLOCK_CYCLES(1);
     } else {
-        uint32_t dst;
-
         SEG_CHECK_READ(cpu_state.ea_seg);
         dst = readmeml(easeg, cpu_state.eaaddr);
         if (cpu_state.abrt)
             return 1;
-        cpu_state.MM[cpu_reg].l[0] = dst;
-        cpu_state.MM[cpu_reg].l[1] = 0;
 
+        op->l[0] = dst;
+        op->l[1] = 0;
         CLOCK_CYCLES(2);
     }
+
+    MMX_SETEXP(cpu_reg);
+
     return 0;
 }
 
 static int
 opMOVD_mm_l_a16(uint32_t fetchdat)
 {
+    MMX_REG *op;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    op = MMX_GETREGP(cpu_reg);
+
     if (cpu_mod == 3) {
-        cpu_state.regs[cpu_rm].l = cpu_state.MM[cpu_reg].l[0];
+        cpu_state.regs[cpu_rm].l = op->l[0];
         CLOCK_CYCLES(1);
     } else {
         SEG_CHECK_WRITE(cpu_state.ea_seg);
         CHECK_WRITE_COMMON(cpu_state.ea_seg, cpu_state.eaaddr, cpu_state.eaaddr + 3);
-        writememl(easeg, cpu_state.eaaddr, cpu_state.MM[cpu_reg].l[0]);
+        writememl(easeg, cpu_state.eaaddr, op->l[0]);
         if (cpu_state.abrt)
             return 1;
+
         CLOCK_CYCLES(2);
     }
+
     return 0;
 }
 static int
 opMOVD_mm_l_a32(uint32_t fetchdat)
 {
+    MMX_REG *op;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    op = MMX_GETREGP(cpu_reg);
+
     if (cpu_mod == 3) {
-        cpu_state.regs[cpu_rm].l = cpu_state.MM[cpu_reg].l[0];
+        cpu_state.regs[cpu_rm].l = op->l[0];
         CLOCK_CYCLES(1);
     } else {
         SEG_CHECK_WRITE(cpu_state.ea_seg);
         CHECK_WRITE_COMMON(cpu_state.ea_seg, cpu_state.eaaddr, cpu_state.eaaddr + 3);
-        writememl(easeg, cpu_state.eaaddr, cpu_state.MM[cpu_reg].l[0]);
+        writememl(easeg, cpu_state.eaaddr, op->l[0]);
         if (cpu_state.abrt)
             return 1;
+
         CLOCK_CYCLES(2);
     }
+
     return 0;
 }
 
@@ -91,45 +115,59 @@ opMOVD_mm_l_a32(uint32_t fetchdat)
 static int
 opMOVD_mm_l_a16_cx(uint32_t fetchdat)
 {
+    MMX_REG *op;
+
     if (in_smm)
         return opSMINT(fetchdat);
 
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    op = MMX_GETREGP(cpu_reg);
+
     if (cpu_mod == 3) {
-        cpu_state.regs[cpu_rm].l = cpu_state.MM[cpu_reg].l[0];
+        cpu_state.regs[cpu_rm].l = op->l[0];
         CLOCK_CYCLES(1);
     } else {
         SEG_CHECK_WRITE(cpu_state.ea_seg);
         CHECK_WRITE_COMMON(cpu_state.ea_seg, cpu_state.eaaddr, cpu_state.eaaddr + 3);
-        writememl(easeg, cpu_state.eaaddr, cpu_state.MM[cpu_reg].l[0]);
+        writememl(easeg, cpu_state.eaaddr, op->l[0]);
         if (cpu_state.abrt)
             return 1;
+
         CLOCK_CYCLES(2);
     }
+
     return 0;
 }
 static int
 opMOVD_mm_l_a32_cx(uint32_t fetchdat)
 {
+    MMX_REG *op;
+
     if (in_smm)
         return opSMINT(fetchdat);
 
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    op = MMX_GETREGP(cpu_reg);
+
     if (cpu_mod == 3) {
-        cpu_state.regs[cpu_rm].l = cpu_state.MM[cpu_reg].l[0];
+        cpu_state.regs[cpu_rm].l = op->l[0];
         CLOCK_CYCLES(1);
     } else {
         SEG_CHECK_WRITE(cpu_state.ea_seg);
         CHECK_WRITE_COMMON(cpu_state.ea_seg, cpu_state.eaaddr, cpu_state.eaaddr + 3);
-        writememl(easeg, cpu_state.eaaddr, cpu_state.MM[cpu_reg].l[0]);
+        writememl(easeg, cpu_state.eaaddr, op->l[0]);
         if (cpu_state.abrt)
             return 1;
+
         CLOCK_CYCLES(2);
     }
+
     return 0;
 }
 #endif
@@ -137,81 +175,121 @@ opMOVD_mm_l_a32_cx(uint32_t fetchdat)
 static int
 opMOVQ_q_mm_a16(uint32_t fetchdat)
 {
+    uint64_t dst;
+    MMX_REG src;
+    MMX_REG *op;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    src = MMX_GETREG(cpu_rm);
+    op = MMX_GETREGP(cpu_reg);
+
     if (cpu_mod == 3) {
-        cpu_state.MM[cpu_reg].q = cpu_state.MM[cpu_rm].q;
+        op->q = src.q;
         CLOCK_CYCLES(1);
     } else {
-        uint64_t dst;
-
         SEG_CHECK_READ(cpu_state.ea_seg);
         dst = readmemq(easeg, cpu_state.eaaddr);
         if (cpu_state.abrt)
             return 1;
-        cpu_state.MM[cpu_reg].q = dst;
+
+        op->q = dst;
         CLOCK_CYCLES(2);
     }
+
+    MMX_SETEXP(cpu_reg);
+
     return 0;
 }
 static int
 opMOVQ_q_mm_a32(uint32_t fetchdat)
 {
+    uint64_t dst;
+    MMX_REG src;
+    MMX_REG *op;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    src = MMX_GETREG(cpu_rm);
+    op = MMX_GETREGP(cpu_reg);
+
     if (cpu_mod == 3) {
-        cpu_state.MM[cpu_reg].q = cpu_state.MM[cpu_rm].q;
+        op->q = src.q;
         CLOCK_CYCLES(1);
     } else {
-        uint64_t dst;
-
         SEG_CHECK_READ(cpu_state.ea_seg);
         dst = readmemq(easeg, cpu_state.eaaddr);
         if (cpu_state.abrt)
             return 1;
-        cpu_state.MM[cpu_reg].q = dst;
+
+        op->q = dst;
         CLOCK_CYCLES(2);
     }
+
+    MMX_SETEXP(cpu_reg);
+
     return 0;
 }
 
 static int
 opMOVQ_mm_q_a16(uint32_t fetchdat)
 {
+    MMX_REG src;
+    MMX_REG *dst;
+
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    src = MMX_GETREG(cpu_reg);
+    dst = MMX_GETREGP(cpu_rm);
+
     if (cpu_mod == 3) {
-        cpu_state.MM[cpu_rm].q = cpu_state.MM[cpu_reg].q;
+        dst->q = src.q;
         CLOCK_CYCLES(1);
+
+        MMX_SETEXP(cpu_rm);
     } else {
         SEG_CHECK_WRITE(cpu_state.ea_seg);
         CHECK_WRITE_COMMON(cpu_state.ea_seg, cpu_state.eaaddr, cpu_state.eaaddr + 7);
-        writememq(easeg, cpu_state.eaaddr, cpu_state.MM[cpu_reg].q);
+        writememq(easeg, cpu_state.eaaddr, src.q);
         if (cpu_state.abrt)
             return 1;
+
         CLOCK_CYCLES(2);
     }
+
     return 0;
 }
 static int
 opMOVQ_mm_q_a32(uint32_t fetchdat)
 {
+    MMX_REG src;
+    MMX_REG *dst;
+
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    src = MMX_GETREG(cpu_reg);
+    dst = MMX_GETREGP(cpu_rm);
+
     if (cpu_mod == 3) {
-        cpu_state.MM[cpu_rm].q = cpu_state.MM[cpu_reg].q;
+        dst->q = src.q;
         CLOCK_CYCLES(1);
+
+        MMX_SETEXP(cpu_rm);
     } else {
         SEG_CHECK_WRITE(cpu_state.ea_seg);
         CHECK_WRITE_COMMON(cpu_state.ea_seg, cpu_state.eaaddr, cpu_state.eaaddr + 7);
-        writememq(easeg, cpu_state.eaaddr, cpu_state.MM[cpu_reg].q);
+        writememq(easeg, cpu_state.eaaddr, src.q);
         if (cpu_state.abrt)
             return 1;
+
         CLOCK_CYCLES(2);
     }
+
     return 0;
 }
diff --git a/src/cpu/x86_ops_mmx_pack.h b/src/cpu/x86_ops_mmx_pack.h
index f0180db91..90590638b 100644
--- a/src/cpu/x86_ops_mmx_pack.h
+++ b/src/cpu/x86_ops_mmx_pack.h
@@ -1,45 +1,61 @@
 static int
 opPUNPCKLDQ_a16(uint32_t fetchdat)
 {
+    uint32_t usrc;
+    MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    src = MMX_GETREG(cpu_rm);
+    dst = MMX_GETREGP(cpu_reg);
+
     if (cpu_mod == 3) {
-        cpu_state.MM[cpu_reg].l[1] = cpu_state.MM[cpu_rm].l[0];
+        dst->l[1] = src.l[0];
         CLOCK_CYCLES(1);
     } else {
-        uint32_t src;
-
         SEG_CHECK_READ(cpu_state.ea_seg);
-        src = readmeml(easeg, cpu_state.eaaddr);
+        usrc = readmeml(easeg, cpu_state.eaaddr);
         if (cpu_state.abrt)
             return 0;
-        cpu_state.MM[cpu_reg].l[1] = src;
+        dst->l[1] = usrc;
 
         CLOCK_CYCLES(2);
     }
+
+    MMX_SETEXP(cpu_reg);
+
     return 0;
 }
 static int
 opPUNPCKLDQ_a32(uint32_t fetchdat)
 {
+    uint32_t usrc;
+    MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    src = MMX_GETREG(cpu_rm);
+    dst = MMX_GETREGP(cpu_reg);
+
     if (cpu_mod == 3) {
-        cpu_state.MM[cpu_reg].l[1] = cpu_state.MM[cpu_rm].l[0];
+        dst->l[1] = src.l[0];
         CLOCK_CYCLES(1);
     } else {
-        uint32_t src;
-
         SEG_CHECK_READ(cpu_state.ea_seg);
-        src = readmeml(easeg, cpu_state.eaaddr);
+        usrc = readmeml(easeg, cpu_state.eaaddr);
         if (cpu_state.abrt)
             return 0;
-        cpu_state.MM[cpu_reg].l[1] = src;
+        dst->l[1] = usrc;
 
         CLOCK_CYCLES(2);
     }
+
+    MMX_SETEXP(cpu_reg);
+
     return 0;
 }
 
@@ -47,13 +63,19 @@ static int
 opPUNPCKHDQ_a16(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].l[0] = cpu_state.MM[cpu_reg].l[1];
-    cpu_state.MM[cpu_reg].l[1] = src.l[1];
+    dst->l[0] = dst->l[1];
+    dst->l[1] = src.l[1];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -61,13 +83,19 @@ static int
 opPUNPCKHDQ_a32(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].l[0] = cpu_state.MM[cpu_reg].l[1];
-    cpu_state.MM[cpu_reg].l[1] = src.l[1];
+    dst->l[0] = dst->l[1];
+    dst->l[1] = src.l[1];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -76,19 +104,25 @@ static int
 opPUNPCKLBW_a16(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].b[7] = src.b[3];
-    cpu_state.MM[cpu_reg].b[6] = cpu_state.MM[cpu_reg].b[3];
-    cpu_state.MM[cpu_reg].b[5] = src.b[2];
-    cpu_state.MM[cpu_reg].b[4] = cpu_state.MM[cpu_reg].b[2];
-    cpu_state.MM[cpu_reg].b[3] = src.b[1];
-    cpu_state.MM[cpu_reg].b[2] = cpu_state.MM[cpu_reg].b[1];
-    cpu_state.MM[cpu_reg].b[1] = src.b[0];
-    cpu_state.MM[cpu_reg].b[0] = cpu_state.MM[cpu_reg].b[0];
+    dst->b[7] = src.b[3];
+    dst->b[6] = dst->b[3];
+    dst->b[5] = src.b[2];
+    dst->b[4] = dst->b[2];
+    dst->b[3] = src.b[1];
+    dst->b[2] = dst->b[1];
+    dst->b[1] = src.b[0];
+    dst->b[0] = dst->b[0];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -96,19 +130,25 @@ static int
 opPUNPCKLBW_a32(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].b[7] = src.b[3];
-    cpu_state.MM[cpu_reg].b[6] = cpu_state.MM[cpu_reg].b[3];
-    cpu_state.MM[cpu_reg].b[5] = src.b[2];
-    cpu_state.MM[cpu_reg].b[4] = cpu_state.MM[cpu_reg].b[2];
-    cpu_state.MM[cpu_reg].b[3] = src.b[1];
-    cpu_state.MM[cpu_reg].b[2] = cpu_state.MM[cpu_reg].b[1];
-    cpu_state.MM[cpu_reg].b[1] = src.b[0];
-    cpu_state.MM[cpu_reg].b[0] = cpu_state.MM[cpu_reg].b[0];
+    dst->b[7] = src.b[3];
+    dst->b[6] = dst->b[3];
+    dst->b[5] = src.b[2];
+    dst->b[4] = dst->b[2];
+    dst->b[3] = src.b[1];
+    dst->b[2] = dst->b[1];
+    dst->b[1] = src.b[0];
+    dst->b[0] = dst->b[0];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -117,19 +157,25 @@ static int
 opPUNPCKHBW_a16(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].b[0] = cpu_state.MM[cpu_reg].b[4];
-    cpu_state.MM[cpu_reg].b[1] = src.b[4];
-    cpu_state.MM[cpu_reg].b[2] = cpu_state.MM[cpu_reg].b[5];
-    cpu_state.MM[cpu_reg].b[3] = src.b[5];
-    cpu_state.MM[cpu_reg].b[4] = cpu_state.MM[cpu_reg].b[6];
-    cpu_state.MM[cpu_reg].b[5] = src.b[6];
-    cpu_state.MM[cpu_reg].b[6] = cpu_state.MM[cpu_reg].b[7];
-    cpu_state.MM[cpu_reg].b[7] = src.b[7];
+    dst->b[0] = dst->b[4];
+    dst->b[1] = src.b[4];
+    dst->b[2] = dst->b[5];
+    dst->b[3] = src.b[5];
+    dst->b[4] = dst->b[6];
+    dst->b[5] = src.b[6];
+    dst->b[6] = dst->b[7];
+    dst->b[7] = src.b[7];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -137,19 +183,25 @@ static int
 opPUNPCKHBW_a32(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].b[0] = cpu_state.MM[cpu_reg].b[4];
-    cpu_state.MM[cpu_reg].b[1] = src.b[4];
-    cpu_state.MM[cpu_reg].b[2] = cpu_state.MM[cpu_reg].b[5];
-    cpu_state.MM[cpu_reg].b[3] = src.b[5];
-    cpu_state.MM[cpu_reg].b[4] = cpu_state.MM[cpu_reg].b[6];
-    cpu_state.MM[cpu_reg].b[5] = src.b[6];
-    cpu_state.MM[cpu_reg].b[6] = cpu_state.MM[cpu_reg].b[7];
-    cpu_state.MM[cpu_reg].b[7] = src.b[7];
+    dst->b[0] = dst->b[4];
+    dst->b[1] = src.b[4];
+    dst->b[2] = dst->b[5];
+    dst->b[3] = src.b[5];
+    dst->b[4] = dst->b[6];
+    dst->b[5] = src.b[6];
+    dst->b[6] = dst->b[7];
+    dst->b[7] = src.b[7];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -158,15 +210,21 @@ static int
 opPUNPCKLWD_a16(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].w[3] = src.w[1];
-    cpu_state.MM[cpu_reg].w[2] = cpu_state.MM[cpu_reg].w[1];
-    cpu_state.MM[cpu_reg].w[1] = src.w[0];
-    cpu_state.MM[cpu_reg].w[0] = cpu_state.MM[cpu_reg].w[0];
+    dst->w[3] = src.w[1];
+    dst->w[2] = dst->w[1];
+    dst->w[1] = src.w[0];
+    dst->w[0] = dst->w[0];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -174,15 +232,21 @@ static int
 opPUNPCKLWD_a32(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].w[3] = src.w[1];
-    cpu_state.MM[cpu_reg].w[2] = cpu_state.MM[cpu_reg].w[1];
-    cpu_state.MM[cpu_reg].w[1] = src.w[0];
-    cpu_state.MM[cpu_reg].w[0] = cpu_state.MM[cpu_reg].w[0];
+    dst->w[3] = src.w[1];
+    dst->w[2] = dst->w[1];
+    dst->w[1] = src.w[0];
+    dst->w[0] = dst->w[0];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -191,15 +255,21 @@ static int
 opPUNPCKHWD_a16(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].w[0] = cpu_state.MM[cpu_reg].w[2];
-    cpu_state.MM[cpu_reg].w[1] = src.w[2];
-    cpu_state.MM[cpu_reg].w[2] = cpu_state.MM[cpu_reg].w[3];
-    cpu_state.MM[cpu_reg].w[3] = src.w[3];
+    dst->w[0] = dst->w[2];
+    dst->w[1] = src.w[2];
+    dst->w[2] = dst->w[3];
+    dst->w[3] = src.w[3];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -207,15 +277,21 @@ static int
 opPUNPCKHWD_a32(uint32_t fetchdat)
 {
     MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSRC();
 
-    cpu_state.MM[cpu_reg].w[0] = cpu_state.MM[cpu_reg].w[2];
-    cpu_state.MM[cpu_reg].w[1] = src.w[2];
-    cpu_state.MM[cpu_reg].w[2] = cpu_state.MM[cpu_reg].w[3];
-    cpu_state.MM[cpu_reg].w[3] = src.w[3];
+    dst->w[0] = dst->w[2];
+    dst->w[1] = src.w[2];
+    dst->w[2] = dst->w[3];
+    dst->w[3] = src.w[3];
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -223,42 +299,52 @@ opPUNPCKHWD_a32(uint32_t fetchdat)
 static int
 opPACKSSWB_a16(uint32_t fetchdat)
 {
-    MMX_REG src, dst;
+    MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
-    MMX_GETSRC();
-    dst = cpu_state.MM[cpu_reg];
 
-    cpu_state.MM[cpu_reg].sb[0] = SSATB(dst.sw[0]);
-    cpu_state.MM[cpu_reg].sb[1] = SSATB(dst.sw[1]);
-    cpu_state.MM[cpu_reg].sb[2] = SSATB(dst.sw[2]);
-    cpu_state.MM[cpu_reg].sb[3] = SSATB(dst.sw[3]);
-    cpu_state.MM[cpu_reg].sb[4] = SSATB(src.sw[0]);
-    cpu_state.MM[cpu_reg].sb[5] = SSATB(src.sw[1]);
-    cpu_state.MM[cpu_reg].sb[6] = SSATB(src.sw[2]);
-    cpu_state.MM[cpu_reg].sb[7] = SSATB(src.sw[3]);
+    dst = MMX_GETREGP(cpu_reg);
+
+    MMX_GETSRC();
+
+    dst->sb[0] = SSATB(dst->sw[0]);
+    dst->sb[1] = SSATB(dst->sw[1]);
+    dst->sb[2] = SSATB(dst->sw[2]);
+    dst->sb[3] = SSATB(dst->sw[3]);
+    dst->sb[4] = SSATB(src.sw[0]);
+    dst->sb[5] = SSATB(src.sw[1]);
+    dst->sb[6] = SSATB(src.sw[2]);
+    dst->sb[7] = SSATB(src.sw[3]);
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
 static int
 opPACKSSWB_a32(uint32_t fetchdat)
 {
-    MMX_REG src, dst;
+    MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
-    MMX_GETSRC();
-    dst = cpu_state.MM[cpu_reg];
 
-    cpu_state.MM[cpu_reg].sb[0] = SSATB(dst.sw[0]);
-    cpu_state.MM[cpu_reg].sb[1] = SSATB(dst.sw[1]);
-    cpu_state.MM[cpu_reg].sb[2] = SSATB(dst.sw[2]);
-    cpu_state.MM[cpu_reg].sb[3] = SSATB(dst.sw[3]);
-    cpu_state.MM[cpu_reg].sb[4] = SSATB(src.sw[0]);
-    cpu_state.MM[cpu_reg].sb[5] = SSATB(src.sw[1]);
-    cpu_state.MM[cpu_reg].sb[6] = SSATB(src.sw[2]);
-    cpu_state.MM[cpu_reg].sb[7] = SSATB(src.sw[3]);
+    dst = MMX_GETREGP(cpu_reg);
+
+    MMX_GETSRC();
+
+    dst->sb[0] = SSATB(dst->sw[0]);
+    dst->sb[1] = SSATB(dst->sw[1]);
+    dst->sb[2] = SSATB(dst->sw[2]);
+    dst->sb[3] = SSATB(dst->sw[3]);
+    dst->sb[4] = SSATB(src.sw[0]);
+    dst->sb[5] = SSATB(src.sw[1]);
+    dst->sb[6] = SSATB(src.sw[2]);
+    dst->sb[7] = SSATB(src.sw[3]);
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -266,42 +352,52 @@ opPACKSSWB_a32(uint32_t fetchdat)
 static int
 opPACKUSWB_a16(uint32_t fetchdat)
 {
-    MMX_REG src, dst;
+    MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
-    MMX_GETSRC();
-    dst = cpu_state.MM[cpu_reg];
 
-    cpu_state.MM[cpu_reg].b[0] = USATB(dst.sw[0]);
-    cpu_state.MM[cpu_reg].b[1] = USATB(dst.sw[1]);
-    cpu_state.MM[cpu_reg].b[2] = USATB(dst.sw[2]);
-    cpu_state.MM[cpu_reg].b[3] = USATB(dst.sw[3]);
-    cpu_state.MM[cpu_reg].b[4] = USATB(src.sw[0]);
-    cpu_state.MM[cpu_reg].b[5] = USATB(src.sw[1]);
-    cpu_state.MM[cpu_reg].b[6] = USATB(src.sw[2]);
-    cpu_state.MM[cpu_reg].b[7] = USATB(src.sw[3]);
+    dst = MMX_GETREGP(cpu_reg);
+
+    MMX_GETSRC();
+
+    dst->b[0] = USATB(dst->sw[0]);
+    dst->b[1] = USATB(dst->sw[1]);
+    dst->b[2] = USATB(dst->sw[2]);
+    dst->b[3] = USATB(dst->sw[3]);
+    dst->b[4] = USATB(src.sw[0]);
+    dst->b[5] = USATB(src.sw[1]);
+    dst->b[6] = USATB(src.sw[2]);
+    dst->b[7] = USATB(src.sw[3]);
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
 static int
 opPACKUSWB_a32(uint32_t fetchdat)
 {
-    MMX_REG src, dst;
+    MMX_REG src;
+    MMX_REG *dst;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
-    MMX_GETSRC();
-    dst = cpu_state.MM[cpu_reg];
 
-    cpu_state.MM[cpu_reg].b[0] = USATB(dst.sw[0]);
-    cpu_state.MM[cpu_reg].b[1] = USATB(dst.sw[1]);
-    cpu_state.MM[cpu_reg].b[2] = USATB(dst.sw[2]);
-    cpu_state.MM[cpu_reg].b[3] = USATB(dst.sw[3]);
-    cpu_state.MM[cpu_reg].b[4] = USATB(src.sw[0]);
-    cpu_state.MM[cpu_reg].b[5] = USATB(src.sw[1]);
-    cpu_state.MM[cpu_reg].b[6] = USATB(src.sw[2]);
-    cpu_state.MM[cpu_reg].b[7] = USATB(src.sw[3]);
+    dst = MMX_GETREGP(cpu_reg);
+
+    MMX_GETSRC();
+
+    dst->b[0] = USATB(dst->sw[0]);
+    dst->b[1] = USATB(dst->sw[1]);
+    dst->b[2] = USATB(dst->sw[2]);
+    dst->b[3] = USATB(dst->sw[3]);
+    dst->b[4] = USATB(src.sw[0]);
+    dst->b[5] = USATB(src.sw[1]);
+    dst->b[6] = USATB(src.sw[2]);
+    dst->b[7] = USATB(src.sw[3]);
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -309,34 +405,48 @@ opPACKUSWB_a32(uint32_t fetchdat)
 static int
 opPACKSSDW_a16(uint32_t fetchdat)
 {
-    MMX_REG src, dst;
+    MMX_REG src;
+    MMX_REG *dst;
+    MMX_REG dst2;
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
-    MMX_GETSRC();
-    dst = cpu_state.MM[cpu_reg];
 
-    cpu_state.MM[cpu_reg].sw[0] = SSATW(dst.sl[0]);
-    cpu_state.MM[cpu_reg].sw[1] = SSATW(dst.sl[1]);
-    cpu_state.MM[cpu_reg].sw[2] = SSATW(src.sl[0]);
-    cpu_state.MM[cpu_reg].sw[3] = SSATW(src.sl[1]);
+    dst = MMX_GETREGP(cpu_reg);
+    dst2 = *dst;
+
+    MMX_GETSRC();
+
+    dst->sw[0] = SSATW(dst2.sl[0]);
+    dst->sw[1] = SSATW(dst2.sl[1]);
+    dst->sw[2] = SSATW(src.sl[0]);
+    dst->sw[3] = SSATW(src.sl[1]);
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
 static int
 opPACKSSDW_a32(uint32_t fetchdat)
 {
-    MMX_REG src, dst;
+    MMX_REG src;
+    MMX_REG *dst;
+    MMX_REG dst2;
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
-    MMX_GETSRC();
-    dst = cpu_state.MM[cpu_reg];
 
-    cpu_state.MM[cpu_reg].sw[0] = SSATW(dst.sl[0]);
-    cpu_state.MM[cpu_reg].sw[1] = SSATW(dst.sl[1]);
-    cpu_state.MM[cpu_reg].sw[2] = SSATW(src.sl[0]);
-    cpu_state.MM[cpu_reg].sw[3] = SSATW(src.sl[1]);
+    dst = MMX_GETREGP(cpu_reg);
+    dst2 = *dst;
+
+    MMX_GETSRC();
+
+    dst->sw[0] = SSATW(dst2.sl[0]);
+    dst->sw[1] = SSATW(dst2.sl[1]);
+    dst->sw[2] = SSATW(src.sl[0]);
+    dst->sw[3] = SSATW(src.sl[1]);
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
diff --git a/src/cpu/x86_ops_mmx_shift.h b/src/cpu/x86_ops_mmx_shift.h
index c9ddc9b93..c0c80e87e 100644
--- a/src/cpu/x86_ops_mmx_shift.h
+++ b/src/cpu/x86_ops_mmx_shift.h
@@ -1,6 +1,6 @@
 #define MMX_GETSHIFT()                             \
     if (cpu_mod == 3) {                            \
-        shift = cpu_state.MM[cpu_rm].b[0];         \
+        shift = (MMX_GETREG(cpu_rm)).b[0];         \
         CLOCK_CYCLES(1);                           \
     } else {                                       \
         SEG_CHECK_READ(cpu_state.ea_seg);          \
@@ -16,37 +16,39 @@ opPSxxW_imm(uint32_t fetchdat)
     int reg   = fetchdat & 7;
     int op    = fetchdat & 0x38;
     int shift = (fetchdat >> 8) & 0xff;
+    MMX_REG *dst;
 
     cpu_state.pc += 2;
     MMX_ENTER();
+    dst = MMX_GETREGP(reg);
 
     switch (op) {
         case 0x10: /*PSRLW*/
             if (shift > 15)
-                cpu_state.MM[reg].q = 0;
+                dst->q = 0;
             else {
-                cpu_state.MM[reg].w[0] >>= shift;
-                cpu_state.MM[reg].w[1] >>= shift;
-                cpu_state.MM[reg].w[2] >>= shift;
-                cpu_state.MM[reg].w[3] >>= shift;
+                dst->w[0] >>= shift;
+                dst->w[1] >>= shift;
+                dst->w[2] >>= shift;
+                dst->w[3] >>= shift;
             }
             break;
         case 0x20: /*PSRAW*/
             if (shift > 15)
                 shift = 15;
-            cpu_state.MM[reg].sw[0] >>= shift;
-            cpu_state.MM[reg].sw[1] >>= shift;
-            cpu_state.MM[reg].sw[2] >>= shift;
-            cpu_state.MM[reg].sw[3] >>= shift;
+            dst->sw[0] >>= shift;
+            dst->sw[1] >>= shift;
+            dst->sw[2] >>= shift;
+            dst->sw[3] >>= shift;
             break;
         case 0x30: /*PSLLW*/
             if (shift > 15)
-                cpu_state.MM[reg].q = 0;
+                dst->q = 0;
             else {
-                cpu_state.MM[reg].w[0] <<= shift;
-                cpu_state.MM[reg].w[1] <<= shift;
-                cpu_state.MM[reg].w[2] <<= shift;
-                cpu_state.MM[reg].w[3] <<= shift;
+                dst->w[0] <<= shift;
+                dst->w[1] <<= shift;
+                dst->w[2] <<= shift;
+                dst->w[3] <<= shift;
             }
             break;
         default:
@@ -55,6 +57,8 @@ opPSxxW_imm(uint32_t fetchdat)
             return 0;
     }
 
+    MMX_SETEXP(reg);
+
     CLOCK_CYCLES(1);
     return 0;
 }
@@ -62,126 +66,162 @@ opPSxxW_imm(uint32_t fetchdat)
 static int
 opPSLLW_a16(uint32_t fetchdat)
 {
+    MMX_REG *dst;
     int shift;
 
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSHIFT();
 
     if (shift > 15)
-        cpu_state.MM[cpu_reg].q = 0;
+        dst->q = 0;
     else {
-        cpu_state.MM[cpu_reg].w[0] <<= shift;
-        cpu_state.MM[cpu_reg].w[1] <<= shift;
-        cpu_state.MM[cpu_reg].w[2] <<= shift;
-        cpu_state.MM[cpu_reg].w[3] <<= shift;
+        dst->w[0] <<= shift;
+        dst->w[1] <<= shift;
+        dst->w[2] <<= shift;
+        dst->w[3] <<= shift;
     }
 
+    MMX_SETEXP(cpu_reg);
+
     return 0;
 }
 static int
 opPSLLW_a32(uint32_t fetchdat)
 {
+    MMX_REG *dst;
     int shift;
 
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSHIFT();
 
     if (shift > 15)
-        cpu_state.MM[cpu_reg].q = 0;
+        dst->q = 0;
     else {
-        cpu_state.MM[cpu_reg].w[0] <<= shift;
-        cpu_state.MM[cpu_reg].w[1] <<= shift;
-        cpu_state.MM[cpu_reg].w[2] <<= shift;
-        cpu_state.MM[cpu_reg].w[3] <<= shift;
+        dst->w[0] <<= shift;
+        dst->w[1] <<= shift;
+        dst->w[2] <<= shift;
+        dst->w[3] <<= shift;
     }
 
+    MMX_SETEXP(cpu_reg);
+
     return 0;
 }
 
 static int
 opPSRLW_a16(uint32_t fetchdat)
 {
+    MMX_REG *dst;
     int shift;
 
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSHIFT();
 
     if (shift > 15)
-        cpu_state.MM[cpu_reg].q = 0;
+        dst->q = 0;
     else {
-        cpu_state.MM[cpu_reg].w[0] >>= shift;
-        cpu_state.MM[cpu_reg].w[1] >>= shift;
-        cpu_state.MM[cpu_reg].w[2] >>= shift;
-        cpu_state.MM[cpu_reg].w[3] >>= shift;
+        dst->w[0] >>= shift;
+        dst->w[1] >>= shift;
+        dst->w[2] >>= shift;
+        dst->w[3] >>= shift;
     }
 
+    MMX_SETEXP(cpu_reg);
+
     return 0;
 }
 static int
 opPSRLW_a32(uint32_t fetchdat)
 {
+    MMX_REG *dst;
     int shift;
 
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSHIFT();
 
     if (shift > 15)
-        cpu_state.MM[cpu_reg].q = 0;
+        dst->q = 0;
     else {
-        cpu_state.MM[cpu_reg].w[0] >>= shift;
-        cpu_state.MM[cpu_reg].w[1] >>= shift;
-        cpu_state.MM[cpu_reg].w[2] >>= shift;
-        cpu_state.MM[cpu_reg].w[3] >>= shift;
+        dst->w[0] >>= shift;
+        dst->w[1] >>= shift;
+        dst->w[2] >>= shift;
+        dst->w[3] >>= shift;
     }
 
+    MMX_SETEXP(cpu_reg);
+
     return 0;
 }
 
 static int
 opPSRAW_a16(uint32_t fetchdat)
 {
+    MMX_REG *dst;
     int shift;
 
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSHIFT();
 
     if (shift > 15)
         shift = 15;
 
-    cpu_state.MM[cpu_reg].sw[0] >>= shift;
-    cpu_state.MM[cpu_reg].sw[1] >>= shift;
-    cpu_state.MM[cpu_reg].sw[2] >>= shift;
-    cpu_state.MM[cpu_reg].sw[3] >>= shift;
+    dst->sw[0] >>= shift;
+    dst->sw[1] >>= shift;
+    dst->sw[2] >>= shift;
+    dst->sw[3] >>= shift;
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
 static int
 opPSRAW_a32(uint32_t fetchdat)
 {
+    MMX_REG *dst;
     int shift;
 
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSHIFT();
 
     if (shift > 15)
         shift = 15;
 
-    cpu_state.MM[cpu_reg].sw[0] >>= shift;
-    cpu_state.MM[cpu_reg].sw[1] >>= shift;
-    cpu_state.MM[cpu_reg].sw[2] >>= shift;
-    cpu_state.MM[cpu_reg].sw[3] >>= shift;
+    dst->sw[0] >>= shift;
+    dst->sw[1] >>= shift;
+    dst->sw[2] >>= shift;
+    dst->sw[3] >>= shift;
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -192,31 +232,34 @@ opPSxxD_imm(uint32_t fetchdat)
     int reg   = fetchdat & 7;
     int op    = fetchdat & 0x38;
     int shift = (fetchdat >> 8) & 0xff;
+    MMX_REG *dst;
 
     cpu_state.pc += 2;
     MMX_ENTER();
 
+    dst = MMX_GETREGP(reg);
+
     switch (op) {
         case 0x10: /*PSRLD*/
             if (shift > 31)
-                cpu_state.MM[reg].q = 0;
+                dst->q = 0;
             else {
-                cpu_state.MM[reg].l[0] >>= shift;
-                cpu_state.MM[reg].l[1] >>= shift;
+                dst->l[0] >>= shift;
+                dst->l[1] >>= shift;
             }
             break;
         case 0x20: /*PSRAD*/
             if (shift > 31)
                 shift = 31;
-            cpu_state.MM[reg].sl[0] >>= shift;
-            cpu_state.MM[reg].sl[1] >>= shift;
+            dst->sl[0] >>= shift;
+            dst->sl[1] >>= shift;
             break;
         case 0x30: /*PSLLD*/
             if (shift > 31)
-                cpu_state.MM[reg].q = 0;
+                dst->q = 0;
             else {
-                cpu_state.MM[reg].l[0] <<= shift;
-                cpu_state.MM[reg].l[1] <<= shift;
+                dst->l[0] <<= shift;
+                dst->l[1] <<= shift;
             }
             break;
         default:
@@ -225,6 +268,8 @@ opPSxxD_imm(uint32_t fetchdat)
             return 0;
     }
 
+    MMX_SETEXP(reg);
+
     CLOCK_CYCLES(1);
     return 0;
 }
@@ -232,114 +277,150 @@ opPSxxD_imm(uint32_t fetchdat)
 static int
 opPSLLD_a16(uint32_t fetchdat)
 {
+    MMX_REG *dst;
     int shift;
 
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSHIFT();
 
     if (shift > 31)
-        cpu_state.MM[cpu_reg].q = 0;
+        dst->q = 0;
     else {
-        cpu_state.MM[cpu_reg].l[0] <<= shift;
-        cpu_state.MM[cpu_reg].l[1] <<= shift;
+        dst->l[0] <<= shift;
+        dst->l[1] <<= shift;
     }
 
+    MMX_SETEXP(cpu_reg);
+
     return 0;
 }
 static int
 opPSLLD_a32(uint32_t fetchdat)
 {
+    MMX_REG *dst;
     int shift;
 
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSHIFT();
 
     if (shift > 31)
-        cpu_state.MM[cpu_reg].q = 0;
+        dst->q = 0;
     else {
-        cpu_state.MM[cpu_reg].l[0] <<= shift;
-        cpu_state.MM[cpu_reg].l[1] <<= shift;
+        dst->l[0] <<= shift;
+        dst->l[1] <<= shift;
     }
 
+    MMX_SETEXP(cpu_reg);
+
     return 0;
 }
 
 static int
 opPSRLD_a16(uint32_t fetchdat)
 {
+    MMX_REG *dst;
     int shift;
 
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSHIFT();
 
     if (shift > 31)
-        cpu_state.MM[cpu_reg].q = 0;
+        dst->q = 0;
     else {
-        cpu_state.MM[cpu_reg].l[0] >>= shift;
-        cpu_state.MM[cpu_reg].l[1] >>= shift;
+        dst->l[0] >>= shift;
+        dst->l[1] >>= shift;
     }
 
+    MMX_SETEXP(cpu_reg);
+
     return 0;
 }
 static int
 opPSRLD_a32(uint32_t fetchdat)
 {
+    MMX_REG *dst;
     int shift;
 
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSHIFT();
 
     if (shift > 31)
-        cpu_state.MM[cpu_reg].q = 0;
+        dst->q = 0;
     else {
-        cpu_state.MM[cpu_reg].l[0] >>= shift;
-        cpu_state.MM[cpu_reg].l[1] >>= shift;
+        dst->l[0] >>= shift;
+        dst->l[1] >>= shift;
     }
 
+    MMX_SETEXP(cpu_reg);
+
     return 0;
 }
 
 static int
 opPSRAD_a16(uint32_t fetchdat)
 {
+    MMX_REG *dst;
     int shift;
 
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSHIFT();
 
     if (shift > 31)
         shift = 31;
 
-    cpu_state.MM[cpu_reg].sl[0] >>= shift;
-    cpu_state.MM[cpu_reg].sl[1] >>= shift;
+    dst->sl[0] >>= shift;
+    dst->sl[1] >>= shift;
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
 static int
 opPSRAD_a32(uint32_t fetchdat)
 {
+    MMX_REG *dst;
     int shift;
 
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSHIFT();
 
     if (shift > 31)
         shift = 31;
 
-    cpu_state.MM[cpu_reg].sl[0] >>= shift;
-    cpu_state.MM[cpu_reg].sl[1] >>= shift;
+    dst->sl[0] >>= shift;
+    dst->sl[1] >>= shift;
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -350,27 +431,32 @@ opPSxxQ_imm(uint32_t fetchdat)
     int reg   = fetchdat & 7;
     int op    = fetchdat & 0x38;
     int shift = (fetchdat >> 8) & 0xff;
+    MMX_REG *dst;
 
     cpu_state.pc += 2;
+
     MMX_ENTER();
 
+    dst = MMX_GETREGP(reg);
+
     switch (op) {
         case 0x10: /*PSRLW*/
             if (shift > 63)
-                cpu_state.MM[reg].q = 0;
+                dst->q = 0;
             else
-                cpu_state.MM[reg].q >>= shift;
+                dst->q >>= shift;
             break;
         case 0x20: /*PSRAW*/
             if (shift > 63)
                 shift = 63;
-            cpu_state.MM[reg].sq >>= shift;
+
+            dst->sq >>= shift;
             break;
         case 0x30: /*PSLLW*/
             if (shift > 63)
-                cpu_state.MM[reg].q = 0;
+                dst->q = 0;
             else
-                cpu_state.MM[reg].q <<= shift;
+                dst->q <<= shift;
             break;
         default:
             cpu_state.pc = cpu_state.oldpc;
@@ -378,6 +464,8 @@ opPSxxQ_imm(uint32_t fetchdat)
             return 0;
     }
 
+    MMX_SETEXP(reg);
+
     CLOCK_CYCLES(1);
     return 0;
 }
@@ -385,34 +473,46 @@ opPSxxQ_imm(uint32_t fetchdat)
 static int
 opPSLLQ_a16(uint32_t fetchdat)
 {
+    MMX_REG *dst;
     int shift;
 
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSHIFT();
 
     if (shift > 63)
-        cpu_state.MM[cpu_reg].q = 0;
+        dst->q = 0;
     else
-        cpu_state.MM[cpu_reg].q <<= shift;
+        dst->q <<= shift;
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
 static int
 opPSLLQ_a32(uint32_t fetchdat)
 {
+    MMX_REG *dst;
     int shift;
 
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSHIFT();
 
     if (shift > 63)
-        cpu_state.MM[cpu_reg].q = 0;
+        dst->q = 0;
     else
-        cpu_state.MM[cpu_reg].q <<= shift;
+        dst->q <<= shift;
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
@@ -420,34 +520,46 @@ opPSLLQ_a32(uint32_t fetchdat)
 static int
 opPSRLQ_a16(uint32_t fetchdat)
 {
+    MMX_REG *dst;
     int shift;
 
     MMX_ENTER();
 
     fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSHIFT();
 
     if (shift > 63)
-        cpu_state.MM[cpu_reg].q = 0;
+        dst->q = 0;
     else
-        cpu_state.MM[cpu_reg].q >>= shift;
+        dst->q >>= shift;
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
 static int
 opPSRLQ_a32(uint32_t fetchdat)
 {
+    MMX_REG *dst;
     int shift;
 
     MMX_ENTER();
 
     fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
     MMX_GETSHIFT();
 
     if (shift > 63)
-        cpu_state.MM[cpu_reg].q = 0;
+        dst->q = 0;
     else
-        cpu_state.MM[cpu_reg].q >>= shift;
+        dst->q >>= shift;
+
+    MMX_SETEXP(cpu_reg);
 
     return 0;
 }
diff --git a/src/cpu/x87.c b/src/cpu/x87.c
index 98ceb105b..181b7b9ca 100644
--- a/src/cpu/x87.c
+++ b/src/cpu/x87.c
@@ -439,6 +439,79 @@ FPU_tagof(const floatx80 reg)
     return X87_TAG_VALID;
 }
 
+uint8_t
+pack_FPU_TW(uint16_t twd)
+{
+    uint8_t tag_byte = 0;
+
+    if ((twd & 0x0003) != 0x0003) tag_byte |= 0x01;
+    if ((twd & 0x000c) != 0x000c) tag_byte |= 0x02;
+    if ((twd & 0x0030) != 0x0030) tag_byte |= 0x04;
+    if ((twd & 0x00c0) != 0x00c0) tag_byte |= 0x08;
+    if ((twd & 0x0300) != 0x0300) tag_byte |= 0x10;
+    if ((twd & 0x0c00) != 0x0c00) tag_byte |= 0x20;
+    if ((twd & 0x3000) != 0x3000) tag_byte |= 0x40;
+    if ((twd & 0xc000) != 0xc000) tag_byte |= 0x80;
+
+    return tag_byte;
+}
+
+uint16_t
+unpack_FPU_TW(uint16_t tag_byte)
+{
+    uint32_t twd = 0;
+
+  /*                                 FTW
+   *
+   * Note that the original format for FTW can be recreated from the stored
+   * FTW valid bits and the stored 80-bit FP data (assuming the stored data
+   * was not the contents of MMX registers) using the following table:
+
+     | Exponent | Exponent | Fraction | J,M bits | FTW valid | x87 FTW |
+     |  all 1s  |  all 0s  |  all 0s  |          |           |         |
+     -------------------------------------------------------------------
+     |    0     |    0     |    0     |    0x    |     1     | S    10 |
+     |    0     |    0     |    0     |    1x    |     1     | V    00 |
+     -------------------------------------------------------------------
+     |    0     |    0     |    1     |    00    |     1     | S    10 |
+     |    0     |    0     |    1     |    10    |     1     | V    00 |
+     -------------------------------------------------------------------
+     |    0     |    1     |    0     |    0x    |     1     | S    10 |
+     |    0     |    1     |    0     |    1x    |     1     | S    10 |
+     -------------------------------------------------------------------
+     |    0     |    1     |    1     |    00    |     1     | Z    01 |
+     |    0     |    1     |    1     |    10    |     1     | S    10 |
+     -------------------------------------------------------------------
+     |    1     |    0     |    0     |    1x    |     1     | S    10 |
+     |    1     |    0     |    0     |    1x    |     1     | S    10 |
+     -------------------------------------------------------------------
+     |    1     |    0     |    1     |    00    |     1     | S    10 |
+     |    1     |    0     |    1     |    10    |     1     | S    10 |
+     -------------------------------------------------------------------
+     |        all combinations above             |     0     | E    11 |
+
+   *
+   * The J-bit is defined to be the 1-bit binary integer to the left of
+   * the decimal place in the significand.
+   *
+   * The M-bit is defined to be the most significant bit of the fractional
+   * portion of the significand (i.e., the bit immediately to the right of
+   * the decimal place). When the M-bit is the most significant bit of the
+   * fractional portion  of the significand, it must be  0 if the fraction
+   * is all 0's.
+   */
+
+    for (int index = 7; index >= 0; index--, twd <<= 2, tag_byte <<= 1) {
+        if (tag_byte & 0x80) {
+            const floatx80 *fpu_reg = &fpu_state.st_space[index & 7];
+            twd |= FPU_tagof(*fpu_reg);
+        } else {
+            twd |= X87_TAG_EMPTY;
+        }
+    }
+
+    return (twd >> 2);
+}
 
 #ifdef ENABLE_808X_LOG
 void
diff --git a/src/cpu/x87.h b/src/cpu/x87.h
index 5d460bc4b..66d51dbd9 100644
--- a/src/cpu/x87.h
+++ b/src/cpu/x87.h
@@ -10,9 +10,14 @@ static __inline void
 x87_set_mmx(void)
 {
     uint64_t *p;
-    cpu_state.TOP   = 0;
-    p               = (uint64_t *) cpu_state.tag;
-    *p              = 0x0101010101010101ull;
+    if (fpu_softfloat) {
+        fpu_state.tag = 0;
+        fpu_state.tos = 0; /* reset FPU Top-Of-Stack */
+    } else {
+        cpu_state.TOP   = 0;
+        p               = (uint64_t *) cpu_state.tag;
+        *p              = 0x0101010101010101ull;
+    }
     cpu_state.ismmx = 1;
 }
 
@@ -20,8 +25,13 @@ static __inline void
 x87_emms(void)
 {
     uint64_t *p;
-    p               = (uint64_t *) cpu_state.tag;
-    *p              = 0;
+    if (fpu_softfloat) {
+        fpu_state.tag = 0xffff;
+        fpu_state.tos = 0; /* reset FPU Top-Of-Stack */
+    } else {
+        p               = (uint64_t *) cpu_state.tag;
+        *p              = 0;
+    }
     cpu_state.ismmx = 0;
 }
 
@@ -141,6 +151,8 @@ void FPU_stack_underflow(uint32_t fetchdat, int stnr, int pop_stack);
 int FPU_handle_NaN32(floatx80 a, float32 b, floatx80 *r, struct float_status_t *status);
 int FPU_handle_NaN64(floatx80 a, float64 b, floatx80 *r, struct float_status_t *status);
 int FPU_tagof(const floatx80 reg);
+uint8_t pack_FPU_TW(uint16_t twd);
+uint16_t unpack_FPU_TW(uint16_t tag_byte);
 
 static __inline uint16_t
 i387_get_control_word(void)
diff --git a/src/ddma.c b/src/ddma.c
index 7623dc0df..098b1c573 100644
--- a/src/ddma.c
+++ b/src/ddma.c
@@ -34,6 +34,7 @@
 #include <86box/pit.h>
 #include <86box/dma.h>
 #include <86box/ddma.h>
+#include <86box/plat_unused.h>
 
 #ifdef ENABLE_DDMA_LOG
 int ddma_do_log = ENABLE_DDMA_LOG;
@@ -54,9 +55,9 @@ ddma_log(const char *fmt, ...)
 #endif
 
 static uint8_t
-ddma_reg_read(uint16_t addr, void *p)
+ddma_reg_read(uint16_t addr, void *priv)
 {
-    ddma_channel_t *dev  = (ddma_channel_t *) p;
+    ddma_channel_t *dev  = (ddma_channel_t *) priv;
     uint8_t         ret  = 0xff;
     int             ch   = dev->channel;
     int             dmab = (ch >= 4) ? 0xc0 : 0x00;
@@ -80,15 +81,18 @@ ddma_reg_read(uint16_t addr, void *p)
         case 0x09:
             ret = inb(dmab + 0x08);
             break;
+
+        default:
+            break;
     }
 
     return ret;
 }
 
 static void
-ddma_reg_write(uint16_t addr, uint8_t val, void *p)
+ddma_reg_write(uint16_t addr, uint8_t val, void *priv)
 {
-    ddma_channel_t *dev          = (ddma_channel_t *) p;
+    ddma_channel_t *dev          = (ddma_channel_t *) priv;
     int             ch           = dev->channel;
     int             page_regs[4] = { 7, 3, 1, 2 };
     int             dmab = (ch >= 4) ? 0xc0 : 0x00;
@@ -138,6 +142,9 @@ ddma_reg_write(uint16_t addr, uint8_t val, void *p)
         case 0x0f:
             outb(dmab + 0x0a, (val << 2) | (ch & 3));
             break;
+
+        default:
+            break;
     }
 }
 
@@ -163,7 +170,7 @@ ddma_close(void *priv)
 }
 
 static void *
-ddma_init(const device_t *info)
+ddma_init(UNUSED(const device_t *info))
 {
     ddma_t *dev;
 
diff --git a/src/device.c b/src/device.c
index 7d8f5e88e..47dae84cf 100644
--- a/src/device.c
+++ b/src/device.c
@@ -88,46 +88,46 @@ device_init(void)
 }
 
 void
-device_set_context(device_context_t *c, const device_t *d, int inst)
+device_set_context(device_context_t *c, const device_t *dev, int inst)
 {
     void *sec;
     void *single_sec;
 
     memset(c, 0, sizeof(device_context_t));
-    c->dev      = d;
+    c->dev      = dev;
     c->instance = inst;
     if (inst) {
-        sprintf(c->name, "%s #%i", d->name, inst);
+        sprintf(c->name, "%s #%i", dev->name, inst);
 
         /* If this is the first instance and a numbered section is not present, but a non-numbered
            section of the same name is, rename the non-numbered section to numbered. */
         if (inst == 1) {
             sec        = config_find_section(c->name);
-            single_sec = config_find_section((char *) d->name);
+            single_sec = config_find_section((char *) dev->name);
             if ((sec == NULL) && (single_sec != NULL))
                 config_rename_section(single_sec, c->name);
         }
     } else
-        sprintf(c->name, "%s", d->name);
+        sprintf(c->name, "%s", dev->name);
 }
 
 static void
-device_context_common(const device_t *d, int inst)
+device_context_common(const device_t *dev, int inst)
 {
     memcpy(&device_prev, &device_current, sizeof(device_context_t));
-    device_set_context(&device_current, d, inst);
+    device_set_context(&device_current, dev, inst);
 }
 
 void
-device_context(const device_t *d)
+device_context(const device_t *dev)
 {
-    device_context_common(d, 0);
+    device_context_common(dev, 0);
 }
 
 void
-device_context_inst(const device_t *d, int inst)
+device_context_inst(const device_t *dev, int inst)
 {
-    device_context_common(d, inst);
+    device_context_common(dev, inst);
 }
 
 void
@@ -137,13 +137,13 @@ device_context_restore(void)
 }
 
 static void *
-device_add_common(const device_t *d, const device_t *cd, void *p, void *params, int inst)
+device_add_common(const device_t *dev, const device_t *cd, void *p, void *params, int inst)
 {
     void *priv = NULL;
     int   c;
 
     for (c = 0; c < 256; c++) {
-        if (!inst && (devices[c] == (device_t *) d)) {
+        if (!inst && (devices[c] == (device_t *) dev)) {
             device_log("DEVICE: device already exists!\n");
             return (NULL);
         }
@@ -157,17 +157,17 @@ device_add_common(const device_t *d, const device_t *cd, void *p, void *params,
 
     /* Do this so that a chained device_add will not identify the same ID
        its master device is already trying to assign. */
-    devices[c] = (device_t *) d;
+    devices[c] = (device_t *) dev;
 
     if (p == NULL) {
         memcpy(&device_prev, &device_current, sizeof(device_context_t));
         device_set_context(&device_current, cd, inst);
 
-        if (d->init != NULL) {
-            priv = (d->flags & DEVICE_EXTPARAMS) ? d->init_ext(d, params) : d->init(d);
+        if (dev->init != NULL) {
+            priv = (dev->flags & DEVICE_EXTPARAMS) ? dev->init_ext(dev, params) : dev->init(dev);
             if (priv == NULL) {
-                if (d->name)
-                    device_log("DEVICE: device '%s' init failed\n", d->name);
+                if (dev->name)
+                    device_log("DEVICE: device '%s' init failed\n", dev->name);
                 else
                     device_log("DEVICE: device init failed\n");
 
@@ -178,8 +178,8 @@ device_add_common(const device_t *d, const device_t *cd, void *p, void *params,
             }
         }
 
-        if (d->name)
-            device_log("DEVICE: device '%s' init successful\n", d->name);
+        if (dev->name)
+            device_log("DEVICE: device '%s' init successful\n", dev->name);
         else
             device_log("DEVICE: device init successful\n");
 
@@ -192,114 +192,114 @@ device_add_common(const device_t *d, const device_t *cd, void *p, void *params,
 }
 
 char *
-device_get_internal_name(const device_t *d)
+device_get_internal_name(const device_t *dev)
 {
-    if (d == NULL)
+    if (dev == NULL)
         return "";
 
-    return (char *) d->internal_name;
+    return (char *) dev->internal_name;
 }
 
 void *
-device_add(const device_t *d)
+device_add(const device_t *dev)
 {
-    return device_add_common(d, d, NULL, NULL, 0);
+    return device_add_common(dev, dev, NULL, NULL, 0);
 }
 
 void *
-device_add_parameters(const device_t *d, void *params)
+device_add_parameters(const device_t *dev, void *params)
 {
-    return device_add_common(d, d, NULL, params, 0);
+    return device_add_common(dev, dev, NULL, params, 0);
 }
 
 /* For devices that do not have an init function (internal video etc.) */
 void
-device_add_ex(const device_t *d, void *priv)
+device_add_ex(const device_t *dev, void *priv)
 {
-    device_add_common(d, d, priv, NULL, 0);
+    device_add_common(dev, dev, priv, NULL, 0);
 }
 
 void
-device_add_ex_parameters(const device_t *d, void *priv, void *params)
+device_add_ex_parameters(const device_t *dev, void *priv, void *params)
 {
-    device_add_common(d, d, priv, params, 0);
+    device_add_common(dev, dev, priv, params, 0);
 }
 
 void *
-device_add_inst(const device_t *d, int inst)
+device_add_inst(const device_t *dev, int inst)
 {
-    return device_add_common(d, d, NULL, NULL, inst);
+    return device_add_common(dev, dev, NULL, NULL, inst);
 }
 
 void *
-device_add_inst_parameters(const device_t *d, int inst, void *params)
+device_add_inst_parameters(const device_t *dev, int inst, void *params)
 {
-    return device_add_common(d, d, NULL, params, inst);
+    return device_add_common(dev, dev, NULL, params, inst);
 }
 
 /* For devices that do not have an init function (internal video etc.) */
 void
-device_add_inst_ex(const device_t *d, void *priv, int inst)
+device_add_inst_ex(const device_t *dev, void *priv, int inst)
 {
-    device_add_common(d, d, priv, NULL, inst);
+    device_add_common(dev, dev, priv, NULL, inst);
 }
 
 void
-device_add_inst_ex_parameters(const device_t *d, void *priv, int inst, void *params)
+device_add_inst_ex_parameters(const device_t *dev, void *priv, int inst, void *params)
 {
-    device_add_common(d, d, priv, params, inst);
+    device_add_common(dev, dev, priv, params, inst);
 }
 
 /* These eight are to add a device with another device's context - will be
    used to add machines' internal devices. */
 void *
-device_cadd(const device_t *d, const device_t *cd)
+device_cadd(const device_t *dev, const device_t *cd)
 {
-    return device_add_common(d, cd, NULL, NULL, 0);
+    return device_add_common(dev, cd, NULL, NULL, 0);
 }
 
 void *
-device_cadd_parameters(const device_t *d, const device_t *cd, void *params)
+device_cadd_parameters(const device_t *dev, const device_t *cd, void *params)
 {
-    return device_add_common(d, cd, NULL, params, 0);
+    return device_add_common(dev, cd, NULL, params, 0);
 }
 
 /* For devices that do not have an init function (internal video etc.) */
 void
-device_cadd_ex(const device_t *d, const device_t *cd, void *priv)
+device_cadd_ex(const device_t *dev, const device_t *cd, void *priv)
 {
-    device_add_common(d, cd, priv, NULL, 0);
+    device_add_common(dev, cd, priv, NULL, 0);
 }
 
 void
-device_cadd_ex_parameters(const device_t *d, const device_t *cd, void *priv, void *params)
+device_cadd_ex_parameters(const device_t *dev, const device_t *cd, void *priv, void *params)
 {
-    device_add_common(d, cd, priv, params, 0);
+    device_add_common(dev, cd, priv, params, 0);
 }
 
 void *
-device_cadd_inst(const device_t *d, const device_t *cd, int inst)
+device_cadd_inst(const device_t *dev, const device_t *cd, int inst)
 {
-    return device_add_common(d, cd, NULL, NULL, inst);
+    return device_add_common(dev, cd, NULL, NULL, inst);
 }
 
 void *
-device_cadd_inst_parameters(const device_t *d, const device_t *cd, int inst, void *params)
+device_cadd_inst_parameters(const device_t *dev, const device_t *cd, int inst, void *params)
 {
-    return device_add_common(d, cd, NULL, params, inst);
+    return device_add_common(dev, cd, NULL, params, inst);
 }
 
 /* For devices that do not have an init function (internal video etc.) */
 void
-device_cadd_inst_ex(const device_t *d, const device_t *cd, void *priv, int inst)
+device_cadd_inst_ex(const device_t *dev, const device_t *cd, void *priv, int inst)
 {
-    device_add_common(d, cd, priv, NULL, inst);
+    device_add_common(dev, cd, priv, NULL, inst);
 }
 
 void
-device_cadd_inst_ex_parameters(const device_t *d, const device_t *cd, void *priv, int inst, void *params)
+device_cadd_inst_ex_parameters(const device_t *dev, const device_t *cd, void *priv, int inst, void *params)
 {
-    device_add_common(d, cd, priv, params, inst);
+    device_add_common(dev, cd, priv, params, inst);
 }
 
 void
@@ -328,11 +328,11 @@ device_reset_all(uint32_t match_flags)
 }
 
 void *
-device_get_priv(const device_t *d)
+device_get_priv(const device_t *dev)
 {
     for (uint16_t c = 0; c < DEVICE_MAX; c++) {
         if (devices[c] != NULL) {
-            if (devices[c] == d)
+            if (devices[c] == dev)
                 return (device_priv[c]);
         }
     }
@@ -341,15 +341,15 @@ device_get_priv(const device_t *d)
 }
 
 int
-device_available(const device_t *d)
+device_available(const device_t *dev)
 {
     device_config_t      *config = NULL;
     device_config_bios_t *bios   = NULL;
     int                   roms_present = 0;
     int                   i = 0;
 
-    if (d != NULL) {
-        config = (device_config_t *) d->config;
+    if (dev != NULL) {
+        config = (device_config_t *) dev->config;
         if (config != NULL) {
             while (config->type != -1) {
                 if (config->type == CONFIG_BIOS) {
@@ -372,8 +372,8 @@ device_available(const device_t *d)
         }
 
         /* No CONFIG_BIOS field present, use the classic available(). */
-        if (d->available != NULL)
-            return (d->available());
+        if (dev->available != NULL)
+            return (dev->available());
         else
             return 1;
     }
@@ -383,13 +383,13 @@ device_available(const device_t *d)
 }
 
 const char *
-device_get_bios_file(const device_t *d, const char *internal_name, int file_no)
+device_get_bios_file(const device_t *dev, const char *internal_name, int file_no)
 {
     device_config_t      *config = NULL;
     device_config_bios_t *bios   = NULL;
 
-    if (d != NULL) {
-        config = (device_config_t *) d->config;
+    if (dev != NULL) {
+        config = (device_config_t *) dev->config;
         if (config != NULL) {
             while (config->type != -1) {
                 if (config->type == CONFIG_BIOS) {
@@ -416,18 +416,18 @@ device_get_bios_file(const device_t *d, const char *internal_name, int file_no)
 }
 
 int
-device_has_config(const device_t *d)
+device_has_config(const device_t *dev)
 {
     int              c = 0;
     device_config_t *config;
 
-    if (d == NULL)
+    if (dev == NULL)
         return 0;
 
-    if (d->config == NULL)
+    if (dev->config == NULL)
         return 0;
 
-    config = (device_config_t *) d->config;
+    config = (device_config_t *) dev->config;
 
     while (config->type != -1) {
         if (config->type != CONFIG_MAC)
@@ -439,11 +439,11 @@ device_has_config(const device_t *d)
 }
 
 int
-device_poll(const device_t *d, int x, int y, int z, int b)
+device_poll(const device_t *dev, int x, int y, int z, int b)
 {
     for (uint16_t c = 0; c < DEVICE_MAX; c++) {
         if (devices[c] != NULL) {
-            if (devices[c] == d) {
+            if (devices[c] == dev) {
                 if (devices[c]->poll)
                     return (devices[c]->poll(x, y, z, b, 0, 0, device_priv[c]));
             }
@@ -454,11 +454,11 @@ device_poll(const device_t *d, int x, int y, int z, int b)
 }
 
 void
-device_register_pci_slot(const device_t *d, int device, int type, int inta, int intb, int intc, int intd)
+device_register_pci_slot(const device_t *dev, int device, int type, int inta, int intb, int intc, int intd)
 {
     for (uint16_t c = 0; c < DEVICE_MAX; c++) {
         if (devices[c] != NULL) {
-            if (devices[c] == d) {
+            if (devices[c] == dev) {
                 if (devices[c]->register_pci_slot)
                     devices[c]->register_pci_slot(device, type, inta, intb, intc, intd, device_priv[c]);
                 return;
@@ -470,38 +470,38 @@ device_register_pci_slot(const device_t *d, int device, int type, int inta, int
 }
 
 void
-device_get_name(const device_t *d, int bus, char *name)
+device_get_name(const device_t *dev, int bus, char *name)
 {
     char *sbus = NULL;
     char *fbus;
     char *tname;
     char  pbus[8] = { 0 };
 
-    if (d == NULL)
+    if (dev == NULL)
         return;
 
     name[0] = 0x00;
 
     if (bus) {
-        if (d->flags & DEVICE_ISA)
-            sbus = (d->flags & DEVICE_AT) ? "ISA16" : "ISA";
-        else if (d->flags & DEVICE_CBUS)
+        if (dev->flags & DEVICE_ISA)
+            sbus = (dev->flags & DEVICE_AT) ? "ISA16" : "ISA";
+        else if (dev->flags & DEVICE_CBUS)
             sbus = "C-BUS";
-        else if (d->flags & DEVICE_MCA)
+        else if (dev->flags & DEVICE_MCA)
             sbus = "MCA";
-        else if (d->flags & DEVICE_EISA)
+        else if (dev->flags & DEVICE_EISA)
             sbus = "EISA";
-        else if (d->flags & DEVICE_VLB)
+        else if (dev->flags & DEVICE_VLB)
             sbus = "VLB";
-        else if (d->flags & DEVICE_PCI)
+        else if (dev->flags & DEVICE_PCI)
             sbus = "PCI";
-        else if (d->flags & DEVICE_AGP)
+        else if (dev->flags & DEVICE_AGP)
             sbus = "AGP";
-        else if (d->flags & DEVICE_AC97)
+        else if (dev->flags & DEVICE_AC97)
             sbus = "AMR";
-        else if (d->flags & DEVICE_COM)
+        else if (dev->flags & DEVICE_COM)
             sbus = "COM";
-        else if (d->flags & DEVICE_LPT)
+        else if (dev->flags & DEVICE_LPT)
             sbus = "LPT";
 
         if (sbus != NULL) {
@@ -515,7 +515,7 @@ device_get_name(const device_t *d, int bus, char *name)
                 sbus = "ISA";
             else if (!strcmp(sbus, "COM") || !strcmp(sbus, "LPT")) {
                 sbus = NULL;
-                strcat(name, d->name);
+                strcat(name, dev->name);
                 return;
             }
 
@@ -525,17 +525,17 @@ device_get_name(const device_t *d, int bus, char *name)
             strcat(pbus, ")");
 
             /* Allocate the temporary device name string and set it to all zeroes. */
-            tname = (char *) malloc(strlen(d->name) + 1);
-            memset(tname, 0x00, strlen(d->name) + 1);
+            tname = (char *) malloc(strlen(dev->name) + 1);
+            memset(tname, 0x00, strlen(dev->name) + 1);
 
             /* First strip the bus string with parentheses. */
-            fbus = strstr(d->name, pbus);
-            if (fbus == d->name)
-                strcat(tname, d->name + strlen(pbus) + 1);
+            fbus = strstr(dev->name, pbus);
+            if (fbus == dev->name)
+                strcat(tname, dev->name + strlen(pbus) + 1);
             else if (fbus == NULL)
-                strcat(tname, d->name);
+                strcat(tname, dev->name);
             else {
-                strncat(tname, d->name, fbus - d->name - 1);
+                strncat(tname, dev->name, fbus - dev->name - 1);
                 strcat(tname, fbus + strlen(pbus));
             }
 
@@ -556,9 +556,9 @@ device_get_name(const device_t *d, int bus, char *name)
             free(tname);
             tname = NULL;
         } else
-            strcat(name, d->name);
+            strcat(name, dev->name);
     } else
-        strcat(name, d->name);
+        strcat(name, dev->name);
 }
 
 void
@@ -771,7 +771,7 @@ device_is_valid(const device_t *device, int m)
     if ((device->flags & DEVICE_AGP) && !machine_has_bus(m, MACHINE_BUS_AGP))
         return 0;
 
-    if ((device->flags & DEVICE_PS2) && !machine_has_bus(m, MACHINE_BUS_PS2))
+    if ((device->flags & DEVICE_PS2) && !machine_has_bus(m, MACHINE_BUS_PS2_PORTS))
         return 0;
 
     if ((device->flags & DEVICE_AC97) && !machine_has_bus(m, MACHINE_BUS_AC97))
diff --git a/src/device/kbc_at.c b/src/device/kbc_at.c
index 7ece0cabc..ee5a6d17e 100644
--- a/src/device/kbc_at.c
+++ b/src/device/kbc_at.c
@@ -81,7 +81,9 @@
 #define KBC_VEN_ACER       0x20
 #define KBC_VEN_NCR        0x24
 #define KBC_VEN_ALI        0x28
-#define KBC_VEN_MASK       0x3c
+#define KBC_VEN_SIEMENS    0x2c
+#define KBC_VEN_COMPAQ     0x30
+#define KBC_VEN_MASK       0x7c
 
 #define FLAG_CLOCK         0x01
 #define FLAG_CACHE         0x02
@@ -747,6 +749,12 @@ write_p2(atkbc_t *dev, uint8_t val)
                 flushmmucache();
                 if (kbc_ven == KBC_VEN_ALI)
                     smbase = 0x00030000;
+                /* Yes, this is a hack, but until someone gets ahold of the real PCD-2L
+                   and can find out what they actually did to make it boot from FFFFF0
+                   correctly despite A20 being gated when the CPU is reset, this will
+                   have to do. */
+                else if (kbc_ven == KBC_VEN_SIEMENS)
+                    loadcs(0xF000);
             }
         }
     }
@@ -974,6 +982,8 @@ write64_generic(void *priv, uint8_t val)
             } else if (((dev->flags & KBC_TYPE_MASK) >= KBC_TYPE_PS2_1) && ((dev->flags & KBC_TYPE_MASK) < KBC_TYPE_GREEN))
                 /* (B0 or F0) | (0x08 or 0x0c) */
                 kbc_delay_to_ob(dev, ((dev->p1 | fixed_bits) & 0xf0) | (((dev->flags & KBC_VEN_MASK) == KBC_VEN_ACER) ? 0x08 : 0x0c), 0, 0x00);
+            else if (kbc_ven == KBC_VEN_COMPAQ)
+                kbc_delay_to_ob(dev, dev->p1 | (hasfpu ? 0x00 : 0x04), 0, 0x00);
             else
                 /* (B0 or F0) | (0x04 or 0x44) */
                 kbc_delay_to_ob(dev, dev->p1 | fixed_bits, 0, 0x00);
@@ -1031,7 +1041,7 @@ write60_ami(void *priv, uint8_t val)
     switch (dev->command) {
         /* 0x40 - 0x5F are aliases for 0x60-0x7F */
         case 0x40 ... 0x5f:
-            kbc_at_log("ATkbc: AMI - alias write to %08X\n", dev->command);
+            kbc_at_log("ATkbc: AMI - alias write to %02X\n", dev->command & 0x1f);
             dev->mem[(dev->command & 0x1f) + 0x20] = val;
             if (dev->command == 0x60)
                 write_cmd(dev, val);
@@ -1170,7 +1180,7 @@ write64_ami(void *priv, uint8_t val)
             break;
 
         case 0xaf: /* set extended controller RAM */
-            if (kbc_ven != KBC_VEN_ALI) {
+            if ((kbc_ven != KBC_VEN_SIEMENS) && (kbc_ven != KBC_VEN_ALI)) {
                 kbc_at_log("ATkbc: set extended controller RAM\n");
                 dev->wantdata      = 1;
                 dev->state         = STATE_KBC_PARAM;
@@ -1282,6 +1292,39 @@ write64_ami(void *priv, uint8_t val)
     return write64_generic(dev, val);
 }
 
+static uint8_t
+write64_siemens(void *priv, uint8_t val)
+{
+    atkbc_t *dev     = (atkbc_t *) priv;
+
+    switch (val) {
+        case 0x92: /*Siemens Award - 92 sent by PCD-2L BIOS*/
+            kbc_at_log("Siemens Award - 92 sent by PCD-2L BIOS\n");
+            return 0;
+
+        case 0x94: /*Siemens Award - 94 sent by PCD-2L BIOS*/
+            kbc_at_log("Siemens Award - 94 sent by PCD-2L BIOS\n");
+            return 0;
+
+        case 0x9a: /*Siemens Award - 9A sent by PCD-2L BIOS*/
+            kbc_at_log("Siemens Award - 9A sent by PCD-2L BIOS\n");
+            return 0;
+
+        case 0x9c: /*Siemens Award - 9C sent by PCD-2L BIOS*/
+            kbc_at_log("Siemens Award - 9C sent by PCD-2L BIOS\n");
+            return 0;
+
+        case 0xa9: /*Siemens Award - A9 sent by PCD-2L BIOS*/
+            kbc_at_log("Siemens Award - A9 sent by PCD-2L BIOS\n");
+            return 0;
+
+        default:
+            break;
+    }
+
+    return write64_ami(dev, val);
+}
+
 static uint8_t
 write60_quadtel(void *priv, UNUSED(uint8_t val))
 {
@@ -1729,13 +1772,15 @@ static void
 kbc_at_write(uint16_t port, uint8_t val, void *priv)
 {
     atkbc_t *dev = (atkbc_t *) priv;
+    uint8_t kbc_ven = dev->flags & KBC_VEN_MASK;
+    uint8_t fast_a20 = (kbc_ven != KBC_VEN_SIEMENS);
 
     kbc_at_log("ATkbc: [%04X:%08X] write(%04X) = %02X\n", CS, cpu_state.pc, port, val);
 
     switch (port) {
         case 0x60:
             dev->status &= ~STAT_CD;
-            if (dev->wantdata && (dev->command == 0xd1)) {
+            if (fast_a20 && dev->wantdata && (dev->command == 0xd1)) {
                 kbc_at_log("ATkbc: write P2\n");
 
 #if 0
@@ -1765,7 +1810,7 @@ kbc_at_write(uint16_t port, uint8_t val, void *priv)
 
         case 0x64:
             dev->status |= STAT_CD;
-            if (val == 0xd1) {
+            if (fast_a20 && (val == 0xd1)) {
                 kbc_at_log("ATkbc: write P2\n");
                 dev->wantdata  = 1;
                 dev->state     = STATE_KBC_PARAM;
@@ -1915,10 +1960,18 @@ kbc_at_init(const device_t *info)
     kbc_award_revision = 0x42;
 
     switch (dev->flags & KBC_VEN_MASK) {
+        case KBC_VEN_SIEMENS:
+            kbc_ami_revision = '8';
+            kbc_award_revision = 0x42;
+            dev->write60_ven = write60_ami;
+            dev->write64_ven = write64_siemens;
+            break;
+
         case KBC_VEN_ACER:
         case KBC_VEN_GENERIC:
         case KBC_VEN_NCR:
         case KBC_VEN_IBM_PS1:
+        case KBC_VEN_COMPAQ:
             dev->write64_ven = write64_generic;
             break;
 
@@ -2008,6 +2061,20 @@ const device_t keyboard_at_device = {
     .config        = NULL
 };
 
+const device_t keyboard_at_siemens_device = {
+    .name          = "PC/AT Keyboard",
+    .internal_name = "keyboard_at",
+    .flags         = DEVICE_KBC,
+    .local         = KBC_TYPE_ISA | KBC_VEN_SIEMENS,
+    .init          = kbc_at_init,
+    .close         = kbc_at_close,
+    .reset         = kbc_at_reset,
+    { .available = NULL },
+    .speed_changed = NULL,
+    .force_redraw  = NULL,
+    .config        = NULL
+};
+
 const device_t keyboard_at_ami_device = {
     .name          = "PC/AT Keyboard (AMI)",
     .internal_name = "keyboard_at_ami",
@@ -2078,6 +2145,20 @@ const device_t keyboard_at_ncr_device = {
     .config        = NULL
 };
 
+const device_t keyboard_at_compaq_device = {
+    .name          = "PC/AT Keyboard (Compaq)",
+    .internal_name = "keyboard_at_compaq",
+    .flags         = DEVICE_KBC,
+    .local         = KBC_TYPE_ISA | KBC_VEN_COMPAQ,
+    .init          = kbc_at_init,
+    .close         = kbc_at_close,
+    .reset         = kbc_at_reset,
+    { .available = NULL },
+    .speed_changed = NULL,
+    .force_redraw  = NULL,
+    .config        = NULL
+};
+
 const device_t keyboard_ps2_device = {
     .name          = "PS/2 Keyboard",
     .internal_name = "keyboard_ps2",
diff --git a/src/device/serial_passthrough.c b/src/device/serial_passthrough.c
index a8c531f07..b703cd67c 100644
--- a/src/device/serial_passthrough.c
+++ b/src/device/serial_passthrough.c
@@ -119,6 +119,8 @@ static void
 serial_passthrough_speed_changed(void *priv)
 {
     serial_passthrough_t *dev = (serial_passthrough_t *) priv;
+    if (!dev)
+        return;
 
     timer_stop(&dev->host_to_serial_timer);
     /* FIXME: do something to dev->baudrate */
@@ -132,9 +134,11 @@ static void
 serial_passthrough_dev_close(void *priv)
 {
     serial_passthrough_t *dev = (serial_passthrough_t *) priv;
+    if (!dev)
+        return;
 
     /* Detach passthrough device from COM port */
-    if (dev && dev->serial && dev->serial->sd)
+    if (dev->serial && dev->serial->sd)
         memset(dev->serial->sd, 0, sizeof(serial_device_t));
 
     plat_serpt_close(dev);
@@ -184,6 +188,10 @@ serial_passthrough_dev_init(const device_t *info)
     /* Attach passthrough device to a COM port */
     dev->serial = serial_attach_ex(dev->port, serial_passthrough_rcr_cb,
                                    serial_passthrough_write, serial_passthrough_transmit_period, serial_passthrough_lcr_callback, dev);
+    if (!dev->serial) {
+        free(dev);
+        return NULL;
+    }
 
     strncpy(dev->host_serial_path, device_get_config_string("host_serial_path"), 1023);
 #ifdef _WIN32
diff --git a/src/disk/hdc_esdi_at.c b/src/disk/hdc_esdi_at.c
index 9fc6b58d5..014212b1d 100644
--- a/src/disk/hdc_esdi_at.c
+++ b/src/disk/hdc_esdi_at.c
@@ -70,7 +70,7 @@
 #define CMD_SET_PARAMETERS  0x91
 #define CMD_READ_PARAMETERS 0xec
 
-typedef struct {
+typedef struct drive_t {
     int cfg_spt;
     int cfg_hpc;
     int current_cylinder;
@@ -81,7 +81,7 @@ typedef struct {
     int hdd_num;
 } drive_t;
 
-typedef struct {
+typedef struct esdi_t {
     uint8_t status;
     uint8_t error;
     int     secount;
@@ -821,9 +821,9 @@ loadhd(esdi_t *esdi, int hdd_num, int d, UNUSED(const char *fn))
 }
 
 static void
-esdi_rom_write(uint32_t addr, uint8_t val, void *p)
+esdi_rom_write(uint32_t addr, uint8_t val, void *priv)
 {
-    rom_t *rom = (rom_t *) p;
+    rom_t *rom = (rom_t *) priv;
 
     addr &= rom->mask;
 
diff --git a/src/disk/hdc_esdi_mca.c b/src/disk/hdc_esdi_mca.c
index 6e7edb1be..712ab900b 100644
--- a/src/disk/hdc_esdi_mca.c
+++ b/src/disk/hdc_esdi_mca.c
@@ -96,7 +96,8 @@
 #define CMD_ADAPTER     0
 
 typedef struct esdi_drive_t {
-    int spt, hpc;
+    int spt;
+    int hpc;
     int tracks;
     int sectors;
     int present;
@@ -139,7 +140,7 @@ typedef struct esdi_t {
 
     uint32_t rba;
 
-    struct {
+    struct cmds {
         int req_in_progress;
     } cmds[3];
 
diff --git a/src/disk/hdc_ide.c b/src/disk/hdc_ide.c
index ea771235e..041d8ffc4 100644
--- a/src/disk/hdc_ide.c
+++ b/src/disk/hdc_ide.c
@@ -110,7 +110,7 @@
 
 #define IDE_TIME                       10.0
 
-typedef struct {
+typedef struct ide_board_t {
     int        bit32;
     int        cur_dev;
     int        irq;
@@ -123,7 +123,7 @@ typedef struct {
     ide_t     *ide[2];
 } ide_board_t;
 
-typedef struct {
+typedef struct ide_bm_t {
     int (*dma)(int channel, uint8_t *data, int transfer_length, int out, void *priv);
     void (*set_irq)(int channel, void *priv);
     void *priv;
@@ -244,6 +244,9 @@ ide_get_xfer_time(ide_t *ide, int size)
                 case 0x10:
                     period = (50.0 / 3.0);
                     break;
+
+                default:
+                    break;
             }
             break;
         case 0x100: /* Single Word DMA */
@@ -257,6 +260,9 @@ ide_get_xfer_time(ide_t *ide, int size)
                 case 0x04:
                     period = (25.0 / 3.0);
                     break;
+
+                default:
+                    break;
             }
             break;
         case 0x200: /* Multiword DMA */
@@ -270,6 +276,9 @@ ide_get_xfer_time(ide_t *ide, int size)
                 case 0x04:
                     period = (50.0 / 3.0);
                     break;
+
+                default:
+                    break;
             }
             break;
         case 0x300: /* Ultra DMA */
@@ -292,8 +301,14 @@ ide_get_xfer_time(ide_t *ide, int size)
                 case 0x20:
                     period = 100.0;
                     break;
+
+                default:
+                    break;
             }
             break;
+
+        default:
+            break;
     }
 
     period = (1.0 / period);         /* get us for 1 byte */
@@ -2576,6 +2591,9 @@ ide_callback(void *priv)
 
         case 0xFF:
             goto abort_cmd;
+
+        default:
+            break;
     }
 
 abort_cmd:
@@ -3043,6 +3061,9 @@ ide_init(const device_t *info)
             if (info->local & 1)
                 ide_board_init(1, 15, 0x170, 0x376, info->local);
             break;
+
+        default:
+            break;
     }
 
     return ide_drives;
diff --git a/src/disk/hdc_ide_cmd640.c b/src/disk/hdc_ide_cmd640.c
index cc642b3bb..6299bf7f4 100644
--- a/src/disk/hdc_ide_cmd640.c
+++ b/src/disk/hdc_ide_cmd640.c
@@ -37,7 +37,7 @@
 #include <86box/zip.h>
 #include <86box/mo.h>
 
-typedef struct {
+typedef struct cmd640_t {
     uint8_t  vlb_idx;
     uint8_t  id;
     uint8_t  in_cfg;
diff --git a/src/disk/hdc_ide_cmd646.c b/src/disk/hdc_ide_cmd646.c
index af5af2557..00f765bf7 100644
--- a/src/disk/hdc_ide_cmd646.c
+++ b/src/disk/hdc_ide_cmd646.c
@@ -37,7 +37,7 @@
 #include <86box/zip.h>
 #include <86box/mo.h>
 
-typedef struct {
+typedef struct cmd646_t {
     uint8_t     vlb_idx;
     uint8_t     single_channel;
     uint8_t     in_cfg;
diff --git a/src/disk/hdc_ide_opti611.c b/src/disk/hdc_ide_opti611.c
index 2f893f07e..4e44c55f8 100644
--- a/src/disk/hdc_ide_opti611.c
+++ b/src/disk/hdc_ide_opti611.c
@@ -30,11 +30,11 @@
 #include <86box/hdc_ide.h>
 #include <86box/plat_unused.h>
 
-typedef struct
-{
-    uint8_t tries,
-        in_cfg, cfg_locked,
-        regs[19];
+typedef struct opti611_t {
+    uint8_t tries;
+    uint8_t in_cfg;
+    uint8_t cfg_locked;
+    uint8_t regs[19];
 } opti611_t;
 
 static void opti611_ide_handler(opti611_t *dev);
diff --git a/src/disk/hdc_ide_sff8038i.c b/src/disk/hdc_ide_sff8038i.c
index e6eb6d6d9..fb16c8788 100644
--- a/src/disk/hdc_ide_sff8038i.c
+++ b/src/disk/hdc_ide_sff8038i.c
@@ -462,29 +462,27 @@ sff_bus_master_reset(sff8038i_t *dev, uint16_t old_base)
 }
 
 static void
-sff_reset(void *p)
+sff_reset(void *priv)
 {
-    int i = 0;
-
 #ifdef ENABLE_SFF_LOG
     sff_log("SFF8038i: Reset\n");
 #endif
 
-    for (i = 0; i < CDROM_NUM; i++) {
+    for (uint8_t i = 0; i < CDROM_NUM; i++) {
         if ((cdrom[i].bus_type == CDROM_BUS_ATAPI) && (cdrom[i].ide_channel < 4) && cdrom[i].priv)
             scsi_cdrom_reset((scsi_common_t *) cdrom[i].priv);
     }
-    for (i = 0; i < ZIP_NUM; i++) {
+    for (uint8_t i = 0; i < ZIP_NUM; i++) {
         if ((zip_drives[i].bus_type == ZIP_BUS_ATAPI) && (zip_drives[i].ide_channel < 4) && zip_drives[i].priv)
             zip_reset((scsi_common_t *) zip_drives[i].priv);
     }
-    for (i = 0; i < MO_NUM; i++) {
+    for (uint8_t i = 0; i < MO_NUM; i++) {
         if ((mo_drives[i].bus_type == MO_BUS_ATAPI) && (mo_drives[i].ide_channel < 4) && mo_drives[i].priv)
             mo_reset((scsi_common_t *) mo_drives[i].priv);
     }
 
-    sff_bus_master_set_irq(0x00, p);
-    sff_bus_master_set_irq(0x01, p);
+    sff_bus_master_set_irq(0x00, priv);
+    sff_bus_master_set_irq(0x01, priv);
 }
 
 void
@@ -543,9 +541,9 @@ sff_set_irq_pin(sff8038i_t *dev, int irq_pin)
 }
 
 static void
-sff_close(void *p)
+sff_close(void *priv)
 {
-    sff8038i_t *dev = (sff8038i_t *) p;
+    sff8038i_t *dev = (sff8038i_t *) priv;
 
     free(dev);
 
@@ -554,9 +552,8 @@ sff_close(void *p)
         next_id = 0;
 }
 
-static void
-    *
-    sff_init(UNUSED(const device_t *info))
+static void *
+sff_init(UNUSED(const device_t *info))
 {
     sff8038i_t *dev = (sff8038i_t *) malloc(sizeof(sff8038i_t));
     memset(dev, 0, sizeof(sff8038i_t));
diff --git a/src/disk/hdc_st506_at.c b/src/disk/hdc_st506_at.c
index eaf9832a5..b202ebe47 100644
--- a/src/disk/hdc_st506_at.c
+++ b/src/disk/hdc_st506_at.c
@@ -74,7 +74,7 @@
 #define CMD_DIAGNOSE       0x90
 #define CMD_SET_PARAMETERS 0x91
 
-typedef struct {
+typedef struct drive_t {
     int8_t  present;  /* drive is present */
     int8_t  hdd_num;  /* drive number in system */
     int8_t  steprate; /* current servo step rate */
@@ -89,7 +89,7 @@ typedef struct {
     int16_t curcyl;   /* current track number */
 } drive_t;
 
-typedef struct {
+typedef struct mfm_t {
     uint8_t  precomp;  /* 1: precomp/error register */
     uint8_t  error;
     uint8_t  secount;  /* 2: sector count register */
@@ -435,6 +435,9 @@ mfm_write(uint16_t port, uint8_t val, void *priv)
             mfm->fdisk = val;
             irq_update(mfm);
             break;
+
+        default:
+            break;
     }
 }
 
diff --git a/src/disk/hdc_st506_xt.c b/src/disk/hdc_st506_xt.c
index 743f25410..c895aab45 100644
--- a/src/disk/hdc_st506_xt.c
+++ b/src/disk/hdc_st506_xt.c
@@ -229,7 +229,7 @@ enum {
     STATE_DONE
 };
 
-typedef struct {
+typedef struct drive_t {
     int8_t  present;
     uint8_t hdd_num;
 
@@ -247,7 +247,7 @@ typedef struct {
     uint16_t cfg_cyl;
 } drive_t;
 
-typedef struct {
+typedef struct hdc_t {
     uint8_t type; /* controller type */
 
     uint8_t spt; /* sectors-per-track for controller */
@@ -290,7 +290,7 @@ typedef struct {
 } hdc_t;
 
 /* Supported drives table for the Xebec controller. */
-typedef struct {
+typedef struct hd_type_t {
     uint16_t tracks;
     uint8_t  hpc;
     uint8_t  spt;
@@ -508,6 +508,9 @@ st506_callback(void *priv)
                 case STATE_DONE:
                     st506_complete(dev);
                     break;
+
+                default:
+                    break;
             }
             break;
 
@@ -534,6 +537,9 @@ st506_callback(void *priv)
                 case STATE_SENT_DATA:
                     st506_complete(dev);
                     break;
+
+                default:
+                    break;
             }
             break;
 
@@ -570,6 +576,9 @@ st506_callback(void *priv)
                     ui_sb_update_icon(SB_HDD | HDD_BUS_MFM, 0);
                     st506_complete(dev);
                     break;
+
+                default:
+                    break;
             }
             break;
 
@@ -602,6 +611,9 @@ st506_callback(void *priv)
 
                     timer_advance_u64(&dev->timer, ST506_TIME);
                     break;
+
+                default:
+                    break;
             }
             break;
 
@@ -645,6 +657,9 @@ st506_callback(void *priv)
                     ui_sb_update_icon(SB_HDD | HDD_BUS_MFM, 0);
                     st506_complete(dev);
                     break;
+
+                default:
+                    break;
             }
             break;
 
@@ -737,6 +752,9 @@ st506_callback(void *priv)
                     }
                     dev->state = STATE_SEND_DATA;
                     break;
+
+                default:
+                    break;
             }
             break;
 
@@ -832,6 +850,9 @@ st506_callback(void *priv)
                     }
                     dev->state = STATE_RECEIVE_DATA;
                     break;
+
+                default:
+                    break;
             }
             break;
 
@@ -887,6 +908,9 @@ st506_callback(void *priv)
                     }
                     st506_complete(dev);
                     break;
+
+                default:
+                    break;
             }
             break;
 
@@ -906,6 +930,9 @@ st506_callback(void *priv)
                 case STATE_SENT_DATA:
                     st506_complete(dev);
                     break;
+
+                default:
+                    break;
             }
             break;
 
@@ -946,6 +973,9 @@ st506_callback(void *priv)
                 case STATE_SENT_DATA:
                     st506_complete(dev);
                     break;
+
+                default:
+                    break;
             }
             break;
 
@@ -987,6 +1017,9 @@ st506_callback(void *priv)
                 case STATE_RECEIVED_DATA:
                     st506_complete(dev);
                     break;
+
+                default:
+                    break;
             }
             break;
 
@@ -1008,6 +1041,9 @@ st506_callback(void *priv)
                     case STATE_SENT_DATA:
                         st506_complete(dev);
                         break;
+
+                    default:
+                        break;
                 }
             else {
                 st506_error(dev, ERR_BAD_COMMAND);
@@ -1120,6 +1156,9 @@ st506_callback(void *priv)
                 case STATE_SENT_DATA:
                     st506_complete(dev);
                     break;
+
+                default:
+                    break;
             }
             break;
 
@@ -1139,6 +1178,9 @@ st506_callback(void *priv)
                     /* FIXME: ignore the results. */
                     st506_complete(dev);
                     break;
+
+                default:
+                    break;
             }
             break;
 
@@ -1161,6 +1203,9 @@ st506_callback(void *priv)
                 case STATE_SENT_DATA:
                     st506_complete(dev);
                     break;
+
+                default:
+                    break;
             }
             break;
 
@@ -1205,6 +1250,9 @@ st506_read(uint16_t port, void *priv)
                         timer_set_delay_u64(&dev->timer, ST506_TIME);
                     }
                     break;
+
+                default:
+                    break;
             }
             break;
 
@@ -1217,6 +1265,9 @@ st506_read(uint16_t port, void *priv)
         case 2: /* read option jumpers */
             ret = dev->switches;
             break;
+
+        default:
+            break;
     }
     st506_xt_log("ST506: read(%04x) = %02x\n", port, ret);
 
@@ -1257,6 +1308,9 @@ st506_write(uint16_t port, uint8_t val, void *priv)
                         timer_set_delay_u64(&dev->timer, ST506_TIME);
                     }
                     break;
+
+                default:
+                    break;
             }
             break;
 
@@ -1282,6 +1336,9 @@ st506_write(uint16_t port, uint8_t val, void *priv)
                 picintc(1 << dev->irq);
             }
             break;
+
+        default:
+            break;
     }
 }
 
@@ -1561,6 +1618,9 @@ st506_init(const device_t *info)
                 case 19: /* v2.0 */
                     fn = ST11_BIOS_FILE_NEW;
                     break;
+
+                default:
+                    break;
             }
             dev->base      = device_get_config_hex16("base");
             dev->irq       = device_get_config_int("irq");
@@ -1665,6 +1725,9 @@ st506_init(const device_t *info)
             dev->base     = 0x01f0;
             dev->switches = 0x0c;
             break;
+
+        default:
+            break;
     }
 
     /* Load the ROM BIOS. */
diff --git a/src/disk/hdc_xta.c b/src/disk/hdc_xta.c
index 7b4fc432f..99e638335 100644
--- a/src/disk/hdc_xta.c
+++ b/src/disk/hdc_xta.c
@@ -182,7 +182,7 @@ enum {
 
 /* The device control block (6 bytes) */
 #pragma pack(push, 1)
-typedef struct {
+typedef struct dcb_t {
     uint8_t cmd; /* [7:5] class, [4:0] opcode    */
 
     uint8_t head   : 5; /* [4:0] head number        */
@@ -202,7 +202,7 @@ typedef struct {
 
 /* The (configured) Drive Parameters. */
 #pragma pack(push, 1)
-typedef struct {
+typedef struct dprm_t {
     uint8_t cyl_high; /* (MSB) number of cylinders */
     uint8_t cyl_low;  /* (LSB) number of cylinders */
     uint8_t heads;    /* number of heads per cylinder */
@@ -215,7 +215,7 @@ typedef struct {
 #pragma pack(pop)
 
 /* Define an attached drive. */
-typedef struct {
+typedef struct drive_t {
     int8_t id;   /* drive ID on bus */
     int8_t present; /* drive is present */
     int8_t hdd_num; /* index to global disk table */
@@ -232,7 +232,7 @@ typedef struct {
     uint16_t cfg_tracks;
 } drive_t;
 
-typedef struct {
+typedef struct hdc_t {
     const char *name; /* controller name */
 
     uint16_t base; /* controller base I/O address */
@@ -436,6 +436,9 @@ do_fmt:
 
             /* This saves us a LOT of code. */
             goto do_fmt;
+
+        default:
+            break;
     }
 
     /* De-activate the status icon. */
@@ -497,6 +500,10 @@ hdc_callback(void *priv)
 
                 case STATE_SDONE:
                     set_intr(dev);
+                    break;
+
+                default:
+                    break;
             }
             break;
 
@@ -606,6 +613,9 @@ do_send:
                     /* This saves us a LOT of code. */
                     dev->state = STATE_SEND;
                     goto do_send;
+
+                default:
+                    break;
             }
             break;
 
@@ -710,6 +720,9 @@ do_recv:
                     /* This saves us a LOT of code. */
                     dev->state = STATE_RECV;
                     goto do_recv;
+
+                default:
+                    break;
             }
             break;
 
@@ -758,6 +771,9 @@ do_recv:
                     dev->status &= ~STAT_REQ;
                     set_intr(dev);
                     break;
+
+                default:
+                    break;
             }
             break;
 
@@ -802,6 +818,9 @@ do_recv:
                                dev->data, dev->buf_len);
                     set_intr(dev);
                     break;
+
+                default:
+                    break;
             }
             break;
 
@@ -815,6 +834,9 @@ do_recv:
                 case STATE_RDONE:
                     set_intr(dev);
                     break;
+
+                default:
+                    break;
             }
             break;
 
@@ -834,6 +856,9 @@ do_recv:
                 case STATE_RDONE:
                     set_intr(dev);
                     break;
+
+                default:
+                    break;
             }
             break;
 
@@ -847,6 +872,9 @@ do_recv:
                 case STATE_RDONE:
                     set_intr(dev);
                     break;
+
+                default:
+                    break;
             }
             break;
 
@@ -900,6 +928,9 @@ hdc_read(uint16_t port, void *priv)
         case 2:         /* "read option jumpers" */
             ret = 0xff; /* all switches off */
             break;
+
+        default:
+            break;
     }
 
     return ret;
@@ -963,6 +994,9 @@ hdc_write(uint16_t port, uint8_t val, void *priv)
 #endif
             dev->intr = val;
             break;
+
+        default:
+            break;
     }
 }
 
@@ -1000,12 +1034,16 @@ xta_init(const device_t *info)
             dev->irq  = 5;
             dev->dma  = 3;
             break;
+
+        default:
+            break;
     }
 
     xta_log("%s: initializing (I/O=%04X, IRQ=%d, DMA=%d",
             dev->name, dev->base, dev->irq, dev->dma);
     if (dev->rom_addr != 0x000000)
         xta_log(", BIOS=%06X", dev->rom_addr);
+
     xta_log(")\n");
 
     /* Load any disks for this device class. */
diff --git a/src/disk/hdc_xtide.c b/src/disk/hdc_xtide.c
index 5caa2337e..91bf5a2e3 100644
--- a/src/disk/hdc_xtide.c
+++ b/src/disk/hdc_xtide.c
@@ -52,7 +52,7 @@
 #define ROM_PATH_PS2AT  "roms/hdd/xtide/ide_at_1_1_5.bin"
 #define ROM_PATH_AT_386 "roms/hdd/xtide/ide_386.bin"
 
-typedef struct {
+typedef struct xtide_t {
     void   *ide_board;
     uint8_t data_high;
     rom_t   bios_rom;
@@ -85,6 +85,9 @@ xtide_write(uint16_t port, uint8_t val, void *priv)
         case 0xe:
             ide_write_devctl(0x0, val, xtide->ide_board);
             return;
+
+        default:
+            break;
     }
 }
 
diff --git a/src/disk/hdd.c b/src/disk/hdd.c
index 751121690..a4b8a6f73 100644
--- a/src/disk/hdd.c
+++ b/src/disk/hdd.c
@@ -107,8 +107,8 @@ hdd_bus_to_string(int bus, UNUSED(int cdrom))
     char *s = "none";
 
     switch (bus) {
-        case HDD_BUS_DISABLED:
         default:
+        case HDD_BUS_DISABLED:
             break;
 
         case HDD_BUS_MFM:
@@ -165,7 +165,7 @@ hdd_seek_get_time(hard_disk_t *hdd, uint32_t dst_addr, uint8_t operation, uint8_
         fatal("hdd_seek_get_time(): hdd->num_zones < 0)\n");
         return 0.0;
     }
-    for (int i = 0; i < hdd->num_zones; i++) {
+    for (uint32_t i = 0; i < hdd->num_zones; i++) {
         zone = &hdd->zones[i];
         if (zone->end_sector >= dst_addr)
             break;
diff --git a/src/disk/hdd_image.c b/src/disk/hdd_image.c
index 0e61271a9..4d3e354f2 100644
--- a/src/disk/hdd_image.c
+++ b/src/disk/hdd_image.c
@@ -40,8 +40,7 @@
 #define HDD_IMAGE_HDX 2
 #define HDD_IMAGE_VHD 3
 
-typedef struct
-{
+typedef struct hdd_image_t {
     FILE     *file; /* Used for HDD_IMAGE_RAW, HDD_IMAGE_HDI, and HDD_IMAGE_HDX. */
     MVHDMeta *vhd;  /* Used for HDD_IMAGE_VHD. */
     uint32_t  base;
diff --git a/src/disk/mo.c b/src/disk/mo.c
index 9cf46afab..d7dbba895 100644
--- a/src/disk/mo.c
+++ b/src/disk/mo.c
@@ -498,6 +498,9 @@ mo_atapi_phase_to_scsi(mo_t *dev)
                 return 1;
             case 3:
                 return 7;
+
+            default:
+                break;
         }
     } else {
         if ((dev->phase & 3) == 3)
@@ -569,6 +572,9 @@ mo_mode_sense_read(mo_t *dev, uint8_t page_control, uint8_t page, uint8_t pos)
             else
                 return mo_mode_sense_pages_default.pages[page][pos];
             break;
+
+        default:
+            break;
     }
 
     return 0;
@@ -1419,6 +1425,9 @@ mo_command(scsi_common_t *sc, uint8_t *cdb)
                     dev->sector_pos = (((uint32_t) cdb[2]) << 24) | (((uint32_t) cdb[3]) << 16) | (((uint32_t) cdb[4]) << 8) | ((uint32_t) cdb[5]);
                     mo_log("MO %i: Length: %i, LBA: %i\n", dev->id, dev->sector_len, dev->sector_pos);
                     break;
+
+                default:
+                    break;
             }
 
             if (!dev->sector_len) {
@@ -1512,6 +1521,9 @@ mo_command(scsi_common_t *sc, uint8_t *cdb)
                     dev->sector_len = (((uint32_t) cdb[6]) << 24) | (((uint32_t) cdb[7]) << 16) | (((uint32_t) cdb[8]) << 8) | ((uint32_t) cdb[9]);
                     dev->sector_pos = (((uint32_t) cdb[2]) << 24) | (((uint32_t) cdb[3]) << 16) | (((uint32_t) cdb[4]) << 8) | ((uint32_t) cdb[5]);
                     break;
+
+                default:
+                    break;
             }
 
             if ((dev->sector_pos >= dev->drv->medium_size) /* ||
@@ -1637,6 +1649,9 @@ mo_command(scsi_common_t *sc, uint8_t *cdb)
                 case 3: /* Load the disk (close tray). */
                     mo_reload(dev->id);
                     break;
+
+                default:
+                    break;
             }
 
             mo_command_complete(dev);
@@ -1744,6 +1759,9 @@ mo_command(scsi_common_t *sc, uint8_t *cdb)
                 case GPCMD_SEEK_10:
                     pos = (cdb[2] << 24) | (cdb[3] << 16) | (cdb[4] << 8) | cdb[5];
                     break;
+
+                default:
+                    break;
             }
             mo_seek(dev, pos);
             mo_command_complete(dev);
@@ -1782,6 +1800,9 @@ mo_command(scsi_common_t *sc, uint8_t *cdb)
                 case GPCMD_ERASE_12:
                     dev->sector_len = (((uint32_t) cdb[6]) << 24) | (((uint32_t) cdb[7]) << 16) | (((uint32_t) cdb[8]) << 8) | ((uint32_t) cdb[9]);
                     break;
+
+                default:
+                    break;
             }
 
             /*Erase all remaining sectors*/
@@ -1804,6 +1825,9 @@ mo_command(scsi_common_t *sc, uint8_t *cdb)
                 case GPCMD_ERASE_12:
                     dev->sector_pos = (((uint32_t) cdb[2]) << 24) | (((uint32_t) cdb[3]) << 16) | (((uint32_t) cdb[4]) << 8) | ((uint32_t) cdb[5]);
                     break;
+
+                default:
+                    break;
             }
 
             dev->sector_pos += previous_pos;
@@ -1958,6 +1982,9 @@ mo_phase_data_out(scsi_common_t *sc)
                 return 0;
             }
             break;
+
+        default:
+            break;
     }
 
     mo_command_stop((scsi_common_t *) dev);
diff --git a/src/disk/zip.c b/src/disk/zip.c
index 768963695..648182cc3 100644
--- a/src/disk/zip.c
+++ b/src/disk/zip.c
@@ -643,6 +643,9 @@ zip_atapi_phase_to_scsi(zip_t *dev)
                 return 1;
             case 3:
                 return 7;
+
+            default:
+                break;
         }
     } else {
         if ((dev->phase & 3) == 3)
@@ -713,13 +716,11 @@ zip_mode_sense_read(zip_t *dev, uint8_t page_control, uint8_t page, uint8_t pos)
             if (dev->drv->is_250 && (page == 5) && (pos == 9) && (dev->drv->medium_size == ZIP_SECTORS))
                 return 0x60;
             return dev->ms_pages_saved.pages[page][pos];
-            break;
         case 1:
             if (dev->drv->is_250)
                 return zip_250_mode_sense_pages_changeable.pages[page][pos];
             else
                 return zip_mode_sense_pages_changeable.pages[page][pos];
-            break;
         case 2:
             if (dev->drv->is_250) {
                 if ((page == 5) && (pos == 9) && (dev->drv->medium_size == ZIP_SECTORS))
@@ -734,6 +735,8 @@ zip_mode_sense_read(zip_t *dev, uint8_t page_control, uint8_t page, uint8_t pos)
                 else
                     return zip_mode_sense_pages_default.pages[page][pos];
             }
+
+        default:
             break;
     }
 
@@ -962,7 +965,7 @@ zip_data_command_finish(zip_t *dev, int len, int block_len, int alloc_len, int d
 }
 
 static void
-zip_sense_clear(zip_t *dev, int command)
+zip_sense_clear(zip_t *dev, UNUSED(int command))
 {
     zip_sense_key = zip_asc = zip_ascq = 0;
 }
@@ -1116,7 +1119,7 @@ zip_data_phase_error(zip_t *dev)
 }
 
 static int
-zip_blocks(zip_t *dev, int32_t *len, int first_batch, int out)
+zip_blocks(zip_t *dev, int32_t *len, UNUSED(int first_batch), int out)
 {
     *len = 0;
 
@@ -1251,7 +1254,9 @@ zip_pre_execution_check(zip_t *dev, uint8_t *cdb)
 static void
 zip_seek(zip_t *dev, uint32_t pos)
 {
-    /* zip_log("ZIP %i: Seek %08X\n", dev->id, pos); */
+#if 0
+    zip_log("ZIP %i: Seek %08X\n", dev->id, pos);
+#endif
     dev->sector_pos = pos;
 }
 
@@ -1513,6 +1518,9 @@ zip_command(scsi_common_t *sc, uint8_t *cdb)
                     dev->sector_len = (((uint32_t) cdb[6]) << 24) | (((uint32_t) cdb[7]) << 16) | (((uint32_t) cdb[8]) << 8) | ((uint32_t) cdb[9]);
                     dev->sector_pos = (((uint32_t) cdb[2]) << 24) | (((uint32_t) cdb[3]) << 16) | (((uint32_t) cdb[4]) << 8) | ((uint32_t) cdb[5]);
                     break;
+
+                default:
+                    break;
             }
 
             if (!dev->sector_len) {
@@ -1597,6 +1605,9 @@ zip_command(scsi_common_t *sc, uint8_t *cdb)
                     dev->sector_len = (((uint32_t) cdb[6]) << 24) | (((uint32_t) cdb[7]) << 16) | (((uint32_t) cdb[8]) << 8) | ((uint32_t) cdb[9]);
                     dev->sector_pos = (((uint32_t) cdb[2]) << 24) | (((uint32_t) cdb[3]) << 16) | (((uint32_t) cdb[4]) << 8) | ((uint32_t) cdb[5]);
                     break;
+
+                default:
+                    break;
             }
 
             if ((dev->sector_pos >= dev->drv->medium_size) /* ||
@@ -1775,6 +1786,9 @@ zip_command(scsi_common_t *sc, uint8_t *cdb)
                 case 3: /* Load the disc (close tray). */
                     zip_reload(dev->id);
                     break;
+
+                default:
+                    break;
             }
 
             zip_command_complete(dev);
@@ -1911,6 +1925,9 @@ atapi_out:
                 case GPCMD_SEEK_10:
                     pos = (cdb[2] << 24) | (cdb[3] << 16) | (cdb[4] << 8) | cdb[5];
                     break;
+
+                default:
+                    break;
             }
             zip_seek(dev, pos);
             zip_command_complete(dev);
@@ -2177,6 +2194,9 @@ zip_phase_data_out(scsi_common_t *sc)
                 return 0;
             }
             break;
+
+        default:
+            break;
     }
 
     zip_command_stop((scsi_common_t *) dev);
diff --git a/src/dma.c b/src/dma.c
index 710ea0abf..9dc984125 100644
--- a/src/dma.c
+++ b/src/dma.c
@@ -31,6 +31,7 @@
 #include <86box/io.h>
 #include <86box/pic.h>
 #include <86box/dma.h>
+#include <86box/plat_unused.h>
 
 dma_t   dma[8];
 uint8_t dma_e;
@@ -50,9 +51,9 @@ static uint16_t dma_sg_base;
 static uint16_t dma16_buffer[65536];
 static uint32_t dma_mask;
 
-static struct {
-    int xfr_command,
-        xfr_channel;
+static struct dma_ps2_t {
+    int xfr_command;
+    int xfr_channel;
     int byte_ptr;
 
     int is_ps2;
@@ -228,6 +229,9 @@ dma_sg_write(uint16_t port, uint8_t val, void *priv)
             dev->ptr = (dev->ptr & 0x00fffffc) | (val << 24);
             dev->ptr %= (mem_size * 1024);
             break;
+
+        default:
+            break;
     }
 }
 
@@ -258,6 +262,9 @@ dma_sg_writew(uint16_t port, uint16_t val, void *priv)
             dev->ptr = (dev->ptr & 0x0000fffc) | (val << 16);
             dev->ptr %= (mem_size * 1024);
             break;
+
+        default:
+            break;
     }
 }
 
@@ -284,6 +291,9 @@ dma_sg_writel(uint16_t port, uint32_t val, void *priv)
             dev->ptr %= (mem_size * 1024);
             dev->ptr0 = val & 0xff;
             break;
+
+        default:
+            break;
     }
 }
 
@@ -325,6 +335,9 @@ dma_sg_read(uint16_t port, void *priv)
         case 0x23:
             ret = dev->ptr >> 24;
             break;
+
+        default:
+            break;
     }
 
     dma_log("DMA S/G BYTE  read : %04X       %02X\n", port, ret);
@@ -356,6 +369,9 @@ dma_sg_readw(uint16_t port, void *priv)
         case 0x22:
             ret = dev->ptr >> 16;
             break;
+
+        default:
+            break;
     }
 
     dma_log("DMA S/G WORD  read : %04X     %04X\n", port, ret);
@@ -384,6 +400,9 @@ dma_sg_readl(uint16_t port, void *priv)
         case 0x20:
             ret = dev->ptr0 | (dev->ptr & 0xffffff00);
             break;
+
+        default:
+            break;
     }
 
     dma_log("DMA S/G DWORD read : %04X %08X\n", port, ret);
@@ -392,7 +411,7 @@ dma_sg_readl(uint16_t port, void *priv)
 }
 
 static void
-dma_ext_mode_write(uint16_t addr, uint8_t val, void *priv)
+dma_ext_mode_write(uint16_t addr, uint8_t val, UNUSED(void *priv))
 {
     int channel = (val & 0x03);
 
@@ -416,11 +435,14 @@ dma_ext_mode_write(uint16_t addr, uint8_t val, void *priv)
         case 0x03:
             dma[channel].transfer_mode = 0x0102;
             break;
+
+        default:
+            break;
     }
 }
 
 static uint8_t
-dma_sg_int_status_read(uint16_t addr, void *priv)
+dma_sg_int_status_read(UNUSED(uint16_t addr), UNUSED(void *priv))
 {
     uint8_t ret = 0x00;
 
@@ -433,7 +455,7 @@ dma_sg_int_status_read(uint16_t addr, void *priv)
 }
 
 static uint8_t
-dma_read(uint16_t addr, void *priv)
+dma_read(uint16_t addr, UNUSED(void *priv))
 {
     int     channel = (addr >> 1) & 3;
     uint8_t temp;
@@ -468,13 +490,16 @@ dma_read(uint16_t addr, void *priv)
 
         case 0xd: /*Temporary register*/
             return 0;
+
+        default:
+            break;
     }
 
     return (dmaregs[0][addr & 0xf]);
 }
 
 static void
-dma_write(uint16_t addr, uint8_t val, void *priv)
+dma_write(uint16_t addr, uint8_t val, UNUSED(void *priv))
 {
     int channel = (addr >> 1) & 3;
 
@@ -562,11 +587,14 @@ dma_write(uint16_t addr, uint8_t val, void *priv)
         case 0xf: /*Mask write*/
             dma_m = (dma_m & 0xf0) | (val & 0xf);
             return;
+
+        default:
+            break;
     }
 }
 
 static uint8_t
-dma_ps2_read(uint16_t addr, void *priv)
+dma_ps2_read(uint16_t addr, UNUSED(void *priv))
 {
     dma_t  *dma_c = &dma[dma_ps2.xfr_channel];
     uint8_t temp  = 0xff;
@@ -589,6 +617,9 @@ dma_ps2_read(uint16_t addr, void *priv)
                             temp             = (dma_c->ac >> 16) & 0xff;
                             dma_ps2.byte_ptr = 0;
                             break;
+
+                        default:
+                            break;
                     }
                     break;
 
@@ -626,12 +657,15 @@ dma_ps2_read(uint16_t addr, void *priv)
                     fatal("Bad XFR Read command %i channel %i\n", dma_ps2.xfr_command, dma_ps2.xfr_channel);
             }
             break;
+
+        default:
+            break;
     }
     return temp;
 }
 
 static void
-dma_ps2_write(uint16_t addr, uint8_t val, void *priv)
+dma_ps2_write(uint16_t addr, uint8_t val, UNUSED(void *priv))
 {
     dma_t  *dma_c = &dma[dma_ps2.xfr_channel];
     uint8_t mode;
@@ -654,6 +688,9 @@ dma_ps2_write(uint16_t addr, uint8_t val, void *priv)
                     if (!(dma_m & (1 << dma_ps2.xfr_channel)))
                         dma_ps2_run(dma_ps2.xfr_channel);
                     break;
+
+                default:
+                    break;
             }
             break;
 
@@ -683,6 +720,9 @@ dma_ps2_write(uint16_t addr, uint8_t val, void *priv)
                             dma_c->ac        = (dma_c->ac & 0x00ffff) | (val << 16);
                             dma_ps2.byte_ptr = 0;
                             break;
+
+                        default:
+                            break;
                     }
                     dma_c->ab = dma_c->ac;
                     break;
@@ -719,11 +759,14 @@ dma_ps2_write(uint16_t addr, uint8_t val, void *priv)
                     fatal("Bad XFR command %i channel %i val %02x\n", dma_ps2.xfr_command, dma_ps2.xfr_channel, val);
             }
             break;
+
+        default:
+            break;
     }
 }
 
 static uint8_t
-dma16_read(uint16_t addr, void *priv)
+dma16_read(uint16_t addr, UNUSED(void *priv))
 {
     int     channel = ((addr >> 2) & 3) + 4;
     uint8_t temp;
@@ -760,13 +803,16 @@ dma16_read(uint16_t addr, void *priv)
             temp |= dma_stat >> 4;
             dma_stat &= ~0xf0;
             return temp;
+
+        default:
+            break;
     }
 
     return (dmaregs[1][addr & 0xf]);
 }
 
 static void
-dma16_write(uint16_t addr, uint8_t val, void *priv)
+dma16_write(uint16_t addr, uint8_t val, UNUSED(void *priv))
 {
     int channel = ((addr >> 2) & 3) + 4;
     addr >>= 1;
@@ -855,6 +901,9 @@ dma16_write(uint16_t addr, uint8_t val, void *priv)
         case 0xf: /*Mask write*/
             dma_m = (dma_m & 0x0f) | ((val & 0xf) << 4);
             return;
+
+        default:
+            break;
     }
 }
 
@@ -864,7 +913,7 @@ dma16_write(uint16_t addr, uint8_t val, void *priv)
     }
 
 static void
-dma_page_write(uint16_t addr, uint8_t val, void *priv)
+dma_page_write(uint16_t addr, uint8_t val, UNUSED(void *priv))
 {
     uint8_t convert[8] = CHANNELS;
 
@@ -897,7 +946,7 @@ dma_page_write(uint16_t addr, uint8_t val, void *priv)
 }
 
 static uint8_t
-dma_page_read(uint16_t addr, void *priv)
+dma_page_read(uint16_t addr, UNUSED(void *priv))
 {
     uint8_t convert[8] = CHANNELS;
     uint8_t ret        = 0xff;
@@ -917,7 +966,7 @@ dma_page_read(uint16_t addr, void *priv)
 }
 
 static void
-dma_high_page_write(uint16_t addr, uint8_t val, void *priv)
+dma_high_page_write(uint16_t addr, uint8_t val, UNUSED(void *priv))
 {
     uint8_t convert[8] = CHANNELS;
 
@@ -937,7 +986,7 @@ dma_high_page_write(uint16_t addr, uint8_t val, void *priv)
 }
 
 static uint8_t
-dma_high_page_read(uint16_t addr, void *priv)
+dma_high_page_read(uint16_t addr, UNUSED(void *priv))
 {
     uint8_t convert[8] = CHANNELS;
     uint8_t ret        = 0xff;
diff --git a/src/floppy/fdc.c b/src/floppy/fdc.c
index e29d5d2f8..99767d857 100644
--- a/src/floppy/fdc.c
+++ b/src/floppy/fdc.c
@@ -203,9 +203,9 @@ fdc_get_current_drive(void)
 }
 
 void
-fdc_ctrl_reset(void *p)
+fdc_ctrl_reset(void *priv)
 {
-    fdc_t *fdc = (fdc_t *) p;
+    fdc_t *fdc = (fdc_t *) priv;
 
     fdc->stat = 0x80;
     fdc->pnum = fdc->ptot = 0;
@@ -227,8 +227,8 @@ int
 fdc_get_compare_condition(fdc_t *fdc)
 {
     switch (fdc->interrupt) {
-        case 0x11:
         default:
+        case 0x11:
             return 0;
         case 0x19:
             return 1;
@@ -517,6 +517,9 @@ fdc_update_rate(fdc_t *fdc, int drive)
                     case 2:
                         fdc->bit_rate = 2000;
                         break;
+
+                    default:
+                        break;
                 }
                 break;
             case 2: /*Double density*/
@@ -525,6 +528,9 @@ fdc_update_rate(fdc_t *fdc, int drive)
             case 3: /*Extended density*/
                 fdc->bit_rate = 1000;
                 break;
+
+            default:
+                break;
         }
 
     fdc->bitcell_period = (1000000 / fdc->bit_rate) * 2; /*Bitcell period in ns*/
@@ -544,8 +550,9 @@ fdc_get_bit_rate(fdc_t *fdc)
             return 2;
         case 1000:
             return 3;
+
         default:
-            return 2;
+            break;
     }
     return 2;
 }
@@ -566,6 +573,9 @@ fdc_get_densel(fdc_t *fdc, int drive)
                 return 0;
             case 2:
                 return 1;
+
+            default:
+                break;
         }
     }
 
@@ -575,6 +585,9 @@ fdc_get_densel(fdc_t *fdc, int drive)
                 return 1;
             case 3:
                 return 0;
+
+            default:
+                break;
         }
     } else {
         switch (fdc->densel_force) {
@@ -582,6 +595,9 @@ fdc_get_densel(fdc_t *fdc, int drive)
                 return 0;
             case 1:
                 return 1;
+
+            default:
+                break;
         }
     }
 
@@ -592,6 +608,9 @@ fdc_get_densel(fdc_t *fdc, int drive)
         case 1:
         case 2:
             return fdc->densel_polarity ? 0 : 1;
+
+        default:
+            break;
     }
 
     return 0;
@@ -1230,6 +1249,9 @@ fdc_write(uint16_t addr, uint8_t val, void *priv)
                                 fdc->perp |= (fdc->params[0] & 0x03);
                             }
                             return;
+
+                        default:
+                            break;
                     }
                 } else
                     fdc->stat = 0x90 | (fdc->stat & 0xf);
@@ -1242,6 +1264,9 @@ fdc_write(uint16_t addr, uint8_t val, void *priv)
             if (fdc->flags & FDC_FLAG_PS1)
                 fdc->noprec = !!(val & 0x04);
             return;
+
+        default:
+            break;
     }
 }
 
@@ -1300,6 +1325,9 @@ fdc_read(uint16_t addr, void *priv)
                     case 3:
                         ret |= 0x61;
                         break;
+
+                    default:
+                        break;
                 }
             } else {
                 if (is486 || !fdc->enable_3f1)
@@ -1684,6 +1712,9 @@ fdc_callback(void *priv)
                         fdc->stat = 0x90;
                     }
                     break;
+
+                default:
+                    break;
             }
             fdc->inread = 1;
             return;
@@ -1787,6 +1818,9 @@ fdc_callback(void *priv)
             fdc->paramstogo     = 1;
             fdc->interrupt      = 0;
             return;
+
+        default:
+            break;
     }
 }
 
diff --git a/src/floppy/fdc_magitronic.c b/src/floppy/fdc_magitronic.c
index 5ebd9de42..084ce8c81 100644
--- a/src/floppy/fdc_magitronic.c
+++ b/src/floppy/fdc_magitronic.c
@@ -35,7 +35,7 @@
 #define ROM_ADDR     (uint32_t)(device_get_config_hex20("bios_addr") & 0x000fffff)
 
 #define DRIVE_SELECT (int) (real_drive(dev->fdc_controller, i))
-typedef struct {
+typedef struct b215_t {
     fdc_t *fdc_controller;
     rom_t  rom;
 } b215_t;
@@ -58,7 +58,7 @@ b215_read(UNUSED(uint16_t addr), void *priv)
 */
     int drive_spec[2];
 
-    for (int i = 0; i <= 1; i++) {
+    for (uint8_t i = 0; i <= 1; i++) {
         if (fdd_is_525(DRIVE_SELECT)) {
             if (!fdd_is_dd(DRIVE_SELECT))
                 drive_spec[i] = 1;
diff --git a/src/floppy/fdc_monster.c b/src/floppy/fdc_monster.c
index 91598b897..1629ac1c4 100644
--- a/src/floppy/fdc_monster.c
+++ b/src/floppy/fdc_monster.c
@@ -37,7 +37,7 @@
 #define BIOS_ADDR       (uint32_t)(device_get_config_hex20("bios_addr") & 0x000fffff)
 #define ROM_MONSTER_FDC "roms/floppy/monster-fdc/floppy_bios.bin"
 
-typedef struct {
+typedef struct monster_fdc_t {
     rom_t  bios_rom;
     fdc_t *fdc_pri;
     fdc_t *fdc_sec;
diff --git a/src/floppy/fdc_pii15xb.c b/src/floppy/fdc_pii15xb.c
index 8bbda7991..5fd38d250 100644
--- a/src/floppy/fdc_pii15xb.c
+++ b/src/floppy/fdc_pii15xb.c
@@ -80,7 +80,7 @@ MiniMicro 4 also won't work with the XT FDC which the Zilog claims to be.
 #define ROM_PII_151B "roms/floppy/dtk/pii-151b.rom"
 #define ROM_PII_158B "roms/floppy/dtk/pii-158b.rom"
 
-typedef struct {
+typedef struct pii_t {
     rom_t bios_rom;
 } pii_t;
 
diff --git a/src/floppy/fdd_86f.c b/src/floppy/fdd_86f.c
index c59e04052..f2c0ffbbc 100644
--- a/src/floppy/fdd_86f.c
+++ b/src/floppy/fdd_86f.c
@@ -125,30 +125,30 @@ enum {
     FMT_POSTTRK_GAP4
 };
 
-typedef struct {
+typedef struct sliding_buffer_t {
     uint8_t  buffer[10];
     uint32_t pos;
     uint32_t len;
 } sliding_buffer_t;
 
-typedef struct {
+typedef struct find_t {
     uint32_t bits_obtained;
     uint16_t bytes_obtained;
     uint16_t sync_marks;
     uint32_t sync_pos;
 } find_t;
 
-typedef struct {
+typedef struct split_byte_t {
     unsigned nibble0 : 4;
     unsigned nibble1 : 4;
 } split_byte_t;
 
-typedef union {
+typedef union decoded_t {
     uint8_t      byte;
     split_byte_t nibbles;
 } decoded_t;
 
-typedef struct {
+typedef struct sector_t {
     uint8_t c;
     uint8_t h;
     uint8_t r;
@@ -179,7 +179,7 @@ typedef struct {
  *      If bits 6, 5 are 0, and bit 7 is 1, the extra bitcell count
  *      specifies the entire bitcell count
  */
-typedef struct {
+typedef struct d86f_t {
     FILE   *f;
     uint8_t   state;
     uint8_t   fill;
@@ -369,43 +369,43 @@ d86f_index_hole_pos(int drive, int side)
 }
 
 uint32_t
-null_index_hole_pos(int drive, int side)
+null_index_hole_pos(UNUSED(int drive), UNUSED(int side))
 {
     return 0;
 }
 
 uint16_t
-null_disk_flags(int drive)
+null_disk_flags(UNUSED(int drive))
 {
     return 0x09;
 }
 
 uint16_t
-null_side_flags(int drive)
+null_side_flags(UNUSED(int drive))
 {
     return 0x0A;
 }
 
 void
-null_writeback(int drive)
+null_writeback(UNUSED(int drive))
 {
     return;
 }
 
 void
-null_set_sector(int drive, int side, uint8_t c, uint8_t h, uint8_t r, uint8_t n)
+null_set_sector(UNUSED(int drive), UNUSED(int side), UNUSED(uint8_t c), UNUSED(uint8_t h), UNUSED(uint8_t r), UNUSED(uint8_t n))
 {
     return;
 }
 
 void
-null_write_data(int drive, int side, uint16_t pos, uint8_t data)
+null_write_data(UNUSED(int drive), UNUSED(int side), UNUSED(uint16_t pos), UNUSED(uint8_t data))
 {
     return;
 }
 
 int
-null_format_conditions(int drive)
+null_format_conditions(UNUSED(int drive))
 {
     return 0;
 }
@@ -419,7 +419,7 @@ d86f_extra_bit_cells(int drive, int side)
 }
 
 int32_t
-null_extra_bit_cells(int drive, int side)
+null_extra_bit_cells(UNUSED(int drive), UNUSED(int side))
 {
     return 0;
 }
@@ -433,7 +433,7 @@ common_encoded_data(int drive, int side)
 }
 
 void
-common_read_revolution(int drive)
+common_read_revolution(UNUSED(int drive))
 {
     return;
 }
@@ -637,9 +637,9 @@ d86f_get_array_size(int drive, int side, int words)
         array_size = 0;
     else
         switch (hole) {
+            default:
             case 0:
             case 1:
-            default:
                 array_size = 12500;
                 switch (rm) {
                     case 1:
@@ -830,7 +830,7 @@ d86f_has_extra_bit_cells(int drive)
 }
 
 uint32_t
-d86f_header_size(int drive)
+d86f_header_size(UNUSED(int drive))
 {
     return 8;
 }
@@ -902,15 +902,14 @@ d86f_wrong_densel(int drive)
         is_3mode = 1;
 
     switch (d86f_hole(drive)) {
-        case 0:
         default:
+        case 0:
             if (fdd_is_dd(drive))
                 return 0;
             if (fdd_get_densel(drive))
                 return 1;
             else
                 return 0;
-            break;
 
         case 1:
             if (fdd_is_dd(drive))
@@ -923,7 +922,6 @@ d86f_wrong_densel(int drive)
                 else
                     return 1;
             }
-            break;
 
         case 2:
             if (fdd_is_dd(drive) || !fdd_is_ed(drive))
@@ -932,7 +930,6 @@ d86f_wrong_densel(int drive)
                 return 0;
             else
                 return 1;
-            break;
     }
 }
 
@@ -983,6 +980,9 @@ d86f_encode_byte(int drive, int sync, decoded_t b, decoded_t prev_b)
 
                 case 0xfc:
                     return result | d86f_encode_get_clock(0x01);
+
+                default:
+                    break;
             }
         } else {
             switch (b.byte) {
@@ -993,6 +993,9 @@ d86f_encode_byte(int drive, int sync, decoded_t b, decoded_t prev_b)
 
                 case 0xfc:
                     return result | d86f_encode_get_clock(0xd7);
+
+                default:
+                    break;
             }
         }
     }
@@ -1040,6 +1043,9 @@ d86f_get_bitcell_period(int drive)
         case 5:
             rate = 2000.0;
             break;
+
+        default:
+            break;
     }
 
     if (!mfm)
@@ -1194,7 +1200,7 @@ d86f_put_bit(int drive, int side, int bit)
 }
 
 static uint8_t
-decodefm(int drive, uint16_t dat)
+decodefm(UNUSED(int drive), uint16_t dat)
 {
     uint8_t temp = 0;
 
@@ -1542,6 +1548,9 @@ d86f_compare_byte(int drive, uint8_t received_byte, uint8_t disk_byte)
             if ((received_byte >= disk_byte) || (received_byte == 0xFF))
                 dev->satisfying_bytes++;
             break;
+
+        default:
+            break;
     }
 }
 
@@ -1850,7 +1859,7 @@ endian_swap(uint16_t word)
 }
 
 void
-d86f_format_finish(int drive, int side, int mfm, uint16_t sc, uint16_t gap_fill, int do_write)
+d86f_format_finish(int drive, int side, int mfm, UNUSED(uint16_t sc), uint16_t gap_fill, int do_write)
 {
     d86f_t *dev = d86f[drive];
 
@@ -1871,7 +1880,7 @@ d86f_format_finish(int drive, int side, int mfm, uint16_t sc, uint16_t gap_fill,
 }
 
 void
-d86f_format_turbo_finish(int drive, int side, int do_write)
+d86f_format_turbo_finish(int drive, UNUSED(int side), int do_write)
 {
     d86f_t *dev = d86f[drive];
 
@@ -2071,13 +2080,14 @@ d86f_format_track(int drive, int side, int do_write)
                     /* Sector within allotted amount, change state to SECTOR_ID_SYNC. */
                     dev->format_state = FMT_SECTOR_ID_SYNC;
                     fdc_request_next_sector_id(d86f_fdc);
-                    break;
                 } else {
                     dev->format_state = FMT_POSTTRK_GAP4;
                     dev->sector_count = 0;
-                    break;
                 }
                 break;
+
+            default:
+                break;
         }
     }
 }
@@ -2922,8 +2932,8 @@ d86f_read_track(int drive, int track, int thin_track, int side, uint16_t *da, ui
     } else {
         if (!thin_track) {
             switch ((dev->disk_flags >> 1) & 3) {
-                case 0:
                 default:
+                case 0:
                     dev->side_flags[side] = 0x0A;
                     break;
 
@@ -3174,7 +3184,7 @@ d86f_stop(int drive)
 }
 
 int
-d86f_common_command(int drive, int sector, int track, int side, int rate, int sector_size)
+d86f_common_command(int drive, int sector, int track, int side, UNUSED(int rate), int sector_size)
 {
     d86f_t *dev = d86f[drive];
 
@@ -3258,7 +3268,7 @@ d86f_comparesector(int drive, int sector, int track, int side, int rate, int sec
 }
 
 void
-d86f_readaddress(int drive, int side, int rate)
+d86f_readaddress(int drive, UNUSED(int side), UNUSED(int rate))
 {
     d86f_t *dev = d86f[drive];
 
@@ -3308,7 +3318,7 @@ d86f_add_track(int drive, int track, int side)
 }
 
 void
-d86f_common_format(int drive, int side, int rate, uint8_t fill, int proxy)
+d86f_common_format(int drive, int side, UNUSED(int rate), uint8_t fill, int proxy)
 {
     d86f_t  *dev = d86f[drive];
     uint16_t temp;
@@ -3753,8 +3763,8 @@ d86f_load(int drive, char *fn)
         }
     } else {
         switch ((dev->disk_flags >> 1) >> 3) {
-            case 0:
             default:
+            case 0:
                 dev->side_flags[1] = 0x0a;
                 break;
 
diff --git a/src/game/gameport.c b/src/game/gameport.c
index 32dbc332e..413a1ff5a 100644
--- a/src/game/gameport.c
+++ b/src/game/gameport.c
@@ -37,7 +37,7 @@
 #include <86box/joystick_tm_fcs.h>
 #include <86box/plat_unused.h>
 
-typedef struct {
+typedef struct g_axis_t {
     pc_timer_t                  timer;
     int                         axis_nr;
     struct _joystick_instance_ *joystick;
@@ -116,9 +116,10 @@ static uint8_t gameport_pnp_rom[] = {
 };
 static const isapnp_device_config_t gameport_pnp_defaults[] = {
     {.activate = 1,
-     .io       = {
-          { .base = 0x200 },
-      }}
+        .io       = {
+            { .base = 0x200 },
+        }
+    }
 };
 
 const device_t *standalone_gameport_type;
diff --git a/src/game/joystick_sw_pad.c b/src/game/joystick_sw_pad.c
index 27ae94290..5c91ee1e9 100644
--- a/src/game/joystick_sw_pad.c
+++ b/src/game/joystick_sw_pad.c
@@ -67,8 +67,7 @@
 #include <86box/joystick_sw_pad.h>
 #include <86box/plat_unused.h>
 
-typedef struct
-{
+typedef struct sw_data {
     pc_timer_t poll_timer;
     int        poll_left;
     int        poll_clock;
diff --git a/src/include/86box/acpi.h b/src/include/86box/acpi.h
index 4d8359569..11be67278 100644
--- a/src/include/86box/acpi.h
+++ b/src/include/86box/acpi.h
@@ -59,8 +59,7 @@ extern "C" {
 #define VEN_VIA         0x01106
 #define VEN_VIA_596B    0x11106
 
-typedef struct
-{
+typedef struct acpi_regs_t {
     uint8_t  acpitst;
     uint8_t  auxen;
     uint8_t  auxsts;
@@ -111,8 +110,7 @@ typedef struct
     uint32_t pad0;
 } acpi_regs_t;
 
-typedef struct
-{
+typedef struct acpi_t {
     acpi_regs_t regs;
     uint8_t     gpireg2_default;
     uint8_t     pad[3];
diff --git a/src/include/86box/apm.h b/src/include/86box/apm.h
index d9e1a165d..2d5f0f117 100644
--- a/src/include/86box/apm.h
+++ b/src/include/86box/apm.h
@@ -21,8 +21,7 @@
 extern "C" {
 #endif
 
-typedef struct
-{
+typedef struct apm_t {
     uint8_t cmd;
     uint8_t stat;
     uint8_t do_smi;
diff --git a/src/include/86box/cdrom.h b/src/include/86box/cdrom.h
index 82098e54e..69b4e90f3 100644
--- a/src/include/86box/cdrom.h
+++ b/src/include/86box/cdrom.h
@@ -114,7 +114,7 @@ static const struct {
 /* To shut up the GCC compilers. */
 struct cdrom;
 
-typedef struct {
+typedef struct subchannel_t {
     uint8_t attr;
     uint8_t track;
     uint8_t index;
@@ -126,7 +126,7 @@ typedef struct {
     uint8_t rel_f;
 } subchannel_t;
 
-typedef struct {
+typedef struct track_info_t {
     int     number;
     uint8_t attr;
     uint8_t m;
@@ -135,7 +135,7 @@ typedef struct {
 } track_info_t;
 
 /* Define the various CD-ROM drive operations (ops). */
-typedef struct {
+typedef struct cdrom_ops_t {
     void (*get_tracks)(struct cdrom *dev, int *first, int *last);
     void (*get_track_info)(struct cdrom *dev, uint32_t track, int end, track_info_t *ti);
     void (*get_subchannel)(struct cdrom *dev, uint32_t lba, subchannel_t *subc);
diff --git a/src/include/86box/config.h b/src/include/86box/config.h
index 388cf9605..72b0a709f 100644
--- a/src/include/86box/config.h
+++ b/src/include/86box/config.h
@@ -27,7 +27,7 @@ extern "C" {
 #endif
 
 #if 0
-typedef struct {
+typedef struct storage_cfg_t {
     uint8_t     id;
     uint8_t bus_type;    /* Bus type: IDE, SCSI, etc. */
     uint8_t bus, : 4;    /* ID of the bus (for example, for IDE,
@@ -47,7 +47,7 @@ typedef struct {
     uint32_t tracks;
 } storage_cfg_t;
 
-typedef struct {
+typedef struct config_t {
     /* General configuration */
     int vid_resize;           /* Window is resizable or not */
     int vid_renderer;         /* Renderer */
diff --git a/src/include/86box/ddma.h b/src/include/86box/ddma.h
index 48cc9a6ed..01c2e980d 100644
--- a/src/include/86box/ddma.h
+++ b/src/include/86box/ddma.h
@@ -21,15 +21,13 @@
 extern "C" {
 #endif
 
-typedef struct
-{
+typedef struct ddma_channel_t {
     uint16_t io_base;
     int      channel;
     int      enable;
 } ddma_channel_t;
 
-typedef struct
-{
+typedef struct ddma_t {
     ddma_channel_t channels[8];
 } ddma_t;
 
diff --git a/src/include/86box/device.h b/src/include/86box/device.h
index 295126c22..74dd5ee3d 100644
--- a/src/include/86box/device.h
+++ b/src/include/86box/device.h
@@ -86,12 +86,12 @@ enum {
 #define BIOS_INTERLEAVED_INVERT          8
 #define BIOS_HIGH_BIT_INVERT             16
 
-typedef struct {
+typedef struct device_config_selection_t {
     const char *description;
     int         value;
 } device_config_selection_t;
 
-typedef struct {
+typedef struct device_config_bios_t {
     const char  *name;
     const char  *internal_name;
     int          bios_type;
@@ -103,13 +103,13 @@ typedef struct {
     const char  *files[9];
 } device_config_bios_t;
 
-typedef struct {
+typedef struct device_config_spinner_t {
     int16_t min;
     int16_t max;
     int16_t step;
 } device_config_spinner_t;
 
-typedef struct {
+typedef struct device_config_t {
     const char                     *name;
     const char                     *description;
     int                             type;
@@ -144,7 +144,7 @@ typedef struct _device_ {
     const device_config_t *config;
 } device_t;
 
-typedef struct {
+typedef struct device_context_t {
     const device_t *dev;
     char            name[2048];
     int             instance;
diff --git a/src/include/86box/disksizes.h b/src/include/86box/disksizes.h
index a32990bd9..b5cd1bc30 100644
--- a/src/include/86box/disksizes.h
+++ b/src/include/86box/disksizes.h
@@ -22,7 +22,7 @@
 extern "C" {
 #endif
 
-typedef struct {
+typedef struct disk_size_t {
     int hole;
     int sides;
     int data_rate;
diff --git a/src/include/86box/dma.h b/src/include/86box/dma.h
index 7e46cdd56..7ead53ba0 100644
--- a/src/include/86box/dma.h
+++ b/src/include/86box/dma.h
@@ -43,7 +43,7 @@
 #define DMA_OVER   0x10000
 #define DMA_VERIFY 0x20000
 
-typedef struct {
+typedef struct dma_t {
     uint8_t  m;
     uint8_t  mode;
     uint8_t  page;
diff --git a/src/include/86box/fdc.h b/src/include/86box/fdc.h
index b0e530754..9441bdfc4 100644
--- a/src/include/86box/fdc.h
+++ b/src/include/86box/fdc.h
@@ -56,7 +56,7 @@ extern int fdc_type;
 #define FDC_FLAG_TER            0x2000 /* Is Tertiary */
 #define FDC_FLAG_QUA            0x3000 /* Is Quaternary */
 
-typedef struct {
+typedef struct fdc_t {
     uint8_t dor;
     uint8_t stat;
     uint8_t command;
diff --git a/src/include/86box/fdd.h b/src/include/86box/fdd.h
index 7e9edd624..e21cbb73e 100644
--- a/src/include/86box/fdd.h
+++ b/src/include/86box/fdd.h
@@ -65,7 +65,7 @@ extern int   fdd_get_from_internal_name(char *s);
 
 extern int fdd_current_track(int drive);
 
-typedef struct {
+typedef struct DRIVE {
     int id;
 
     void (*seek)(int drive, int track);
diff --git a/src/include/86box/fifo8.h b/src/include/86box/fifo8.h
index 811f0522f..9f88ec408 100644
--- a/src/include/86box/fifo8.h
+++ b/src/include/86box/fifo8.h
@@ -1,7 +1,7 @@
 #ifndef EMU_FIFO8_H
 #define EMU_FIFO8_H
 
-typedef struct {
+typedef struct Fifo8 {
     /* All fields are private */
     uint8_t *data;
     uint32_t capacity;
diff --git a/src/include/86box/gameport.h b/src/include/86box/gameport.h
index 7556fa0d5..3b779e4e3 100644
--- a/src/include/86box/gameport.h
+++ b/src/include/86box/gameport.h
@@ -34,7 +34,7 @@
 
 #define GAMEPORT_SIO        0x1000000
 
-typedef struct {
+typedef struct plat_joystick_t {
     char name[260];
 
     int a[8];
@@ -68,7 +68,7 @@ typedef struct {
     int nr_sliders;
 } plat_joystick_t;
 
-typedef struct {
+typedef struct joystick_t {
     int axis[8];
     int button[32];
     int pov[4];
@@ -79,7 +79,7 @@ typedef struct {
     int pov_mapping[4][2];
 } joystick_t;
 
-typedef struct {
+typedef struct joystick_if_t {
     const char *name;
     const char *internal_name;
 
diff --git a/src/include/86box/hdc_ide_sff8038i.h b/src/include/86box/hdc_ide_sff8038i.h
index 8f0a0896a..490a13a42 100644
--- a/src/include/86box/hdc_ide_sff8038i.h
+++ b/src/include/86box/hdc_ide_sff8038i.h
@@ -20,8 +20,7 @@
 #ifndef EMU_HDC_IDE_SFF8038I_H
 #define EMU_HDC_IDE_SFF8038I_H
 
-typedef struct
-{
+typedef struct sff8038i_t {
     uint8_t  command;
     uint8_t  status;
     uint8_t  ptr0;
diff --git a/src/include/86box/hdd.h b/src/include/86box/hdd.h
index 6f6b085f2..9874fb4aa 100644
--- a/src/include/86box/hdd.h
+++ b/src/include/86box/hdd.h
@@ -87,7 +87,7 @@ enum {
 #define HDD_MAX_ZONES     16
 #define HDD_MAX_CACHE_SEG 16
 
-typedef struct {
+typedef struct hdd_preset_t {
     const char *name;
     const char *internal_name;
     uint32_t    zones;
@@ -101,7 +101,7 @@ typedef struct {
     double      track_seek_ms;
 } hdd_preset_t;
 
-typedef struct {
+typedef struct hdd_cache_seg_t {
     uint32_t id;
     uint32_t lba_addr;
     uint32_t ra_addr;
@@ -110,7 +110,7 @@ typedef struct {
     uint8_t  valid;
 } hdd_cache_seg_t;
 
-typedef struct {
+typedef struct hdd_cache_t {
     // Read cache
     hdd_cache_seg_t segments[HDD_MAX_CACHE_SEG];
     uint32_t        num_segments;
@@ -126,7 +126,7 @@ typedef struct {
     uint64_t write_start_time;
 } hdd_cache_t;
 
-typedef struct {
+typedef struct hdd_zone_t {
     uint32_t cylinders;
     uint32_t sectors_per_track;
     double   sector_time_usec;
@@ -136,7 +136,7 @@ typedef struct {
 } hdd_zone_t;
 
 /* Define the virtual Hard Disk. */
-typedef struct {
+typedef struct hard_disk_t {
     uint8_t id;
     union {
         uint8_t channel; /* Needed for Settings to reduce the number of if's */
diff --git a/src/include/86box/isapnp.h b/src/include/86box/isapnp.h
index 7259b5710..71c1bb29a 100644
--- a/src/include/86box/isapnp.h
+++ b/src/include/86box/isapnp.h
@@ -31,25 +31,25 @@ enum {
     ISAPNP_CARD_NO_KEY       = 3  /* cheat code for Crystal CS423x */
 };
 
-typedef struct {
+typedef struct isapnp_device_config_t {
     uint8_t activate;
-    struct {
+    struct pnp_mem_t {
         uint32_t base : 24;
         uint32_t size : 24;
     } mem[4];
-    struct {
+    struct pnp_mem32_t {
         uint32_t base;
         uint32_t size;
     } mem32[4];
-    struct {
+    struct pnp_io_t {
         uint16_t base;
     } io[8];
-    struct {
+    struct pnp_irq_t {
         uint8_t irq : 4;
         uint8_t level : 1;
         uint8_t type : 1;
     } irq[2];
-    struct {
+    struct pnp_dma_t {
         uint8_t dma : 3;
     } dma[2];
 } isapnp_device_config_t;
diff --git a/src/include/86box/keyboard.h b/src/include/86box/keyboard.h
index 500bb2843..f09507601 100644
--- a/src/include/86box/keyboard.h
+++ b/src/include/86box/keyboard.h
@@ -39,7 +39,7 @@ enum {
 };
 
 /* Used by the AT / PS/2 keyboard controller, common device, keyboard, and mouse. */
-typedef struct {
+typedef struct kbc_at_port_t {
     uint8_t wantcmd;
     uint8_t dat;
 
@@ -51,7 +51,7 @@ typedef struct {
 } kbc_at_port_t;
 
 /* Used by the AT / PS/2 common device, keyboard, and mouse. */
-typedef struct {
+typedef struct atkbc_dev_t {
     const char *name; /* name of this device */
 
     uint8_t type;
@@ -88,7 +88,7 @@ typedef struct {
     kbc_at_port_t *port;
 } atkbc_dev_t;
 
-typedef struct {
+typedef struct scancode {
     const uint8_t mk[4];
     const uint8_t brk[4];
 } scancode;
@@ -228,11 +228,13 @@ extern const device_t keyboard_xt_olivetti_device;
 extern const device_t keyboard_xt_zenith_device;
 extern const device_t keyboard_xtclone_device;
 extern const device_t keyboard_at_device;
+extern const device_t keyboard_at_siemens_device;
 extern const device_t keyboard_at_ami_device;
 extern const device_t keyboard_at_tg_ami_device;
 extern const device_t keyboard_at_toshiba_device;
 extern const device_t keyboard_at_olivetti_device;
 extern const device_t keyboard_at_ncr_device;
+extern const device_t keyboard_at_compaq_device;
 extern const device_t keyboard_ps2_device;
 extern const device_t keyboard_ps2_ps1_device;
 extern const device_t keyboard_ps2_ps1_pci_device;
diff --git a/src/include/86box/lpt.h b/src/include/86box/lpt.h
index a22644b48..4c1c793ae 100644
--- a/src/include/86box/lpt.h
+++ b/src/include/86box/lpt.h
@@ -17,8 +17,7 @@
 #define LPT6_IRQ  5
 #endif
 
-typedef struct
-{
+typedef struct lpt_device_t {
     const char *name;
     const char *internal_name;
 
diff --git a/src/include/86box/machine.h b/src/include/86box/machine.h
index d695ac357..7e17c270f 100644
--- a/src/include/86box/machine.h
+++ b/src/include/86box/machine.h
@@ -30,14 +30,16 @@
 #define MACHINE_BUS_CARTRIDGE 0x00000004 /* sys has two cartridge bays */
 #define MACHINE_BUS_ISA16     0x00000008 /* sys has ISA16 bus - PC/AT architecture */
 #define MACHINE_BUS_CBUS      0x00000010 /* sys has C-BUS bus */
-#define MACHINE_BUS_PS2       0x00000020 /* system has PS/2 keyboard and mouse ports */
-#define MACHINE_BUS_EISA      0x00000040 /* sys has EISA bus */
-#define MACHINE_BUS_VLB       0x00000080 /* sys has VL bus */
-#define MACHINE_BUS_MCA       0x00000100 /* sys has MCA bus */
-#define MACHINE_BUS_PCI       0x00000200 /* sys has PCI bus */
-#define MACHINE_BUS_PCMCIA    0x00000400 /* sys has PCMCIA bus */
-#define MACHINE_BUS_AGP       0x00000800 /* sys has AGP bus */
-#define MACHINE_BUS_AC97      0x00001000 /* sys has AC97 bus (ACR/AMR/CNR slot) */
+#define MACHINE_BUS_PS2_LATCH 0x00000020 /* system has PS/2 keyboard controller IRQ latch */
+#define MACHINE_BUS_PS2_PORTS 0x00000040 /* system has PS/2 keyboard and mouse ports */
+#define MACHINE_BUS_PS2       (MACHINE_BUS_PS2_LATCH | MACHINE_BUS_PS2_PORTS)
+#define MACHINE_BUS_EISA      0x00000080 /* sys has EISA bus */
+#define MACHINE_BUS_VLB       0x00000100 /* sys has VL bus */
+#define MACHINE_BUS_MCA       0x00000200 /* sys has MCA bus */
+#define MACHINE_BUS_PCI       0x00000400 /* sys has PCI bus */
+#define MACHINE_BUS_PCMCIA    0x00000800 /* sys has PCMCIA bus */
+#define MACHINE_BUS_AGP       0x00001000 /* sys has AGP bus */
+#define MACHINE_BUS_AC97      0x00002000 /* sys has AC97 bus (ACR/AMR/CNR slot) */
 /* Aliases. */
 #define MACHINE_CASSETTE  (MACHINE_BUS_CASSETTE)  /* sys has cassette port */
 #define MACHINE_CARTRIDGE (MACHINE_BUS_CARTRIDGE) /* sys has two cartridge bays */
@@ -98,19 +100,20 @@
 #define MACHINE_AV            (MACHINE_VIDEO | MACHINE_SOUND)    /* sys has video and sound */
 #define MACHINE_AG            (MACHINE_SOUND | MACHINE_GAMEPORT) /* sys has sound and game port */
 /* Feature flags for internal storage controllers. */
-#define MACHINE_HDC      0x03FE0000 /* sys has int HDC */
-#define MACHINE_MFM      0x00020000 /* sys has int MFM/RLL */
-#define MACHINE_XTA      0x00040000 /* sys has int XTA */
-#define MACHINE_ESDI     0x00080000 /* sys has int ESDI */
-#define MACHINE_IDE_PRI  0x00100000 /* sys has int pri IDE/ATAPI */
-#define MACHINE_IDE_SEC  0x00200000 /* sys has int sec IDE/ATAPI */
-#define MACHINE_IDE_TER  0x00400000 /* sys has int ter IDE/ATAPI */
-#define MACHINE_IDE_QUA  0x00800000 /* sys has int qua IDE/ATAPI */
-#define MACHINE_SCSI_PRI 0x01000000 /* sys has int pri SCSI */
-#define MACHINE_SCSI_SEC 0x02000000 /* sys has int sec SCSI */
-#define MACHINE_USB_PRI  0x04000000 /* sys has int pri USB */
-#define MACHINE_USB_SEC  0x08000000 /* sys has int sec USB */
-#define MACHINE_COREBOOT 0x10000000 /* sys has coreboot BIOS */
+#define MACHINE_HDC            0x03FE0000 /* sys has int HDC */
+#define MACHINE_MFM            0x00020000 /* sys has int MFM/RLL */
+#define MACHINE_XTA            0x00040000 /* sys has int XTA */
+#define MACHINE_ESDI           0x00080000 /* sys has int ESDI */
+#define MACHINE_IDE_PRI        0x00100000 /* sys has int pri IDE/ATAPI */
+#define MACHINE_IDE_SEC        0x00200000 /* sys has int sec IDE/ATAPI */
+#define MACHINE_IDE_TER        0x00400000 /* sys has int ter IDE/ATAPI */
+#define MACHINE_IDE_QUA        0x00800000 /* sys has int qua IDE/ATAPI */
+#define MACHINE_SCSI_PRI       0x01000000 /* sys has int pri SCSI */
+#define MACHINE_SCSI_SEC       0x02000000 /* sys has int sec SCSI */
+#define MACHINE_USB_PRI        0x04000000 /* sys has int pri USB */
+#define MACHINE_USB_SEC        0x08000000 /* sys has int sec USB */
+#define MACHINE_COREBOOT       0x10000000 /* sys has coreboot BIOS */
+#define MACHINE_SOFTFLOAT_ONLY 0x20000000 /* sys requires softfloat FPU */
 /* Combined flags. */
 #define MACHINE_IDE        (MACHINE_IDE_PRI)                   /* sys has int single IDE/ATAPI - mark as pri IDE/ATAPI */
 #define MACHINE_IDE_DUAL   (MACHINE_IDE_PRI | MACHINE_IDE_SEC) /* sys has int dual IDE/ATAPI - mark as both pri and sec IDE/ATAPI */
@@ -541,9 +544,8 @@ extern int machine_at_cmdpc_init(const machine_t *);
 extern int machine_at_portableii_init(const machine_t *);
 extern int machine_at_portableiii_init(const machine_t *);
 extern int machine_at_portableiii386_init(const machine_t *);
-#if defined(DEV_BRANCH) && defined(USE_DESKPRO386)
 extern int machine_at_deskpro386_init(const machine_t *);
-#endif
+extern int machine_at_deskpro386_01_1988_init(const machine_t *);
 
 /* m_at_socket4.c */
 extern void machine_at_premiere_common_init(const machine_t *, int);
diff --git a/src/include/86box/machine_status.h b/src/include/86box/machine_status.h
index 6948dc556..e7c57881b 100644
--- a/src/include/86box/machine_status.h
+++ b/src/include/86box/machine_status.h
@@ -1,20 +1,20 @@
 #ifndef EMU_MACHINE_STATUS_H
 #define EMU_MACHINE_STATUS_H
 
-typedef struct {
+typedef struct dev_status_empty_active_t {
     atomic_bool_t empty;
     atomic_bool_t active;
 } dev_status_empty_active_t;
 
-typedef struct {
+typedef struct dev_status_active_t {
     atomic_bool_t active;
 } dev_status_active_t;
 
-typedef struct {
+typedef struct dev_status_empty_t {
     atomic_bool_t empty;
 } dev_status_empty_t;
 
-typedef struct {
+typedef struct machine_status_t {
     dev_status_empty_active_t fdd[FDD_NUM];
     dev_status_empty_active_t cdrom[CDROM_NUM];
     dev_status_empty_active_t zip[ZIP_NUM];
diff --git a/src/include/86box/mem.h b/src/include/86box/mem.h
index 18def5e76..57ab009e3 100644
--- a/src/include/86box/mem.h
+++ b/src/include/86box/mem.h
@@ -158,20 +158,21 @@
 #define mem_set_access_smram_bus(smm, base, size, is_smram) \
     mem_set_access((smm ? ACCESS_BUS_SMM : ACCESS_BUS), 1, base, size, is_smram)
 
-typedef struct {
+typedef struct state_t {
     uint16_t x : 5;
     uint16_t w : 5;
     uint16_t r : 5;
     uint16_t pad : 1;
 } state_t;
 
-typedef union {
+typedef union mem_state_t {
     uint16_t vals[4];
     state_t  states[4];
 } mem_state_t;
 
 typedef struct _mem_mapping_ {
-    struct _mem_mapping_ *prev, *next;
+    struct _mem_mapping_ *prev;
+    struct _mem_mapping_ *next;
 
     int enable;
 
diff --git a/src/include/86box/mo.h b/src/include/86box/mo.h
index 28312e2f5..a11c4d636 100644
--- a/src/include/86box/mo.h
+++ b/src/include/86box/mo.h
@@ -27,7 +27,7 @@
 
 #define MO_TIME  10.0
 
-typedef struct {
+typedef struct mo_type_t {
     uint32_t sectors;
     uint16_t bytes_per_sector;
 } mo_type_t;
@@ -48,8 +48,7 @@ static const mo_type_t mo_types[KNOWN_MO_TYPES] = {
     { 637041,  1024},
 };
 
-typedef struct
-{
+typedef struct mo_drive_type_t {
     const char vendor[9];
     const char model[16];
     const char revision[5];
@@ -89,7 +88,7 @@ enum {
     MO_BUS_USB      = 7
 };
 
-typedef struct {
+typedef struct mo_drive_t {
     uint8_t id;
 
     union {
@@ -121,7 +120,7 @@ typedef struct {
 
 } mo_drive_t;
 
-typedef struct {
+typedef struct mo_t {
     mode_sense_pages_t ms_pages_saved;
 
     mo_drive_t *drv;
diff --git a/src/include/86box/net_dp8390.h b/src/include/86box/net_dp8390.h
index bc3220c69..e9e1e6c71 100644
--- a/src/include/86box/net_dp8390.h
+++ b/src/include/86box/net_dp8390.h
@@ -36,7 +36,7 @@
 #define DP8390_FLAG_CHECK_CR   0x02
 #define DP8390_FLAG_CLEAR_IRQ  0x04
 
-typedef struct {
+typedef struct dp8390_t {
     /* Page 0 */
 
     /* Command Register - 00h read/write */
diff --git a/src/include/86box/net_event.h b/src/include/86box/net_event.h
index 61eaad166..48580a359 100644
--- a/src/include/86box/net_event.h
+++ b/src/include/86box/net_event.h
@@ -1,7 +1,7 @@
 #ifndef EMU_NET_EVENT_H
 #define EMU_NET_EVENT_H
 
-typedef struct {
+typedef struct net_evt_t {
 #ifdef _WIN32
     HANDLE handle;
 #else
diff --git a/src/include/86box/network.h b/src/include/86box/network.h
index 19ec5fa37..d7958ee9a 100644
--- a/src/include/86box/network.h
+++ b/src/include/86box/network.h
@@ -93,7 +93,7 @@ enum {
     NET_QUEUE_TX_HOST = 2
 };
 
-typedef struct {
+typedef struct netcard_conf_t {
     uint16_t device_num;
     int      net_type;
     char     host_dev_name[128];
@@ -111,7 +111,7 @@ typedef struct netpkt {
     int      len;
 } netpkt_t;
 
-typedef struct {
+typedef struct netqueue_t {
     netpkt_t packets[NET_QUEUE_LEN];
     int      head;
     int      tail;
diff --git a/src/include/86box/pci.h b/src/include/86box/pci.h
index 3fcb74f6f..60c0d8742 100644
--- a/src/include/86box/pci.h
+++ b/src/include/86box/pci.h
@@ -91,13 +91,30 @@ enum {
 
 typedef union {
     uint32_t addr;
-    uint8_t  addr_regs[4];
+    uint8_t addr_regs[4];
 } bar_t;
 
+
+#define PCI_IO_ON   0x01
+#define PCI_IO_DEV0 0x02
+
+
 extern int pci_burst_time;
 extern int agp_burst_time;
 extern int pci_nonburst_time;
 extern int agp_nonburst_time;
+extern int pci_take_over_io;
+
+extern uint32_t pci_base;
+extern uint32_t pci_size;
+
+
+extern void     pci_type2_write(uint16_t port, uint8_t val, void *priv);
+extern void     pci_type2_writew(uint16_t port, uint16_t val, void *priv);
+extern void     pci_type2_writel(uint16_t port, uint32_t val, void *priv);
+extern uint8_t  pci_type2_read(uint16_t port, void *priv);
+extern uint16_t pci_type2_readw(uint16_t port, void *priv);
+extern uint32_t pci_type2_readl(uint16_t port, void *priv);
 
 extern void pci_set_irq_routing(int pci_int, int irq);
 extern void pci_set_irq_level(int pci_int, int level);
diff --git a/src/include/86box/pit.h b/src/include/86box/pit.h
index 044f20dad..bb6c26969 100644
--- a/src/include/86box/pit.h
+++ b/src/include/86box/pit.h
@@ -19,7 +19,7 @@
 #ifndef EMU_PIT_H
 #define EMU_PIT_H
 
-typedef struct {
+typedef struct ctr_t {
     uint8_t m;
     uint8_t ctrl;
     uint8_t read_status;
@@ -44,13 +44,13 @@ typedef struct {
     int do_read_status;
 
     union {
-        int count;
+        int32_t count;
         struct {
-            int units     : 4;
-            int tens      : 4;
-            int hundreds  : 4;
-            int thousands : 4;
-            int myriads   : 4;
+            int32_t units     : 4;
+            int32_t tens      : 4;
+            int32_t hundreds  : 4;
+            int32_t thousands : 4;
+            int32_t myriads   : 4;
         };
     };
 
@@ -77,7 +77,7 @@ enum {
     PIT_8254_FAST = 3
 };
 
-typedef struct {
+typedef struct pit_intf_t {
     uint8_t (*read)(uint16_t addr, void *priv);
     void (*write)(uint16_t addr, uint8_t val, void *priv);
     /* Gets a counter's count. */
diff --git a/src/include/86box/pit_fast.h b/src/include/86box/pit_fast.h
index d1bca1265..a7caeabe7 100644
--- a/src/include/86box/pit_fast.h
+++ b/src/include/86box/pit_fast.h
@@ -19,7 +19,7 @@
 #ifndef EMU_PIT_FAST_H
 #define EMU_PIT_FAST_H
 
-typedef struct {
+typedef struct ctrf_t {
     uint8_t m;
     uint8_t ctrl;
     uint8_t read_status;
@@ -45,13 +45,13 @@ typedef struct {
     int rereadlatch;
 
     union {
-        int count;
+        int32_t count;
         struct {
-            int units     : 4;
-            int tens      : 4;
-            int hundreds  : 4;
-            int thousands : 4;
-            int myriads   : 4;
+            int32_t units     : 4;
+            int32_t tens      : 4;
+            int32_t hundreds  : 4;
+            int32_t thousands : 4;
+            int32_t myriads   : 4;
         };
     };
 
@@ -62,7 +62,7 @@ typedef struct {
     void (*out_func)(int new_out, int old_out);
 } ctrf_t;
 
-typedef struct {
+typedef struct pitf_t {
     int    flags;
     ctrf_t counters[3];
 
diff --git a/src/include/86box/plat_dir.h b/src/include/86box/plat_dir.h
index ba1ac180f..485314f46 100644
--- a/src/include/86box/plat_dir.h
+++ b/src/include/86box/plat_dir.h
@@ -39,7 +39,7 @@ struct dirent {
 };
 #    define d_namlen d_reclen
 
-typedef struct {
+typedef struct DIR_t {
     short flags;  /* internal flags		*/
     short offset; /* offset of entry into dir	*/
     long  handle; /* open handle to Win32 system	*/
diff --git a/src/include/86box/plat_dynld.h b/src/include/86box/plat_dynld.h
index 986e72424..44891d4ac 100644
--- a/src/include/86box/plat_dynld.h
+++ b/src/include/86box/plat_dynld.h
@@ -18,7 +18,7 @@
 #ifndef PLAT_DYNLD_H
 #define PLAT_DYNLD_H
 
-typedef struct {
+typedef struct dllimp_t {
     const char *name;
     void       *func;
 } dllimp_t;
diff --git a/src/include/86box/port_6x.h b/src/include/86box/port_6x.h
index 71633f601..e0c4b0508 100644
--- a/src/include/86box/port_6x.h
+++ b/src/include/86box/port_6x.h
@@ -20,8 +20,7 @@
 #define EMU_PORT_6X_H
 
 #ifdef _TIMER_H_
-typedef struct
-{
+typedef struct port_6x_t {
     uint8_t refresh;
     uint8_t flags;
 
diff --git a/src/include/86box/port_92.h b/src/include/86box/port_92.h
index dd33d0681..2dd4319be 100644
--- a/src/include/86box/port_92.h
+++ b/src/include/86box/port_92.h
@@ -20,8 +20,7 @@
 #define EMU_PORT_92_H
 
 #ifdef _TIMER_H_
-typedef struct
-{
+typedef struct port_92_t {
     uint8_t reg;
     uint8_t flags;
 
diff --git a/src/include/86box/rom.h b/src/include/86box/rom.h
index 91f813a95..5da516aef 100644
--- a/src/include/86box/rom.h
+++ b/src/include/86box/rom.h
@@ -32,7 +32,7 @@
 #define bios_load_interleavedr(a, b, c, d, e)     bios_load(a, b, c, d, e, FLAG_INT | FLAG_REP)
 #define bios_load_aux_interleaved(a, b, c, d, e)  bios_load(a, b, c, d, e, FLAG_INT | FLAG_AUX)
 
-typedef struct {
+typedef struct rom_t {
     uint8_t      *rom;
     int           sz;
     uint32_t      mask;
diff --git a/src/include/86box/row.h b/src/include/86box/row.h
new file mode 100644
index 000000000..9dab7d4c3
--- /dev/null
+++ b/src/include/86box/row.h
@@ -0,0 +1,46 @@
+/*
+ * 86Box	A hypervisor and IBM PC system emulator that specializes in
+ *		running old operating systems and software designed for IBM
+ *		PC systems and compatibles from 1981 through fairly recent
+ *		system designs based on the PCI bus.
+ *
+ *		This file is part of the 86Box distribution.
+ *
+ *		Definitions for the SMRAM interface.
+ *
+ *
+ *
+ * Authors:	Miran Grca, <mgrca8@gmail.com>
+ *
+ *		Copyright 2016-2020 Miran Grca.
+ */
+
+#ifndef EMU_ROW_H
+# define EMU_ROW_H
+
+typedef struct _row_
+{
+    struct _smram_  *prev;
+    struct _smram_  *next;
+
+    uint8_t         *buf;
+
+    mem_mapping_t   mapping;
+
+    uint32_t        host_base;
+    uint32_t        host_size;
+    uint32_t        ram_base;
+    uint32_t        ram_size;
+    uint32_t        ram_mask;
+    uint32_t        boundary;
+} row_t;
+
+
+extern void         row_disable(uint8_t row_id);
+extern void         row_set_boundary(uint8_t row_id, uint32_t boundary);
+
+
+extern device_t     row_device;
+
+
+#endif	/*EMU_ROW_H*/
diff --git a/src/include/86box/scsi_cdrom.h b/src/include/86box/scsi_cdrom.h
index ea7f5f334..64c24c10e 100644
--- a/src/include/86box/scsi_cdrom.h
+++ b/src/include/86box/scsi_cdrom.h
@@ -22,7 +22,7 @@
 #define CDROM_TIME 10.0
 
 #ifdef SCSI_DEVICE_H
-typedef struct {
+typedef struct scsi_cdrom_t {
     /* Common block. */
     mode_sense_pages_t ms_pages_saved;
 
diff --git a/src/include/86box/scsi_device.h b/src/include/86box/scsi_device.h
index 5af0919f9..7256970cd 100644
--- a/src/include/86box/scsi_device.h
+++ b/src/include/86box/scsi_device.h
@@ -338,7 +338,7 @@
 #define MODE_SELECT_PHASE_PAGE_HEADER 3
 #define MODE_SELECT_PHASE_PAGE        4
 
-typedef struct {
+typedef struct mode_sense_pages_t {
     uint8_t pages[0x40][0x40];
 } mode_sense_pages_t;
 
@@ -384,7 +384,7 @@ typedef struct scsi_common_s {
     double callback;
 } scsi_common_t;
 
-typedef struct {
+typedef struct scsi_device_t {
     int32_t buffer_length;
 
     uint8_t  status;
diff --git a/src/include/86box/scsi_disk.h b/src/include/86box/scsi_disk.h
index b8ce9892c..ebf21c073 100644
--- a/src/include/86box/scsi_disk.h
+++ b/src/include/86box/scsi_disk.h
@@ -16,7 +16,7 @@
 #ifndef SCSI_DISK_H
 #define SCSI_DISK_H
 
-typedef struct {
+typedef struct scsi_disk_t {
     mode_sense_pages_t ms_pages_saved;
 
     hard_disk_t *drv;
diff --git a/src/include/86box/scsi_x54x.h b/src/include/86box/scsi_x54x.h
index 77ec136be..213873c5a 100644
--- a/src/include/86box/scsi_x54x.h
+++ b/src/include/86box/scsi_x54x.h
@@ -203,13 +203,13 @@ typedef struct addr24_s {
 } addr24_t;
 
 /* Structure for the INQUIRE_SETUP_INFORMATION reply. */
-typedef struct {
+typedef struct ReplyInquireSetupInformationSynchronousValue_t {
     uint8_t uOffset         : 4;
     uint8_t uTransferPeriod : 3;
     uint8_t fSynchronous    : 1;
 } ReplyInquireSetupInformationSynchronousValue;
 
-typedef struct {
+typedef struct ReplyInquireSetupInformation_t {
     uint8_t fSynchronousInitiationEnabled     : 1;
     uint8_t fParityCheckingEnabled            : 1;
     uint8_t uReserved1                        : 6;
@@ -223,24 +223,24 @@ typedef struct {
     uint8_t                                      VendorSpecificData[28];
 } ReplyInquireSetupInformation;
 
-typedef struct {
+typedef struct MailboxInit_t {
     uint8_t Count;
     addr24_t Address;
 } MailboxInit_t;
 
-typedef struct {
+typedef struct Mailbox_t {
     uint8_t  CmdStatus;
     addr24_t CCBPointer;
 } Mailbox_t;
 
-typedef struct {
+typedef struct Mailbox32_t {
     uint32_t CCBPointer;
     union {
-        struct {
+        struct out_t {
             uint8_t Reserved[3];
             uint8_t ActionCode;
         } out;
-        struct {
+        struct in_t {
             uint8_t HostStatus;
             uint8_t TargetStatus;
             uint8_t Reserved;
@@ -255,7 +255,7 @@ typedef struct {
       Bytes 16 and 17   Reserved (must be 0)
       Bytes 18 through 18+n-1, where n=size of CDB  Command Descriptor Block */
 
-typedef struct {
+typedef struct CCB32_t {
     uint8_t Opcode;
     uint8_t Reserved1 : 3,
         ControlByte   : 2,
@@ -277,7 +277,7 @@ typedef struct {
     uint32_t SensePointer;
 } CCB32;
 
-typedef struct {
+typedef struct CCB_t {
     uint8_t  Opcode;
     uint8_t  Lun         : 3;
     uint8_t  ControlByte : 2;
@@ -294,7 +294,7 @@ typedef struct {
     uint8_t  Cdb[12];
 } CCB;
 
-typedef struct {
+typedef struct CCBC_t {
     uint8_t Opcode;
     uint8_t Pad1    : 3,
         ControlByte : 2,
@@ -309,7 +309,7 @@ typedef struct {
     uint8_t Cdb[12];
 } CCBC;
 
-typedef union {
+typedef union CCBU_t {
     CCB32 new;
     CCB  old;
     CCBC common;
@@ -320,26 +320,25 @@ typedef struct {
     uint8_t *RequestSenseBuffer;
     uint32_t CCBPointer;
     int      Is24bit;
-    uint8_t  TargetID,
-        LUN,
-        HostStatus,
-        TargetStatus,
-        MailboxCompletionCode;
+    uint8_t  TargetID;
+    uint8_t  LUN;
+    uint8_t  HostStatus;
+    uint8_t  TargetStatus;
+    uint8_t  MailboxCompletionCode;
 } Req_t;
 
-typedef struct
-{
+typedef struct BIOSCMD_t {
     uint8_t command;
     uint8_t lun  : 3,
         reserved : 2,
         id       : 3;
     union {
-        struct {
+        struct chs_t {
             uint16_t cyl;
             uint8_t  head;
             uint8_t  sec;
         } chs;
-        struct {
+        struct lba_t {
             uint8_t lba0; /* MSB */
             uint8_t lba1;
             uint8_t lba2;
@@ -350,12 +349,12 @@ typedef struct
     addr24_t  dma_address;
 } BIOSCMD;
 
-typedef struct {
+typedef struct SGE32_t {
     uint32_t Segment;
     uint32_t SegmentPointer;
 } SGE32;
 
-typedef struct {
+typedef struct SGE_t {
     addr24_t Segment;
     addr24_t SegmentPointer;
 } SGE;
@@ -368,92 +367,111 @@ typedef struct {
 #define X54X_MBX_24BIT         16
 #define X54X_ISAPNP            32
 
-typedef struct {
+typedef struct x54x_t {
     /* 32 bytes */
-    char vendor[16], /* name of device vendor */
-        name[16];    /* name of device */
+    char vendor[16]; /* name of device vendor */
+    char name[16];   /* name of device */
 
     /* 24 bytes */
-    int8_t type, /* type of device */
-        IrqEnabled, Irq,
-        DmaChannel,
-        HostID;
+    int8_t type; /* type of device */
+    int8_t IrqEnabled;
+    int8_t Irq;
+    int8_t DmaChannel;
+    int8_t HostID;
 
-    uint8_t callback_phase : 4,
-        callback_sub_phase : 4,
-        scsi_cmd_phase, bus,
-        sync,
-        parity, shram_mode,
-        Geometry, Control,
-        Command, CmdParam,
-        BusOnTime, BusOffTime,
-        ATBusSpeed, setup_info_len,
-        max_id, pci_slot,
-        temp_cdb[12];
+    uint8_t callback_phase : 4;
+    uint8_t callback_sub_phase : 4;
+    uint8_t scsi_cmd_phase;
+    uint8_t bus;
+    uint8_t sync;
+    uint8_t parity;
+    uint8_t shram_mode;
+    uint8_t Geometry;
+    uint8_t Control;
+    uint8_t Command;
+    uint8_t CmdParam;
+    uint8_t BusOnTime;
+    uint8_t BusOffTime;
+    uint8_t ATBusSpeed;
+    uint8_t setup_info_len;
+    uint8_t max_id;
+    uint8_t pci_slot;
+    uint8_t temp_cdb[12];
 
-    volatile uint8_t /* for multi-threading, keep */
-        Status,
-        Interrupt, /* these volatile */
-        MailboxIsBIOS, ToRaise,
-        flags;
+    /* for multi-threading, keep these volatile */
+    volatile uint8_t Status;
+    volatile uint8_t Interrupt; 
+    volatile uint8_t MailboxIsBIOS;
+    volatile uint8_t ToRaise;
+    volatile uint8_t flags;
 
     /* 65928 bytes */
-    uint8_t pos_regs[8], /* MCA */
-        CmdBuf[128],
-        DataBuf[65536],
-        shadow_ram[128],
-        dma_buffer[128],
-        cmd_33_buf[4096];
+    uint8_t pos_regs[8]; /* MCA */
+    uint8_t CmdBuf[128];
+    uint8_t DataBuf[65536];
+    uint8_t shadow_ram[128];
+    uint8_t dma_buffer[128];
+    uint8_t cmd_33_buf[4096];
 
     /* 16 bytes */
     char *fw_rev; /* The 4 bytes of the revision command information + 2 extra bytes for BusLogic */
 
-    uint8_t *rom1, /* main BIOS image */
-        *rom2,     /* SCSI-Select image */
-        *nvr;      /* EEPROM buffer */
+    uint8_t *rom1; /* main BIOS image */
+    uint8_t *rom2; /* SCSI-Select image */
+    uint8_t *nvr;  /* EEPROM buffer */
 
     /* 6 words = 12 bytes */
-    uint16_t DataReply, DataReplyLeft,
-        rom_ioaddr,    /* offset in BIOS of I/O addr */
-        rom_shram,     /* index to shared RAM */
-        rom_shramsz,   /* size of shared RAM */
-        rom_fwhigh,    /* offset in BIOS of ver ID */
-        pnp_len,       /* length of the PnP ROM */
-        pnp_offset,    /* offset in the microcode ROM of the PnP ROM */
-        cmd_33_len,    /* length of the SCSISelect code decompressor program */
-        cmd_33_offset; /* offset in the microcode ROM of the SCSISelect code decompressor program */
+    uint16_t DataReply;
+    uint16_t DataReplyLeft;
+    uint16_t rom_ioaddr;    /* offset in BIOS of I/O addr */
+    uint16_t rom_shram;     /* index to shared RAM */
+    uint16_t rom_shramsz;   /* size of shared RAM */
+    uint16_t rom_fwhigh;    /* offset in BIOS of ver ID */
+    uint16_t pnp_len;       /* length of the PnP ROM */
+    uint16_t pnp_offset;    /* offset in the microcode ROM of the PnP ROM */
+    uint16_t cmd_33_len;    /* length of the SCSISelect code decompressor program */
+    uint16_t cmd_33_offset; /* offset in the microcode ROM of the SCSISelect code decompressor program */
 
     /* 16 + 20 + 52 = 88 bytes */
-    volatile int
-        MailboxOutInterrupts,
-        PendingInterrupt, Lock,
-        target_data_len, pad0;
+    volatile int MailboxOutInterrupts;
+    volatile int PendingInterrupt;
+    volatile int Lock;
+    volatile int target_data_len;
+    volatile int pad0;
 
-    uint32_t Base, fdc_address, rom_addr, /* address of BIOS ROM */
-        CmdParamLeft, Outgoing,
-        transfer_size;
+    uint32_t Base;
+    uint32_t fdc_address;
+    uint32_t rom_addr; /* address of BIOS ROM */
+    uint32_t CmdParamLeft;
+    uint32_t Outgoing;
+    uint32_t transfer_size;
 
-    volatile uint32_t
-        MailboxInit,
-        MailboxCount,
-        MailboxOutAddr, MailboxOutPosCur,
-        MailboxInAddr, MailboxInPosCur,
-        MailboxReq,
-        BIOSMailboxInit, BIOSMailboxCount,
-        BIOSMailboxOutAddr, BIOSMailboxOutPosCur,
-        BIOSMailboxReq,
-        Residue, card_bus; /* Basically a copy of device flags */
+    volatile uint32_t MailboxInit;
+    volatile uint32_t MailboxCount;
+    volatile uint32_t MailboxOutAddr;
+    volatile uint32_t MailboxOutPosCur;
+    volatile uint32_t MailboxInAddr;
+    volatile uint32_t MailboxInPosCur;
+    volatile uint32_t MailboxReq;
+    volatile uint32_t BIOSMailboxInit;
+    volatile uint32_t BIOSMailboxCount;
+    volatile uint32_t BIOSMailboxOutAddr;
+    volatile uint32_t BIOSMailboxOutPosCur;
+    volatile uint32_t BIOSMailboxReq;
+    volatile uint32_t Residue;
+    volatile uint32_t card_bus; /* Basically a copy of device flags */
 
     /* 8 bytes */
     uint64_t temp_period;
 
     /* 16 bytes */
-    double media_period, ha_bps; /* bytes per second */
+    double media_period;
+    double ha_bps; /* bytes per second */
 
     /* 8 bytes */
-    char *bios_path, /* path to BIOS image file */
-        *mcode_path, /* path to microcode image file, needed by the AHA-1542CP */
-        *nvr_path;   /* path to NVR image file */
+    char *bios_path;  /* path to BIOS image file */
+    char *mcode_path; /* path to microcode image file, needed by the AHA-1542CP */
+    char *nvr_path;   /* path to NVR image file */
 
     /* 56 bytes */
     /* Pointer to a structure of vendor-specific data that only the vendor-specific code can understand */
@@ -486,12 +504,13 @@ typedef struct {
     /* Pointer to a function that resets vendor-specific data */
     void (*ven_reset)(void *p);
 
-    rom_t bios,   /* BIOS memory descriptor */
-        uppersck; /* BIOS memory descriptor */
+    rom_t bios;     /* BIOS memory descriptor */
+    rom_t uppersck; /* BIOS memory descriptor */
 
     mem_mapping_t mmio_mapping;
 
-    pc_timer_t timer, ResetCB;
+    pc_timer_t timer;
+    pc_timer_t ResetCB;
 
     Req_t Req;
 
diff --git a/src/include/86box/serial.h b/src/include/86box/serial.h
index 5402cd77b..b27e54c5e 100644
--- a/src/include/86box/serial.h
+++ b/src/include/86box/serial.h
@@ -107,7 +107,7 @@ typedef struct serial_device_s {
     serial_t *serial;
 } serial_device_t;
 
-typedef struct {
+typedef struct serial_port_s {
     uint8_t enabled;
 } serial_port_t;
 
diff --git a/src/include/86box/smbus.h b/src/include/86box/smbus.h
index 0f7c3fa92..340d1d00e 100644
--- a/src/include/86box/smbus.h
+++ b/src/include/86box/smbus.h
@@ -29,7 +29,7 @@ enum {
     SMBUS_VIA   = 1
 };
 
-typedef struct {
+typedef struct smbus_piix4_t {
     uint32_t   local;
     uint16_t   io_base;
     int        clock;
@@ -47,7 +47,7 @@ typedef struct {
     void      *i2c;
 } smbus_piix4_t;
 
-typedef struct {
+typedef struct smbus_ali7101_t {
     uint32_t   local;
     uint16_t   io_base;
     uint8_t    stat;
diff --git a/src/include/86box/snd_ac97.h b/src/include/86box/snd_ac97.h
index fcd6f251f..ee75d56c4 100644
--- a/src/include/86box/snd_ac97.h
+++ b/src/include/86box/snd_ac97.h
@@ -103,21 +103,21 @@ enum {
     AC97_CODEC_AK4540   = 7
 };
 
-typedef struct {
+typedef struct ac97_vendor_reg_t {
     const uint16_t index;
     const uint16_t value;
     const uint16_t write_mask;
 } ac97_vendor_reg_t;
 
-typedef struct {
-    uint32_t vendor_id;
-    uint32_t min_rate;
-    uint32_t max_rate;
-    uint32_t misc_flags;
-    uint16_t reset_flags;
-    uint16_t extid_flags;
-    uint16_t powerdown_mask;
-    uint16_t regs[64];
+typedef struct ac97_codec_t {
+    uint32_t                 vendor_id;
+    uint32_t                 min_rate;
+    uint32_t                 max_rate;
+    uint32_t                 misc_flags;
+    uint16_t                 reset_flags;
+    uint16_t                 extid_flags;
+    uint16_t                 powerdown_mask;
+    uint16_t                 regs[64];
     uint8_t                  codec_id;
     uint8_t                  vendor_reg_page_max;
     const ac97_vendor_reg_t *vendor_regs;
diff --git a/src/include/86box/snd_ad1848.h b/src/include/86box/snd_ad1848.h
index 049999da0..be7831fb4 100644
--- a/src/include/86box/snd_ad1848.h
+++ b/src/include/86box/snd_ad1848.h
@@ -30,7 +30,7 @@ enum {
     AD1848_TYPE_CS4236  = 4
 };
 
-typedef struct {
+typedef struct ad1848_t {
     uint8_t type;
     uint8_t index;
     uint8_t xindex;
diff --git a/src/include/86box/snd_opl.h b/src/include/86box/snd_opl.h
index ff4e1f4a8..3a3f93ad5 100644
--- a/src/include/86box/snd_opl.h
+++ b/src/include/86box/snd_opl.h
@@ -31,7 +31,7 @@ enum fm_driver {
     FM_DRV_MAX   = 2
 };
 
-typedef struct {
+typedef struct fm_drv_t {
     uint8_t (*read)(uint16_t port, void *priv);
     void (*write)(uint16_t port, uint8_t val, void *priv);
     int32_t *(*update)(void *priv);
diff --git a/src/include/86box/usb.h b/src/include/86box/usb.h
index 8fc844363..7c2e13ee9 100644
--- a/src/include/86box/usb.h
+++ b/src/include/86box/usb.h
@@ -56,16 +56,14 @@ typedef struct usb_params_t
     void* parent_priv;
 } usb_params_t;
 
-typedef union
-{
+typedef union {
     uint32_t l;
     uint16_t w[2];
     uint8_t  b[4];
 } ohci_mmio_t;
 
 /* USB Host Controller device struct */
-typedef struct usb_t
-{
+typedef struct usb_t {
     uint8_t       uhci_io[32];
     ohci_mmio_t   ohci_mmio[1024];
     uint16_t      uhci_io_base;
@@ -88,14 +86,12 @@ typedef struct usb_t
 #pragma pack(push, 1)
 
 /* Base USB descriptor struct. */
-typedef struct usb_desc_base_t
-{
+typedef struct usb_desc_base_t {
     uint8_t bLength;
     uint8_t bDescriptorType;
 } usb_desc_base_t;
 
-enum usb_desc_setup_req_types
-{
+enum usb_desc_setup_req_types {
     USB_SETUP_TYPE_DEVICE    = 0x0,
     USB_SETUP_TYPE_INTERFACE = 0x1,
     USB_SETUP_TYPE_ENDPOING  = 0x2,
@@ -106,8 +102,7 @@ enum usb_desc_setup_req_types
 
 #define USB_SETUP_DEV_TO_HOST 0x80
 
-typedef struct usb_desc_setup_t
-{
+typedef struct usb_desc_setup_t {
     uint8_t  bmRequestType;
     uint8_t  bRequest;
     uint16_t wValue;
@@ -115,8 +110,7 @@ typedef struct usb_desc_setup_t
     uint16_t wLength;
 } usb_desc_setup_t;
 
-typedef struct usb_desc_endpoint_t
-{
+typedef struct usb_desc_endpoint_t {
     usb_desc_base_t base;
     uint8_t         bEndpointAddress;
     uint8_t         bmAttributes;
@@ -124,8 +118,7 @@ typedef struct usb_desc_endpoint_t
     uint8_t         bInterval;
 } usb_desc_endpoint_t;
 
-typedef struct usb_desc_hid_t
-{
+typedef struct usb_desc_hid_t {
     usb_desc_base_t base;
 
     uint16_t bcdHID;
@@ -135,8 +128,7 @@ typedef struct usb_desc_hid_t
     uint16_t wDescriptorLength;
 } usb_desc_hid_t;
 
-typedef struct usb_desc_interface_t
-{
+typedef struct usb_desc_interface_t {
     usb_desc_base_t base;
 
     uint8_t bInterfaceNumber;
@@ -148,14 +140,12 @@ typedef struct usb_desc_interface_t
     uint8_t iInterface;
 } usb_desc_interface_t;
 
-typedef struct usb_desc_string_t
-{
+typedef struct usb_desc_string_t {
     usb_desc_base_t base;
     uint16_t        bString[];
 } usb_desc_string_t;
 
-typedef struct usb_desc_conf_t
-{
+typedef struct usb_desc_conf_t {
     usb_desc_base_t base;
 
     uint16_t wTotalLength;
@@ -166,8 +156,7 @@ typedef struct usb_desc_conf_t
     uint8_t  bMaxPower;
 } usb_desc_conf_t;
 
-typedef struct usb_desc_device_t
-{
+typedef struct usb_desc_device_t {
     usb_desc_base_t base;
 
     uint16_t bcdUSB;
@@ -187,8 +176,7 @@ typedef struct usb_desc_device_t
 #pragma pack(pop)
 
 /* USB endpoint device struct. Incomplete and unused. */
-typedef struct usb_device_t
-{
+typedef struct usb_device_t {
     usb_desc_device_t device_desc;
     struct {
         usb_desc_conf_t  conf_desc;
diff --git a/src/include/86box/vid_8514a.h b/src/include/86box/vid_8514a.h
index e6510e3f1..b21da9fba 100644
--- a/src/include/86box/vid_8514a.h
+++ b/src/include/86box/vid_8514a.h
@@ -24,6 +24,8 @@ typedef struct ibm8514_t {
 
     int force_old_addr;
     int type;
+    int local;
+    int bpp;
 
     uint32_t vram_size;
     uint32_t vram_mask;
@@ -32,6 +34,7 @@ typedef struct ibm8514_t {
     uint8_t   dac_mask, dac_status;
     uint32_t *map8;
     int       dac_addr, dac_pos, dac_r, dac_g;
+    int       internal_pitch;
 
     struct {
         uint16_t subsys_cntl;
@@ -58,7 +61,7 @@ typedef struct ibm8514_t {
         uint8_t  pix_trans[2];
         int      poly_draw;
         int      ssv_state;
-        int      x1, x2, y1, y2;
+        int16_t  x1, x2, x3, y1, y2;
         int      sys_cnt, sys_cnt2;
         int      temp_cnt;
         int16_t  cx, cy, oldcy;
@@ -80,6 +83,7 @@ typedef struct ibm8514_t {
 
         uint16_t scratch;
         int      fill_state, xdir, ydir;
+        uint32_t ge_offset;
     } accel;
 
     uint16_t test;
@@ -90,7 +94,7 @@ typedef struct ibm8514_t {
         dispon, hdisp_on, linecountff,
         vc, linepos, oddeven, cursoron, blink, scrollcache,
         firstline, lastline, firstline_draw, lastline_draw,
-        displine, fullchange, x_add, y_add;
+        displine, fullchange;
     uint32_t ma, maback;
 
     uint8_t *vram, *changedvram, linedbl;
@@ -103,11 +107,13 @@ typedef struct ibm8514_t {
     int     disp_cntl, interlace;
     uint8_t subsys_cntl, subsys_stat;
 
-    volatile int force_busy, force_busy2;
+    atomic_int force_busy, force_busy2;
 
     int      blitter_busy;
     uint64_t blitter_time;
     uint64_t status_time;
     int pitch;
+    int ext_pitch;
+    int ext_crt_pitch;
 } ibm8514_t;
 #endif /*VIDEO_8514A_H*/
diff --git a/src/include/86box/vid_ati_eeprom.h b/src/include/86box/vid_ati_eeprom.h
index 1fa083eaa..7de5170d8 100644
--- a/src/include/86box/vid_ati_eeprom.h
+++ b/src/include/86box/vid_ati_eeprom.h
@@ -43,6 +43,7 @@ typedef struct ati_eeprom_t {
 } ati_eeprom_t;
 
 void ati_eeprom_load(ati_eeprom_t *eeprom, char *fn, int type);
+void ati_eeprom_load_mach8(ati_eeprom_t *eeprom, char *fn);
 void ati_eeprom_write(ati_eeprom_t *eeprom, int ena, int clk, int dat);
 int  ati_eeprom_read(ati_eeprom_t *eeprom);
 
diff --git a/src/include/86box/vid_svga.h b/src/include/86box/vid_svga.h
index 93fb9f17c..f8dafdd20 100644
--- a/src/include/86box/vid_svga.h
+++ b/src/include/86box/vid_svga.h
@@ -31,8 +31,9 @@
 #    define FLAG_NOSKEW       16
 #    define FLAG_ADDR_BY16    32
 #    define FLAG_RAMDAC_SHIFT 64
-#    define FLAG_128K_MASK    128
-
+#    define FLAG_ATI      	  128
+#    define FLAG_S3_911_16BIT 256
+#    define FLAG_512K_MASK    512
 struct monitor_t;
 
 typedef struct {
diff --git a/src/include/86box/vid_xga.h b/src/include/86box/vid_xga.h
index 069718b9f..550ae7cd6 100644
--- a/src/include/86box/vid_xga.h
+++ b/src/include/86box/vid_xga.h
@@ -36,6 +36,7 @@ typedef struct xga_t {
     mem_mapping_t  linear_mapping;
     mem_mapping_t  video_mapping;
     rom_t          bios_rom;
+    rom_t		   vga_bios_rom;
     xga_hwcursor_t hwcursor, hwcursor_latch;
     PALETTE        extpal;
 
@@ -57,7 +58,7 @@ typedef struct xga_t {
     uint8_t  clk_sel_1, clk_sel_2;
     uint8_t  hwc_control;
     uint8_t  bus_arb;
-    uint8_t  select_pos_isa;
+    uint8_t  isa_pos_enable;
     uint8_t  hwcursor_oddeven;
     uint8_t  cfg_reg_instance;
     uint8_t  rowcount;
@@ -70,6 +71,8 @@ typedef struct xga_t {
     uint8_t  sprite_data[1024];
     uint8_t  scrollcache;
     uint8_t  direct_color;
+    uint8_t  dma_channel;
+    uint8_t  instance_isa, instance_num, ext_mem_addr;
     uint8_t *vram, *changedvram;
 
     int16_t hwc_pos_x;
diff --git a/src/include/86box/vid_xga_device.h b/src/include/86box/vid_xga_device.h
index a92c6d3c6..7aa274d30 100644
--- a/src/include/86box/vid_xga_device.h
+++ b/src/include/86box/vid_xga_device.h
@@ -17,6 +17,11 @@
 
 #ifndef VIDEO_XGA_DEVICE_H
 #define VIDEO_XGA_DEVICE_H
+extern int xga_has_vga;
+
+#ifdef EMU_DEVICE_H
 extern const device_t xga_device;
 extern const device_t xga_isa_device;
+extern const device_t inmos_isa_device;
+#endif
 #endif /*VIDEO_XGA_DEVICE_H*/
diff --git a/src/include/86box/video.h b/src/include/86box/video.h
index 4c94fabfd..00925b31a 100644
--- a/src/include/86box/video.h
+++ b/src/include/86box/video.h
@@ -60,50 +60,59 @@ enum {
 #define VIDEO_FLAG_TYPE_CGA     0
 #define VIDEO_FLAG_TYPE_MDA     1
 #define VIDEO_FLAG_TYPE_SPECIAL 2
-#define VIDEO_FLAG_TYPE_NONE    3
-#define VIDEO_FLAG_TYPE_MASK    3
+#define VIDEO_FLAG_TYPE_8514    3
+#define VIDEO_FLAG_TYPE_XGA     4
+#define VIDEO_FLAG_TYPE_NONE    5
+#define VIDEO_FLAG_TYPE_MASK    7
 
-typedef struct {
+typedef struct video_timings_t {
     int type;
-    int write_b, write_w, write_l;
-    int read_b, read_w, read_l;
+    int write_b;
+    int write_w;
+    int write_l;
+    int read_b;
+    int read_w;
+    int read_l;
 } video_timings_t;
 
-typedef struct {
-    int       w, h;
+typedef struct bitmap_t {
+    int       w;
+    int       h;
     uint32_t *dat;
     uint32_t *line[2112];
 } bitmap_t;
 
-typedef struct {
-    uint8_t r, g, b;
+typedef struct rgb_t {
+    uint8_t r;
+    uint8_t g;
+    uint8_t b;
 } rgb_t;
 
-typedef struct {
+typedef struct dbcs_font_t {
     uint8_t chr[32];
 } dbcs_font_t;
 
 struct blit_data_struct;
 
 typedef struct monitor_t {
-    char      name[512];
-    int       mon_xsize;
-    int       mon_ysize;
-    int       mon_scrnsz_x;
-    int       mon_scrnsz_y;
-    int       mon_efscrnsz_y;
-    int       mon_unscaled_size_x;
-    int       mon_unscaled_size_y;
-    int       mon_res_x;
-    int       mon_res_y;
-    int       mon_bpp;
-    bitmap_t *target_buffer;
-    int       mon_video_timing_read_b,
-        mon_video_timing_read_w,
-        mon_video_timing_read_l;
-    int mon_video_timing_write_b,
-        mon_video_timing_write_w,
-        mon_video_timing_write_l;
+    char                     name[512];
+    int                      mon_xsize;
+    int                      mon_ysize;
+    int                      mon_scrnsz_x;
+    int                      mon_scrnsz_y;
+    int                      mon_efscrnsz_y;
+    int                      mon_unscaled_size_x;
+    int                      mon_unscaled_size_y;
+    int                      mon_res_x;
+    int                      mon_res_y;
+    int                      mon_bpp;
+    bitmap_t                *target_buffer;
+    int                      mon_video_timing_read_b;
+    int                      mon_video_timing_read_w;
+    int                      mon_video_timing_read_l;
+    int                      mon_video_timing_write_b;
+    int                      mon_video_timing_write_w;
+    int                      mon_video_timing_write_l;
     int                      mon_overscan_x;
     int                      mon_overscan_y;
     int                      mon_force_resize;
@@ -137,10 +146,14 @@ extern int                video_fullscreen_scale_maximized;
 
 typedef rgb_t PALETTE[256];
 
-// extern int	changeframecount;
+#if 0
+extern int changeframecount;
+#endif
 
 extern volatile int screenshots;
-// extern bitmap_t	*buffer32;
+#if 0
+extern bitmap_t	*buffer32;
+#endif
 #define buffer32             (monitors[monitor_index_global].target_buffer)
 #define pal_lookup           (monitors[monitor_index_global].mon_pal_lookup)
 #define overscan_x           (monitors[monitor_index_global].mon_overscan_x)
@@ -163,12 +176,14 @@ extern volatile int screenshots;
 #define efscrnsz_y           (monitors[monitor_index_global].mon_efscrnsz_y)
 #define unscaled_size_x      (monitors[monitor_index_global].mon_unscaled_size_x)
 #define unscaled_size_y      (monitors[monitor_index_global].mon_unscaled_size_y)
-extern PALETTE cgapal;
-extern PALETTE cgapal_mono[6];
-// extern uint32_t	pal_lookup[256];
-extern int video_fullscreen;
-extern int video_fullscreen_scale;
-extern int video_fullscreen_first;
+extern PALETTE      cgapal;
+extern PALETTE      cgapal_mono[6];
+#if 0
+extern uint32_t	    pal_lookup[256];
+#endif
+extern int          video_fullscreen;
+extern int          video_fullscreen_scale;
+extern int          video_fullscreen_first;
 extern uint8_t      fontdat[2048][8];
 extern uint8_t      fontdatm[2048][16];
 extern uint8_t      fontdatw[512][32];
@@ -176,23 +191,24 @@ extern uint8_t      fontdat8x12[256][16];
 extern uint8_t      fontdat12x18[256][36];
 extern dbcs_font_t *fontdatksc5601;
 extern dbcs_font_t *fontdatksc5601_user;
-extern uint32_t    *video_6to8,
-    *video_8togs,
-    *video_8to32,
-    *video_15to32,
-    *video_16to32;
-extern int enable_overscan;
-extern int force_43;
-extern int vid_resize;
-extern int herc_blend;
-extern int vid_cga_contrast;
-extern int video_grayscale;
-extern int video_graytype;
+extern uint32_t    *video_6to8;
+extern uint32_t    *video_8togs;
+extern uint32_t    *video_8to32;
+extern uint32_t    *video_15to32;
+extern uint32_t    *video_16to32;
+extern int          enable_overscan;
+extern int          force_43;
+extern int          vid_resize;
+extern int          herc_blend;
+extern int          vid_cga_contrast;
+extern int          video_grayscale;
+extern int          video_graytype;
 
 extern double cpuclock;
 extern int    emu_fps;
 extern int    frames;
 extern int    readflash;
+extern int    ibm8514_has_vga;
 
 /* Function handler pointers. */
 extern void (*video_recalctimings)(void);
@@ -219,6 +235,8 @@ extern int   video_card_get_flags(int card);
 extern int   video_is_mda(void);
 extern int   video_is_cga(void);
 extern int   video_is_ega_vga(void);
+extern int   video_is_8514(void);
+extern int   video_is_xga(void);
 extern void  video_inform_monitor(int type, const video_timings_t *ptr, int monitor_index);
 extern int   video_get_type_monitor(int monitor_index);
 
@@ -277,8 +295,12 @@ extern uint32_t video_color_transform(uint32_t color);
 /* IBM XGA */
 extern void xga_device_add(void);
 
-/* IBM 8514/A and generic clones*/
+/* IBM 8514/A and clones*/
 extern void ibm8514_device_add(void);
+extern const device_t mach8_isa_device;
+extern const device_t mach32_isa_device;
+extern const device_t mach32_vlb_device;
+extern const device_t mach32_pci_device;
 
 /* ATi Mach64 */
 extern const device_t mach64gx_isa_device;
diff --git a/src/ini.c b/src/ini.c
index 0921b61ff..4ea2b6477 100644
--- a/src/ini.c
+++ b/src/ini.c
@@ -42,7 +42,7 @@ typedef struct _list_ {
     struct _list_ *next;
 } list_t;
 
-typedef struct {
+typedef struct section_t {
     list_t list;
 
     char name[128];
@@ -50,7 +50,7 @@ typedef struct {
     list_t entry_head;
 } section_t;
 
-typedef struct {
+typedef struct entry_t {
     list_t list;
 
     char    name[128];
diff --git a/src/io.c b/src/io.c
index 99b4954f8..173e6707b 100644
--- a/src/io.c
+++ b/src/io.c
@@ -29,6 +29,7 @@
 #include <86box/timer.h>
 #include "cpu.h"
 #include <86box/m_amstrad.h>
+#include <86box/pci.h>
 
 #define NPORTS 65536 /* PC/AT supports 64K ports */
 
@@ -47,10 +48,11 @@ typedef struct _io_ {
 } io_t;
 
 typedef struct {
-    uint8_t  enable;
-    uint16_t base, size;
-    void (*func)(int size, uint16_t addr, uint8_t write, uint8_t val, void *priv),
-        *priv;
+    uint8_t   enable;
+    uint16_t  base;
+    uint16_t  size;
+    void    (*func)(int size, uint16_t addr, uint8_t write, uint8_t val, void *priv);
+    void     *priv;
 } io_trap_t;
 
 int   initialized = 0;
@@ -286,15 +288,25 @@ inb(uint16_t port)
     int     found  = 0;
     int     qfound = 0;
 
-    p = io[port];
-    while (p) {
-        q = p->next;
-        if (p->inb) {
-            ret &= p->inb(port, p->priv);
-            found |= 1;
-            qfound++;
+    if ((pci_take_over_io & PCI_IO_ON) && (port >= pci_base) && (port < (pci_base + pci_size))) {
+        ret = pci_type2_read(port, NULL);
+        found = 1;
+        qfound = 1;
+    } else if ((pci_take_over_io & PCI_IO_DEV0) && (port >= 0xc000) && (port < 0xc100)) {
+        ret = pci_type2_read(port, NULL);
+        found = 1;
+        qfound = 1;
+    } else {
+        p = io[port];
+        while (p) {
+            q = p->next;
+            if (p->inb) {
+                ret &= p->inb(port, p->priv);
+                found |= 1;
+                qfound++;
+            }
+            p = q;
         }
-        p = q;
     }
 
     if (amstrad_latch & 0x80000000) {
@@ -310,8 +322,10 @@ inb(uint16_t port)
         cycles -= io_delay;
 
     /* TriGem 486-BIOS MHz output. */
-    /* if (port == 0x1ed)
-        ret = 0xfe; */
+#if 0
+    if (port == 0x1ed)
+        ret = 0xfe;
+#endif
 
     io_log("[%04X:%08X] (%i, %i, %04i) in b(%04X) = %02X\n", CS, cpu_state.pc, in_smm, found, qfound, port, ret);
 
@@ -326,15 +340,25 @@ outb(uint16_t port, uint8_t val)
     int   found  = 0;
     int   qfound = 0;
 
-    p = io[port];
-    while (p) {
-        q = p->next;
-        if (p->outb) {
-            p->outb(port, val, p->priv);
-            found |= 1;
-            qfound++;
+    if ((pci_take_over_io & PCI_IO_ON) && (port >= pci_base) && (port < (pci_base + pci_size))) {
+        pci_type2_write(port, val, NULL);
+        found = 1;
+        qfound = 1;
+    } else if ((pci_take_over_io & PCI_IO_DEV0) && (port >= 0xc000) && (port < 0xc100)) {
+        pci_type2_write(port, val, NULL);
+        found = 1;
+        qfound = 1;
+    } else {
+        p = io[port];
+        while (p) {
+            q = p->next;
+            if (p->outb) {
+                p->outb(port, val, p->priv);
+                found |= 1;
+                qfound++;
+            }
+            p = q;
         }
-        p = q;
     }
 
     if (!found) {
@@ -360,32 +384,42 @@ inw(uint16_t port)
     int      qfound = 0;
     uint8_t  ret8[2];
 
-    p = io[port];
-    while (p) {
-        q = p->next;
-        if (p->inw) {
-            ret &= p->inw(port, p->priv);
-            found |= 2;
-            qfound++;
-        }
-        p = q;
-    }
-
-    ret8[0] = ret & 0xff;
-    ret8[1] = (ret >> 8) & 0xff;
-    for (uint8_t i = 0; i < 2; i++) {
-        p = io[(port + i) & 0xffff];
+    if ((pci_take_over_io & PCI_IO_ON) && (port >= pci_base) && (port < (pci_base + pci_size))) {
+        ret = pci_type2_readw(port, NULL);
+        found = 2;
+        qfound = 1;
+    } else if ((pci_take_over_io & PCI_IO_DEV0) && (port >= 0xc000) && (port < 0xc100)) {
+        ret = pci_type2_readw(port, NULL);
+        found = 2;
+        qfound = 1;
+    } else {
+        p = io[port];
         while (p) {
             q = p->next;
-            if (p->inb && !p->inw) {
-                ret8[i] &= p->inb(port + i, p->priv);
-                found |= 1;
+            if (p->inw) {
+                ret &= p->inw(port, p->priv);
+                found |= 2;
                 qfound++;
             }
             p = q;
         }
+
+        ret8[0] = ret & 0xff;
+        ret8[1] = (ret >> 8) & 0xff;
+        for (uint8_t i = 0; i < 2; i++) {
+            p = io[(port + i) & 0xffff];
+            while (p) {
+                q = p->next;
+                if (p->inb && !p->inw) {
+                    ret8[i] &= p->inb(port + i, p->priv);
+                    found |= 1;
+                    qfound++;
+                }
+                p = q;
+            }
+        }
+        ret = (ret8[1] << 8) | ret8[0];
     }
-    ret = (ret8[1] << 8) | ret8[0];
 
     if (amstrad_latch & 0x80000000) {
         if (port & 0x80)
@@ -412,28 +446,38 @@ outw(uint16_t port, uint16_t val)
     int   found  = 0;
     int   qfound = 0;
 
-    p = io[port];
-    while (p) {
-        q = p->next;
-        if (p->outw) {
-            p->outw(port, val, p->priv);
-            found |= 2;
-            qfound++;
-        }
-        p = q;
-    }
-
-    for (uint8_t i = 0; i < 2; i++) {
-        p = io[(port + i) & 0xffff];
+    if ((pci_take_over_io & PCI_IO_ON) && (port >= pci_base) && (port < (pci_base + pci_size))) {
+        pci_type2_writew(port, val, NULL);
+        found = 2;
+        qfound = 1;
+    } else if ((pci_take_over_io & PCI_IO_DEV0) && (port >= 0xc000) && (port < 0xc100)) {
+        pci_type2_writew(port, val, NULL);
+        found = 2;
+        qfound = 1;
+    } else {
+        p = io[port];
         while (p) {
             q = p->next;
-            if (p->outb && !p->outw) {
-                p->outb(port + i, val >> (i << 3), p->priv);
-                found |= 1;
+            if (p->outw) {
+                p->outw(port, val, p->priv);
+                found |= 2;
                 qfound++;
             }
             p = q;
         }
+
+        for (uint8_t i = 0; i < 2; i++) {
+            p = io[(port + i) & 0xffff];
+            while (p) {
+                q = p->next;
+                if (p->outb && !p->outw) {
+                    p->outb(port + i, val >> (i << 3), p->priv);
+                    found |= 1;
+                    qfound++;
+                }
+                p = q;
+            }
+        }
     }
 
     if (!found) {
@@ -460,59 +504,69 @@ inl(uint16_t port)
     int      found  = 0;
     int      qfound = 0;
 
-    p = io[port];
-    while (p) {
-        q = p->next;
-        if (p->inl) {
-            ret &= p->inl(port, p->priv);
-            found |= 4;
-            qfound++;
-        }
-        p = q;
-    }
-
-    ret16[0] = ret & 0xffff;
-    ret16[1] = (ret >> 16) & 0xffff;
-    p        = io[port & 0xffff];
-    while (p) {
-        q = p->next;
-        if (p->inw && !p->inl) {
-            ret16[0] &= p->inw(port, p->priv);
-            found |= 2;
-            qfound++;
-        }
-        p = q;
-    }
-
-    p = io[(port + 2) & 0xffff];
-    while (p) {
-        q = p->next;
-        if (p->inw && !p->inl) {
-            ret16[1] &= p->inw(port + 2, p->priv);
-            found |= 2;
-            qfound++;
-        }
-        p = q;
-    }
-    ret = (ret16[1] << 16) | ret16[0];
-
-    ret8[0] = ret & 0xff;
-    ret8[1] = (ret >> 8) & 0xff;
-    ret8[2] = (ret >> 16) & 0xff;
-    ret8[3] = (ret >> 24) & 0xff;
-    for (uint8_t i = 0; i < 4; i++) {
-        p = io[(port + i) & 0xffff];
+    if ((pci_take_over_io & PCI_IO_ON) && (port >= pci_base) && (port < (pci_base + pci_size))) {
+        ret = pci_type2_readl(port, NULL);
+        found = 4;
+        qfound = 1;
+    } else if ((pci_take_over_io & PCI_IO_DEV0) && (port >= 0xc000) && (port < 0xc100)) {
+        ret = pci_type2_readl(port, NULL);
+        found = 4;
+        qfound = 1;
+    } else {
+        p = io[port];
         while (p) {
             q = p->next;
-            if (p->inb && !p->inw && !p->inl) {
-                ret8[i] &= p->inb(port + i, p->priv);
-                found |= 1;
+            if (p->inl) {
+                ret &= p->inl(port, p->priv);
+                found |= 4;
                 qfound++;
             }
             p = q;
         }
+
+        ret16[0] = ret & 0xffff;
+        ret16[1] = (ret >> 16) & 0xffff;
+        p        = io[port & 0xffff];
+        while (p) {
+            q = p->next;
+            if (p->inw && !p->inl) {
+                ret16[0] &= p->inw(port, p->priv);
+                found |= 2;
+                qfound++;
+            }
+            p = q;
+        }
+
+        p = io[(port + 2) & 0xffff];
+        while (p) {
+            q = p->next;
+            if (p->inw && !p->inl) {
+                ret16[1] &= p->inw(port + 2, p->priv);
+                found |= 2;
+                qfound++;
+            }
+            p = q;
+        }
+        ret = (ret16[1] << 16) | ret16[0];
+
+        ret8[0] = ret & 0xff;
+        ret8[1] = (ret >> 8) & 0xff;
+        ret8[2] = (ret >> 16) & 0xff;
+        ret8[3] = (ret >> 24) & 0xff;
+        for (uint8_t i = 0; i < 4; i++) {
+            p = io[(port + i) & 0xffff];
+            while (p) {
+                q = p->next;
+                if (p->inb && !p->inw && !p->inl) {
+                    ret8[i] &= p->inb(port + i, p->priv);
+                    found |= 1;
+                    qfound++;
+                }
+                p = q;
+            }
+        }
+        ret = (ret8[3] << 24) | (ret8[2] << 16) | (ret8[1] << 8) | ret8[0];
     }
-    ret = (ret8[3] << 24) | (ret8[2] << 16) | (ret8[1] << 8) | ret8[0];
 
     if (amstrad_latch & 0x80000000) {
         if (port & 0x80)
@@ -540,42 +594,52 @@ outl(uint16_t port, uint32_t val)
     int   qfound = 0;
     int   i      = 0;
 
-    p = io[port];
-    if (p) {
-        while (p) {
-            q = p->next;
-            if (p->outl) {
-                p->outl(port, val, p->priv);
-                found |= 4;
-                qfound++;
+    if ((pci_take_over_io & PCI_IO_ON) && (port >= pci_base) && (port < (pci_base + pci_size))) {
+        pci_type2_writel(port, val, NULL);
+        found = 4;
+        qfound = 1;
+    } else if ((pci_take_over_io & PCI_IO_DEV0) && (port >= 0xc000) && (port < 0xc100)) {
+        pci_type2_writel(port, val, NULL);
+        found = 4;
+        qfound = 1;
+    } else {
+        p = io[port];
+        if (p) {
+            while (p) {
+                q = p->next;
+                if (p->outl) {
+                    p->outl(port, val, p->priv);
+                    found |= 4;
+                    qfound++;
+                }
+                p = q;
             }
-            p = q;
         }
-    }
 
-    for (i = 0; i < 4; i += 2) {
-        p = io[(port + i) & 0xffff];
-        while (p) {
-            q = p->next;
-            if (p->outw && !p->outl) {
-                p->outw(port + i, val >> (i << 3), p->priv);
-                found |= 2;
-                qfound++;
+        for (i = 0; i < 4; i += 2) {
+            p = io[(port + i) & 0xffff];
+            while (p) {
+                q = p->next;
+                if (p->outw && !p->outl) {
+                    p->outw(port + i, val >> (i << 3), p->priv);
+                    found |= 2;
+                    qfound++;
+                }
+                p = q;
             }
-            p = q;
         }
-    }
 
-    for (i = 0; i < 4; i++) {
-        p = io[(port + i) & 0xffff];
-        while (p) {
-            q = p->next;
-            if (p->outb && !p->outw && !p->outl) {
-                p->outb(port + i, val >> (i << 3), p->priv);
-                found |= 1;
-                qfound++;
+        for (i = 0; i < 4; i++) {
+            p = io[(port + i) & 0xffff];
+            while (p) {
+                q = p->next;
+                if (p->outb && !p->outw && !p->outl) {
+                    p->outb(port + i, val >> (i << 3), p->priv);
+                    found |= 1;
+                    qfound++;
+                }
+                p = q;
             }
-            p = q;
         }
     }
 
diff --git a/src/ioapic.c b/src/ioapic.c
index 1fed64220..c3939f249 100644
--- a/src/ioapic.c
+++ b/src/ioapic.c
@@ -27,8 +27,9 @@
 #include <86box/machine.h>
 #include <86box/mem.h>
 #include <86box/chipset.h>
+#include <86box/plat_unused.h>
 
-typedef struct {
+typedef struct ioapic_t {
     uint8_t dummy;
 } ioapic_t;
 
@@ -51,7 +52,7 @@ ioapic_log(const char *fmt, ...)
 #endif
 
 static void
-ioapic_write(uint16_t port, uint8_t val, void *priv)
+ioapic_write(UNUSED(uint16_t port), uint8_t val, UNUSED(void *priv))
 {
     uint32_t pcmp;
 
@@ -83,8 +84,9 @@ ioapic_write(uint16_t port, uint8_t val, void *priv)
 }
 
 static void
-ioapic_reset(ioapic_t *dev)
+ioapic_reset(UNUSED(ioapic_t *dev))
 {
+    //
 }
 
 static void
@@ -99,7 +101,7 @@ ioapic_close(void *priv)
 }
 
 static void *
-ioapic_init(const device_t *info)
+ioapic_init(UNUSED(const device_t *info))
 {
     ioapic_t *dev = (ioapic_t *) malloc(sizeof(ioapic_t));
     memset(dev, 0, sizeof(ioapic_t));
diff --git a/src/log.c b/src/log.c
index 02bb6132e..b5267d70b 100644
--- a/src/log.c
+++ b/src/log.c
@@ -35,10 +35,11 @@
 #include <86box/log.h>
 
 #ifndef RELEASE_BUILD
-typedef struct
-{
-    char buff[1024], *dev_name;
-    int  seen, suppr_seen;
+typedef struct log_t {
+    char  buff[1024];
+    char *dev_name;
+    int   seen;
+    int   suppr_seen;
 } log_t;
 
 extern FILE *stdlog; /* file to log output to */
diff --git a/src/lpt.c b/src/lpt.c
index f4d2ad67c..f4fc7192f 100644
--- a/src/lpt.c
+++ b/src/lpt.c
@@ -124,6 +124,9 @@ lpt_write(uint16_t port, uint8_t val, void *priv)
             dev->ctrl       = val;
             dev->enable_irq = val & 0x10;
             break;
+
+        default:
+            break;
     }
 }
 
@@ -154,6 +157,9 @@ lpt_read(uint16_t port, void *priv)
             else
                 ret = 0xe0 | dev->ctrl | dev->enable_irq;
             break;
+
+        default:
+            break;
     }
 
     return ret;
diff --git a/src/machine/m_at.c b/src/machine/m_at.c
index 92f8c0f5f..9fc53ebcd 100644
--- a/src/machine/m_at.c
+++ b/src/machine/m_at.c
@@ -246,7 +246,14 @@ machine_at_siemens_init(const machine_t *model)
     if (bios_only || !ret)
         return ret;
 
-    machine_at_ibm_common_init(model);
+    machine_at_common_init_ex(model, 1);
+
+    device_add(&keyboard_at_siemens_device);
+
+    mem_remap_top(384);
+
+    if (fdc_type == FDC_INTERNAL)
+        device_add(&fdc_at_device);
 
     return ret;
 }
diff --git a/src/machine/m_at_compaq.c b/src/machine/m_at_compaq.c
index dd8370a4f..3676a9d92 100644
--- a/src/machine/m_at_compaq.c
+++ b/src/machine/m_at_compaq.c
@@ -31,6 +31,7 @@
 #include <86box/mem.h>
 #include <86box/rom.h>
 #include <86box/device.h>
+#include <86box/keyboard.h>
 #include <86box/fdd.h>
 #include <86box/fdc.h>
 #include <86box/fdc_ext.h>
@@ -41,22 +42,28 @@
 #include <86box/vid_cga.h>
 #include <86box/vid_cga_comp.h>
 
+
+static video_timings_t timing_compaq_plasma = { .type = VIDEO_ISA, .write_b = 8, .write_w = 16, .write_l = 32, .read_b = 8, .read_w = 16, .read_l = 32 };
+
 enum {
     COMPAQ_PORTABLEII = 0,
     COMPAQ_PORTABLEIII,
     COMPAQ_PORTABLEIII386,
-    COMPAQ_DESKPRO386
+    COMPAQ_DESKPRO386,
+    COMPAQ_DESKPRO386_01_1988
 };
 
 #define CGA_RGB       0
 #define CGA_COMPOSITE 1
 
-#define COMPOSITE_OLD 0
-#define COMPOSITE_NEW 1
-
 /*Very rough estimate*/
 #define VID_CLOCK (double) (651 * 416 * 60)
 
+static uint8_t cga_crtcmask[32] = {
+    0xff, 0xff, 0xff, 0xff, 0x7f, 0x1f, 0x7f, 0x7f, 0xf3, 0x1f, 0x7f, 0x1f, 0x3f, 0xff, 0x3f, 0xff,
+    0xff, 0xff, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00
+};
+
 /* Mapping of attributes to colours */
 static uint32_t amber;
 static uint32_t black;
@@ -72,6 +79,8 @@ static uint32_t normcols[256][2];
  */
 static int8_t cpq_st_display_internal = -1;
 
+static uint8_t mdaattr[256][2][2];
+
 static void
 compaq_plasma_display_set(uint8_t internal)
 {
@@ -85,21 +94,13 @@ compaq_plasma_display_get(void)
 }
 
 typedef struct compaq_plasma_t {
-    mem_mapping_t plasma_mapping;
     cga_t         cga;
     uint8_t       port_23c6;
     uint8_t       internal_monitor;
-    uint8_t       attrmap; /* Attribute mapping register */
-    int           linepos, displine;
-    uint8_t      *vram;
-    uint64_t      dispontime, dispofftime;
-    int           dispon, fullchange;
+    uint8_t       attrmap;
 } compaq_plasma_t;
 
-static uint8_t cga_crtcmask[32] = {
-    0xff, 0xff, 0xff, 0xff, 0x7f, 0x1f, 0x7f, 0x7f, 0xf3, 0x1f, 0x7f, 0x1f, 0x3f, 0xff, 0x3f, 0xff,
-    0xff, 0xff, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00
-};
+static int compaq_machine_type = 0;
 
 /* Compaq Deskpro 386 remaps RAM from 0xA0000-0xFFFFF to 0xFA0000-0xFFFFFF */
 static mem_mapping_t ram_mapping;
@@ -121,8 +122,18 @@ compaq_plasma_recalctimings(compaq_plasma_t *self)
     disptime          = 651;
     _dispontime       = 640;
     _dispofftime      = disptime - _dispontime;
-    self->dispontime  = (uint64_t) (_dispontime * (cpuclock / VID_CLOCK) * (double) (1ULL << 32));
-    self->dispofftime = (uint64_t) (_dispofftime * (cpuclock / VID_CLOCK) * (double) (1ULL << 32));
+    self->cga.dispontime  = (uint64_t) (_dispontime * (cpuclock / VID_CLOCK) * (double) (1ULL << 32));
+    self->cga.dispofftime = (uint64_t) (_dispofftime * (cpuclock / VID_CLOCK) * (double) (1ULL << 32));
+}
+
+static void
+compaq_plasma_waitstates(void *p)
+{
+    int ws_array[16] = { 3, 4, 5, 6, 7, 8, 4, 5, 6, 7, 8, 4, 5, 6, 7, 8 };
+    int ws;
+
+    ws = ws_array[cycles & 0xf];
+    sub_cycles(ws);
 }
 
 static void
@@ -130,7 +141,8 @@ compaq_plasma_write(uint32_t addr, uint8_t val, void *priv)
 {
     compaq_plasma_t *self = (compaq_plasma_t *) priv;
 
-    self->vram[addr & 0x7fff] = val;
+    self->cga.vram[addr & 0x7fff] = val;
+    compaq_plasma_waitstates(&self->cga);
 }
 
 static uint8_t
@@ -139,206 +151,12 @@ compaq_plasma_read(uint32_t addr, void *priv)
     compaq_plasma_t *self = (compaq_plasma_t *) priv;
     uint8_t          ret;
 
-    ret = (self->vram[addr & 0x7fff]);
+    compaq_plasma_waitstates(&self->cga);
+    ret = (self->cga.vram[addr & 0x7fff]);
 
     return ret;
 }
 
-/* Draw a row of text in 80-column mode */
-static void
-compaq_plasma_text80(compaq_plasma_t *self)
-{
-    uint32_t cols[2];
-    int      c;
-    uint8_t  chr;
-    uint8_t  attr;
-    int      drawcursor;
-    int      cursorline;
-    int      blink;
-    uint16_t addr;
-    uint8_t  sc;
-    uint16_t ma = (self->cga.crtc[13] | (self->cga.crtc[12] << 8)) & 0x7fff;
-    uint16_t ca = (self->cga.crtc[15] | (self->cga.crtc[14] << 8)) & 0x7fff;
-
-    sc   = (self->displine) & 15;
-    addr = ((ma & ~1) + (self->displine >> 4) * 80) * 2;
-    ma += (self->displine >> 4) * 80;
-
-    if ((self->cga.crtc[10] & 0x60) == 0x20)
-        cursorline = 0;
-    else
-        cursorline = ((self->cga.crtc[10] & 0x0F) * 2 <= sc) && ((self->cga.crtc[11] & 0x0F) * 2 >= sc);
-
-    for (uint8_t x = 0; x < 80; x++) {
-        chr        = self->vram[(addr + 2 * x) & 0x7FFF];
-        attr       = self->vram[(addr + 2 * x + 1) & 0x7FFF];
-        drawcursor = ((ma == ca) && cursorline && (self->cga.cgamode & 8) && (self->cga.cgablink & 16));
-
-        blink = ((self->cga.cgablink & 16) && (self->cga.cgamode & 0x20) && (attr & 0x80) && !drawcursor);
-
-        if (self->cga.cgamode & 0x20) { /* Blink */
-            cols[1] = blinkcols[attr][1];
-            cols[0] = blinkcols[attr][0];
-            if (blink)
-                cols[1] = cols[0];
-        } else {
-            cols[1] = normcols[attr][1];
-            cols[0] = normcols[attr][0];
-        }
-        if (drawcursor) {
-            for (c = 0; c < 8; c++)
-                ((uint32_t *) buffer32->line[self->displine])[(x << 3) + c] = cols[(fontdatm[chr][sc] & (1 << (c ^ 7))) ? 1 : 0] ^ (amber ^ black);
-        } else {
-            for (c = 0; c < 8; c++)
-                ((uint32_t *) buffer32->line[self->displine])[(x << 3) + c] = cols[(fontdatm[chr][sc] & (1 << (c ^ 7))) ? 1 : 0];
-        }
-        ++ma;
-    }
-}
-
-/* Draw a row of text in 40-column mode */
-static void
-compaq_plasma_text40(compaq_plasma_t *self)
-{
-    uint32_t cols[2];
-    int      c;
-    uint8_t  chr;
-    uint8_t  attr;
-    int      drawcursor;
-    int      cursorline;
-    int      blink;
-    uint16_t addr;
-    uint8_t  sc;
-    uint16_t ma = (self->cga.crtc[13] | (self->cga.crtc[12] << 8)) & 0x7fff;
-    uint16_t ca = (self->cga.crtc[15] | (self->cga.crtc[14] << 8)) & 0x7fff;
-
-    sc   = (self->displine) & 15;
-    addr = ((ma & ~1) + (self->displine >> 4) * 40) * 2;
-    ma += (self->displine >> 4) * 40;
-
-    if ((self->cga.crtc[10] & 0x60) == 0x20)
-        cursorline = 0;
-    else
-        cursorline = ((self->cga.crtc[10] & 0x0F) * 2 <= sc) && ((self->cga.crtc[11] & 0x0F) * 2 >= sc);
-
-    for (uint8_t x = 0; x < 40; x++) {
-        chr        = self->vram[(addr + 2 * x) & 0x7FFF];
-        attr       = self->vram[(addr + 2 * x + 1) & 0x7FFF];
-        drawcursor = ((ma == ca) && cursorline && (self->cga.cgamode & 8) && (self->cga.cgablink & 16));
-
-        blink = ((self->cga.cgablink & 16) && (self->cga.cgamode & 0x20) && (attr & 0x80) && !drawcursor);
-
-        if (self->cga.cgamode & 0x20) { /* Blink */
-            cols[1] = blinkcols[attr][1];
-            cols[0] = blinkcols[attr][0];
-            if (blink)
-                cols[1] = cols[0];
-        } else {
-            cols[1] = normcols[attr][1];
-            cols[0] = normcols[attr][0];
-        }
-        if (drawcursor) {
-            for (c = 0; c < 8; c++) {
-                ((uint32_t *) buffer32->line[self->displine])[(x << 4) + c * 2] = ((uint32_t *) buffer32->line[self->displine])[(x << 4) + c * 2 + 1] = cols[(fontdatm[chr][sc] & (1 << (c ^ 7))) ? 1 : 0] ^ (amber ^ black);
-            }
-        } else {
-            for (c = 0; c < 8; c++) {
-                ((uint32_t *) buffer32->line[self->displine])[(x << 4) + c * 2] = ((uint32_t *) buffer32->line[self->displine])[(x << 4) + c * 2 + 1] = cols[(fontdatm[chr][sc] & (1 << (c ^ 7))) ? 1 : 0];
-            }
-        }
-        ++ma;
-    }
-}
-
-/* Draw a line in CGA 640x200 or Compaq Plasma 640x400 mode */
-static void
-compaq_plasma_cgaline6(compaq_plasma_t *self)
-{
-    uint8_t  dat;
-    uint32_t ink = 0;
-    uint16_t addr;
-    uint32_t fg = (self->cga.cgacol & 0x0F) ? amber : black;
-    uint32_t bg = black;
-
-    uint16_t ma = (self->cga.crtc[13] | (self->cga.crtc[12] << 8)) & 0x7fff;
-
-    if ((self->cga.crtc[9] == 3) || (self->port_23c6 & 1)) /* 640*400 */ {
-        addr = ((self->displine) & 1) * 0x2000 + ((self->displine >> 1) & 1) * 0x4000 + (self->displine >> 2) * 80 + ((ma & ~1) << 1);
-    } else {
-        addr = ((self->displine >> 1) & 1) * 0x2000 + (self->displine >> 2) * 80 + ((ma & ~1) << 1);
-    }
-    for (uint8_t x = 0; x < 80; x++) {
-        dat = self->vram[addr & 0x7FFF];
-        addr++;
-
-        for (uint8_t c = 0; c < 8; c++) {
-            ink = (dat & 0x80) ? fg : bg;
-            if (!(self->cga.cgamode & 8))
-                ink = black;
-            ((uint32_t *) buffer32->line[self->displine])[x * 8 + c] = ink;
-            dat <<= 1;
-        }
-    }
-}
-
-/* Draw a line in CGA 320x200 mode. Here the CGA colours are converted to
- * dither patterns: colour 1 to 25% grey, colour 2 to 50% grey */
-static void
-compaq_plasma_cgaline4(compaq_plasma_t *self)
-{
-    uint8_t  dat;
-    uint8_t  pattern;
-    uint32_t ink0 = 0;
-    uint32_t ink1 = 0;
-    uint16_t addr;
-
-    uint16_t ma = (self->cga.crtc[13] | (self->cga.crtc[12] << 8)) & 0x7fff;
-
-    /* 320*200 */
-    addr = ((self->displine >> 1) & 1) * 0x2000 + (self->displine >> 2) * 80 + ((ma & ~1) << 1);
-
-    for (uint8_t x = 0; x < 80; x++) {
-        dat = self->vram[addr & 0x7FFF];
-        addr++;
-
-        for (uint8_t c = 0; c < 4; c++) {
-            pattern = (dat & 0xC0) >> 6;
-            if (!(self->cga.cgamode & 8))
-                pattern = 0;
-
-            switch (pattern & 3) {
-                case 0:
-                    ink0 = ink1 = black;
-                    break;
-                case 1:
-                    if (self->displine & 1) {
-                        ink0 = black;
-                        ink1 = black;
-                    } else {
-                        ink0 = amber;
-                        ink1 = black;
-                    }
-                    break;
-                case 2:
-                    if (self->displine & 1) {
-                        ink0 = black;
-                        ink1 = amber;
-                    } else {
-                        ink0 = amber;
-                        ink1 = black;
-                    }
-                    break;
-                case 3:
-                    ink0 = ink1 = amber;
-                    break;
-            }
-            ((uint32_t *) buffer32->line[self->displine])[x * 8 + 2 * c]     = ink0;
-            ((uint32_t *) buffer32->line[self->displine])[x * 8 + 2 * c + 1] = ink1;
-            dat <<= 2;
-        }
-    }
-}
-
 static void
 compaq_plasma_out(uint16_t addr, uint8_t val, void *priv)
 {
@@ -348,7 +166,7 @@ compaq_plasma_out(uint16_t addr, uint8_t val, void *priv)
     switch (addr) {
         /* Emulated CRTC, register select */
         case 0x3d4:
-            self->cga.crtcreg = val & 31;
+            cga_out(addr, val, &self->cga);
             break;
 
         /* Emulated CRTC, value */
@@ -366,33 +184,26 @@ compaq_plasma_out(uint16_t addr, uint8_t val, void *priv)
 
             if (old != val) {
                 if (self->cga.crtcreg < 0xe || self->cga.crtcreg > 0x10) {
-                    self->fullchange = changeframecount;
+                    self->cga.fullchange = changeframecount;
                     compaq_plasma_recalctimings(self);
                 }
             }
             break;
-
         case 0x3d8:
-            self->cga.cgamode = val;
-            break;
-
         case 0x3d9:
-            self->cga.cgacol = val;
+            cga_out(addr, val, &self->cga);
             break;
 
         case 0x13c6:
-            if (val & 8)
-                compaq_plasma_display_set(1);
-            else
-                compaq_plasma_display_set(0);
+            compaq_plasma_display_set((val & 8) ? 1 : 0);
             break;
 
         case 0x23c6:
             self->port_23c6 = val;
             if (val & 8) /* Disable internal CGA */
-                mem_mapping_disable(&self->plasma_mapping);
+                mem_mapping_disable(&self->cga.mapping);
             else
-                mem_mapping_enable(&self->plasma_mapping);
+                mem_mapping_enable(&self->cga.mapping);
             break;
     }
 }
@@ -405,31 +216,38 @@ compaq_plasma_in(uint16_t addr, void *priv)
 
     switch (addr) {
         case 0x3d4:
-            ret = self->cga.crtcreg;
+        case 0x3da:
+            ret = cga_in(addr, &self->cga);
             break;
 
         case 0x3d5:
             if (self->cga.crtcreg == 0x12) {
-                ret = self->attrmap & 0x0F;
+                ret = self->attrmap & 0x0f;
                 if (self->internal_monitor)
                     ret |= 0x30; /* Plasma / CRT */
             } else
-                ret = self->cga.crtc[self->cga.crtcreg];
-            break;
-
-        case 0x3da:
-            ret = self->cga.cgastat;
+                ret = cga_in(addr, &self->cga);
             break;
 
         case 0x13c6:
-            if (compaq_plasma_display_get())
-                ret = 8;
-            else
-                ret = 0;
+            ret = compaq_plasma_display_get() ? 8 : 0;
+            ret |= 4;
+            break;
+
+        case 0x1bc6:
+            ret = 0;
+            if (compaq_plasma_display_get()) {
+                if ((self->cga.cgamode & 0x12) == 0x12) {
+                    if (self->port_23c6 & 8)
+                        ret |= 0x40;
+                    else
+                        ret |= 0x20;
+                }
+            }
             break;
 
         case 0x23c6:
-            ret = self->port_23c6;
+            ret = 0;
             break;
     }
 
@@ -440,56 +258,242 @@ static void
 compaq_plasma_poll(void *p)
 {
     compaq_plasma_t *self = (compaq_plasma_t *) p;
+    uint8_t  chr, attr;
+    uint8_t  sc;
+    uint16_t ma  = (self->cga.crtc[13] | (self->cga.crtc[12] << 8)) & 0x7fff;
+    uint16_t ca  = (self->cga.crtc[15] | (self->cga.crtc[14] << 8)) & 0x7fff;
+    uint16_t addr;
+    int      drawcursor;
+    int      x, c;
+    int      cursorline;
+    int      blink     = 0;
+    int      underline = 0;
+    uint32_t ink = 0;
+    uint32_t fg = (self->cga.cgacol & 0x0f) ? amber : black;
+    uint32_t bg = black;
+    uint32_t cols[2];
+    uint8_t  dat2, pattern;
+    uint32_t ink0 = 0, ink1 = 0;
 
     /* Switch between internal plasma and external CRT display. */
-    if (cpq_st_display_internal != -1 && cpq_st_display_internal != self->internal_monitor) {
+    if ((cpq_st_display_internal != -1) && (cpq_st_display_internal != self->internal_monitor)) {
         self->internal_monitor = cpq_st_display_internal;
         compaq_plasma_recalctimings(self);
     }
 
+    /* graphic mode and not mode 40h */
     if (!self->internal_monitor && !(self->port_23c6 & 1)) {
         cga_poll(&self->cga);
         return;
     }
 
-    if (!self->linepos) {
-        timer_advance_u64(&self->cga.timer, self->dispofftime);
+    /* mode 40h or text mode */
+    if (!self->cga.linepos) {
+        timer_advance_u64(&self->cga.timer, self->cga.dispofftime);
         self->cga.cgastat |= 1;
-        self->linepos = 1;
-        if (self->dispon) {
-            if (self->displine == 0)
+        self->cga.linepos = 1;
+        if (self->cga.cgadispon) {
+            if (self->cga.displine == 0) {
                 video_wait_for_buffer();
+            }
+            if (self->cga.cgamode & 2) {
+                if (self->cga.cgamode & 0x10) {
+                    /* 640x400 mode */
+                    if (self->port_23c6 & 1) /* 640*400 */ {
+                        addr = ((self->cga.displine) & 1) * 0x2000 + ((self->cga.displine >> 1) & 1) * 0x4000 + (self->cga.displine >> 2) * 80 + ((ma & ~1) << 1);
+                    } else {
+                        addr = ((self->cga.displine >> 1) & 1) * 0x2000 + (self->cga.displine >> 2) * 80 + ((ma & ~1) << 1);
+                    }
+                    for (x = 0; x < 80; x++) {
+                        dat2 = self->cga.vram[(addr & 0x7FFF)];
+                        addr++;
 
-            /* Graphics */
-            if (self->cga.cgamode & 0x02) {
-                if (self->cga.cgamode & 0x10)
-                    compaq_plasma_cgaline6(self);
+                        for (c = 0; c < 8; c++) {
+                            ink = (dat2 & 0x80) ? fg : bg;
+                            if (!(self->cga.cgamode & 8))
+                                ink = black;
+                            ((uint32_t *) buffer32->line[self->cga.displine])[x * 8 + c] = ink;
+                            dat2 <<= 1;
+                        }
+                    }
+                } else {
+                    addr = ((self->cga.displine >> 1) & 1) * 0x2000 + (self->cga.displine >> 2) * 80 + ((ma & ~1) << 1);
+                    for (x = 0; x < 80; x++) {
+                        dat2 = self->cga.vram[(addr & 0x7fff)];
+                        addr++;
+
+                        for (c = 0; c < 4; c++) {
+                            pattern = (dat2 & 0xC0) >> 6;
+                            if (!(self->cga.cgamode & 8))
+                                pattern = 0;
+
+                            switch (pattern & 3) {
+                                case 0:
+                                    ink0 = ink1 = black;
+                                    break;
+                                case 1:
+                                    if (self->cga.displine & 1) {
+                                        ink0 = black;
+                                        ink1 = black;
+                                    } else {
+                                        ink0 = amber;
+                                        ink1 = black;
+                                    }
+                                    break;
+                                case 2:
+                                    if (self->cga.displine & 1) {
+                                        ink0 = black;
+                                        ink1 = amber;
+                                    } else {
+                                        ink0 = amber;
+                                        ink1 = black;
+                                    }
+                                    break;
+                                case 3:
+                                    ink0 = ink1 = amber;
+                                    break;
+                            }
+                            buffer32->line[self->cga.displine][x * 8 + 2 * c]     = ink0;
+                            buffer32->line[self->cga.displine][x * 8 + 2 * c + 1] = ink1;
+                            dat2 <<= 2;
+                        }
+                    }
+                }
+            } else if (self->cga.cgamode & 1) {
+                /* 80-col */
+                sc = self->cga.displine & 0x0f;
+                addr = ((ma & ~1) + (self->cga.displine >> 4) * 80) * 2;
+                ma += (self->cga.displine >> 4) * 80;
+
+                if ((self->cga.crtc[0x0a] & 0x60) == 0x20)
+                    cursorline = 0;
                 else
-                    compaq_plasma_cgaline4(self);
-            } else if (self->cga.cgamode & 0x01) /* High-res text */
-                compaq_plasma_text80(self);
-            else
-                compaq_plasma_text40(self);
+                    cursorline = ((self->cga.crtc[0x0a] & 0x0f) * 2 <= sc) && ((self->cga.crtc[0x0b] & 0x0F) * 2 >= sc);
+
+                /* for each text column */
+                for (x = 0; x < 80; x++) {
+                    /* video output enabled */
+                    if (self->cga.cgamode & 8) {
+                        chr        = self->cga.vram[(addr + 2 * x) & 0x7fff];
+                        attr       = self->cga.vram[(addr + 2 * x + 1) & 0x7fff];
+                    } else
+                        chr = attr = 0;
+                    /* check if cursor has to be drawn */
+                    drawcursor = ((ma == ca) && cursorline && self->cga.cursoron);
+                    underline = ((self->port_23c6 & 0x40) && (attr & 0x1) && !(attr & 0x6));
+                    if (underline) {
+                        /* set forecolor to white */
+                        attr |= 7;
+                    }
+                    blink = 0;
+                    /* blink active */
+                    if (self->cga.cgamode & 0x20) {
+                        cols[1] = blinkcols[attr][1];
+                        cols[0] = blinkcols[attr][0];
+                        /* attribute 7 active and not cursor */
+                        if ((self->cga.cgablink & 8) && (attr & 0x80) && !self->cga.drawcursor) {
+                            /* set blinking */
+                            cols[1] = cols[0];
+                            blink   = 1;
+                        }
+                    } else {
+                        /* Set intensity bit */
+                        cols[1] = normcols[attr][1];
+                        cols[0] = normcols[attr][0];
+                        blink   = (attr & 0x80) * 8 + 7 + 16;
+                    }
+                    /* character underline active and 7th row of pixels in character height being drawn */
+                    if (underline && (sc == 7)) {
+                        /* for each pixel in character width */
+                        for (c = 0; c < 8; c++)
+                            buffer32->line[self->cga.displine][(x << 3) + c] = mdaattr[attr][blink][1];
+                    } else if (drawcursor) {
+                        for (c = 0; c < 8; c++)
+                            buffer32->line[self->cga.displine][(x << 3) + c] = cols[(fontdatm[chr + self->cga.fontbase][sc] & (1 << (c ^ 7))) ? 1 : 0] ^ (amber ^ black);
+                    } else {
+                        for (c = 0; c < 8; c++)
+                            buffer32->line[self->cga.displine][(x << 3) + c] = cols[(fontdatm[chr + self->cga.fontbase][sc] & (1 << (c ^ 7))) ? 1 : 0];
+                    }
+
+                    ++ma;
+                }
+            } else { /* 40-col */
+                sc = self->cga.displine & 0x0f;
+                addr = ((ma & ~1) + (self->cga.displine >> 4) * 40) * 2;
+                ma += (self->cga.displine >> 4) * 40;
+
+                if ((self->cga.crtc[0x0a] & 0x60) == 0x20)
+                    cursorline = 0;
+                else
+                    cursorline = ((self->cga.crtc[0x0a] & 0x0f) * 2 <= sc) && ((self->cga.crtc[0x0b] & 0x0F) * 2 >= sc);
+
+                for (x = 0; x < 40; x++) {
+                    if (self->cga.cgamode & 8) {
+                        chr        = self->cga.vram[(addr + 2 * x) & 0x7fff];
+                        attr       = self->cga.vram[(addr + 2 * x + 1) & 0x7fff];
+                    } else {
+                        chr = attr = 0;
+                    }
+                    drawcursor = ((ma == ca) && cursorline && self->cga.cursoron);
+                    underline = ((self->port_23c6 & 0x40) && (attr & 0x1) && !(attr & 0x6));
+                    if (underline) {
+                        /* set forecolor to white */
+                        attr |= 7;
+                    }
+                    blink = 0;
+                    /* blink active */
+                    if (self->cga.cgamode & 0x20) {
+                        cols[1] = blinkcols[attr][1];
+                        cols[0] = blinkcols[attr][0];
+                        /* attribute 7 active and not cursor */
+                        if ((self->cga.cgablink & 8) && (attr & 0x80) && !self->cga.drawcursor) {
+                            /* set blinking */
+                            cols[1] = cols[0];
+                            blink   = 1;
+                        }
+                    } else {
+                        /* Set intensity bit */
+                        cols[1] = normcols[attr][1];
+                        cols[0] = normcols[attr][0];
+                        blink   = (attr & 0x80) * 8 + 7 + 16;
+                    }
+                    /* character underline active and 7th row of pixels in character height being drawn */
+                    if (underline && (sc == 7)) {
+                        /* for each pixel in character width */
+                        for (c = 0; c < 8; c++)
+                            buffer32->line[self->cga.displine][(x << 4) + (c * 2)] = buffer32->line[self->cga.displine][(x << 4) + (c * 2) + 1] = mdaattr[attr][blink][1];
+                    } else if (drawcursor) {
+                        for (c = 0; c < 8; c++) {
+                            buffer32->line[self->cga.displine][(x << 4) + c * 2] = buffer32->line[self->cga.displine][(x << 4) + c * 2 + 1] = cols[(fontdatm[chr][sc] & (1 << (c ^ 7))) ? 1 : 0] ^ (amber ^ black);
+                        }
+                    } else {
+                        for (c = 0; c < 8; c++) {
+                            buffer32->line[self->cga.displine][(x << 4) + c * 2] = buffer32->line[self->cga.displine][(x << 4) + c * 2 + 1] = cols[(fontdatm[chr][sc] & (1 << (c ^ 7))) ? 1 : 0];
+                        }
+                    }
+                    ++ma;
+                }
+            }
         }
-        self->displine++;
+        self->cga.displine++;
         /* Hardcode a fixed refresh rate and VSYNC timing */
-        if (self->displine == 400) { /* Start of VSYNC */
+        if (self->cga.displine == 400) { /* Start of VSYNC */
             self->cga.cgastat |= 8;
-            self->dispon = 0;
+            self->cga.cgadispon = 0;
         }
-        if (self->displine == 416) { /* End of VSYNC */
-            self->displine = 0;
+        if (self->cga.displine == 416) { /* End of VSYNC */
+            self->cga.displine = 0;
             self->cga.cgastat &= ~8;
-            self->dispon = 1;
+            self->cga.cgadispon = 1;
         }
     } else {
-        if (self->dispon)
+        if (self->cga.cgadispon)
             self->cga.cgastat &= ~1;
 
-        timer_advance_u64(&self->cga.timer, self->dispontime);
-        self->linepos = 0;
+        timer_advance_u64(&self->cga.timer, self->cga.dispontime);
+        self->cga.linepos = 0;
 
-        if (self->displine == 400) {
+        if (self->cga.displine == 400) {
             /* Hardcode 640x400 window size */
             if ((640 != xsize) || (400 != ysize) || video_force_resize_get()) {
                 xsize = 640;
@@ -515,14 +519,41 @@ compaq_plasma_poll(void *p)
                     video_bpp = 1;
                 else
                     video_bpp = 2;
-
             } else
                 video_bpp = 0;
+
             self->cga.cgablink++;
         }
     }
 }
 
+static void
+compaq_plasma_mdaattr_rebuild(void)
+{
+    int c;
+
+    for (c = 0; c < 256; c++) {
+        mdaattr[c][0][0] = mdaattr[c][1][0] = mdaattr[c][1][1] = 16;
+        if (c & 8)
+            mdaattr[c][0][1] = 15 + 16;
+        else
+            mdaattr[c][0][1] = 7 + 16;
+    }
+
+    mdaattr[0x70][0][1] = 16;
+    mdaattr[0x70][0][0] = mdaattr[0x70][1][0] = mdaattr[0x70][1][1] = 16 + 15;
+    mdaattr[0xF0][0][1]                                             = 16;
+    mdaattr[0xF0][0][0] = mdaattr[0xF0][1][0] = mdaattr[0xF0][1][1] = 16 + 15;
+    mdaattr[0x78][0][1]                                             = 16 + 7;
+    mdaattr[0x78][0][0] = mdaattr[0x78][1][0] = mdaattr[0x78][1][1] = 16 + 15;
+    mdaattr[0xF8][0][1]                                             = 16 + 7;
+    mdaattr[0xF8][0][0] = mdaattr[0xF8][1][0] = mdaattr[0xF8][1][1] = 16 + 15;
+    mdaattr[0x00][0][1] = mdaattr[0x00][1][1] = 16;
+    mdaattr[0x08][0][1] = mdaattr[0x08][1][1] = 16;
+    mdaattr[0x80][0][1] = mdaattr[0x80][1][1] = 16;
+    mdaattr[0x88][0][1] = mdaattr[0x88][1][1] = 16;
+}
+
 static void
 compaq_plasma_recalcattrs(compaq_plasma_t *self)
 {
@@ -537,10 +568,10 @@ compaq_plasma_recalcattrs(compaq_plasma_t *self)
      *            are bold */
 
     /* Set up colours */
-    amber = makecol(0xff, 0x7D, 0x00);
-    black = makecol(0x64, 0x19, 0x00);
+    amber = makecol(0xff, 0x7d, 0x00);
+    black = makecol(0x64, 0x0c, 0x00);
 
-    /* Initialise the attribute mapping. Start by defaulting everything
+    /* Initialize the attribute mapping. Start by defaulting everything
      * to black on amber, and with bold set by bit 3 */
     for (n = 0; n < 256; n++) {
         blinkcols[n][0] = normcols[n][0] = amber;
@@ -607,36 +638,37 @@ compaq_plasma_recalcattrs(compaq_plasma_t *self)
 static void *
 compaq_plasma_init(const device_t *info)
 {
-    int              display_type;
     compaq_plasma_t *self = malloc(sizeof(compaq_plasma_t));
     memset(self, 0, sizeof(compaq_plasma_t));
 
-    display_type        = device_get_config_int("display_type");
-    self->cga.composite = (display_type != CGA_RGB);
-    self->cga.revision  = device_get_config_int("composite_type");
+    video_inform(VIDEO_FLAG_TYPE_CGA, &timing_compaq_plasma);
+    loadfont_ex("roms/machines/portableiii/K Combined.bin", 1, 0x4bb2);
 
-    self->vram             = malloc(0x8000);
+    self->cga.composite = 0;
+    self->cga.revision  = 0;
+
+    self->cga.vram             = malloc(0x8000);
     self->internal_monitor = 1;
 
     cga_comp_init(self->cga.revision);
     timer_add(&self->cga.timer, compaq_plasma_poll, self, 1);
-    mem_mapping_add(&self->plasma_mapping, 0xb8000, 0x08000, compaq_plasma_read, NULL, NULL, compaq_plasma_write, NULL, NULL, NULL /*self->cga.vram*/, MEM_MAPPING_EXTERNAL, self);
+    mem_mapping_add(&self->cga.mapping, 0xb8000, 0x08000, compaq_plasma_read, NULL, NULL, compaq_plasma_write, NULL, NULL, NULL /*self->cga.vram*/, MEM_MAPPING_EXTERNAL, self);
     io_sethandler(0x03d0, 0x0010, compaq_plasma_in, NULL, NULL, compaq_plasma_out, NULL, NULL, self);
     io_sethandler(0x13c6, 0x0001, compaq_plasma_in, NULL, NULL, compaq_plasma_out, NULL, NULL, self);
+    io_sethandler(0x1bc6, 0x0001, compaq_plasma_in, NULL, NULL, compaq_plasma_out, NULL, NULL, self);
     io_sethandler(0x23c6, 0x0001, compaq_plasma_in, NULL, NULL, compaq_plasma_out, NULL, NULL, self);
 
     /* Default attribute mapping is 4 */
     self->attrmap = 4;
     compaq_plasma_recalcattrs(self);
 
-    self->cga.cgastat = 0xF4;
-    self->cga.vram    = self->vram;
-
+    self->cga.cgastat = 0xf4;
     overscan_x = overscan_y = 16;
 
     self->cga.rgb_type = device_get_config_int("rgb_type");
     cga_palette        = (self->cga.rgb_type << 1);
     cgapal_rebuild();
+    compaq_plasma_mdaattr_rebuild();
 
     return self;
 }
@@ -646,8 +678,7 @@ compaq_plasma_close(void *p)
 {
     compaq_plasma_t *self = (compaq_plasma_t *) p;
 
-    free(self->vram);
-
+    free(self->cga.vram);
     free(self);
 }
 
@@ -661,38 +692,10 @@ compaq_plasma_speed_changed(void *p)
 
 const device_config_t compaq_plasma_config[] = {
   // clang-format off
-    {
-        .name = "display_type",
-        .description = "Display type",
-        .type = CONFIG_SELECTION,
-        .default_string = "",
-        .default_int = CGA_RGB,
-        .file_filter = "",
-        .spinner = { 0 },
-        .selection = {
-            { .description = "RGB",       .value = CGA_RGB       },
-            { .description = "Composite", .value = CGA_COMPOSITE },
-            { .description = ""                                  }
-        }
-    },
-    {
-        .name = "composite_type",
-        .description = "Composite type",
-        .type = CONFIG_SELECTION,
-        .default_string = "",
-        .default_int = COMPOSITE_OLD,
-        .file_filter = "",
-        .spinner = { 0 },
-        {
-            { .description = "Old", .value = COMPOSITE_OLD },
-            { .description = "New", .value = COMPOSITE_NEW },
-            { .description = ""                            }
-        }
-    },
     {
         .name = "rgb_type",
-        .description = "RGB type",
-        .type = CONFIG_SELECTION,
+		.description = "RGB type",
+		.type = CONFIG_SELECTION,
         .default_string = "",
         .default_int = 0,
         .file_filter = "",
@@ -702,7 +705,6 @@ const device_config_t compaq_plasma_config[] = {
             { .description = "Green Monochrome", .value = 1 },
             { .description = "Amber Monochrome", .value = 2 },
             { .description = "Gray Monochrome",  .value = 3 },
-            { .description = "Color (no brown)", .value = 4 },
             { .description = ""                             }
         }
     },
@@ -781,26 +783,36 @@ write_raml(uint32_t addr, uint32_t val, void *priv)
 static void
 machine_at_compaq_init(const machine_t *model, int type)
 {
-    if (type != COMPAQ_DESKPRO386)
+    compaq_machine_type = type;
+
+    if ((type != COMPAQ_DESKPRO386) && (type != COMPAQ_DESKPRO386_01_1988))
         mem_remap_top(384);
 
     if (fdc_type == FDC_INTERNAL)
         device_add(&fdc_at_device);
 
-    mem_mapping_add(&ram_mapping, 0xfa0000, 0x60000,
-                    read_ram, read_ramw, read_raml,
-                    write_ram, write_ramw, write_raml,
-                    0xa0000 + ram, MEM_MAPPING_INTERNAL, NULL);
+    if ((type == COMPAQ_DESKPRO386) || (type == COMPAQ_DESKPRO386_01_1988) || (type == COMPAQ_PORTABLEIII386))
+        mem_mapping_add(&ram_mapping, 0xfa0000, 0x60000,
+                        read_ram, read_ramw, read_raml,
+                        write_ram, write_ramw, write_raml,
+                        0xa0000 + ram, MEM_MAPPING_EXTERNAL, NULL);
+    else
+        mem_mapping_add(&ram_mapping, 0xfa0000, 0x60000,
+                        read_ram, read_ramw, read_raml,
+                        write_ram, write_ramw, write_raml,
+                        0xa0000 + ram, MEM_MAPPING_INTERNAL, NULL);
 
     video_reset(gfxcard[0]);
 
     switch (type) {
         case COMPAQ_PORTABLEII:
+			machine_at_init(model);
             break;
 
         case COMPAQ_PORTABLEIII:
             if (gfxcard[0] == VID_INTERNAL)
                 device_add(&compaq_plasma_device);
+			machine_at_init(model);
             break;
 
         case COMPAQ_PORTABLEIII386:
@@ -808,15 +820,17 @@ machine_at_compaq_init(const machine_t *model, int type)
                 device_add(&ide_isa_device);
             if (gfxcard[0] == VID_INTERNAL)
                 device_add(&compaq_plasma_device);
+			machine_at_init(model);
             break;
 
         case COMPAQ_DESKPRO386:
+        case COMPAQ_DESKPRO386_01_1988:
             if (hdc_current == 1)
                 device_add(&ide_isa_device);
+            machine_at_common_init(model);
+            device_add(&keyboard_at_compaq_device);
             break;
     }
-
-    machine_at_init(model);
 }
 
 int
@@ -841,9 +855,9 @@ machine_at_portableiii_init(const machine_t *model)
 {
     int ret;
 
-    ret = bios_load_interleavedr("roms/machines/portableiii/Compaq Portable III - BIOS - 106779-002 - Even.bin",
-                                 "roms/machines/portableiii/Compaq Portable III - BIOS - 106778-002 - Odd.bin",
-                                 0x000f8000, 65536, 0);
+    ret = bios_load_linearr("roms/machines/portableiii/K Combined.bin",
+                                0x000f8000, 65536, 0);
+
 
     if (bios_only || !ret)
         return ret;
@@ -858,9 +872,8 @@ machine_at_portableiii386_init(const machine_t *model)
 {
     int ret;
 
-    ret = bios_load_interleavedr("roms/machines/portableiii/Compaq Portable III - BIOS - 106779-002 - Even.bin",
-                                 "roms/machines/portableiii/Compaq Portable III - BIOS - 106778-002 - Odd.bin",
-                                 0x000f8000, 65536, 0);
+    ret = bios_load_linearr("roms/machines/portableiii/K Combined.bin",
+                                0x000f8000, 65536, 0);
 
     if (bios_only || !ret)
         return ret;
@@ -870,14 +883,13 @@ machine_at_portableiii386_init(const machine_t *model)
     return ret;
 }
 
-#if defined(DEV_BRANCH) && defined(USE_DESKPRO386)
 int
 machine_at_deskpro386_init(const machine_t *model)
 {
     int ret;
 
     ret = bios_load_linearr("roms/machines/deskpro386/1986-09-04-HI.json.bin",
-                            0x000fc000, 65536, 0);
+                            0x000f8000, 65536, 0);
 
     if (bios_only || !ret)
         return ret;
@@ -886,4 +898,19 @@ machine_at_deskpro386_init(const machine_t *model)
 
     return ret;
 }
-#endif
+
+int
+machine_at_deskpro386_01_1988_init(const machine_t *model)
+{
+    int ret;
+
+    ret = bios_load_linearr("roms/machines/deskpro386/1988-01-28.json.bin",
+                            0x000f8000, 65536, 0);
+
+    if (bios_only || !ret)
+        return ret;
+
+    machine_at_compaq_init(model, COMPAQ_DESKPRO386_01_1988);
+
+    return ret;
+}
diff --git a/src/machine/m_at_slot1.c b/src/machine/m_at_slot1.c
index 3cedb669a..f13fe72cd 100644
--- a/src/machine/m_at_slot1.c
+++ b/src/machine/m_at_slot1.c
@@ -200,8 +200,8 @@ machine_at_p2bls_init(const machine_t *model)
     pci_init(PCI_CONFIG_TYPE_1);
     pci_register_slot(0x00, PCI_CARD_NORTHBRIDGE, 0, 0, 0, 0);
     pci_register_slot(0x04, PCI_CARD_SOUTHBRIDGE, 1, 2, 3, 4);
-    pci_register_slot(0x06, PCI_CARD_NORMAL,      4, 1, 2, 3); /* SCSI */
-    pci_register_slot(0x07, PCI_CARD_NORMAL,      3, 4, 1, 2); /* LAN */
+    pci_register_slot(0x06, PCI_CARD_SCSI,        4, 1, 2, 3);
+    pci_register_slot(0x07, PCI_CARD_NETWORK,     3, 4, 1, 2);
     pci_register_slot(0x0B, PCI_CARD_NORMAL,      2, 3, 4, 1);
     pci_register_slot(0x0C, PCI_CARD_NORMAL,      1, 2, 3, 4);
     pci_register_slot(0x09, PCI_CARD_NORMAL,      4, 1, 2, 3);
diff --git a/src/machine/machine.c b/src/machine/machine.c
index 893700c7e..d48cbdf21 100644
--- a/src/machine/machine.c
+++ b/src/machine/machine.c
@@ -41,6 +41,7 @@
 #include <86box/video.h>
 #include <86box/machine.h>
 #include <86box/isamem.h>
+#include <86box/pci.h>
 
 int bios_only = 0;
 int machine;
@@ -103,6 +104,8 @@ machine_init_ex(int m)
 
         /* Reset the fast off stuff. */
         cpu_fast_off_reset();
+
+        pci_take_over_io = 0x00000000;
     }
 
     /* All good, boot the machine! */
diff --git a/src/machine/machine_table.c b/src/machine/machine_table.c
index ea5d06a70..b9945b865 100644
--- a/src/machine/machine_table.c
+++ b/src/machine/machine_table.c
@@ -2587,8 +2587,8 @@ const machine_t machines[] = {
         .cpu = {
             .package = CPU_PKG_286,
             .block = CPU_BLOCK_NONE,
-            .min_bus = 6000000,
-            .max_bus = 8000000,
+            .min_bus = 0,
+            .max_bus = 0,
             .min_voltage = 0,
             .max_voltage = 0,
             .min_multi = 0,
@@ -2904,8 +2904,8 @@ const machine_t machines[] = {
         .cpu = {
             .package = CPU_PKG_286,
             .block = CPU_BLOCK_NONE,
-            .min_bus = 6000000,
-            .max_bus = 8000000,
+            .min_bus = 0,
+            .max_bus = 0,
             .min_voltage = 0,
             .max_voltage = 0,
             .min_multi = 0,
@@ -2943,8 +2943,8 @@ const machine_t machines[] = {
         .cpu = {
             .package = CPU_PKG_286,
             .block = CPU_BLOCK_NONE,
-            .min_bus = 6000000,
-            .max_bus = 8000000,
+            .min_bus = 0,
+            .max_bus = 0,
             .min_voltage = 0,
             .max_voltage = 0,
             .min_multi = 0,
@@ -4670,9 +4670,8 @@ const machine_t machines[] = {
         .net_device = NULL
     },
     /* Uses Compaq KBC firmware. */
-#if defined(DEV_BRANCH) && defined(USE_DESKPRO386)
     {
-        .name = "[ISA] Compaq Deskpro 386",
+        .name = "[ISA] Compaq Deskpro 386 (September 1986)",
         .internal_name = "deskpro386",
         .type = MACHINE_TYPE_386DX,
         .chipset = MACHINE_CHIPSET_DISCRETE,
@@ -4692,13 +4691,51 @@ const machine_t machines[] = {
             .max_multi = 0
         },
         .bus_flags = MACHINE_AT,
-        .flags = MACHINE_IDE | MACHINE_APM,
+        .flags = MACHINE_IDE,
         .ram = {
-            .min = 1024,
-            .max = 14336,
-            .step = 1024
+            .min = 640,
+            .max = 16384,
+            .step = 128
         },
-        .nvrmask = 127,
+        .nvrmask = 63,
+        .kbc_device = NULL,
+        .kbc_p1 = 0,
+        .gpio = 0,
+        .device = NULL,
+        .fdc_device = NULL,
+        .sio_device = NULL,
+        .vid_device = NULL,
+        .snd_device = NULL,
+        .net_device = NULL
+    },
+    {
+        .name = "[ISA] Compaq Deskpro 386 (January 1988)",
+        .internal_name = "deskpro386_01_1988",
+        .type = MACHINE_TYPE_386DX,
+        .chipset = MACHINE_CHIPSET_DISCRETE,
+        .init = machine_at_deskpro386_01_1988_init,
+        .pad = 0,
+        .pad0 = 0,
+        .pad1 = MACHINE_AVAILABLE,
+        .pad2 = 0,
+        .cpu = {
+            .package = CPU_PKG_386DX,
+            .block = CPU_BLOCK_NONE,
+            .min_bus = 0,
+            .max_bus = 0,
+            .min_voltage = 0,
+            .max_voltage = 0,
+            .min_multi = 0,
+            .max_multi = 0
+        },
+        .bus_flags = MACHINE_AT,
+        .flags = MACHINE_IDE,
+        .ram = {
+            .min = 640,
+            .max = 16384,
+            .step = 128
+        },
+        .nvrmask = 63,
         .kbc_device = NULL,
         .kbc_p1 = 0,
         .gpio = 0,
@@ -4709,7 +4746,6 @@ const machine_t machines[] = {
         .snd_device = NULL,
         .net_device = NULL
     },
-#endif /* defined(DEV_BRANCH) && defined(USE_DESKPRO386) */
     {
         .name = "[ISA] Compaq Portable III (386)",
         .internal_name = "portableiii386",
@@ -4731,7 +4767,7 @@ const machine_t machines[] = {
             .max_multi = 0
         },
         .bus_flags = MACHINE_AT,
-        .flags = MACHINE_IDE | MACHINE_VIDEO | MACHINE_APM,
+        .flags = MACHINE_IDE | MACHINE_VIDEO,
         .ram = {
             .min = 1024,
             .max = 14336,
@@ -5492,7 +5528,7 @@ const machine_t machines[] = {
             .max_multi = 0
         },
         .bus_flags = MACHINE_PS2_MCA,
-        .flags = MACHINE_VIDEO,
+        .flags = MACHINE_VIDEO | MACHINE_SOFTFLOAT_ONLY,
         .ram = {
             .min = 2048,
             .max = 65536,
@@ -6799,7 +6835,7 @@ const machine_t machines[] = {
             .min_multi = 0,
             .max_multi = 0
         },
-        .bus_flags = MACHINE_PCI,
+        .bus_flags = MACHINE_BUS_PS2_LATCH | MACHINE_PCI,
         .flags = MACHINE_IDE_DUAL | MACHINE_APM,
         .ram = {
             .min = 1024,
@@ -6877,7 +6913,7 @@ const machine_t machines[] = {
             .min_multi = 0,
             .max_multi = 0
         },
-        .bus_flags = MACHINE_PCI,
+        .bus_flags = MACHINE_BUS_PS2_LATCH | MACHINE_PCI,
         .flags = MACHINE_IDE_DUAL | MACHINE_APM,
         .ram = {
             .min = 1024,
diff --git a/src/mem/CMakeLists.txt b/src/mem/CMakeLists.txt
index 6aad80544..001e7b5bc 100644
--- a/src/mem/CMakeLists.txt
+++ b/src/mem/CMakeLists.txt
@@ -14,4 +14,4 @@
 #
 
 add_library(mem OBJECT catalyst_flash.c i2c_eeprom.c intel_flash.c mem.c rom.c
-    smram.c spd.c sst_flash.c)
+    row.c smram.c spd.c sst_flash.c)
diff --git a/src/mem/catalyst_flash.c b/src/mem/catalyst_flash.c
index 5e473f540..46294a60a 100644
--- a/src/mem/catalyst_flash.c
+++ b/src/mem/catalyst_flash.c
@@ -55,19 +55,22 @@ enum {
 };
 
 typedef struct flash_t {
-    uint8_t command, pad,
-        pad0, pad1,
-        *array;
+    uint8_t command;
+    uint8_t pad;
+    uint8_t pad0;
+    uint8_t pad1;
+    uint8_t *array;
 
-    mem_mapping_t mapping, mapping_h[2];
+    mem_mapping_t mapping;
+    mem_mapping_t mapping_h[2];
 } flash_t;
 
 static char flash_path[1024];
 
 static uint8_t
-flash_read(uint32_t addr, void *p)
+flash_read(uint32_t addr, void *priv)
 {
-    flash_t *dev = (flash_t *) p;
+    flash_t *dev = (flash_t *) priv;
     uint8_t  ret = 0xff;
 
     addr &= biosmask;
@@ -86,6 +89,9 @@ flash_read(uint32_t addr, void *p)
             else if (addr == 0x00001)
                 ret = 0xB4; /* 28F010 */
             break;
+
+        default:
+            break;
     }
 
     return ret;
@@ -141,13 +147,15 @@ flash_write(uint32_t addr, uint8_t val, void *p)
 }
 
 static void
-flash_writew(uint32_t addr, uint16_t val, void *p)
+flash_writew(UNUSED(uint32_t addr), UNUSED(uint16_t val), UNUSED(void *priv))
 {
+    //
 }
 
 static void
-flash_writel(uint32_t addr, uint32_t val, void *p)
+flash_writel(UNUSED(uint32_t addr), UNUSED(uint32_t val), UNUSED(void *priv))
 {
+    //
 }
 
 static void
@@ -179,7 +187,7 @@ catalyst_flash_reset(void *priv)
 }
 
 static void *
-catalyst_flash_init(const device_t *info)
+catalyst_flash_init(UNUSED(const device_t *info))
 {
     FILE    *f;
     flash_t *dev;
diff --git a/src/mem/i2c_eeprom.c b/src/mem/i2c_eeprom.c
index e29d16139..7d83bbe88 100644
--- a/src/mem/i2c_eeprom.c
+++ b/src/mem/i2c_eeprom.c
@@ -23,13 +23,18 @@
 #include <wchar.h>
 #include <86box/86box.h>
 #include <86box/i2c.h>
+#include <86box/plat_unused.h>
 
-typedef struct {
+typedef struct i2c_eeprom_t {
     void   *i2c;
-    uint8_t addr, *data, writable;
+    uint8_t addr;
+    uint8_t *data;
+    uint8_t writable;
 
-    uint32_t addr_mask, addr_register;
-    uint8_t  addr_len, addr_pos;
+    uint32_t addr_mask;
+    uint32_t addr_register;
+    uint8_t  addr_len;
+    uint8_t  addr_pos;
 } i2c_eeprom_t;
 
 #ifdef ENABLE_I2C_EEPROM_LOG
@@ -51,7 +56,7 @@ i2c_eeprom_log(const char *fmt, ...)
 #endif
 
 static uint8_t
-i2c_eeprom_start(void *bus, uint8_t addr, uint8_t read, void *priv)
+i2c_eeprom_start(UNUSED(void *bus), uint8_t addr, uint8_t read, void *priv)
 {
     i2c_eeprom_t *dev = (i2c_eeprom_t *) priv;
 
@@ -66,7 +71,7 @@ i2c_eeprom_start(void *bus, uint8_t addr, uint8_t read, void *priv)
 }
 
 static uint8_t
-i2c_eeprom_read(void *bus, uint8_t addr, void *priv)
+i2c_eeprom_read(UNUSED(void *bus), UNUSED(uint8_t addr), void *priv)
 {
     i2c_eeprom_t *dev = (i2c_eeprom_t *) priv;
     uint8_t       ret = dev->data[dev->addr_register];
@@ -79,7 +84,7 @@ i2c_eeprom_read(void *bus, uint8_t addr, void *priv)
 }
 
 static uint8_t
-i2c_eeprom_write(void *bus, uint8_t addr, uint8_t data, void *priv)
+i2c_eeprom_write(UNUSED(void *bus), uint8_t addr, uint8_t data, void *priv)
 {
     i2c_eeprom_t *dev = (i2c_eeprom_t *) priv;
 
@@ -104,7 +109,7 @@ i2c_eeprom_write(void *bus, uint8_t addr, uint8_t data, void *priv)
 }
 
 static void
-i2c_eeprom_stop(void *bus, uint8_t addr, void *priv)
+i2c_eeprom_stop(UNUSED(void *bus), UNUSED(uint8_t addr), void *priv)
 {
     i2c_eeprom_t *dev = (i2c_eeprom_t *) priv;
 
diff --git a/src/mem/intel_flash.c b/src/mem/intel_flash.c
index af7a71eea..ce529670e 100644
--- a/src/mem/intel_flash.c
+++ b/src/mem/intel_flash.c
@@ -58,25 +58,30 @@ enum {
 };
 
 typedef struct flash_t {
-    uint8_t command, status,
-        pad, flags,
-        *array;
+    uint8_t  command;
+    uint8_t  status;
+    uint8_t  pad;
+    uint8_t  flags;
+    uint8_t *array;
 
-    uint16_t flash_id, pad16;
+    uint16_t flash_id;
+    uint16_t pad16;
 
-    uint32_t program_addr,
-        block_start[BLOCKS_NUM], block_end[BLOCKS_NUM],
-        block_len[BLOCKS_NUM];
+    uint32_t program_addr;
+    uint32_t block_start[BLOCKS_NUM];
+    uint32_t block_end[BLOCKS_NUM];
+    uint32_t block_len[BLOCKS_NUM];
 
-    mem_mapping_t mapping[4], mapping_h[16];
+    mem_mapping_t mapping[4];
+    mem_mapping_t mapping_h[16];
 } flash_t;
 
 static char flash_path[1024];
 
 static uint8_t
-flash_read(uint32_t addr, void *p)
+flash_read(uint32_t addr, void *priv)
 {
-    flash_t *dev = (flash_t *) p;
+    flash_t *dev = (flash_t *) priv;
     uint8_t  ret = 0xff;
 
     if (dev->flags & FLAG_INV_A16)
@@ -84,8 +89,8 @@ flash_read(uint32_t addr, void *p)
     addr &= biosmask;
 
     switch (dev->command) {
-        case CMD_READ_ARRAY:
         default:
+        case CMD_READ_ARRAY:
             ret = dev->array[addr];
             break;
 
@@ -105,9 +110,9 @@ flash_read(uint32_t addr, void *p)
 }
 
 static uint16_t
-flash_readw(uint32_t addr, void *p)
+flash_readw(uint32_t addr, void *priv)
 {
-    flash_t  *dev = (flash_t *) p;
+    flash_t  *dev = (flash_t *) priv;
     uint16_t *q;
     uint16_t  ret = 0xffff;
 
@@ -123,8 +128,8 @@ flash_readw(uint32_t addr, void *p)
 
     if (dev->flags & FLAG_WORD)
         switch (dev->command) {
-            case CMD_READ_ARRAY:
             default:
+            case CMD_READ_ARRAY:
                 break;
 
             case CMD_IID:
@@ -143,9 +148,9 @@ flash_readw(uint32_t addr, void *p)
 }
 
 static uint32_t
-flash_readl(uint32_t addr, void *p)
+flash_readl(uint32_t addr, void *priv)
 {
-    flash_t  *dev = (flash_t *) p;
+    flash_t  *dev = (flash_t *) priv;
     uint32_t *q;
 
     if (dev->flags & FLAG_INV_A16)
@@ -158,10 +163,9 @@ flash_readl(uint32_t addr, void *p)
 }
 
 static void
-flash_write(uint32_t addr, uint8_t val, void *p)
+flash_write(uint32_t addr, uint8_t val, void *priv)
 {
-    flash_t *dev = (flash_t *) p;
-    int      i;
+    flash_t *dev = (flash_t *) priv;
     uint32_t bb_mask = biosmask & 0xffffe000;
     if (biosmask == 0x7ffff)
         bb_mask &= 0xffff8000;
@@ -175,7 +179,7 @@ flash_write(uint32_t addr, uint8_t val, void *p)
     switch (dev->command) {
         case CMD_ERASE_SETUP:
             if (val == CMD_ERASE_CONFIRM) {
-                for (i = 0; i < 6; i++) {
+                for (uint8_t i = 0; i < 6; i++) {
                     if ((i == dev->program_addr) && (addr >= dev->block_start[i]) && (addr <= dev->block_end[i]))
                         memset(&(dev->array[dev->block_start[i]]), 0xff, dev->block_len[i]);
                 }
@@ -200,7 +204,7 @@ flash_write(uint32_t addr, uint8_t val, void *p)
                     dev->status = 0;
                     break;
                 case CMD_ERASE_SETUP:
-                    for (i = 0; i < 7; i++) {
+                    for (uint8_t i = 0; i < 7; i++) {
                         if ((addr >= dev->block_start[i]) && (addr <= dev->block_end[i]))
                             dev->program_addr = i;
                     }
@@ -209,6 +213,9 @@ flash_write(uint32_t addr, uint8_t val, void *p)
                 case CMD_PROGRAM_SETUP_ALT:
                     dev->program_addr = addr;
                     break;
+
+                default:
+                    break;
             }
     }
 }
@@ -266,16 +273,19 @@ flash_writew(uint32_t addr, uint16_t val, void *p)
                     case CMD_PROGRAM_SETUP_ALT:
                         dev->program_addr = addr;
                         break;
+
+                    default:
+                        break;
                 }
         }
 }
 
 static void
-flash_writel(uint32_t addr, uint32_t val, void *p)
+flash_writel(UNUSED(uint32_t addr), UNUSED(uint32_t val), UNUSED(void *priv))
 {
 #if 0
-    flash_writew(addr, val & 0xffff, p);
-    flash_writew(addr + 2, (val >> 16) & 0xffff, p);
+    flash_writew(addr, val & 0xffff, priv);
+    flash_writew(addr + 2, (val >> 16) & 0xffff, priv);
 #endif
 }
 
@@ -522,10 +532,10 @@ intel_flash_init(const device_t *info)
 }
 
 static void
-intel_flash_close(void *p)
+intel_flash_close(void *priv)
 {
     FILE    *f;
-    flash_t *dev = (flash_t *) p;
+    flash_t *dev = (flash_t *) priv;
 
     f = nvr_fopen(flash_path, "wb");
     fwrite(&(dev->array[dev->block_start[BLOCK_MAIN1]]), dev->block_len[BLOCK_MAIN1], 1, f);
diff --git a/src/mem/mem.c b/src/mem/mem.c
index 3008e658c..b23ab3253 100644
--- a/src/mem/mem.c
+++ b/src/mem/mem.c
@@ -1776,7 +1776,7 @@ mem_write_phys(void *src, uint32_t addr, int transfer_size)
 }
 
 uint8_t
-mem_read_ram(uint32_t addr, void *priv)
+mem_read_ram(uint32_t addr, UNUSED(void *priv))
 {
 #ifdef ENABLE_MEM_LOG
     if ((addr >= 0xa0000) && (addr <= 0xbffff))
@@ -1790,7 +1790,7 @@ mem_read_ram(uint32_t addr, void *priv)
 }
 
 uint16_t
-mem_read_ramw(uint32_t addr, void *priv)
+mem_read_ramw(uint32_t addr, UNUSED(void *priv))
 {
 #ifdef ENABLE_MEM_LOG
     if ((addr >= 0xa0000) && (addr <= 0xbffff))
@@ -1804,7 +1804,7 @@ mem_read_ramw(uint32_t addr, void *priv)
 }
 
 uint32_t
-mem_read_raml(uint32_t addr, void *priv)
+mem_read_raml(uint32_t addr, UNUSED(void *priv))
 {
 #ifdef ENABLE_MEM_LOG
     if ((addr >= 0xa0000) && (addr <= 0xbffff))
@@ -1818,7 +1818,7 @@ mem_read_raml(uint32_t addr, void *priv)
 }
 
 uint8_t
-mem_read_ram_2gb(uint32_t addr, void *priv)
+mem_read_ram_2gb(uint32_t addr, UNUSED(void *priv))
 {
 #ifdef ENABLE_MEM_LOG
     if ((addr >= 0xa0000) && (addr <= 0xbffff))
@@ -1831,7 +1831,7 @@ mem_read_ram_2gb(uint32_t addr, void *priv)
 }
 
 uint16_t
-mem_read_ram_2gbw(uint32_t addr, void *priv)
+mem_read_ram_2gbw(uint32_t addr, UNUSED(void *priv))
 {
 #ifdef ENABLE_MEM_LOG
     if ((addr >= 0xa0000) && (addr <= 0xbffff))
@@ -1844,7 +1844,7 @@ mem_read_ram_2gbw(uint32_t addr, void *priv)
 }
 
 uint32_t
-mem_read_ram_2gbl(uint32_t addr, void *priv)
+mem_read_ram_2gbl(uint32_t addr, UNUSED(void *priv))
 {
 #ifdef ENABLE_MEM_LOG
     if ((addr >= 0xa0000) && (addr <= 0xbffff))
@@ -1858,71 +1858,71 @@ mem_read_ram_2gbl(uint32_t addr, void *priv)
 
 #ifdef USE_NEW_DYNAREC
 static inline int
-page_index(page_t *p)
+page_index(page_t *page)
 {
-    return ((uintptr_t) p - (uintptr_t) pages) / sizeof(page_t);
+    return ((uintptr_t) page - (uintptr_t) pages) / sizeof(page_t);
 }
 
 void
-page_add_to_evict_list(page_t *p)
+page_add_to_evict_list(page_t *page)
 {
-    pages[purgable_page_list_head].evict_prev = page_index(p);
-    p->evict_next                             = purgable_page_list_head;
-    p->evict_prev                             = 0;
+    pages[purgable_page_list_head].evict_prev = page_index(page);
+    page->evict_next                          = purgable_page_list_head;
+    page->evict_prev                          = 0;
     purgable_page_list_head                   = pages[purgable_page_list_head].evict_prev;
     purgeable_page_count++;
 }
 
 void
-page_remove_from_evict_list(page_t *p)
+page_remove_from_evict_list(page_t *page)
 {
-    if (!page_in_evict_list(p))
+    if (!page_in_evict_list(page))
         fatal("page_remove_from_evict_list: not in evict list!\n");
-    if (p->evict_prev)
-        pages[p->evict_prev].evict_next = p->evict_next;
+    if (page->evict_prev)
+        pages[page->evict_prev].evict_next = page->evict_next;
     else
-        purgable_page_list_head = p->evict_next;
-    if (p->evict_next)
-        pages[p->evict_next].evict_prev = p->evict_prev;
-    p->evict_prev = EVICT_NOT_IN_LIST;
+        purgable_page_list_head = page->evict_next;
+    if (page->evict_next)
+        pages[page->evict_next].evict_prev = page->evict_prev;
+    page->evict_prev = EVICT_NOT_IN_LIST;
     purgeable_page_count--;
 }
 
 void
-mem_write_ramb_page(uint32_t addr, uint8_t val, page_t *p)
+mem_write_ramb_page(uint32_t addr, uint8_t val, page_t *page)
 {
-    if (p == NULL)
+    if (page == NULL)
         return;
 
 #    ifdef USE_DYNAREC
-    if ((p->mem == NULL) || (p->mem == page_ff) || (val != p->mem[addr & 0xfff]) || codegen_in_recompile) {
+    if ((page->mem == NULL) || (page->mem == page_ff) || (val != page->mem[addr & 0xfff]) || codegen_in_recompile) {
 #    else
-    if ((p->mem == NULL) || (p->mem == page_ff) || (val != p->mem[addr & 0xfff])) {
+    if ((page->mem == NULL) || (page->mem == page_ff) || (val != page->mem[addr & 0xfff])) {
 #    endif
         uint64_t mask        = (uint64_t) 1 << ((addr >> PAGE_MASK_SHIFT) & PAGE_MASK_MASK);
         int      byte_offset = (addr >> PAGE_BYTE_MASK_SHIFT) & PAGE_BYTE_MASK_OFFSET_MASK;
         uint64_t byte_mask   = (uint64_t) 1 << (addr & PAGE_BYTE_MASK_MASK);
 
-        p->mem[addr & 0xfff] = val;
-        p->dirty_mask |= mask;
-        if ((p->code_present_mask & mask) && !page_in_evict_list(p))
-            page_add_to_evict_list(p);
-        p->byte_dirty_mask[byte_offset] |= byte_mask;
-        if ((p->byte_code_present_mask[byte_offset] & byte_mask) && !page_in_evict_list(p))
-            page_add_to_evict_list(p);
+        page->mem[addr & 0xfff] = val;
+        page->dirty_mask |= mask;
+        if ((page->code_present_mask & mask) && !page_in_evict_list(page))
+            page_add_to_evict_list(page);
+        page->byte_dirty_mask[byte_offset] |= byte_mask;
+        if ((page->byte_code_present_mask[byte_offset] & byte_mask) && !page_in_evict_list(page))
+            page_add_to_evict_list(page);
     }
 }
 
 void
-mem_write_ramw_page(uint32_t addr, uint16_t val, page_t *p)
+mem_write_ramw_page(uint32_t addr, uint16_t val, page_t *page)
 {
-    if (p == NULL)
+    if (page == NULL)
         return;
 
 #    ifdef USE_DYNAREC
-    if ((p->mem == NULL) || (p->mem == page_ff) || (val != *(uint16_t *) &p->mem[addr & 0xfff]) || codegen_in_recompile) {
+    if ((page->mem == NULL) || (page->mem == page_ff) || (val != *(uint16_t *) &page->mem[addr & 0xfff]) || codegen_in_recompile) {
 #    else
-    if ((p->mem == NULL) || (p->mem == page_ff) || (val != *(uint16_t *) &p->mem[addr & 0xfff])) {
+    if ((page->mem == NULL) || (page->mem == page_ff) || (val != *(uint16_t *) &page->mem[addr & 0xfff])) {
 #    endif
         uint64_t mask        = (uint64_t) 1 << ((addr >> PAGE_MASK_SHIFT) & PAGE_MASK_MASK);
         int      byte_offset = (addr >> PAGE_BYTE_MASK_SHIFT) & PAGE_BYTE_MASK_OFFSET_MASK;
@@ -1930,34 +1930,34 @@ mem_write_ramw_page(uint32_t addr, uint16_t val, page_t *p)
 
         if ((addr & 0xf) == 0xf)
             mask |= (mask << 1);
-        *(uint16_t *) &p->mem[addr & 0xfff] = val;
-        p->dirty_mask |= mask;
-        if ((p->code_present_mask & mask) && !page_in_evict_list(p))
-            page_add_to_evict_list(p);
+        *(uint16_t *) &page->mem[addr & 0xfff] = val;
+        page->dirty_mask |= mask;
+        if ((page->code_present_mask & mask) && !page_in_evict_list(page))
+            page_add_to_evict_list(page);
         if ((addr & PAGE_BYTE_MASK_MASK) == PAGE_BYTE_MASK_MASK) {
-            p->byte_dirty_mask[byte_offset + 1] |= 1;
-            if ((p->byte_code_present_mask[byte_offset + 1] & 1) && !page_in_evict_list(p))
-                page_add_to_evict_list(p);
+            page->byte_dirty_mask[byte_offset + 1] |= 1;
+            if ((page->byte_code_present_mask[byte_offset + 1] & 1) && !page_in_evict_list(page))
+                page_add_to_evict_list(page);
         } else
             byte_mask |= (byte_mask << 1);
 
-        p->byte_dirty_mask[byte_offset] |= byte_mask;
+        page->byte_dirty_mask[byte_offset] |= byte_mask;
 
-        if ((p->byte_code_present_mask[byte_offset] & byte_mask) && !page_in_evict_list(p))
-            page_add_to_evict_list(p);
+        if ((page->byte_code_present_mask[byte_offset] & byte_mask) && !page_in_evict_list(page))
+            page_add_to_evict_list(page);
     }
 }
 
 void
-mem_write_raml_page(uint32_t addr, uint32_t val, page_t *p)
+mem_write_raml_page(uint32_t addr, uint32_t val, page_t *page)
 {
-    if (p == NULL)
+    if (page == NULL)
         return;
 
 #    ifdef USE_DYNAREC
-    if ((p->mem == NULL) || (p->mem == page_ff) || (val != *(uint32_t *) &p->mem[addr & 0xfff]) || codegen_in_recompile) {
+    if ((page->mem == NULL) || (page->mem == page_ff) || (val != *(uint32_t *) &page->mem[addr & 0xfff]) || codegen_in_recompile) {
 #    else
-    if ((p->mem == NULL) || (p->mem == page_ff) || (val != *(uint32_t *) &p->mem[addr & 0xfff])) {
+    if ((page->mem == NULL) || (page->mem == page_ff) || (val != *(uint32_t *) &page->mem[addr & 0xfff])) {
 #    endif
         uint64_t mask        = (uint64_t) 1 << ((addr >> PAGE_MASK_SHIFT) & PAGE_MASK_MASK);
         int      byte_offset = (addr >> PAGE_BYTE_MASK_SHIFT) & PAGE_BYTE_MASK_OFFSET_MASK;
@@ -1965,79 +1965,79 @@ mem_write_raml_page(uint32_t addr, uint32_t val, page_t *p)
 
         if ((addr & 0xf) >= 0xd)
             mask |= (mask << 1);
-        *(uint32_t *) &p->mem[addr & 0xfff] = val;
-        p->dirty_mask |= mask;
-        p->byte_dirty_mask[byte_offset] |= byte_mask;
-        if (!page_in_evict_list(p) && ((p->code_present_mask & mask) || (p->byte_code_present_mask[byte_offset] & byte_mask)))
-            page_add_to_evict_list(p);
+        *(uint32_t *) &page->mem[addr & 0xfff] = val;
+        page->dirty_mask |= mask;
+        page->byte_dirty_mask[byte_offset] |= byte_mask;
+        if (!page_in_evict_list(page) && ((page->code_present_mask & mask) || (page->byte_code_present_mask[byte_offset] & byte_mask)))
+            page_add_to_evict_list(page);
         if ((addr & PAGE_BYTE_MASK_MASK) > (PAGE_BYTE_MASK_MASK - 3)) {
             uint32_t byte_mask_2 = 0xf >> (4 - (addr & 3));
 
-            p->byte_dirty_mask[byte_offset + 1] |= byte_mask_2;
-            if ((p->byte_code_present_mask[byte_offset + 1] & byte_mask_2) && !page_in_evict_list(p))
-                page_add_to_evict_list(p);
+            page->byte_dirty_mask[byte_offset + 1] |= byte_mask_2;
+            if ((page->byte_code_present_mask[byte_offset + 1] & byte_mask_2) && !page_in_evict_list(page))
+                page_add_to_evict_list(page);
         }
     }
 }
 #else
 void
-mem_write_ramb_page(uint32_t addr, uint8_t val, page_t *p)
+mem_write_ramb_page(uint32_t addr, uint8_t val, page_t *page)
 {
-    if (p == NULL)
+    if (page == NULL)
         return;
 
 #    ifdef USE_DYNAREC
-    if ((p->mem == NULL) || (p->mem == page_ff) || (val != p->mem[addr & 0xfff]) || codegen_in_recompile) {
+    if ((page->mem == NULL) || (page->mem == page_ff) || (val != page->mem[addr & 0xfff]) || codegen_in_recompile) {
 #    else
-    if ((p->mem == NULL) || (p->mem == page_ff) || (val != p->mem[addr & 0xfff])) {
+    if ((page->mem == NULL) || (page->mem == page_ff) || (val != page->mem[addr & 0xfff])) {
 #    endif
         uint64_t mask = (uint64_t) 1 << ((addr >> PAGE_MASK_SHIFT) & PAGE_MASK_MASK);
-        p->dirty_mask[(addr >> PAGE_MASK_INDEX_SHIFT) & PAGE_MASK_INDEX_MASK] |= mask;
-        p->mem[addr & 0xfff] = val;
+        page->dirty_mask[(addr >> PAGE_MASK_INDEX_SHIFT) & PAGE_MASK_INDEX_MASK] |= mask;
+        page->mem[addr & 0xfff] = val;
     }
 }
 
 void
-mem_write_ramw_page(uint32_t addr, uint16_t val, page_t *p)
+mem_write_ramw_page(uint32_t addr, uint16_t val, page_t *page)
 {
-    if (p == NULL)
+    if (page == NULL)
         return;
 
 #    ifdef USE_DYNAREC
-    if ((p->mem == NULL) || (p->mem == page_ff) || (val != *(uint16_t *) &p->mem[addr & 0xfff]) || codegen_in_recompile) {
+    if ((page->mem == NULL) || (page->mem == page_ff) || (val != *(uint16_t *) &page->mem[addr & 0xfff]) || codegen_in_recompile) {
 #    else
-    if ((p->mem == NULL) || (p->mem == page_ff) || (val != *(uint16_t *) &p->mem[addr & 0xfff])) {
+    if ((page->mem == NULL) || (page->mem == page_ff) || (val != *(uint16_t *) &page->mem[addr & 0xfff])) {
 #    endif
         uint64_t mask = (uint64_t) 1 << ((addr >> PAGE_MASK_SHIFT) & PAGE_MASK_MASK);
         if ((addr & 0xf) == 0xf)
             mask |= (mask << 1);
-        p->dirty_mask[(addr >> PAGE_MASK_INDEX_SHIFT) & PAGE_MASK_INDEX_MASK] |= mask;
-        *(uint16_t *) &p->mem[addr & 0xfff] = val;
+        page->dirty_mask[(addr >> PAGE_MASK_INDEX_SHIFT) & PAGE_MASK_INDEX_MASK] |= mask;
+        *(uint16_t *) &page->mem[addr & 0xfff] = val;
     }
 }
 
 void
-mem_write_raml_page(uint32_t addr, uint32_t val, page_t *p)
+mem_write_raml_page(uint32_t addr, uint32_t val, page_t *page)
 {
-    if (p == NULL)
+    if (page == NULL)
         return;
 
 #    ifdef USE_DYNAREC
-    if ((p->mem == NULL) || (p->mem == page_ff) || (val != *(uint32_t *) &p->mem[addr & 0xfff]) || codegen_in_recompile) {
+    if ((page->mem == NULL) || (page->mem == page_ff) || (val != *(uint32_t *) &page->mem[addr & 0xfff]) || codegen_in_recompile) {
 #    else
-    if ((p->mem == NULL) || (p->mem == page_ff) || (val != *(uint32_t *) &p->mem[addr & 0xfff])) {
+    if ((page->mem == NULL) || (page->mem == page_ff) || (val != *(uint32_t *) &page->mem[addr & 0xfff])) {
 #    endif
         uint64_t mask = (uint64_t) 1 << ((addr >> PAGE_MASK_SHIFT) & PAGE_MASK_MASK);
         if ((addr & 0xf) >= 0xd)
             mask |= (mask << 1);
-        p->dirty_mask[(addr >> PAGE_MASK_INDEX_SHIFT) & PAGE_MASK_INDEX_MASK] |= mask;
-        *(uint32_t *) &p->mem[addr & 0xfff] = val;
+        page->dirty_mask[(addr >> PAGE_MASK_INDEX_SHIFT) & PAGE_MASK_INDEX_MASK] |= mask;
+        *(uint32_t *) &page->mem[addr & 0xfff] = val;
     }
 }
 #endif
 
 void
-mem_write_ram(uint32_t addr, uint8_t val, void *priv)
+mem_write_ram(uint32_t addr, uint8_t val, UNUSED(void *priv))
 {
 #ifdef ENABLE_MEM_LOG
     if ((addr >= 0xa0000) && (addr <= 0xbffff))
@@ -2065,7 +2065,7 @@ mem_write_ramw(uint32_t addr, uint16_t val, void *priv)
 }
 
 void
-mem_write_raml(uint32_t addr, uint32_t val, void *priv)
+mem_write_raml(uint32_t addr, uint32_t val, UNUSED(void *priv))
 {
 #ifdef ENABLE_MEM_LOG
     if ((addr >= 0xa0000) && (addr <= 0xbffff))
@@ -2079,7 +2079,7 @@ mem_write_raml(uint32_t addr, uint32_t val, void *priv)
 }
 
 static uint8_t
-mem_read_remapped(uint32_t addr, void *priv)
+mem_read_remapped(uint32_t addr, UNUSED(void *priv))
 {
     addr = 0xA0000 + (addr - remap_start_addr);
     if (is286)
@@ -2088,7 +2088,7 @@ mem_read_remapped(uint32_t addr, void *priv)
 }
 
 static uint16_t
-mem_read_remappedw(uint32_t addr, void *priv)
+mem_read_remappedw(uint32_t addr, UNUSED(void *priv))
 {
     addr = 0xA0000 + (addr - remap_start_addr);
     if (is286)
@@ -2097,7 +2097,7 @@ mem_read_remappedw(uint32_t addr, void *priv)
 }
 
 static uint32_t
-mem_read_remappedl(uint32_t addr, void *priv)
+mem_read_remappedl(uint32_t addr, UNUSED(void *priv))
 {
     addr = 0xA0000 + (addr - remap_start_addr);
     if (is286)
@@ -2106,7 +2106,7 @@ mem_read_remappedl(uint32_t addr, void *priv)
 }
 
 static uint8_t
-mem_read_remapped2(uint32_t addr, void *priv)
+mem_read_remapped2(uint32_t addr, UNUSED(void *priv))
 {
     addr = 0xD0000 + (addr - remap_start_addr2);
     if (is286)
@@ -2115,7 +2115,7 @@ mem_read_remapped2(uint32_t addr, void *priv)
 }
 
 static uint16_t
-mem_read_remappedw2(uint32_t addr, void *priv)
+mem_read_remappedw2(uint32_t addr, UNUSED(void *priv))
 {
     addr = 0xD0000 + (addr - remap_start_addr2);
     if (is286)
@@ -2124,7 +2124,7 @@ mem_read_remappedw2(uint32_t addr, void *priv)
 }
 
 static uint32_t
-mem_read_remappedl2(uint32_t addr, void *priv)
+mem_read_remappedl2(uint32_t addr, UNUSED(void *priv))
 {
     addr = 0xD0000 + (addr - remap_start_addr2);
     if (is286)
@@ -2133,7 +2133,7 @@ mem_read_remappedl2(uint32_t addr, void *priv)
 }
 
 static void
-mem_write_remapped(uint32_t addr, uint8_t val, void *priv)
+mem_write_remapped(uint32_t addr, uint8_t val, UNUSED(void *priv))
 {
     uint32_t oldaddr = addr;
     addr             = 0xA0000 + (addr - remap_start_addr);
@@ -2157,7 +2157,7 @@ mem_write_remappedw(uint32_t addr, uint16_t val, void *priv)
 }
 
 static void
-mem_write_remappedl(uint32_t addr, uint32_t val, void *priv)
+mem_write_remappedl(uint32_t addr, uint32_t val, UNUSED(void *priv))
 {
     uint32_t oldaddr = addr;
     addr             = 0xA0000 + (addr - remap_start_addr);
@@ -2169,7 +2169,7 @@ mem_write_remappedl(uint32_t addr, uint32_t val, void *priv)
 }
 
 static void
-mem_write_remapped2(uint32_t addr, uint8_t val, void *priv)
+mem_write_remapped2(uint32_t addr, uint8_t val, UNUSED(void *priv))
 {
     uint32_t oldaddr = addr;
     addr             = 0xD0000 + (addr - remap_start_addr2);
@@ -2193,7 +2193,7 @@ mem_write_remappedw2(uint32_t addr, uint16_t val, void *priv)
 }
 
 static void
-mem_write_remappedl2(uint32_t addr, uint32_t val, void *priv)
+mem_write_remappedl2(uint32_t addr, uint32_t val, UNUSED(void *priv))
 {
     uint32_t oldaddr = addr;
     addr             = 0xD0000 + (addr - remap_start_addr2);
@@ -2208,7 +2208,7 @@ void
 mem_invalidate_range(uint32_t start_addr, uint32_t end_addr)
 {
 #ifdef USE_NEW_DYNAREC
-    page_t *p;
+    page_t *page;
 
     start_addr &= ~PAGE_MASK_MASK;
     end_addr = (end_addr + PAGE_MASK_MASK) & ~PAGE_MASK_MASK;
@@ -2217,15 +2217,15 @@ mem_invalidate_range(uint32_t start_addr, uint32_t end_addr)
         if ((start_addr >> 12) >= pages_sz)
             continue;
 
-        p = &pages[start_addr >> 12];
-        if (p) {
-            p->dirty_mask = 0xffffffffffffffffULL;
+        page = &pages[start_addr >> 12];
+        if (page) {
+            page->dirty_mask = 0xffffffffffffffffULL;
 
-            if ((p->mem != page_ff) && p->byte_dirty_mask)
-                memset(p->byte_dirty_mask, 0xff, 64 * sizeof(uint64_t));
+            if ((page->mem != page_ff) && page->byte_dirty_mask)
+                memset(page->byte_dirty_mask, 0xff, 64 * sizeof(uint64_t));
 
-            if (!page_in_evict_list(p))
-                page_add_to_evict_list(p);
+            if (!page_in_evict_list(page))
+                page_add_to_evict_list(page);
         }
     }
 #else
@@ -2598,7 +2598,9 @@ mem_a20_init(void)
         if (is6117)
             rammask |= 0x03000000;
         flushmmucache();
-        // mem_a20_state = mem_a20_key | mem_a20_alt;
+#if 0
+        mem_a20_state = mem_a20_key | mem_a20_alt;
+#endif
     } else {
         rammask = 0xfffff;
         flushmmucache();
diff --git a/src/mem/rom.c b/src/mem/rom.c
index 80881c723..2ace87de7 100644
--- a/src/mem/rom.c
+++ b/src/mem/rom.c
@@ -384,7 +384,7 @@ rom_reset(uint32_t addr, int sz)
 }
 
 uint8_t
-bios_read(uint32_t addr, void *priv)
+bios_read(uint32_t addr, UNUSED(void *priv))
 {
     uint8_t ret = 0xff;
 
@@ -397,7 +397,7 @@ bios_read(uint32_t addr, void *priv)
 }
 
 uint16_t
-bios_readw(uint32_t addr, void *priv)
+bios_readw(uint32_t addr, UNUSED(void *priv))
 {
     uint16_t ret = 0xffff;
 
@@ -410,7 +410,7 @@ bios_readw(uint32_t addr, void *priv)
 }
 
 uint32_t
-bios_readl(uint32_t addr, void *priv)
+bios_readl(uint32_t addr, UNUSED(void *priv))
 {
     uint32_t ret = 0xffffffff;
 
@@ -442,7 +442,7 @@ bios_add(void)
         mem_mapping_add(&bios_mapping, 0xe0000, 0x20000,
                         bios_read, bios_readw, bios_readl,
                         NULL, NULL, NULL,
-                        &rom[0x20000], MEM_MAPPING_EXTERNAL | MEM_MAPPING_ROM | MEM_MAPPING_ROMCS, 0);
+                        &rom[biosmask + 1 - 0x20000], MEM_MAPPING_EXTERNAL | MEM_MAPPING_ROM | MEM_MAPPING_ROMCS, 0);
 
         mem_set_mem_state_both(0x0e0000, 0x20000,
                                MEM_READ_ROMCS | MEM_WRITE_ROMCS);
@@ -524,7 +524,7 @@ bios_load(const char *fn1, const char *fn2, uint32_t addr, int sz, int off, int
 }
 
 int
-bios_load_linear_combined(const char *fn1, const char *fn2, int sz, int off)
+bios_load_linear_combined(const char *fn1, const char *fn2, int sz, UNUSED(int off))
 {
     uint8_t ret = 0;
 
diff --git a/src/mem/row.c b/src/mem/row.c
new file mode 100644
index 000000000..2152c1f54
--- /dev/null
+++ b/src/mem/row.c
@@ -0,0 +1,343 @@
+/*
+ * 86Box	A hypervisor and IBM PC system emulator that specializes in
+ *		running old operating systems and software designed for IBM
+ *		PC systems and compatibles from 1981 through fairly recent
+ *		system designs based on the PCI bus.
+ *
+ *		This file is part of the 86Box distribution.
+ *
+ *		DRAM row handling.
+ *
+ * Authors:	Miran Grca, <mgrca8@gmail.com>
+ *
+ *		Copyright 2016-2020 Miran Grca.
+ */
+#include <inttypes.h>
+#include <stdarg.h>
+#include <stdio.h>
+#include <stdint.h>
+#include <string.h>
+#include <stdlib.h>
+#include <wchar.h>
+#define HAVE_STDARG_H
+#include <86box/86box.h>
+#include "cpu.h"
+#include "x86_ops.h"
+#include "x86.h"
+#include <86box/config.h>
+#include <86box/device.h>
+#include <86box/io.h>
+#include <86box/mem.h>
+#include <86box/spd.h>
+#include <86box/row.h>
+
+
+/*  0   1   2   3   4   5   6   7   */
+static uint8_t  rows_num, rows_default,
+                rows_bits;
+static uint32_t row_unit;
+static uint8_t  drb_defaults[16];
+static row_t    *rows;
+
+
+static uint8_t
+row_read(uint32_t addr, void *priv)
+{
+    row_t *dev = (row_t *) priv;
+    uint32_t new_addr = ((addr - dev->host_base) & dev->ram_mask) + dev->ram_base;
+
+    addreadlookup(mem_logical_addr, new_addr);
+
+    return dev->buf[new_addr];
+}
+
+
+static uint16_t
+row_readw(uint32_t addr, void *priv)
+{
+    row_t *dev = (row_t *) priv;
+    uint32_t new_addr = ((addr - dev->host_base) & dev->ram_mask) + dev->ram_base;
+
+    addreadlookup(mem_logical_addr, new_addr);
+
+    return *(uint16_t *) &(dev->buf[new_addr]);
+}
+
+
+static uint32_t
+row_readl(uint32_t addr, void *priv)
+{
+    row_t *dev = (row_t *) priv;
+    uint32_t new_addr = ((addr - dev->host_base) & dev->ram_mask) + dev->ram_base;
+
+    addreadlookup(mem_logical_addr, new_addr);
+
+    return *(uint32_t *) &(dev->buf[new_addr]);
+}
+
+
+static void
+row_write(uint32_t addr, uint8_t val, void *priv)
+{
+    row_t *dev = (row_t *) priv;
+    uint32_t new_addr = ((addr - dev->host_base) & dev->ram_mask) + dev->ram_base;
+
+    addwritelookup(mem_logical_addr, new_addr);
+    mem_write_ramb_page(new_addr, val, &pages[addr >> 12]);
+}
+
+
+static void
+row_writew(uint32_t addr, uint16_t val, void *priv)
+{
+    row_t *dev = (row_t *) priv;
+    uint32_t new_addr = ((addr - dev->host_base) & dev->ram_mask) + dev->ram_base;
+
+    addwritelookup(mem_logical_addr, new_addr);
+    mem_write_ramw_page(new_addr, val, &pages[addr >> 12]);
+}
+
+
+static void
+row_writel(uint32_t addr, uint32_t val, void *priv)
+{
+    row_t *dev = (row_t *) priv;
+    uint32_t new_addr = ((addr - dev->host_base) & dev->ram_mask) + dev->ram_base;
+
+    addwritelookup(mem_logical_addr, new_addr);
+    mem_write_raml_page(new_addr, val, &pages[addr >> 12]);
+}
+
+
+void
+row_allocate(uint8_t row_id, uint8_t set)
+{
+    uint32_t c, offset;
+
+    /* Do nothing if size is either zero or invalid. */
+    if ((rows[row_id].host_size == 0x00000000) || (rows[row_id].host_size == 0xffffffff))
+        return;
+
+    if (rows[row_id].ram_size == 0x00000000)
+        return;
+
+    for (c = (rows[row_id].host_base >> 12); c < ((rows[row_id].host_base + rows[row_id].host_size) >> 12); c++) {
+        offset = c - (rows[row_id].host_base >> 12);
+
+        pages[c].mem = set ? (rows[row_id].buf + rows[row_id].ram_base + ((offset << 12) & rows[row_id].ram_mask)) : page_ff;
+        pages[c].write_b = set ? mem_write_ramb_page : NULL;
+        pages[c].write_w = set ? mem_write_ramw_page : NULL;
+        pages[c].write_l = set ? mem_write_raml_page : NULL;
+#ifdef USE_NEW_DYNAREC
+	pages[c].evict_prev = EVICT_NOT_IN_LIST;
+	pages[c].byte_dirty_mask = &byte_dirty_mask[offset * 64];
+	pages[c].byte_code_present_mask = &byte_code_present_mask[offset * 64];
+#endif
+    }
+
+    if (rows[row_id].host_base >= 0x00100000) {
+	mem_set_mem_state_both(rows[row_id].host_base, rows[row_id].host_base + rows[row_id].host_size,
+                               set ? (MEM_READ_INTERNAL | MEM_WRITE_INTERNAL) : (MEM_READ_EXTERNAL | MEM_WRITE_EXTERNAL));
+    } else {
+        if (0x000a0000 > rows[row_id].host_base) {
+	    mem_set_mem_state_both(rows[row_id].host_base, 0x000a0000 - rows[row_id].host_base,
+                                   set ? (MEM_READ_INTERNAL | MEM_WRITE_INTERNAL) : (MEM_READ_EXTERNAL | MEM_WRITE_EXTERNAL));
+        }
+        if ((rows[row_id].host_base + rows[row_id].host_size) > 0x00100000) {
+	    mem_set_mem_state_both(0x00100000, (rows[row_id].host_base + rows[row_id].host_size) - 0x00100000,
+                                   set ? (MEM_READ_INTERNAL | MEM_WRITE_INTERNAL) : (MEM_READ_EXTERNAL | MEM_WRITE_EXTERNAL));
+        }
+    }
+
+    if (set) {
+        mem_mapping_set_addr(&rows[row_id].mapping, rows[row_id].host_base, rows[row_id].host_size);
+        mem_mapping_set_exec(&rows[row_id].mapping, rows[row_id].buf + rows[row_id].ram_base);
+        mem_mapping_set_mask(&rows[row_id].mapping, rows[row_id].ram_mask);
+        if ((rows[row_id].host_base == rows[row_id].ram_base) && (rows[row_id].host_size == rows[row_id].ram_size)) {
+#if (defined __amd64__ || defined _M_X64 || defined __aarch64__ || defined _M_ARM64)
+            mem_mapping_set_handler(&rows[row_id].mapping, mem_read_ram,mem_read_ramw,mem_read_raml,
+                                    mem_write_ram,mem_write_ramw,mem_write_raml);
+#else
+            if (rows[row_id].buf == ram2) {
+                mem_mapping_set_handler(&rows[row_id].mapping, mem_read_ram_2gb,mem_read_ram_2gbw,mem_read_ram_2gbl,
+                                        mem_write_ram,mem_write_ramw,mem_write_raml);
+            } else {
+                mem_mapping_set_handler(&rows[row_id].mapping, mem_read_ram,mem_read_ramw,mem_read_raml,
+                                        mem_write_ram,mem_write_ramw,mem_write_raml);
+            }
+#endif
+        } else {
+            mem_mapping_set_handler(&rows[row_id].mapping, row_read, row_readw, row_readl,
+                                    row_write, row_writew, row_writel);
+        }
+    } else
+        mem_mapping_disable(&rows[row_id].mapping);
+}
+
+
+void
+row_disable(uint8_t row_id)
+{
+    row_allocate(row_id, 0);
+}
+
+
+void
+row_set_boundary(uint8_t row_id, uint32_t boundary)
+{
+    if (row_id >= rows_num)
+        return;
+
+    boundary &= ((1 << rows_bits) - 1);
+
+    rows[row_id].host_size = boundary * row_unit;
+    if (row_id == 0)
+        rows[row_id].host_base = 0x00000000;
+    else {
+        rows[row_id].host_base = rows[row_id - 1].boundary * row_unit;
+        if (rows[row_id - 1].boundary > boundary)
+            rows[row_id].host_size = 0x00000000;
+        else
+            rows[row_id].host_size -= rows[row_id].host_base;
+    }
+
+    rows[row_id].boundary = boundary;
+
+    row_allocate(row_id, 1);
+}
+
+
+void
+row_reset(void *priv)
+{
+    int i;
+    uint32_t boundary, shift;
+
+    for (i = (rows_num - 1); i >= 0; i--)
+        row_disable(i);
+
+    for (i = 0; i < rows_num; i++) {
+        shift = (i & 1) << 2;
+        boundary = ((uint32_t) drb_defaults[i]) + (((((uint32_t) drb_defaults[(i >> 1) + 8]) >> shift) & 0xf) << 8);
+        row_set_boundary(i, boundary);
+    }
+}
+
+
+void
+row_close(void *priv)
+{
+    free(rows);
+    rows = NULL;
+}
+
+
+void *
+row_init(const device_t *info)
+{
+    uint32_t cur_drb = 0, cur_drbe = 0;
+    uint32_t last_drb = 0, last_drbe = 0;
+    uint8_t phys_drbs[16];
+    int i, max = info->local & 0xff;
+    int c;
+    uint32_t shift, drb;
+    uint32_t boundary, mask;
+    row_t *new_rows = NULL;
+
+    rows_bits = ((info->local >> 24) & 0xff);
+    mask = (1 << rows_bits) - 1;
+    row_unit = ((info->local >> 8) & 0xff);
+    memset(phys_drbs, 0x00, 16);
+    spd_write_drbs(phys_drbs, 0x00, max, row_unit);
+    row_unit <<= 20;
+    rows_default = (info->local >> 16) & 0xff;
+    memset(drb_defaults, 0x00, 16);
+    for (i = 0; i < 8; i++)
+        drb_defaults[i] = rows_default;
+
+    new_rows = calloc(max + 1, sizeof(row_t));
+    rows_num = max + 1;
+
+    rows = new_rows;
+
+    mem_mapping_disable(&ram_low_mapping);
+    mem_mapping_disable(&ram_mid_mapping);
+    mem_mapping_disable(&ram_high_mapping);
+#if (!(defined __amd64__ || defined _M_X64 || defined __aarch64__ || defined _M_ARM64))
+    if (mem_size > 1048576)
+        mem_mapping_disable(&ram_2gb_mapping);
+#endif
+
+    for (c = 0; c < pages_sz; c++) {
+        pages[c].mem = page_ff;
+        pages[c].write_b = NULL;
+        pages[c].write_w = NULL;
+        pages[c].write_l = NULL;
+#ifdef USE_NEW_DYNAREC
+	pages[c].evict_prev = EVICT_NOT_IN_LIST;
+	pages[c].byte_dirty_mask = &byte_dirty_mask[c * 64];
+	pages[c].byte_code_present_mask = &byte_code_present_mask[c * 64];
+#endif
+    }
+
+    /* Set all memory space above the default allocated area to external. */
+    boundary = ((uint32_t) rows_default) * row_unit;
+    mem_set_mem_state_both(boundary, (mem_size << 10) - boundary, MEM_READ_EXTERNAL | MEM_WRITE_EXTERNAL);
+
+    for (i = 0; i <= max; i++) {
+        cur_drb = phys_drbs[i];
+        cur_drbe = phys_drbs[(i >> 1) + 8];
+
+        shift = (i & 1) << 2;
+        drb = (cur_drb & mask) + (((cur_drbe >> shift) & 0x03) << 8);
+        rows[i].ram_size = drb * row_unit;
+
+        shift = ((i - 1) & 1) << 2;
+        drb = (last_drb & mask) + (((last_drbe >> shift) & 0x03) << 8);
+        rows[i].ram_base = drb * row_unit;
+        rows[i].ram_size -= rows[i].ram_base;
+
+        rows[i].buf = ram;
+#if (!(defined __amd64__ || defined _M_X64 || defined __aarch64__ || defined _M_ARM64))
+        if (rows[i].ram_base >= (1 << 30)) {
+            rows[i].ram_base -= (1 << 30);
+            rows[i].buf = ram2;
+        }
+#endif
+
+        rows[i].ram_mask = rows[i].ram_size - 1;
+
+        mem_mapping_add(&rows[i].mapping, rows[i].ram_base, rows[i].ram_size,
+                        row_read, row_readw, row_readl,
+                        row_write, row_writew, row_writel,
+                        rows[i].buf + rows[i].ram_base, MEM_MAPPING_INTERNAL, &(rows[i]));
+        mem_mapping_disable(&rows[i].mapping);
+
+        shift = (i & 1) << 2;
+        boundary = ((uint32_t) drb_defaults[i]) + ((((uint32_t) drb_defaults[(i >> 1) + 8]) >> shift) << 8);
+        row_set_boundary(i, boundary);
+
+        last_drb = cur_drb;
+        last_drbe = cur_drbe;
+    }
+
+    flushmmucache();
+
+    return new_rows;
+}
+
+
+/* NOTE: NOT const, so that we can patch it at init. */
+device_t row_device = {
+    .name = "DRAM Rows",
+    .internal_name = "dram_rows",
+    .flags = DEVICE_AT,
+    .local = 0x0000,
+    .init = row_init,
+    .close = row_close,
+    .reset = row_reset,
+    { .available = NULL },
+    .speed_changed = NULL,
+    .force_redraw = NULL,
+    .config = NULL
+};
diff --git a/src/mem/spd.c b/src/mem/spd.c
index c39a3bffd..6d3e88a93 100644
--- a/src/mem/spd.c
+++ b/src/mem/spd.c
@@ -27,6 +27,7 @@
 #include <86box/spd.h>
 #include <86box/version.h>
 #include <86box/machine.h>
+#include <86box/plat_unused.h>
 
 #define SPD_ROLLUP(x) ((x) >= 16 ? ((x) -15) : (x))
 
@@ -54,7 +55,7 @@ spd_log(const char *fmt, ...)
 #endif
 
 static void
-spd_close(void *priv)
+spd_close(UNUSED(void *priv))
 {
     spd_log("SPD: close()\n");
 
@@ -67,7 +68,7 @@ spd_close(void *priv)
 }
 
 static void *
-spd_init(const device_t *info)
+spd_init(UNUSED(const device_t *info))
 {
     spd_log("SPD: init()\n");
 
@@ -181,7 +182,6 @@ spd_register(uint8_t ram_type, uint8_t slot_mask, uint16_t max_module_size)
     uint8_t      slot;
     uint8_t      slot_count;
     uint8_t      row;
-    uint8_t      i;
     uint16_t     min_module_size;
     uint16_t     rows[SPD_MAX_SLOTS];
     uint16_t     asym;
@@ -262,7 +262,7 @@ spd_register(uint8_t ram_type, uint8_t slot_mask, uint16_t max_module_size)
                 edo_data->dram_width     = 8;
 
                 edo_data->spd_rev = 0x12;
-                for (i = spd_write_part_no(edo_data->part_no, (ram_type == SPD_TYPE_FPM) ? "FPM" : "EDO", rows[row]);
+                for (int i = spd_write_part_no(edo_data->part_no, (ram_type == SPD_TYPE_FPM) ? "FPM" : "EDO", rows[row]);
                      i < sizeof(edo_data->part_no); i++)
                     edo_data->part_no[i] = ' '; /* part number should be space-padded */
                 edo_data->rev_code[0] = BCD8(EMU_VERSION_MAJ);
@@ -270,9 +270,9 @@ spd_register(uint8_t ram_type, uint8_t slot_mask, uint16_t max_module_size)
                 edo_data->mfg_year    = 20;
                 edo_data->mfg_week    = 17;
 
-                for (i = 0; i < 63; i++)
+                for (uint8_t i = 0; i < 63; i++)
                     edo_data->checksum += spd_modules[slot]->data[i];
-                for (i = 0; i < 129; i++)
+                for (uint8_t i = 0; i < 129; i++)
                     edo_data->checksum2 += spd_modules[slot]->data[i];
                 break;
 
@@ -316,7 +316,7 @@ spd_register(uint8_t ram_type, uint8_t slot_mask, uint16_t max_module_size)
                 sdram_data->ca_hold = sdram_data->data_hold = 0x08;
 
                 sdram_data->spd_rev = 0x12;
-                for (i = spd_write_part_no(sdram_data->part_no, "SDR", rows[row]);
+                for (int i = spd_write_part_no(sdram_data->part_no, "SDR", rows[row]);
                      i < sizeof(sdram_data->part_no); i++)
                     sdram_data->part_no[i] = ' '; /* part number should be space-padded */
                 sdram_data->rev_code[0] = BCD8(EMU_VERSION_MAJ);
@@ -327,11 +327,14 @@ spd_register(uint8_t ram_type, uint8_t slot_mask, uint16_t max_module_size)
                 sdram_data->freq     = 100;
                 sdram_data->features = 0xFF;
 
-                for (i = 0; i < 63; i++)
+                for (uint8_t i = 0; i < 63; i++)
                     sdram_data->checksum += spd_modules[slot]->data[i];
-                for (i = 0; i < 129; i++)
+                for (uint8_t i = 0; i < 129; i++)
                     sdram_data->checksum2 += spd_modules[slot]->data[i];
                 break;
+
+            default:
+                break;
         }
 
         row++;
@@ -388,7 +391,7 @@ spd_write_drbs(uint8_t *regs, uint8_t reg_min, uint8_t reg_max, uint8_t drb_unit
         /* Write DRB register, adding the previous DRB's value. */
         if (row == 0)
             regs[drb] = 0;
-        else if ((apollo) && (drb == apollo))
+        else if (apollo && (drb == apollo))
             regs[drb] = regs[drb | 0xf]; /* 5F comes before 56 */
         else
             regs[drb] = regs[drb - 1];
@@ -398,7 +401,7 @@ spd_write_drbs(uint8_t *regs, uint8_t reg_min, uint8_t reg_max, uint8_t drb_unit
     }
 }
 
-/* Needed for 430LX. */
+/* Needed for 430NX. */
 void
 spd_write_drbs_with_ext(uint8_t *regs, uint8_t reg_min, uint8_t reg_max, uint8_t drb_unit)
 {
@@ -444,8 +447,9 @@ spd_write_drbs_with_ext(uint8_t *regs, uint8_t reg_min, uint8_t reg_max, uint8_t
             row_val += size / drb_unit; /* this will intentionally overflow on 440GX with 2 GB */
         regs[drb] = row_val & 0xff;
         drb       = reg_min + 8 + (row >> 1);
-        shift     = (row & 0x01) << 3;
-        regs[drb] = (((row_val & 0xfff) >> 8) << shift);
+        shift = (row & 0x01) << 2;
+        /* Limit to 1 GB space, per the 430NX datasheet. */
+        regs[drb] = (regs[drb] & ~(0xf << shift)) | (((row_val >> 8) & 3) << shift);
         spd_log("SPD: DRB[%d] = %d MB (%02Xh raw)\n", row, size, regs[drb]);
     }
 }
diff --git a/src/mem/sst_flash.c b/src/mem/sst_flash.c
index d9be436e9..dbb91d358 100644
--- a/src/mem/sst_flash.c
+++ b/src/mem/sst_flash.c
@@ -35,21 +35,31 @@
 #include <86box/m_xt_xi8088.h>
 
 typedef struct sst_t {
-    uint8_t manufacturer, id, has_bbp, is_39,
-        page_bytes, sdp, bbp_first_8k, bbp_last_8k;
+    uint8_t manufacturer;
+    uint8_t id;
+    uint8_t has_bbp;
+    uint8_t is_39;
+    uint8_t page_bytes;
+    uint8_t sdp;
+    uint8_t bbp_first_8k;
+    uint8_t bbp_last_8k;
 
-    int command_state, id_mode,
-        dirty;
+    int command_state;
+    int id_mode;
+    int dirty;
 
-    uint32_t size, mask,
-        page_mask, page_base,
-        last_addr;
+    uint32_t size;
+    uint32_t mask;
+    uint32_t page_mask;
+    uint32_t page_base;
+    uint32_t last_addr;
 
-    uint8_t page_buffer[128],
-        page_dirty[128];
+    uint8_t  page_buffer[128];
+    uint8_t  page_dirty[128];
     uint8_t *array;
 
-    mem_mapping_t mapping[8], mapping_h[8];
+    mem_mapping_t mapping[8];
+    mem_mapping_t mapping_h[8];
 
     pc_timer_t page_write_timer;
 } sst_t;
@@ -252,9 +262,9 @@ sst_page_write(void *priv)
 }
 
 static uint8_t
-sst_read_id(uint32_t addr, void *p)
+sst_read_id(uint32_t addr, void *priv)
 {
-    sst_t  *dev = (sst_t *) p;
+    sst_t  *dev = (sst_t *) priv;
     uint8_t ret = 0x00;
 
     if ((addr & 0xffff) == 0)
@@ -291,9 +301,9 @@ sst_buf_write(sst_t *dev, uint32_t addr, uint8_t val)
 }
 
 static void
-sst_write(uint32_t addr, uint8_t val, void *p)
+sst_write(uint32_t addr, uint8_t val, void *priv)
 {
-    sst_t *dev = (sst_t *) p;
+    sst_t *dev = (sst_t *) priv;
 
     switch (dev->command_state) {
         case 0:
@@ -359,19 +369,22 @@ sst_write(uint32_t addr, uint8_t val, void *p)
                 dev->bbp_last_8k = 0xff;
             dev->command_state = 0;
             break;
+
+        default:
+            break;
     }
 }
 
 static uint8_t
-sst_read(uint32_t addr, void *p)
+sst_read(uint32_t addr, void *priv)
 {
-    sst_t  *dev = (sst_t *) p;
+    sst_t  *dev = (sst_t *) priv;
     uint8_t ret = 0xff;
 
     addr &= 0x000fffff;
 
     if (dev->id_mode)
-        ret = sst_read_id(addr, p);
+        ret = sst_read_id(addr, priv);
     else {
         if ((addr >= biosaddr) && (addr <= (biosaddr + biosmask)))
             ret = dev->array[addr - biosaddr];
@@ -381,15 +394,15 @@ sst_read(uint32_t addr, void *p)
 }
 
 static uint16_t
-sst_readw(uint32_t addr, void *p)
+sst_readw(uint32_t addr, void *priv)
 {
-    sst_t   *dev = (sst_t *) p;
+    sst_t   *dev = (sst_t *) priv;
     uint16_t ret = 0xffff;
 
     addr &= 0x000fffff;
 
     if (dev->id_mode)
-        ret = sst_read(addr, p) | (sst_read(addr + 1, p) << 8);
+        ret = sst_read(addr, priv) | (sst_read(addr + 1, priv) << 8);
     else {
         if ((addr >= biosaddr) && (addr <= (biosaddr + biosmask)))
             ret = *(uint16_t *) &dev->array[addr - biosaddr];
@@ -399,15 +412,15 @@ sst_readw(uint32_t addr, void *p)
 }
 
 static uint32_t
-sst_readl(uint32_t addr, void *p)
+sst_readl(uint32_t addr, void *priv)
 {
-    sst_t   *dev = (sst_t *) p;
+    sst_t   *dev = (sst_t *) priv;
     uint32_t ret = 0xffffffff;
 
     addr &= 0x000fffff;
 
     if (dev->id_mode)
-        ret = sst_readw(addr, p) | (sst_readw(addr + 2, p) << 16);
+        ret = sst_readw(addr, priv) | (sst_readw(addr + 2, priv) << 16);
     else {
         if ((addr >= biosaddr) && (addr <= (biosaddr + biosmask)))
             ret = *(uint32_t *) &dev->array[addr - biosaddr];
@@ -499,10 +512,10 @@ sst_init(const device_t *info)
 }
 
 static void
-sst_close(void *p)
+sst_close(void *priv)
 {
     FILE  *f;
-    sst_t *dev = (sst_t *) p;
+    sst_t *dev = (sst_t *) priv;
 
     if (dev->dirty) {
         f = nvr_fopen(flash_path, "wb");
diff --git a/src/network/CMakeLists.txt b/src/network/CMakeLists.txt
index 54ec1edce..9f6d49a17 100644
--- a/src/network/CMakeLists.txt
+++ b/src/network/CMakeLists.txt
@@ -16,7 +16,7 @@ set(net_sources)
 list(APPEND net_sources network.c net_pcap.c net_slirp.c net_dp8390.c net_3c501.c
     net_3c503.c net_ne2000.c net_pcnet.c net_wd8003.c net_plip.c net_event.c net_null.c)
 
-option(SLIRP_EXTERNAL "Link against the system-provided libslirp library" OFF)
+option(SLIRP_EXTERNAL "Link against the system-provided libslirp library" ON)
 mark_as_advanced(SLIRP_EXTERNAL)
 
 if(SLIRP_EXTERNAL)
diff --git a/src/nmi.c b/src/nmi.c
index 4fde00765..fbe3bee19 100644
--- a/src/nmi.c
+++ b/src/nmi.c
@@ -7,11 +7,12 @@
 #include <wchar.h>
 #include <86box/io.h>
 #include <86box/nmi.h>
+#include <86box/plat_unused.h>
 
 int nmi_mask;
 
 void
-nmi_write(uint16_t port, uint8_t val, void *p)
+nmi_write(UNUSED(uint16_t port), uint8_t val, UNUSED(void *priv))
 {
     nmi_mask = val & 0x80;
 }
diff --git a/src/nvr_at.c b/src/nvr_at.c
index 52bc7c9b5..7d0b18ae4 100644
--- a/src/nvr_at.c
+++ b/src/nvr_at.c
@@ -297,23 +297,30 @@
 #define FLAG_P6RP4_HACK    0x10
 #define FLAG_PIIX4         0x20
 
-typedef struct {
+typedef struct local_t {
     int8_t stat;
 
-    uint8_t cent, def,
-        flags, read_addr,
-        wp_0d, wp_32,
-        pad, pad0;
+    uint8_t cent;
+    uint8_t def;
+    uint8_t flags;
+    uint8_t read_addr;
+    uint8_t wp_0d;
+    uint8_t wp_32;
+    uint8_t pad;
+    uint8_t pad0;
 
-    uint8_t addr[8], wp[2],
-        bank[8], *lock;
+    uint8_t  addr[8];
+    uint8_t  wp[2];
+    uint8_t  bank[8];
+    uint8_t *lock;
 
-    int16_t count, state;
+    int16_t count;
+    int16_t state;
 
-    uint64_t ecount,
-        rtc_time;
-    pc_timer_t update_timer,
-        rtc_timer;
+    uint64_t   ecount;
+    uint64_t   rtc_time;
+    pc_timer_t update_timer;
+    pc_timer_t rtc_timer;
 } local_t;
 
 static uint8_t nvr_at_inited = 0;
@@ -650,8 +657,10 @@ nvr_write(uint16_t addr, uint8_t val, void *priv)
         return;
 
     if (addr & 1) {
-        // if (local->bank[addr_id] == 0xff)
-        // return;
+#if 0
+        if (local->bank[addr_id] == 0xff)
+            return;
+#endif
         nvr_reg_write(local->addr[addr_id], val, priv);
     } else {
         local->addr[addr_id] = (val & (nvr->size - 1));
@@ -1065,6 +1074,9 @@ nvr_at_init(const device_t *info)
             nvr->irq    = -1;
             local->cent = RTC_CENTURY_ELT;
             break;
+
+        default:
+            break;
     }
 
     local->read_addr = 1;
diff --git a/src/nvr_ps2.c b/src/nvr_ps2.c
index dd87cbb80..a06ffb7bb 100644
--- a/src/nvr_ps2.c
+++ b/src/nvr_ps2.c
@@ -49,7 +49,7 @@
 #include <86box/nvr_ps2.h>
 #include <86box/rom.h>
 
-typedef struct {
+typedef struct ps2_nvr_t {
     int addr;
 
     uint8_t *ram;
@@ -76,6 +76,9 @@ ps2_nvr_read(uint16_t port, void *priv)
         case 0x76:
             ret = nvr->ram[nvr->addr];
             break;
+
+        default:
+            break;
     }
 
     return ret;
@@ -98,6 +101,9 @@ ps2_nvr_write(uint16_t port, uint8_t val, void *priv)
         case 0x76:
             nvr->ram[nvr->addr] = val;
             break;
+
+        default:
+            break;
     }
 }
 
diff --git a/src/pci.c b/src/pci.c
index 5b6a40138..5b4282c5a 100644
--- a/src/pci.c
+++ b/src/pci.c
@@ -34,9 +34,12 @@
 #include <86box/dma.h>
 #include <86box/pci.h>
 #include <86box/keyboard.h>
+#include <86box/plat_unused.h>
 
-typedef struct {
-    uint8_t bus, id, type;
+typedef struct pci_card_t {
+    uint8_t bus;
+    uint8_t id;
+    uint8_t type;
     uint8_t irq_routing[4];
 
     void *priv;
@@ -44,7 +47,7 @@ typedef struct {
     uint8_t (*read)(int func, int addr, void *priv);
 } pci_card_t;
 
-typedef struct {
+typedef struct pci_mirq_t {
     uint8_t enabled;
     uint8_t irq_line;
 } pci_mirq_t;
@@ -53,6 +56,10 @@ int pci_burst_time;
 int agp_burst_time;
 int pci_nonburst_time;
 int agp_nonburst_time;
+int pci_take_over_io;
+
+uint32_t pci_base = 0xc000;
+uint32_t pci_size = 0x1000;
 
 static pci_card_t pci_cards[32];
 static uint8_t    pci_pmc = 0;
@@ -74,8 +81,6 @@ static int        pci_bus;
 static int        pci_enable;
 static int        pci_key;
 static int        trc_reg = 0;
-static uint32_t   pci_base = 0xc000;
-static uint32_t   pci_size = 0x1000;
 
 static void pci_reset_regs(void);
 
@@ -141,7 +146,7 @@ pci_relocate_slot(int type, int new_slot)
 }
 
 static void
-pci_cf8_write(uint16_t port, uint32_t val, void *priv)
+pci_cf8_write(UNUSED(uint16_t port), uint32_t val, UNUSED(void *priv))
 {
     pci_log("cf8 write: %08X\n", val);
     pci_index  = val & 0xff;
@@ -152,13 +157,13 @@ pci_cf8_write(uint16_t port, uint32_t val, void *priv)
 }
 
 static uint32_t
-pci_cf8_read(uint16_t port, void *priv)
+pci_cf8_read(UNUSED(uint16_t port), UNUSED(void *priv))
 {
     return pci_index | (pci_func << 8) | (pci_card << 11) | (pci_bus << 16) | (pci_enable << 31);
 }
 
 static void
-pci_write(uint16_t port, uint8_t val, void *priv)
+pci_write(uint16_t port, uint8_t val, UNUSED(void *priv))
 {
     uint8_t slot = 0;
 
@@ -191,11 +196,14 @@ pci_write(uint16_t port, uint8_t val, void *priv)
 #endif
 
             break;
+
+        default:
+            break;
     }
 }
 
 static void
-pci_writew(uint16_t port, uint16_t val, void *priv)
+pci_writew(uint16_t port, uint16_t val, UNUSED(void *priv))
 {
     uint8_t slot = 0;
 
@@ -229,11 +237,14 @@ pci_writew(uint16_t port, uint16_t val, void *priv)
 #endif
 
             break;
+
+        default:
+            break;
     }
 }
 
 static void
-pci_writel(uint16_t port, uint32_t val, void *priv)
+pci_writel(uint16_t port, uint32_t val, UNUSED(void *priv))
 {
     uint8_t slot = 0;
 
@@ -269,11 +280,14 @@ pci_writel(uint16_t port, uint32_t val, void *priv)
 #endif
 
             break;
+
+        default:
+            break;
     }
 }
 
 static uint8_t
-pci_read(uint16_t port, void *priv)
+pci_read(uint16_t port, UNUSED(void *priv))
 {
     uint8_t slot = 0;
     uint8_t ret  = 0xff;
@@ -302,6 +316,10 @@ pci_read(uint16_t port, void *priv)
             else
                 pci_log("Reading from unasisgned PCI card on slot %02X (pci_cards[%i]) (%02X:%02X)...\n", pci_card, slot, pci_func, pci_index | (port & 3));
 #endif
+            break;
+
+        default:
+            break;
     }
 
     pci_log("Reading %02X, from PCI card on bus %i, slot %02X (pci_cards[%i]) (%02X:%02X)...\n", ret, pci_bus, pci_card, slot, pci_func, pci_index | (port & 3));
@@ -310,7 +328,7 @@ pci_read(uint16_t port, void *priv)
 }
 
 static uint16_t
-pci_readw(uint16_t port, void *priv)
+pci_readw(uint16_t port, UNUSED(void *priv))
 {
     uint8_t  slot = 0;
     uint16_t ret  = 0xffff;
@@ -341,6 +359,10 @@ pci_readw(uint16_t port, void *priv)
             else
                 pci_log("Reading from unasisgned PCI card on slot %02X (pci_cards[%i]) (%02X:%02X)...\n", pci_card, slot, pci_func, pci_index | (port & 3));
 #endif
+            break;
+
+        default:
+            break;
     }
 
     pci_log("Reading %04X, from PCI card on bus %i, slot %02X (pci_cards[%i]) (%02X:%02X)...\n", ret, pci_bus, pci_card, slot, pci_func, pci_index | (port & 3));
@@ -349,7 +371,7 @@ pci_readw(uint16_t port, void *priv)
 }
 
 static uint32_t
-pci_readl(uint16_t port, void *priv)
+pci_readl(uint16_t port, UNUSED(void *priv))
 {
     uint8_t  slot = 0;
     uint32_t ret  = 0xffffffff;
@@ -382,6 +404,10 @@ pci_readl(uint16_t port, void *priv)
             else
                 pci_log("Reading from unasisgned PCI card on slot %02X (pci_cards[%i]) (%02X:%02X)...\n", pci_card, slot, pci_func, pci_index | (port & 3));
 #endif
+            break;
+
+        default:
+            break;
     }
 
     pci_log("Reading %08X, from PCI card on bus %i, slot %02X (pci_cards[%i]) (%02X:%02X)...\n", ret, pci_bus, pci_card, slot, pci_func, pci_index | (port & 3));
@@ -389,19 +415,17 @@ pci_readl(uint16_t port, void *priv)
     return ret;
 }
 
-static void    pci_type2_write(uint16_t port, uint8_t val, void *priv);
-static uint8_t pci_type2_read(uint16_t port, void *priv);
-
 void
 pci_set_pmc(uint8_t pmc)
 {
     pci_log("pci_set_pmc(%02X)\n", pmc);
-    // pci_reset_regs();
+#if 0
+    pci_reset_regs();
+#endif
 
     if (!pci_pmc && (pmc & 0x01)) {
-        io_removehandler(pci_base, pci_size,
-                         pci_type2_read, NULL, NULL,
-                         pci_type2_write, NULL, NULL, NULL);
+        pci_log("PMC: Dellocating ports %04X-%04X...\n", pci_base, pci_base + pci_size - 1);
+        pci_take_over_io &= ~PCI_IO_ON;
 
         io_removehandler(0x0cf8, 1,
                          pci_type2_read, NULL, NULL, pci_type2_write, NULL, NULL, NULL);
@@ -414,13 +438,11 @@ pci_set_pmc(uint8_t pmc)
         io_sethandler(0x0cfc, 4,
                       pci_read, pci_readw, pci_readl, pci_write, pci_writew, pci_writel, NULL);
     } else if (pci_pmc && !(pmc & 0x01)) {
-        io_removehandler(pci_base, pci_size,
-                         pci_type2_read, NULL, NULL,
-                         pci_type2_write, NULL, NULL, NULL);
+        pci_log("PMC: Redellocating ports %04X-%04X...\n", pci_base, pci_base + pci_size - 1);
+        pci_take_over_io &= ~PCI_IO_ON;
         if (pci_key) {
-            io_sethandler(pci_base, pci_size,
-                          pci_type2_read, NULL, NULL,
-                          pci_type2_write, NULL, NULL, NULL);
+            pci_log("PMC: Allocating ports %04X-%04X...\n", pci_base, pci_base + pci_size - 1);
+            pci_take_over_io |= PCI_IO_ON;
         }
 
         io_removehandler(0x0cf8, 1,
@@ -439,67 +461,97 @@ pci_set_pmc(uint8_t pmc)
 }
 
 static void
-pci_type2_write(uint16_t port, uint8_t val, void *priv)
+pci_type2_write_reg(uint16_t port, uint8_t val)
 {
     uint8_t slot = 0;
 
-    if (port == 0xcf8) {
-        pci_func = (val >> 1) & 7;
+    pci_card  = (port >> 8) & 0xf;
+    pci_index = port & 0xff;
 
-        if (!pci_key && (val & 0xf0)) {
-            io_removehandler(pci_base, pci_size,
-                             pci_type2_read, NULL, NULL,
-                             pci_type2_write, NULL, NULL, NULL);
-            io_sethandler(pci_base, pci_size,
-                          pci_type2_read, NULL, NULL,
-                          pci_type2_write, NULL, NULL, NULL);
-        } else if (pci_key && !(val & 0xf0))
-            io_removehandler(pci_base, pci_size,
-                             pci_type2_read, NULL, NULL,
-                             pci_type2_write, NULL, NULL, NULL);
-
-        pci_key = val & 0xf0;
-    } else if (port == 0xcfa) {
-        pci_bus = val;
-
-        pci_log("Allocating ports %04X-%04X...\n", pci_base, pci_base + pci_size - 1);
-
-        /* Evidently, writing here, we should also enable the
-           configuration space. */
-        io_removehandler(pci_base, pci_size,
-                         pci_type2_read, NULL, NULL,
-                         pci_type2_write, NULL, NULL, NULL);
-        io_sethandler(pci_base, pci_size,
-                      pci_type2_read, NULL, NULL,
-                      pci_type2_write, NULL, NULL, NULL);
-
-        /* Mark as enabled. */
-        pci_key |= 0x100;
-    } else if (port == 0xcfb) {
-        pci_log("Write %02X to port 0CFB\n", val);
-        pci_set_pmc(val);
-    } else {
-        pci_card  = (port >> 8) & 0xf;
-        pci_index = port & 0xff;
-
-        slot = pci_card_to_slot_mapping[pci_bus_number_to_index_mapping[pci_bus]][pci_card];
-        if (slot != 0xff) {
-            if (pci_cards[slot].write)
-                pci_cards[slot].write(pci_func, pci_index | (port & 3), val, pci_cards[slot].priv);
-#ifdef ENABLE_PCI_LOG
-            else
-                pci_log("Writing to empty PCI card on slot %02X (pci_cards[%i]) (%02X:%02X)...\n", pci_card, slot, pci_func, pci_index);
-#endif
-        }
+    slot = pci_card_to_slot_mapping[pci_bus_number_to_index_mapping[pci_bus]][pci_card];
+    if (slot != 0xff) {
+        if (pci_cards[slot].write)
+            pci_cards[slot].write(pci_func, pci_index | (port & 3), val, pci_cards[slot].priv);
 #ifdef ENABLE_PCI_LOG
         else
-            pci_log("Writing to unassigned PCI card on slot %02X (pci_cards[%i]) (%02X:%02X)...\n", pci_card, slot, pci_func, pci_index);
+            pci_log("Writing to empty PCI card on slot %02X:%02X (pci_cards[%i]) (%02X:%02X)...\n", pci_bus, pci_card, slot, pci_func, pci_index);
 #endif
     }
+#ifdef ENABLE_PCI_LOG
+    else
+        pci_log("Writing to unassigned PCI card on slot %02X:%02X (pci_cards[%i]) (%02X:%02X)...\n", pci_bus, pci_card, slot, pci_func, pci_index);
+#endif
+}
+
+void
+pci_type2_write(uint16_t port, uint8_t val, UNUSED(void *priv))
+{
+    switch (port) {
+        case 0xcf8:
+            pci_func = (val >> 1) & 7;
+
+            if (val & 0xf0) {
+                pci_log("CF8: Allocating ports %04X-%04X...\n", pci_base, pci_base + pci_size - 1);
+                pci_take_over_io |= PCI_IO_ON;
+            } else {
+                pci_log("CF8: Dellocating ports %04X-%04X...\n", pci_base, pci_base + pci_size - 1);
+                pci_take_over_io &= ~PCI_IO_ON;
+            }
+
+            pci_key = val & 0xf0;
+            break;
+        case 0xcfa:
+            pci_bus = val;
+
+            pci_log("CFA: Allocating ports %04X-%04X...\n", pci_base, pci_base + pci_size - 1);
+
+            /* Evidently, writing here, we should also enable the
+               configuration space. */
+            pci_take_over_io |= PCI_IO_ON;
+
+            /* Mark as enabled. */
+            pci_key |= 0x100;
+            break;
+        case 0xcfb:
+            pci_log("Write %02X to port 0CFB\n", val);
+            pci_set_pmc(val);
+            break;
+
+        case 0xc000 ... 0xc0ff:
+            if (pci_take_over_io == 0x00000000)
+                break;
+
+            pci_type2_write_reg(port, val);
+            break;
+
+        case 0xc100 ... 0xcfff:
+            if (!(pci_take_over_io & PCI_IO_ON))
+                break;
+
+            pci_type2_write_reg(port, val);
+            break;
+
+        default:
+            break;
+    }
 }
 
-static void
+void
+pci_type2_writew(uint16_t port, uint16_t val, void *priv)
+{
+    pci_type2_write(port, val & 0xff, priv);
+    pci_type2_write(port + 1, val >> 8, priv);
+}
+
+void
 pci_type2_writel(uint16_t port, uint32_t val, void *priv)
+{
+    pci_type2_writew(port, val & 0xffff, priv);
+    pci_type2_writew(port + 2, val >> 16, priv);
+}
+
+static void
+pci_type2_cfb_writel(uint16_t port, uint32_t val, void *priv)
 {
     for (uint8_t i = 0; i < 4; i++) {
         /* Make sure to have the DWORD write not pass through to PMC if mechanism 1 is in use,
@@ -510,41 +562,88 @@ pci_type2_writel(uint16_t port, uint32_t val, void *priv)
 }
 
 static uint8_t
-pci_type2_read(uint16_t port, void *priv)
+pci_type2_read_reg(uint16_t port)
 {
     uint8_t slot = 0;
     uint8_t ret  = 0xff;
 
-    if (port == 0xcf8)
-        ret = pci_key | (pci_func << 1);
-    else if (port == 0xcfa)
-        ret = pci_bus;
-    else if (port == 0xcfb)
-        ret = pci_pmc;
-    else {
-        pci_card  = (port >> 8) & 0xf;
-        pci_index = port & 0xff;
+    pci_card  = (port >> 8) & 0xf;
+    pci_index = port & 0xff;
 
-        slot = pci_card_to_slot_mapping[pci_bus_number_to_index_mapping[pci_bus]][pci_card];
-        if (slot != 0xff) {
-            if (pci_cards[slot].read)
-                ret = pci_cards[slot].read(pci_func, pci_index | (port & 3), pci_cards[slot].priv);
-#ifdef ENABLE_PCI_LOG
-            else
-                pci_log("Reading from empty PCI card on slot %02X (pci_cards[%i]) (%02X:%02X)...\n", pci_card, slot, pci_func, pci_index);
-#endif
-        }
+    slot = pci_card_to_slot_mapping[pci_bus_number_to_index_mapping[pci_bus]][pci_card];
+    if (slot != 0xff) {
+        if (pci_cards[slot].read)
+            ret = pci_cards[slot].read(pci_func, pci_index | (port & 3), pci_cards[slot].priv);
 #ifdef ENABLE_PCI_LOG
         else
-            pci_log("Reading from unasisgned PCI card on slot %02X (pci_cards[%i]) (%02X:%02X)...\n", pci_card, slot, pci_func, pci_index);
+            pci_log("Reading from empty PCI card on slot %02X:%02X (pci_cards[%i]) (%02X:%02X)...\n", pci_bus, pci_card, slot, pci_func, pci_index);
+#endif
+    }
+#ifdef ENABLE_PCI_LOG
+    else
+        pci_log("Reading from unasisgned PCI card on slot %02X:%02X (pci_cards[%i]) (%02X:%02X)...\n", pci_bus, pci_card, slot, pci_func, pci_index);
 #endif
 
-        pci_log("Reading %02X at PCI register %02X at bus %02X, card %02X, function %02X\n", ret, pci_index, pci_bus, pci_card, pci_func);
+    pci_log("Reading %02X at PCI register %02X at bus %02X, card %02X, function %02X\n", ret, pci_index, pci_bus, pci_card, pci_func);
+
+    return ret;
+}
+
+uint8_t
+pci_type2_read(uint16_t port, UNUSED(void *priv))
+{
+    uint8_t ret  = 0xff;
+
+    switch (port) {
+        case 0xcf8:
+            ret = pci_key | (pci_func << 1);
+            break;
+        case 0xcfa:
+            ret = pci_bus;
+            break;
+       case 0xcfb:
+            ret = pci_pmc;
+            break;
+
+        case 0xc000 ... 0xc0ff:
+            if (pci_take_over_io == 0x00000000)
+                break;
+
+            ret = pci_type2_read_reg(port);
+            break;
+
+        case 0xc100 ... 0xcfff:
+            if (!(pci_take_over_io & PCI_IO_ON))
+                break;
+
+            ret = pci_type2_read_reg(port);
+            break;
+
+        default:
+            break;
     }
 
     return ret;
 }
 
+uint16_t
+pci_type2_readw(uint16_t port, void *priv)
+{
+    uint16_t ret = pci_type2_read(port, priv);
+    ret |= ((uint16_t) pci_type2_read(port + 1, priv)) << 8;
+
+    return ret;
+}
+
+uint32_t
+pci_type2_readl(uint16_t port, void *priv)
+{
+    uint32_t ret = pci_type2_readw(port, priv);
+    ret |= ((uint32_t) pci_type2_readw(port + 2, priv)) << 16;
+
+    return ret;
+}
+
 void
 pci_set_irq_routing(int pci_int, int irq)
 {
@@ -826,9 +925,7 @@ pci_reset_regs(void)
 {
     pci_index = pci_card = pci_func = pci_bus = pci_key = 0;
 
-    io_removehandler(pci_base, pci_size,
-                     pci_type2_read, NULL, NULL,
-                     pci_type2_write, NULL, NULL, NULL);
+    pci_take_over_io &= ~PCI_IO_ON;
 }
 
 void
@@ -899,19 +996,19 @@ pci_slots_clear(void)
 }
 
 uint32_t
-trc_readl(uint16_t port, void *priv)
+trc_readl(UNUSED(uint16_t port), UNUSED(void *priv))
 {
     return 0xffffffff;
 }
 
 uint16_t
-trc_readw(uint16_t port, void *priv)
+trc_readw(UNUSED(uint16_t port), UNUSED(void *priv))
 {
     return 0xffff;
 }
 
 uint8_t
-trc_read(uint16_t port, void *priv)
+trc_read(UNUSED(uint16_t port), UNUSED(void *priv))
 {
     return trc_reg & 0xfb;
 }
@@ -939,17 +1036,19 @@ trc_reset(uint8_t val)
 }
 
 void
-trc_writel(uint16_t port, uint32_t val, void *priv)
+trc_writel(UNUSED(uint16_t port), UNUSED(uint32_t val), UNUSED(void *priv))
 {
+    //
 }
 
 void
-trc_writew(uint16_t port, uint16_t val, void *priv)
+trc_writew(UNUSED(uint16_t port), UNUSED(uint16_t val), UNUSED(void *priv))
 {
+    //
 }
 
 void
-trc_write(uint16_t port, uint8_t val, void *priv)
+trc_write(UNUSED(uint16_t port), uint8_t val, UNUSED(void *priv))
 {
     pci_log("TRC Write: %02X\n", val);
 
@@ -993,7 +1092,7 @@ pci_init(int type)
         pci_pmc = 0x00;
 
         io_sethandler(0x0cfb, 1,
-                      pci_type2_read, NULL, NULL, pci_type2_write, NULL, pci_type2_writel, NULL);
+                      pci_type2_read, NULL, NULL, pci_type2_write, NULL, pci_type2_cfb_writel, NULL);
     }
 
     if (type & PCI_NO_IRQ_STEERING) {
@@ -1004,6 +1103,8 @@ pci_init(int type)
         pic_elcr_set_enabled(1);
     }
 
+    pci_take_over_io = 0x00000000;
+
     if ((type & PCI_CONFIG_TYPE_MASK) == PCI_CONFIG_TYPE_1) {
         pci_log("PCI: Configuration mechanism #1\n");
         io_sethandler(0x0cf8, 1,
@@ -1024,9 +1125,7 @@ pci_init(int type)
             pci_base = 0xc100;
             pci_size = 0x0f00;
 
-            io_sethandler(0xc000, 0x0100,
-                          pci_type2_read, NULL, NULL,
-                          pci_type2_write, NULL, NULL, NULL);
+            pci_take_over_io |= PCI_IO_DEV0;
         }
     }
 
diff --git a/src/pci_dummy.c b/src/pci_dummy.c
index b85df7ab9..3d437247d 100644
--- a/src/pci_dummy.c
+++ b/src/pci_dummy.c
@@ -9,14 +9,15 @@
 #include <86box/io.h>
 #include <86box/pci.h>
 #include <86box/pci_dummy.h>
+#include <86box/plat_unused.h>
 
-typedef struct
-{
+typedef struct pci_dummy_t {
     uint8_t pci_regs[256];
 
     bar_t pci_bar[2];
 
-    uint8_t card, interrupt_on;
+    uint8_t card;
+    uint8_t interrupt_on;
 } pci_dummy_t;
 
 static void
@@ -29,12 +30,12 @@ pci_dummy_interrupt(int set, pci_dummy_t *dev)
 }
 
 static uint8_t
-pci_dummy_read(uint16_t Port, void *p)
+pci_dummy_read(uint16_t port, void *priv)
 {
-    pci_dummy_t *dev = (pci_dummy_t *) p;
+    pci_dummy_t *dev = (pci_dummy_t *) priv;
     uint8_t      ret = 0xff;
 
-    switch (Port & 0x20) {
+    switch (port & 0x20) {
         case 0x00:
             ret = 0x1a;
             break;
@@ -60,48 +61,54 @@ pci_dummy_read(uint16_t Port, void *p)
                 dev->interrupt_on = 0;
             }
             break;
+
+        default:
+            break;
     }
 
     return ret;
 }
 
 static uint16_t
-pci_dummy_readw(uint16_t Port, void *p)
+pci_dummy_readw(uint16_t port, void *priv)
 {
-    return pci_dummy_read(Port, p);
+    return pci_dummy_read(port, priv);
 }
 
 static uint32_t
-pci_dummy_readl(uint16_t Port, void *p)
+pci_dummy_readl(uint16_t port, void *priv)
 {
-    return pci_dummy_read(Port, p);
+    return pci_dummy_read(port, priv);
 }
 
 static void
-pci_dummy_write(uint16_t Port, uint8_t Val, void *p)
+pci_dummy_write(uint16_t port, UNUSED(uint8_t val), void *priv)
 {
-    pci_dummy_t *dev = (pci_dummy_t *) p;
+    pci_dummy_t *dev = (pci_dummy_t *) priv;
 
-    switch (Port & 0x20) {
+    switch (port & 0x20) {
         case 0x06:
             if (!dev->interrupt_on) {
                 dev->interrupt_on = 1;
                 pci_dummy_interrupt(1, dev);
             }
             break;
+
+        default:
+            break;
     }
 }
 
 static void
-pci_dummy_writew(uint16_t Port, uint16_t Val, void *p)
+pci_dummy_writew(uint16_t port, uint16_t val, void *priv)
 {
-    pci_dummy_write(Port, Val & 0xFF, p);
+    pci_dummy_write(port, val & 0xFF, priv);
 }
 
 static void
-pci_dummy_writel(uint16_t Port, uint32_t Val, void *p)
+pci_dummy_writel(uint16_t port, uint32_t val, void *priv)
 {
-    pci_dummy_write(Port, Val & 0xFF, p);
+    pci_dummy_write(port, val & 0xFF, priv);
 }
 
 static void
@@ -172,7 +179,9 @@ pci_dummy_pci_read(int func, int addr, void *priv)
                 break;
         }
 
-    // pclog("AB0B:071A: PCI_Read(%d, %04X) = %02X\n", func, addr, ret);
+#if 0
+    pclog("AB0B:071A: PCI_Read(%d, %04X) = %02X\n", func, addr, ret);
+#endif
 
     return ret;
 }
@@ -183,7 +192,9 @@ pci_dummy_pci_write(int func, int addr, uint8_t val, void *priv)
     pci_dummy_t *dev = (pci_dummy_t *) priv;
     uint8_t      valxor;
 
-    // pclog("AB0B:071A: PCI_Write(%d, %04X, %02X)\n", func, addr, val);
+#if 0
+    pclog("AB0B:071A: PCI_Write(%d, %04X, %02X)\n", func, addr, val);
+#endif
 
     if (func == 0x00)
         switch (addr) {
@@ -225,6 +236,9 @@ pci_dummy_pci_write(int func, int addr, uint8_t val, void *priv)
                 pclog("AB0B:071A Device %02X: IRQ now: %i\n", dev->card, val);
                 dev->pci_regs[addr] = val;
                 return;
+
+            default:
+                break;
         }
 }
 
@@ -252,7 +266,7 @@ pci_dummy_close(void *priv)
 }
 
 static void *
-pci_dummy_card_init(const device_t *info)
+pci_dummy_card_init(UNUSED(const device_t *info))
 {
     pci_dummy_t *dev = (pci_dummy_t *) calloc(1, sizeof(pci_dummy_t));
 
diff --git a/src/pic.c b/src/pic.c
index ffb370ce2..8a57d15b9 100644
--- a/src/pic.c
+++ b/src/pic.c
@@ -37,6 +37,7 @@
 #include <86box/apm.h>
 #include <86box/nvr.h>
 #include <86box/acpi.h>
+#include <86box/plat_unused.h>
 
 enum {
     STATE_NONE = 0,
@@ -140,7 +141,7 @@ pic_elcr_write(uint16_t port, uint8_t val, void *priv)
 }
 
 uint8_t
-pic_elcr_read(uint16_t port, void *priv)
+pic_elcr_read(UNUSED(uint16_t port), void *priv)
 {
     pic_t *dev = (pic_t *) priv;
 
@@ -191,8 +192,8 @@ find_best_interrupt(pic_t *dev)
 {
     uint8_t b;
     uint8_t intr;
-    int     j;
-    int     ret = -1;
+    uint8_t j;
+    int8_t  ret = -1;
 
     for (uint8_t i = 0; i < 8; i++) {
         j = (i + dev->priority) & 7;
@@ -334,7 +335,7 @@ pic_acknowledge(pic_t *dev)
 static uint8_t
 pic_non_specific_find(pic_t *dev)
 {
-    int     j;
+    uint8_t j;
     uint8_t b;
     uint8_t irq = 0xff;
 
@@ -398,7 +399,7 @@ pic_command(pic_t *dev)
 }
 
 uint8_t
-pic_latch_read(uint16_t addr, void *priv)
+pic_latch_read(UNUSED(uint16_t addr), UNUSED(void *priv))
 {
     uint8_t ret = 0xff;
 
@@ -495,6 +496,9 @@ pic_write(uint16_t addr, uint8_t val, void *priv)
                 dev->imr = val;
                 update_pending();
                 break;
+
+            default:
+                break;
         }
     } else {
         if (val & 0x10) {
@@ -531,14 +535,12 @@ pic_write(uint16_t addr, uint8_t val, void *priv)
 void
 pic_set_pci(void)
 {
-    int i;
-
-    for (i = 0x0024; i < 0x0040; i += 4) {
+    for (uint8_t i = 0x0024; i < 0x0040; i += 4) {
         io_sethandler(i, 0x0002, pic_read, NULL, NULL, pic_write, NULL, NULL, &pic);
         io_sethandler(i + 0x0080, 0x0002, pic_read, NULL, NULL, pic_write, NULL, NULL, &pic2);
     }
 
-    for (i = 0x1120; i < 0x1140; i += 4) {
+    for (uint16_t i = 0x1120; i < 0x1140; i += 4) {
         io_sethandler(i, 0x0002, pic_read, NULL, NULL, pic_write, NULL, NULL, &pic);
         io_sethandler(i + 0x0080, 0x0002, pic_read, NULL, NULL, pic_write, NULL, NULL, &pic2);
     }
@@ -584,7 +586,7 @@ pic_reset_hard(void)
     /* The situation is as follows: There is a giant mess when it comes to these latches on real hardware,
        to the point that there's even boards with board-level latched that get used in place of the latches
        on the chipset, therefore, I'm just doing this here for the sake of simplicity. */
-    if (machine_has_bus(machine, MACHINE_BUS_PS2)) {
+    if (machine_has_bus(machine, MACHINE_BUS_PS2_LATCH)) {
         pic_kbd_latch(0x01);
         pic_mouse_latch(0x01);
     } else {
@@ -628,7 +630,7 @@ picint_common(uint16_t num, int level, int set)
     /* Make sure to ignore all slave IRQ's, and in case of AT+,
        translate IRQ 2 to IRQ 9. */
     for (uint8_t i = 0; i < 8; i++) {
-        b     = (1 << i);
+        b     = (uint8_t) (1 << i);
         raise = num & b;
 
         if (pic.icw3 & b) {
@@ -779,7 +781,7 @@ pic_irq_ack_read(pic_t *dev, int phase)
 uint8_t
 pic_irq_ack(void)
 {
-    int ret;
+    uint8_t ret;
 
     /* Needed for Xi8088. */
     if ((pic.ack_bytes == 0) && pic.int_pending && pic_slave_on(&pic, pic.interrupt)) {
diff --git a/src/pit.c b/src/pit.c
index cf339607d..3fe959c7f 100644
--- a/src/pit.c
+++ b/src/pit.c
@@ -41,6 +41,7 @@
 #include <86box/sound.h>
 #include <86box/snd_speaker.h>
 #include <86box/video.h>
+#include <86box/plat_unused.h>
 
 pit_intf_t pit_devs[2];
 
@@ -171,6 +172,9 @@ ctr_tick(ctr_t *ctr)
                 case 3:
                     ctr_decrease_count(ctr);
                     break;
+
+                default:
+                    break;
             }
             break;
         case 1:
@@ -189,6 +193,9 @@ ctr_tick(ctr_t *ctr)
                 case 6:
                     ctr_decrease_count(ctr);
                     break;
+
+                default:
+                    break;
             }
             break;
         case 2:
@@ -211,6 +218,9 @@ ctr_tick(ctr_t *ctr)
                         }
                     }
                     break;
+
+                default:
+                    break;
             }
             break;
         case 3:
@@ -254,6 +264,9 @@ ctr_tick(ctr_t *ctr)
                             ctr->newcount = 0;
                     }
                     break;
+
+                default:
+                    break;
             }
             break;
         case 4:
@@ -279,6 +292,9 @@ ctr_tick(ctr_t *ctr)
                         ctr->state = 0;
                         ctr_set_out(ctr, 1);
                         break;
+
+                    default:
+                        break;
                 }
             }
             break;
@@ -369,6 +385,9 @@ ctr_latch_count(ctr_t *ctr)
             ctr->rl      = count;
             ctr->latched = 2;
             break;
+
+        default:
+            break;
     }
 
     pit_log("latched counter = %04X\n", ctr->rl & 0xffff);
@@ -439,6 +458,9 @@ pit_ctr_set_gate(void *data, int counter_id, int gate)
                     ctr_set_out(ctr, 1);
             }
             break;
+
+        default:
+            break;
     }
 }
 
@@ -487,9 +509,9 @@ pit_ctr_set_using_timer(void *data, int counter_id, int using_timer)
 }
 
 static void
-pit_timer_over(void *p)
+pit_timer_over(void *priv)
 {
-    pit_t *dev = (pit_t *) p;
+    pit_t *dev = (pit_t *) priv;
 
     dev->clock ^= 1;
 
@@ -602,8 +624,14 @@ pit_write(uint16_t addr, uint8_t val, void *priv)
                     else
                         ctr->wm |= 0x80;
                     break;
+
+                default:
+                    break;
             }
             break;
+
+        default:
+            break;
     }
 }
 
@@ -675,8 +703,14 @@ pit_read(uint16_t addr, void *priv)
                         else
                             ctr->rm |= 0x80;
                         break;
+
+                    default:
+                        break;
                 }
             break;
+
+        default:
+            break;
     }
 
     pit_log("[%04X:%08X] pit_read(%04X, %08X) = %02X\n", CS, cpu_state.pc, addr, priv, ret);
@@ -714,7 +748,7 @@ pit_refresh_timer_at(int new_out, int old_out)
 }
 
 void
-pit_speaker_timer(int new_out, int old_out)
+pit_speaker_timer(int new_out, UNUSED(int old_out))
 {
     int l;
 
@@ -734,7 +768,7 @@ pit_speaker_timer(int new_out, int old_out)
 }
 
 void
-pit_nmi_timer_ps2(int new_out, int old_out)
+pit_nmi_timer_ps2(int new_out, UNUSED(int old_out))
 {
     nmi = new_out;
 
@@ -893,8 +927,8 @@ pit_common_init(int type, void (*out0)(int new_out, int old_out), void (*out1)(i
     pit_intf_t *pit_intf = &pit_devs[0];
 
     switch (type) {
-        case PIT_8253:
         default:
+        case PIT_8253:
             pit       = device_add(&i8253_device);
             *pit_intf = pit_classic_intf;
             break;
@@ -937,8 +971,8 @@ pit_ps2_init(int type)
     pit_intf_t *ps2_pit = &pit_devs[1];
 
     switch (type) {
-        case PIT_8254:
         default:
+        case PIT_8254:
             pit      = device_add(&i8254_ps2_device);
             *ps2_pit = pit_classic_intf;
             break;
diff --git a/src/pit_fast.c b/src/pit_fast.c
index 4e7bc77c7..1bd75f2aa 100644
--- a/src/pit_fast.c
+++ b/src/pit_fast.c
@@ -207,6 +207,9 @@ pitf_ctr_load(ctrf_t *ctr)
         case 5: /*Hardware triggered stobe*/
             ctr->enabled = 1;
             break;
+
+        default:
+            break;
     }
 
     if (ctr->load_func != NULL)
@@ -266,6 +269,9 @@ pitf_set_gate_no_timer(ctrf_t *ctr, int gate)
             }
             ctr->enabled = gate;
             break;
+
+        default:
+            break;
     }
     ctr->gate    = gate;
     ctr->running = ctr->enabled && ctr->using_timer && !ctr->disabled;
@@ -327,7 +333,10 @@ pitf_over(ctrf_t *ctr)
                 if (ctr->using_timer)
                     timer_advance_u64(&ctr->timer, (uint64_t) (((l + 1) >> 1) * PITCONST));
             }
-            //                if (!t) pclog("pit_over: square wave mode c=%x  %lli  %f\n", pit.c[t], tsc, PITCONST);
+#if 0
+            if (!t)
+                pclog("pit_over: square wave mode c=%x  %lli  %f\n", pit.c[t], tsc, PITCONST);
+#endif
             break;
         case 4: /*Software triggered strove*/
             if (!ctr->thit) {
@@ -356,6 +365,9 @@ pitf_over(ctrf_t *ctr)
             if (ctr->using_timer)
                 timer_advance_u64(&ctr->timer, (uint64_t) (0xffff * PITCONST));
             break;
+
+        default:
+            break;
     }
     ctr->running = ctr->enabled && ctr->using_timer && !ctr->disabled;
     if (ctr->using_timer && !ctr->running)
@@ -366,8 +378,10 @@ static __inline void
 pitf_ctr_latch_count(ctrf_t *ctr)
 {
     ctr->rl = pitf_read_timer(ctr);
-    //                        pclog("Timer latch %f %04X %04X\n",pit->c[0],pit->rl[0],pit->l[0]);
-    // pit->ctrl |= 0x30;
+#if 0
+    pclog("Timer latch %f %04X %04X\n",pit->c[0],pit->rl[0],pit->l[0]);
+    pit->ctrl |= 0x30;
+#endif
     ctr->rereadlatch = 0;
     ctr->rm          = 3;
     ctr->latched     = 1;
@@ -472,8 +486,14 @@ pitf_write(uint16_t addr, uint8_t val, void *priv)
                     ctr->l |= val;
                     ctr->wm = 0;
                     break;
+
+                default:
+                    break;
             }
             break;
+
+        default:
+            break;
     }
 }
 
@@ -530,8 +550,14 @@ pitf_read(uint16_t addr, void *priv)
                     else
                         ctr->rm = 0;
                     break;
+
+                default:
+                    break;
             }
             break;
+
+        default:
+            break;
     }
 
     pit_log("[%04X:%08X] pit_read(%04X, %08X) = %02X\n", CS, cpu_state.pc, addr, priv, ret);
@@ -540,9 +566,9 @@ pitf_read(uint16_t addr, void *priv)
 }
 
 static void
-pitf_timer_over(void *p)
+pitf_timer_over(void *priv)
 {
-    ctrf_t *ctr = (ctrf_t *) p;
+    ctrf_t *ctr = (ctrf_t *) priv;
     pitf_over(ctr);
 }
 
diff --git a/src/port_6x.c b/src/port_6x.c
index 2bfebcee0..323d8e0c8 100644
--- a/src/port_6x.c
+++ b/src/port_6x.c
@@ -36,6 +36,7 @@
 #include <86box/ppi.h>
 #include <86box/video.h>
 #include <86box/port_6x.h>
+#include <86box/plat_unused.h>
 
 #define PS2_REFRESH_TIME (16 * TIMER_USEC)
 
@@ -68,11 +69,14 @@ port_6x_write(uint16_t port, uint8_t val, void *priv)
             if (dev->flags & PORT_6X_TURBO)
                 xi8088_turbo_set(!!(val & 0x04));
             break;
+
+        default:
+            break;
     }
 }
 
 static uint8_t
-port_61_read_simple(uint16_t port, void *priv)
+port_61_read_simple(UNUSED(uint16_t port), UNUSED(void *priv))
 {
     uint8_t ret = ppi.pb & 0x1f;
 
@@ -83,7 +87,7 @@ port_61_read_simple(uint16_t port, void *priv)
 }
 
 static uint8_t
-port_61_read(uint16_t port, void *priv)
+port_61_read(UNUSED(uint16_t port), void *priv)
 {
     port_6x_t *dev = (port_6x_t *) priv;
     uint8_t    ret = 0xff;
@@ -106,7 +110,7 @@ port_61_read(uint16_t port, void *priv)
 }
 
 static uint8_t
-port_62_read(uint16_t port, void *priv)
+port_62_read(UNUSED(uint16_t port), UNUSED(void *priv))
 {
     uint8_t ret = 0xff;
 
diff --git a/src/port_92.c b/src/port_92.c
index fd0471b0e..80afe9e58 100644
--- a/src/port_92.c
+++ b/src/port_92.c
@@ -29,6 +29,7 @@
 #include <86box/mem.h>
 #include <86box/pit.h>
 #include <86box/port_92.h>
+#include <86box/plat_unused.h>
 
 #define PORT_92_INV   1
 #define PORT_92_WORD  2
@@ -74,7 +75,7 @@ port_92_readw(uint16_t port, void *priv)
    TODO: ALi M1543(c) behavior.
  */
 static void
-port_92_pulse(void *priv)
+port_92_pulse(UNUSED(void *priv))
 {
     softresetx86(); /* Pulse reset! */
     cpu_set_edx();
@@ -174,7 +175,7 @@ port_92_remove(void *priv)
 }
 
 static void
-port_92_reset(void *priv)
+port_92_reset(UNUSED(void *priv))
 {
     cpu_alt_reset = 0;
 
diff --git a/src/printer/CMakeLists.txt b/src/printer/CMakeLists.txt
index c774258e2..dea0c7fbe 100644
--- a/src/printer/CMakeLists.txt
+++ b/src/printer/CMakeLists.txt
@@ -20,4 +20,15 @@ if(APPLE)
     if (NOT GHOSTSCRIPT_LIB)
         message(WARNING "Could not find ghostscript. The library will not be bundled and any related features will not work.")
     endif()
-endif ()
\ No newline at end of file
+endif()
+
+find_package(PkgConfig REQUIRED)
+pkg_check_modules(FREETYPE REQUIRED IMPORTED_TARGET freetype2)
+target_link_libraries(86Box PkgConfig::FREETYPE)
+if(STATIC_BUILD)
+    if(QT)
+        # Qt provides its own version of harfbuzz which leads to duplicated symbols.
+        target_link_options(86Box PRIVATE "LINKER:--allow-multiple-definition")
+    endif()
+    target_link_libraries(86Box -static ${FREETYPE_STATIC_LIBRARIES})
+endif()
diff --git a/src/printer/png.c b/src/printer/png.c
index 49da1ee18..2134f3343 100644
--- a/src/printer/png.c
+++ b/src/printer/png.c
@@ -92,13 +92,13 @@ png_log(const char *fmt, ...)
 #endif
 
 static void
-error_handler(png_structp arg, const char *str)
+error_handler(UNUSED(png_structp arg), UNUSED(const char *str))
 {
     png_log("PNG: stream 0x%08lx error '%s'\n", arg, str);
 }
 
 static void
-warning_handler(png_structp arg, const char *str)
+warning_handler(UNUSED(png_structp arg), UNUSED(const char *str))
 {
     png_log("PNG: stream 0x%08lx warning '%s'\n", arg, str);
 }
diff --git a/src/printer/prt_cpmap.c b/src/printer/prt_cpmap.c
index ba9e71085..f2916eef8 100644
--- a/src/printer/prt_cpmap.c
+++ b/src/printer/prt_cpmap.c
@@ -572,7 +572,7 @@ static const struct {
 void
 select_codepage(uint16_t code, uint16_t *curmap)
 {
-    int             i = 0;
+    uint16_t        i = 0;
     const uint16_t *map_to_use;
 
     map_to_use = maps[0].map;
diff --git a/src/printer/prt_escp.c b/src/printer/prt_escp.c
index f1697e074..b4099bc63 100644
--- a/src/printer/prt_escp.c
+++ b/src/printer/prt_escp.c
@@ -65,7 +65,6 @@
 #include <86box/pit.h>
 #include <86box/path.h>
 #include <86box/plat.h>
-#include <86box/plat_dynld.h>
 #include <86box/ui.h>
 #include <86box/lpt.h>
 #include <86box/video.h>
@@ -85,45 +84,8 @@
 #define PAGE_CPI     10.0 /* standard 10 cpi */
 #define PAGE_LPI     6.0  /* standard 6 lpi */
 
-#ifdef _WIN32
-#    define PATH_FREETYPE_DLL "freetype.dll"
-#elif defined __APPLE__
-#    define PATH_FREETYPE_DLL "libfreetype.6.dylib"
-#else
-#    define PATH_FREETYPE_DLL "libfreetype.so.6"
-#endif
-
 /* FreeType library handles - global so they can be shared. */
-FT_Library ft_lib    = NULL;
-void      *ft_handle = NULL;
-
-static int (*ft_Init_FreeType)(FT_Library *alibrary);
-static int (*ft_Done_Face)(FT_Face face);
-static int (*ft_New_Face)(FT_Library library, const char *filepathname,
-                          FT_Long face_index, FT_Face *aface);
-static int (*ft_Set_Char_Size)(FT_Face face, FT_F26Dot6 char_width,
-                               FT_F26Dot6 char_height,
-                               FT_UInt    horz_resolution,
-                               FT_UInt    vert_resolution);
-static int (*ft_Set_Transform)(FT_Face face, FT_Matrix *matrix,
-                               FT_Vector *delta);
-static int (*ft_Get_Char_Index)(FT_Face face, FT_ULong charcode);
-static int (*ft_Load_Glyph)(FT_Face face, FT_UInt glyph_index,
-                            FT_Int32 load_flags);
-static int (*ft_Render_Glyph)(FT_GlyphSlot   slot,
-                              FT_Render_Mode render_mode);
-
-static dllimp_t ft_imports[] = {
-    {"FT_Init_FreeType",   &ft_Init_FreeType },
-    { "FT_New_Face",       &ft_New_Face      },
-    { "FT_Done_Face",      &ft_Done_Face     },
-    { "FT_Set_Char_Size",  &ft_Set_Char_Size },
-    { "FT_Set_Transform",  &ft_Set_Transform },
-    { "FT_Get_Char_Index", &ft_Get_Char_Index},
-    { "FT_Load_Glyph",     &ft_Load_Glyph    },
-    { "FT_Render_Glyph",   &ft_Render_Glyph  },
-    { NULL,                NULL              }
-};
+FT_Library ft_lib = NULL;
 
 /* The fonts. */
 #define FONT_DEFAULT   0
@@ -181,7 +143,7 @@ static dllimp_t ft_imports[] = {
 #define PIXX       ((unsigned) floor(dev->curr_x * dev->dpi + 0.5))
 #define PIXY       ((unsigned) floor(dev->curr_y * dev->dpi + 0.5))
 
-typedef struct {
+typedef struct psurface_t {
     int8_t dirty; /* has the page been printed on? */
     char   pad;
 
@@ -192,7 +154,7 @@ typedef struct {
     uint8_t *pixels; /* grayscale pixel data */
 } psurface_t;
 
-typedef struct {
+typedef struct escp_t {
     const char *name;
 
     void *lpt;
@@ -204,12 +166,12 @@ typedef struct {
     uint8_t color;
 
     /* page data (TODO: make configurable) */
-    double page_width, /* all in inches */
-        page_height,
-        left_margin,
-        top_margin,
-        right_margin,
-        bottom_margin;
+    double   page_width; /* all in inches */
+    double   page_height;
+    double   left_margin;
+    double   top_margin;
+    double   right_margin;
+    double   bottom_margin;
     uint16_t dpi;
     double   cpi; /* defined chars per inch */
     double   lpi; /* defined lines per inch */
@@ -255,7 +217,8 @@ typedef struct {
     char        fontpath[1024];
     char        pagepath[1024];
     psurface_t *page;
-    double      curr_x, curr_y; /* print head position (inch) */
+    double      curr_x; /* print head position (x, inch) */
+    double      curr_y; /* print head position (y, inch) */
     uint16_t    current_font;
     FT_Face     fontface;
     int8_t      lq_typeface;
@@ -543,7 +506,7 @@ update_font(escp_t *dev)
 
     /* Release current font if we have one. */
     if (dev->fontface)
-        ft_Done_Face(dev->fontface);
+        FT_Done_Face(dev->fontface);
 
     if (dev->print_quality == QUALITY_DRAFT)
         fn = FONT_FILE_DOTMATRIX;
@@ -579,7 +542,7 @@ update_font(escp_t *dev)
     escp_log("Temp file=%s\n", path);
 
     /* Load the new font. */
-    if (ft_New_Face(ft_lib, path, 0, &dev->fontface)) {
+    if (FT_New_Face(ft_lib, path, 0, &dev->fontface)) {
         escp_log("ESC/P: unable to load font '%s'\n", path);
         dev->fontface = NULL;
     }
@@ -625,7 +588,7 @@ update_font(escp_t *dev)
         dev->actual_cpi /= 2.0 / 3.0;
     }
 
-    ft_Set_Char_Size(dev->fontface,
+    FT_Set_Char_Size(dev->fontface,
                      (uint16_t) (hpoints * 64), (uint16_t) (vpoints * 64),
                      dev->dpi, dev->dpi);
 
@@ -635,7 +598,7 @@ update_font(escp_t *dev)
         matrix.xy = (FT_Fixed) (0.20 * 0x10000L);
         matrix.yx = 0;
         matrix.yy = 0x10000L;
-        ft_Set_Transform(dev->fontface, &matrix, 0);
+        FT_Set_Transform(dev->fontface, &matrix, 0);
     }
 }
 
@@ -1610,9 +1573,9 @@ handle_char(escp_t *dev, uint8_t ch)
 
     /* ok, so we need to print the character now */
     if (ft_lib) {
-        char_index = ft_Get_Char_Index(dev->fontface, dev->curr_cpmap[ch]);
-        ft_Load_Glyph(dev->fontface, char_index, FT_LOAD_DEFAULT);
-        ft_Render_Glyph(dev->fontface->glyph, FT_RENDER_MODE_NORMAL);
+        char_index = FT_Get_Char_Index(dev->fontface, dev->curr_cpmap[ch]);
+        FT_Load_Glyph(dev->fontface, char_index, FT_LOAD_DEFAULT);
+        FT_Render_Glyph(dev->fontface->glyph, FT_RENDER_MODE_NORMAL);
     }
 
     pen_x = PIXX + dev->fontface->glyph->bitmap_left;
@@ -1985,23 +1948,12 @@ read_status(void *priv)
 static void *
 escp_init(void *lpt)
 {
-    const char *fn = PATH_FREETYPE_DLL;
-    escp_t     *dev;
-
-    /* Dynamically load FreeType. */
-    if (ft_handle == NULL) {
-        ft_handle = dynld_module(fn, ft_imports);
-        if (ft_handle == NULL) {
-            ui_msgbox_header(MBX_ERROR, (wchar_t *) IDS_2111, (wchar_t *) IDS_2132);
-            return (NULL);
-        }
-    }
+    escp_t *dev;
 
     /* Initialize FreeType. */
     if (ft_lib == NULL) {
-        if (ft_Init_FreeType(&ft_lib)) {
+        if (FT_Init_FreeType(&ft_lib)) {
             ui_msgbox_header(MBX_ERROR, (wchar_t *) IDS_2111, (wchar_t *) IDS_2132);
-            dynld_close(ft_lib);
             ft_lib = NULL;
             return (NULL);
         }
diff --git a/src/printer/prt_ps.c b/src/printer/prt_ps.c
index 6c72e48de..caa166a10 100644
--- a/src/printer/prt_ps.c
+++ b/src/printer/prt_ps.c
@@ -51,7 +51,9 @@
 #elif defined __APPLE__
 #    define PATH_GHOSTSCRIPT_DLL "libgs.dylib"
 #else
-#    define PATH_GHOSTSCRIPT_DLL "libgs.so.9"
+#    define PATH_GHOSTSCRIPT_DLL      "libgs.so.9"
+#    define PATH_GHOSTSCRIPT_DLL_ALT1 "libgs.so.10"
+#    define PATH_GHOSTSCRIPT_DLL_ALT2 "libgs.so"
 #endif
 
 #define POSTSCRIPT_BUFFER_LENGTH 65536
@@ -236,9 +238,9 @@ timeout_timer(void *priv)
 }
 
 static void
-ps_write_data(uint8_t val, void *p)
+ps_write_data(uint8_t val, void *priv)
 {
-    ps_t *dev = (ps_t *) p;
+    ps_t *dev = (ps_t *) priv;
 
     if (dev == NULL)
         return;
@@ -285,9 +287,9 @@ process_data(ps_t *dev)
 }
 
 static void
-ps_write_ctrl(uint8_t val, void *p)
+ps_write_ctrl(uint8_t val, void *priv)
 {
-    ps_t *dev = (ps_t *) p;
+    ps_t *dev = (ps_t *) priv;
 
     if (dev == NULL)
         return;
@@ -317,9 +319,9 @@ ps_write_ctrl(uint8_t val, void *p)
 }
 
 static uint8_t
-ps_read_status(void *p)
+ps_read_status(void *priv)
 {
-    ps_t   *dev = (ps_t *) p;
+    ps_t   *dev = (ps_t *) priv;
     uint8_t ret = 0x9f;
 
     if (!dev->ack)
@@ -341,12 +343,21 @@ ps_init(void *lpt)
 
     /* Try loading the DLL. */
     ghostscript_handle = dynld_module(PATH_GHOSTSCRIPT_DLL, ghostscript_imports);
-    if (ghostscript_handle == NULL)
+#ifdef PATH_GHOSTSCRIPT_DLL_ALT1
+    if (ghostscript_handle == NULL) {
+        ghostscript_handle = dynld_module(PATH_GHOSTSCRIPT_DLL_ALT1, ghostscript_imports);
+#    ifdef PATH_GHOSTSCRIPT_DLL_ALT2
+        if (ghostscript_handle == NULL)
+            ghostscript_handle = dynld_module(PATH_GHOSTSCRIPT_DLL_ALT2, ghostscript_imports);
+#    endif
+    }
+#endif
+    if (ghostscript_handle == NULL) {
         ui_msgbox_header(MBX_ERROR, (wchar_t *) IDS_2115, (wchar_t *) IDS_2133);
-    else {
-        if (gsapi_revision(&rev, sizeof(rev)) == 0)
+    } else {
+        if (gsapi_revision(&rev, sizeof(rev)) == 0) {
             pclog("Loaded %s, rev %ld (%ld)\n", rev.product, rev.revision, rev.revisiondate);
-        else {
+        } else {
             dynld_close(ghostscript_handle);
             ghostscript_handle = NULL;
         }
@@ -368,9 +379,9 @@ ps_init(void *lpt)
 }
 
 static void
-ps_close(void *p)
+ps_close(void *priv)
 {
-    ps_t *dev = (ps_t *) p;
+    ps_t *dev = (ps_t *) priv;
 
     if (dev == NULL)
         return;
diff --git a/src/printer/prt_text.c b/src/printer/prt_text.c
index 54e5fe6d5..6fe0f531a 100644
--- a/src/printer/prt_text.c
+++ b/src/printer/prt_text.c
@@ -81,7 +81,7 @@
 #define PAGE_CPI 10.0 /* standard 10 cpi */
 #define PAGE_LPI 6.0  /* standard 6 lpi */
 
-typedef struct {
+typedef struct psurface_t {
     int8_t dirty; /* has the page been printed on? */
     char   pad;
 
@@ -91,7 +91,7 @@ typedef struct {
     char *chars; /* character data */
 } psurface_t;
 
-typedef struct {
+typedef struct prnt_t {
     const char *name;
 
     void *lpt;
@@ -104,23 +104,23 @@ typedef struct {
     pc_timer_t timeout_timer;
 
     /* page data (TODO: make configurable) */
-    double page_width, /* all in inches */
-        page_height,
-        left_margin,
-        top_margin,
-        right_margin,
-        bot_margin;
+    double page_width; /* all in inches */
+    double page_height;
+    double left_margin;
+    double top_margin;
+    double right_margin;
+    double bot_margin;
 
     /* internal page data */
     psurface_t *page;
-    uint8_t     max_chars,
-        max_lines;
-    uint8_t curr_x, /* print head position (chars) */
-        curr_y;
+    uint8_t     max_chars;
+    uint8_t     max_lines;
+    uint8_t     curr_x; /* print head position (x, chars) */
+    uint8_t     curr_y; /* print head position (y, chars) */
 
     /* font data */
-    double cpi, /* defined chars per inch */
-        lpi;    /* defined lines per inch */
+    double cpi; /* defined chars per inch */
+    double lpi; /* defined lines per inch */
 
     /* handshake data */
     uint8_t data;
diff --git a/src/qt/cocoa_keyboard.hpp b/src/qt/cocoa_keyboard.hpp
index eaf0cdfe0..da3161bb2 100644
--- a/src/qt/cocoa_keyboard.hpp
+++ b/src/qt/cocoa_keyboard.hpp
@@ -127,3 +127,9 @@ static std::array<uint32_t, 127> cocoa_keycodes = { /* key names in parentheses
     0x150, /* DownArrow */
     0x148, /* UpArrow */
 };
+
+// https://developer.apple.com/documentation/appkit/nseventmodifierflags/
+qint32 NSEventModifierFlagCommand = 1 << 20;
+
+qint32 nvk_Delete = 0x75;
+qint32 nvk_Insert = 0x72;
\ No newline at end of file
diff --git a/src/qt/qt_filefield.cpp b/src/qt/qt_filefield.cpp
index b6db5e611..9272e31a1 100644
--- a/src/qt/qt_filefield.cpp
+++ b/src/qt/qt_filefield.cpp
@@ -31,6 +31,12 @@ FileField::FileField(QWidget *parent)
         fileName_ = ui->label->text();
         emit fileSelected(ui->label->text(), true);
     });
+
+    connect(ui->label, &QLineEdit::textChanged, this, [this]() {
+        fileName_ = ui->label->text();
+        emit fileTextEntered(ui->label->text(), true);
+    });
+
     this->setFixedWidth(this->sizeHint().width() + ui->pushButton->sizeHint().width());
 }
 
diff --git a/src/qt/qt_filefield.hpp b/src/qt/qt_filefield.hpp
index ee011a38b..e3569fd31 100644
--- a/src/qt/qt_filefield.hpp
+++ b/src/qt/qt_filefield.hpp
@@ -19,12 +19,14 @@ public:
 
     void    setFilter(const QString &filter) { filter_ = filter; }
     QString selectedFilter() const { return selectedFilter_; }
+    void    setselectedFilter(const QString &selectedFilter) { selectedFilter_ = selectedFilter; }
 
     void setCreateFile(bool createFile) { createFile_ = createFile; }
     bool createFile() { return createFile_; }
 
 signals:
     void fileSelected(const QString &fileName, bool precheck = false);
+    void fileTextEntered(const QString &fileName, bool precheck = false);
 
 private slots:
     void on_pushButton_clicked();
diff --git a/src/qt/qt_harddiskdialog.cpp b/src/qt/qt_harddiskdialog.cpp
index 6489b9c77..b640c086e 100644
--- a/src/qt/qt_harddiskdialog.cpp
+++ b/src/qt/qt_harddiskdialog.cpp
@@ -84,6 +84,13 @@ HarddiskDialog::HarddiskDialog(bool existing, QWidget *parent)
     ui->lineEditSize->setValidator(new QIntValidator());
     ui->buttonBox->button(QDialogButtonBox::Ok)->setEnabled(false);
 
+    filters = QStringList({ tr("Raw image") % util::DlgFilter({ "img" }, true),
+                          tr("HDI image") % util::DlgFilter({ "hdi" }, true),
+                          tr("HDX image") % util::DlgFilter({ "hdx" }, true),
+                          tr("Fixed-size VHD") % util::DlgFilter({ "vhd" }, true),
+                          tr("Dynamic-size VHD") % util::DlgFilter({ "vhd" }, true),
+                          tr("Differencing VHD") % util::DlgFilter({ "vhd" }, true) });
+
     if (existing) {
         ui->fileField->setFilter(tr("Hard disk images") % util::DlgFilter({ "hd?", "im?", "vhd" }) % tr("All files") % util::DlgFilter({ "*" }, true));
 
@@ -99,24 +106,26 @@ HarddiskDialog::HarddiskDialog(bool existing, QWidget *parent)
 
         connect(ui->fileField, &FileField::fileSelected, this, &HarddiskDialog::onExistingFileSelected);
     } else {
-        QStringList filters({ tr("Raw image") % util::DlgFilter({ "img" }, true),
-                              tr("HDI image") % util::DlgFilter({ "hdi" }, true),
-                              tr("HDX image") % util::DlgFilter({ "hdx" }, true),
-                              tr("Fixed-size VHD") % util::DlgFilter({ "vhd" }, true),
-                              tr("Dynamic-size VHD") % util::DlgFilter({ "vhd" }, true),
-                              tr("Differencing VHD") % util::DlgFilter({ "vhd" }, true) });
-
         ui->fileField->setFilter(filters.join(";;"));
 
         setWindowTitle(tr("Add New Hard Disk"));
         ui->fileField->setCreateFile(true);
 
-        connect(ui->fileField, &FileField::fileSelected, this, [this, filters] {
+        // Enable the OK button as long as the filename length is non-zero
+        connect(ui->fileField, &FileField::fileTextEntered, this, [this] {
+            ui->buttonBox->button(QDialogButtonBox::Ok)->setEnabled((this->fileName().length() > 0));
+        });
+
+        connect(ui->fileField, &FileField::fileSelected, this, [this] {
             int filter = filters.indexOf(ui->fileField->selectedFilter());
             if (filter > -1)
                 ui->comboBoxFormat->setCurrentIndex(filter);
             ui->buttonBox->button(QDialogButtonBox::Ok)->setEnabled(true);
         });
+        // Set the default format to Dynamic-size VHD. Do it last after everything is set up
+        // so the currentIndexChanged signal can do what is needed
+        ui->comboBoxFormat->setCurrentIndex(DEFAULT_DISK_FORMAT);
+        ui->fileField->setselectedFilter(filters.value(DEFAULT_DISK_FORMAT));
     }
 }
 
@@ -179,6 +188,7 @@ HarddiskDialog::on_comboBoxFormat_currentIndexChanged(int index)
         ui->comboBoxBlockSize->show();
         ui->labelBlockSize->show();
     }
+    ui->fileField->setselectedFilter(filters.value(index));
 }
 
 /* If the disk geometry requested in the 86Box GUI is not compatible with the internal VHD geometry,
diff --git a/src/qt/qt_harddiskdialog.hpp b/src/qt/qt_harddiskdialog.hpp
index 0d5fa13bc..9de61c51b 100644
--- a/src/qt/qt_harddiskdialog.hpp
+++ b/src/qt/qt_harddiskdialog.hpp
@@ -52,6 +52,11 @@ private:
 
     bool disallowSizeModifications = false;
 
+    QStringList filters;
+    // "Dynamic-size VHD" is number 4 in the `filters` list and the
+    // comboBoxFormat model
+    const uint8_t DEFAULT_DISK_FORMAT = 4;
+
     bool checkAndAdjustCylinders();
     bool checkAndAdjustHeads();
     bool checkAndAdjustSectors();
diff --git a/src/qt/qt_mainwindow.cpp b/src/qt/qt_mainwindow.cpp
index 7d37c0c84..04317800d 100644
--- a/src/qt/qt_mainwindow.cpp
+++ b/src/qt/qt_mainwindow.cpp
@@ -1092,6 +1092,11 @@ MainWindow::processMacKeyboardInput(bool down, const QKeyEvent *event)
             if (mac_iso_swap)
                 nvk = (nvk == 0x0a) ? 0x32 : 0x0a;
         }
+        // Special case for command + forward delete to send insert.
+        if ((event->nativeModifiers() & NSEventModifierFlagCommand) &&
+            ((event->nativeVirtualKey() == nvk_Delete) || event->key() == Qt::Key_Delete)) {
+            nvk = nvk_Insert; // Qt::Key_Help according to event->key()
+        }
 
         processKeyboardInput(down, nvk);
     }
@@ -1294,7 +1299,7 @@ MainWindow::keyReleaseEvent(QKeyEvent *event)
         fs_on_signal = false;
     }
 
-    if (!send_keyboard_input)
+    if (!send_keyboard_input || event->isAutoRepeat())
         return;
 
 #ifdef Q_OS_MACOS
diff --git a/src/qt/qt_settingsdisplay.cpp b/src/qt/qt_settingsdisplay.cpp
index e64dc74dd..5f2f81c58 100644
--- a/src/qt/qt_settingsdisplay.cpp
+++ b/src/qt/qt_settingsdisplay.cpp
@@ -36,7 +36,7 @@ SettingsDisplay::SettingsDisplay(QWidget *parent)
 {
     ui->setupUi(this);
 
-    videoCard[0]   = gfxcard[0];
+    videoCard[0] = gfxcard[0];
     videoCard[1] = gfxcard[1];
     onCurrentMachineChanged(machine);
 }
@@ -102,6 +102,11 @@ SettingsDisplay::onCurrentMachineChanged(int machineId)
         ui->comboBoxVideoSecondary->setEnabled(true);
         ui->pushButtonConfigureSecondary->setEnabled(true);
     }
+	if (video_card_get_flags(gfxcard[0]) != VIDEO_FLAG_TYPE_8514)
+		ibm8514_has_vga = 0;
+	if (video_card_get_flags(gfxcard[0]) != VIDEO_FLAG_TYPE_XGA)
+		xga_has_vga = 0;
+
     ui->comboBoxVideo->setCurrentIndex(selectedRow);
     if (gfxcard[1] == 0)
         ui->pushButtonConfigureSecondary->setEnabled(false);
@@ -123,10 +128,12 @@ SettingsDisplay::on_pushButtonConfigureVoodoo_clicked()
 void
 SettingsDisplay::on_pushButtonConfigureXga_clicked()
 {
-    if (machine_has_bus(machineId, MACHINE_BUS_MCA) > 0) {
-        DeviceConfig::ConfigureDevice(&xga_device, 0, qobject_cast<Settings *>(Settings::settings));
-    } else {
-        DeviceConfig::ConfigureDevice(&xga_isa_device, 0, qobject_cast<Settings *>(Settings::settings));
+    if (!xga_has_vga) {
+        if (machine_has_bus(machineId, MACHINE_BUS_MCA) > 0) {
+            DeviceConfig::ConfigureDevice(&xga_device, 0, qobject_cast<Settings *>(Settings::settings));
+        } else {
+            DeviceConfig::ConfigureDevice(&xga_isa_device, 0, qobject_cast<Settings *>(Settings::settings));
+        }
     }
 }
 
@@ -139,7 +146,6 @@ SettingsDisplay::on_comboBoxVideo_currentIndexChanged(int index)
     auto curVideoCard_2 = videoCard[1];
     videoCard[0] = ui->comboBoxVideo->currentData().toInt();
     ui->pushButtonConfigure->setEnabled(video_card_has_config(videoCard[0]) > 0);
-
     bool machineHasPci = machine_has_bus(machineId, MACHINE_BUS_PCI) > 0;
     ui->checkBoxVoodoo->setEnabled(machineHasPci);
     if (machineHasPci) {
@@ -149,16 +155,16 @@ SettingsDisplay::on_comboBoxVideo_currentIndexChanged(int index)
 
     bool hasIsa16 = machine_has_bus(machineId, MACHINE_BUS_ISA16) > 0;
     bool has_MCA  = machine_has_bus(machineId, MACHINE_BUS_MCA) > 0;
-    ui->checkBox8514->setEnabled(hasIsa16 || has_MCA);
+    ui->checkBox8514->setEnabled((hasIsa16 || has_MCA) && !ibm8514_has_vga);
     if (hasIsa16 || has_MCA) {
         ui->checkBox8514->setChecked(ibm8514_enabled);
     }
 
-    ui->checkBoxXga->setEnabled(hasIsa16 || has_MCA);
+    ui->checkBoxXga->setEnabled((hasIsa16 || has_MCA) && !xga_has_vga);
     if (hasIsa16 || has_MCA)
         ui->checkBoxXga->setChecked(xga_enabled);
 
-    ui->pushButtonConfigureXga->setEnabled((hasIsa16 || has_MCA) && ui->checkBoxXga->isChecked());
+    ui->pushButtonConfigureXga->setEnabled((hasIsa16 || has_MCA) && ui->checkBoxXga->isChecked() && !xga_has_vga);
 
     int c = 2;
 
@@ -187,7 +193,7 @@ SettingsDisplay::on_comboBoxVideo_currentIndexChanged(int index)
         c++;
     }
 
-    if (videoCard[1] == 0 || (machine_has_flags(machineId, MACHINE_VIDEO_ONLY) > 0)) {
+    if ((videoCard[1] == 0) || (machine_has_flags(machineId, MACHINE_VIDEO_ONLY) > 0)) {
         ui->comboBoxVideoSecondary->setCurrentIndex(0);
         ui->pushButtonConfigureSecondary->setEnabled(false);
     }
@@ -202,7 +208,7 @@ SettingsDisplay::on_checkBoxVoodoo_stateChanged(int state)
 void
 SettingsDisplay::on_checkBoxXga_stateChanged(int state)
 {
-    ui->pushButtonConfigureXga->setEnabled(state == Qt::Checked);
+    ui->pushButtonConfigureXga->setEnabled((state == Qt::Checked) && !xga_has_vga);
 }
 
 void
diff --git a/src/qt/qt_settingsmachine.cpp b/src/qt/qt_settingsmachine.cpp
index 76c89c6c1..569d1f6c3 100644
--- a/src/qt/qt_settingsmachine.cpp
+++ b/src/qt/qt_settingsmachine.cpp
@@ -110,11 +110,7 @@ SettingsMachine::save()
     cpu             = ui->comboBoxSpeed->currentData().toInt();
     fpu_type        = ui->comboBoxFPU->currentData().toInt();
     cpu_use_dynarec = ui->checkBoxDynamicRecompiler->isChecked() ? 1 : 0;
-    fpu_softfloat   = (ui->checkBoxFPUSoftfloat->isChecked() && !cpu_use_dynarec) ? 1 : 0;
-    if (!strcmp(machines[machine].internal_name, "ibmps2_m70_type4")) {
-        cpu_use_dynarec = 0;
-        fpu_softfloat = 1;
-    }
+    fpu_softfloat   = ui->checkBoxFPUSoftfloat->isChecked() ? 1 : 0;
 
     int64_t temp_mem_size;
     if (machine_get_ram_granularity(machine) < 1024) {
@@ -281,25 +277,17 @@ SettingsMachine::on_comboBoxSpeed_currentIndexChanged(int index)
     if (!(flags & CPU_SUPPORTS_DYNAREC)) {
         ui->checkBoxDynamicRecompiler->setChecked(false);
         ui->checkBoxDynamicRecompiler->setEnabled(false);
-        ui->checkBoxFPUSoftfloat->setChecked(fpu_softfloat);
-        ui->checkBoxFPUSoftfloat->setEnabled(cpu_use_dynarec ? false : true);
     } else if (flags & CPU_REQUIRES_DYNAREC) {
         ui->checkBoxDynamicRecompiler->setChecked(true);
         ui->checkBoxDynamicRecompiler->setEnabled(false);
-        ui->checkBoxFPUSoftfloat->setChecked(false);
-        ui->checkBoxFPUSoftfloat->setEnabled(false);
     } else {
         ui->checkBoxDynamicRecompiler->setChecked(cpu_use_dynarec);
         ui->checkBoxDynamicRecompiler->setEnabled(true);
-        ui->checkBoxFPUSoftfloat->setChecked(fpu_softfloat);
-        ui->checkBoxFPUSoftfloat->setEnabled(cpu_use_dynarec ? false : true);
     }
-#else
-    ui->checkBoxFPUSoftfloat->setChecked(fpu_softfloat);
-    ui->checkBoxFPUSoftfloat->setEnabled(true);
 #endif
 
     // win_settings_machine_recalc_fpu
+    int   machineId  = ui->comboBoxMachine->currentData().toInt();
     auto *modelFpu   = ui->comboBoxFPU->model();
     int   removeRows = modelFpu->rowCount();
 
@@ -317,6 +305,10 @@ SettingsMachine::on_comboBoxSpeed_currentIndexChanged(int index)
     ui->comboBoxFPU->setEnabled(modelFpu->rowCount() > 1);
     ui->comboBoxFPU->setCurrentIndex(-1);
     ui->comboBoxFPU->setCurrentIndex(selectedFpuRow);
+
+    ui->checkBoxFPUSoftfloat->setChecked(machine_has_flags(machineId, MACHINE_SOFTFLOAT_ONLY) ? true : fpu_softfloat);
+    ui->checkBoxFPUSoftfloat->setEnabled(machine_has_flags(machineId, MACHINE_SOFTFLOAT_ONLY) ? false : true);
+
 }
 
 void
diff --git a/src/scsi/scsi_cdrom.c b/src/scsi/scsi_cdrom.c
index b20a4b4ab..d100902d2 100644
--- a/src/scsi/scsi_cdrom.c
+++ b/src/scsi/scsi_cdrom.c
@@ -976,7 +976,9 @@ scsi_cdrom_command_common(scsi_cdrom_t *dev)
                 break;
             case 0xc6:
             case 0xc7:
-                if ((!strcmp(cdrom_drive_types[dev->drv->type].internal_name, "TOSHIBA_CD-ROM_DRIVEXM_3433"))) {
+                if (!strcmp(cdrom_drive_types[dev->drv->type].internal_name, "TOSHIBA_CD-ROM_DRIVEXM_3433") ||
+                    !strcmp(cdrom_drive_types[dev->drv->type].internal_name, "TOSHIBA_CD-ROM_XM-3301TA_0272") ||
+                    !strcmp(cdrom_drive_types[dev->drv->type].internal_name, "TOSHIBA_CD-ROM_XM-5701TA_3136")) {
                     bytes_per_second = 176.0 * 1024.0;
                     bytes_per_second *= (double) dev->drv->cur_speed;
                 }
@@ -1000,7 +1002,15 @@ scsi_cdrom_command_common(scsi_cdrom_t *dev)
                 break;
             case 0xc3:
                 if (!strcmp(cdrom_drive_types[dev->drv->type].internal_name, "SONY_CD-ROM_CDU-541_1.0i") ||
-                    !strcmp(cdrom_drive_types[dev->drv->type].internal_name, "SONY_CD-ROM_CDU-76S_1.00")) {
+                    !strcmp(cdrom_drive_types[dev->drv->type].internal_name, "SONY_CD-ROM_CDU-76S_1.00") ||
+                    !strcmp(cdrom_drive_types[dev->drv->type].internal_name, "PIONEER_CD-ROM_DRM-604X_2403")) {
+                    bytes_per_second = 176.0 * 1024.0;
+                    bytes_per_second *= (double) dev->drv->cur_speed;
+                }
+                break;
+            case 0xde:
+                if (!strcmp(cdrom_drive_types[dev->drv->type].internal_name, "NEC_CD-ROM_DRIVE74_1.00") ||
+                    !strcmp(cdrom_drive_types[dev->drv->type].internal_name, "NEC_CD-ROM_DRIVE464_1.05")) {
                     bytes_per_second = 176.0 * 1024.0;
                     bytes_per_second *= (double) dev->drv->cur_speed;
                 }
diff --git a/src/scsi/scsi_pcscsi.c b/src/scsi/scsi_pcscsi.c
index f3dc6ebaf..c0244ed4f 100644
--- a/src/scsi/scsi_pcscsi.c
+++ b/src/scsi/scsi_pcscsi.c
@@ -379,9 +379,9 @@ esp_get_cmd(esp_t *dev, uint32_t maxlen)
             dma_set_drq(dev->DmaChannel, 0);
         } else {
             esp_pci_dma_memory_rw(dev, buf, dmalen, WRITE_TO_DEVICE);
-            dmalen = MIN(fifo8_num_free(&dev->cmdfifo), dmalen);
-            fifo8_push_all(&dev->cmdfifo, buf, dmalen);
         }
+        dmalen = MIN(fifo8_num_free(&dev->cmdfifo), dmalen);
+        fifo8_push_all(&dev->cmdfifo, buf, dmalen);
     } else {
         dmalen = MIN(fifo8_num_used(&dev->fifo), maxlen);
         esp_log("ESP Get command, dmalen = %i\n", dmalen);
@@ -545,6 +545,8 @@ esp_hard_reset(esp_t *dev)
     dev->do_cmd          = 0;
     dev->rregs[ESP_CFG1] = dev->mca ? dev->HostID : 7;
     esp_log("ESP Reset\n");
+    for (uint8_t i = 0; i < 16; i++)
+        scsi_device_reset(&scsi_devices[dev->bus][i]);
     timer_stop(&dev->timer);
 }
 
@@ -569,7 +571,6 @@ esp_do_nodma(esp_t *dev, scsi_device_t *sd)
             esp_do_cmd(dev);
         } else {
             dev->cmdfifo_cdb_offset = fifo8_num_used(&dev->cmdfifo);
-            ;
             esp_log("CDB offset = %i used\n", dev->cmdfifo_cdb_offset);
 
             dev->rregs[ESP_RSTAT] = STAT_TC | STAT_CD;
@@ -665,7 +666,7 @@ esp_do_dma(esp_t *dev, scsi_device_t *sd)
 
     count = tdbc = esp_get_tc(dev);
 
-    if (dev->mca) { /*See the comment in the esp_do_busid_cmd() function.*/
+    if (dev->mca) {
         if (sd->buffer_length < 0) {
             if (dev->dma_enabled)
                 goto done;
@@ -713,7 +714,7 @@ esp_do_dma(esp_t *dev, scsi_device_t *sd)
         return;
     }
 
-    esp_log("ESP SCSI dmaleft = %d, async_len = %i, buffer length = %d\n", esp_get_tc(dev), sd->buffer_length);
+    esp_log("ESP SCSI dmaleft = %d, buffer length = %d\n", esp_get_tc(dev), sd->buffer_length);
 
     /* Make sure count is never bigger than buffer_length. */
     if (count > dev->xfer_counter)
@@ -1082,6 +1083,9 @@ esp_reg_write(esp_t *dev, uint32_t saddr, uint32_t val)
                         esp_pci_soft_reset(dev);
                     break;
                 case CMD_BUSRESET:
+                    for (uint8_t i = 0; i < 16; i++)
+                        scsi_device_reset(&scsi_devices[dev->bus][i]);
+
                     if (!(dev->wregs[ESP_CFG1] & CFG1_RESREPT)) {
                         dev->rregs[ESP_RINTR] |= INTR_RST;
                         esp_log("ESP Bus Reset with IRQ\n");
diff --git a/src/sio/sio_82091aa.c b/src/sio/sio_82091aa.c
index 4011f7ecf..800bf1b1a 100644
--- a/src/sio/sio_82091aa.c
+++ b/src/sio/sio_82091aa.c
@@ -35,9 +35,10 @@
 #include <86box/fdc.h>
 #include <86box/sio.h>
 
-typedef struct {
-    uint8_t cur_reg, has_ide,
-        regs[81];
+typedef struct i82091aa_t {
+    uint8_t   cur_reg;
+    uint8_t   has_ide;
+    uint8_t   regs[81];
     uint16_t  base_address;
     fdc_t    *fdc;
     serial_t *uart[2];
@@ -71,6 +72,9 @@ lpt1_handler(i82091aa_t *dev)
         case 3:
             lpt_port = 0x000;
             break;
+
+        default:
+            break;
     }
 
     if ((dev->regs[0x20] & 0x01) && lpt_port)
@@ -112,6 +116,9 @@ serial_handler(i82091aa_t *dev, int uart)
         case 0x07:
             uart_port = COM3_ADDR;
             break;
+
+        default:
+            break;
     }
 
     if (dev->regs[reg] & 0x01)
@@ -193,6 +200,9 @@ i82091aa_write(uint16_t port, uint8_t val, void *priv)
             if (dev->has_ide && (valxor & 0x03))
                 ide_handler(dev);
             break;
+
+        default:
+            break;
     }
 }
 
diff --git a/src/sio/sio_acc3221.c b/src/sio/sio_acc3221.c
index f5c671c3a..10250d602 100644
--- a/src/sio/sio_acc3221.c
+++ b/src/sio/sio_acc3221.c
@@ -31,6 +31,7 @@
 #include <86box/fdd.h>
 #include <86box/fdc.h>
 #include <86box/sio.h>
+#include <86box/plat_unused.h>
 
 typedef struct acc3221_t {
     int       reg_idx;
@@ -344,9 +345,9 @@ acc3221_serial2_handler(acc3221_t *dev)
 }
 
 static void
-acc3221_write(uint16_t addr, uint8_t val, void *p)
+acc3221_write(uint16_t addr, uint8_t val, void *priv)
 {
-    acc3221_t *dev = (acc3221_t *) p;
+    acc3221_t *dev = (acc3221_t *) priv;
     uint8_t    old;
 
     if (!(addr & 1))
@@ -405,14 +406,17 @@ acc3221_write(uint16_t addr, uint8_t val, void *p)
                         ide_pri_enable();
                 }
                 break;
+
+            default:
+                break;
         }
     }
 }
 
 static uint8_t
-acc3221_read(uint16_t addr, void *p)
+acc3221_read(uint16_t addr, void *priv)
 {
-    acc3221_t *dev = (acc3221_t *) p;
+    acc3221_t *dev = (acc3221_t *) priv;
 
     if (!(addr & 1))
         return dev->reg_idx;
@@ -448,7 +452,7 @@ acc3221_close(void *priv)
 }
 
 static void *
-acc3221_init(const device_t *info)
+acc3221_init(UNUSED(const device_t *info))
 {
     acc3221_t *dev = (acc3221_t *) malloc(sizeof(acc3221_t));
     memset(dev, 0, sizeof(acc3221_t));
diff --git a/src/sio/sio_ali5123.c b/src/sio/sio_ali5123.c
index 68d674b11..25fb1aa65 100644
--- a/src/sio/sio_ali5123.c
+++ b/src/sio/sio_ali5123.c
@@ -37,13 +37,14 @@
 
 #define AB_RST 0x80
 
-typedef struct {
-    uint8_t chip_id, is_apm,
-        tries,
-        regs[48],
-        ld_regs[13][256];
-    int locked,
-        cur_reg;
+typedef struct ali5123_t {
+    uint8_t   chip_id;
+    uint8_t   is_apm;
+    uint8_t   tries;
+    uint8_t   regs[48];
+    uint8_t   ld_regs[13][256];
+    int       locked;
+    int       cur_reg;
     fdc_t    *fdc;
     serial_t *uart[3];
 } ali5123_t;
@@ -125,6 +126,9 @@ ali5123_serial_handler(ali5123_t *dev, int uart)
         case 0x05:
             serial_set_clock_src(dev->uart[uart], 2000000.0);
             break;
+
+        default:
+            break;
     }
 }
 
@@ -256,6 +260,10 @@ ali5123_write(uint16_t port, uint8_t val, void *priv)
                         case 0x07:
                             if (dev->cur_reg == 0xf0)
                                 val &= 0xbf;
+                            break;
+
+                        default:
+                            break;
                     }
                 dev->ld_regs[cur_ld][dev->cur_reg] = val;
             }
@@ -288,6 +296,9 @@ ali5123_write(uint16_t port, uint8_t val, void *priv)
                     ali5123_serial_handler(dev, 2);
                 }
                 break;
+
+            default:
+                break;
         }
 
         return;
@@ -332,6 +343,9 @@ ali5123_write(uint16_t port, uint8_t val, void *priv)
                     if (valxor & 0x08)
                         fdc_update_drvrate(dev->fdc, 3, (val & 0x08) >> 3);
                     break;
+
+                default:
+                    break;
             }
             break;
         case 3:
@@ -346,6 +360,9 @@ ali5123_write(uint16_t port, uint8_t val, void *priv)
                     if (valxor)
                         ali5123_lpt_handler(dev);
                     break;
+
+                default:
+                    break;
             }
             break;
         case 4:
@@ -361,6 +378,9 @@ ali5123_write(uint16_t port, uint8_t val, void *priv)
                     if (valxor)
                         ali5123_serial_handler(dev, 0);
                     break;
+
+                default:
+                    break;
             }
             break;
         case 5:
@@ -376,6 +396,9 @@ ali5123_write(uint16_t port, uint8_t val, void *priv)
                     if (valxor)
                         ali5123_serial_handler(dev, (dev->regs[0x2d] & 0x20) ? 2 : 1);
                     break;
+
+                default:
+                    break;
             }
             break;
         case 0x0b:
@@ -391,8 +414,14 @@ ali5123_write(uint16_t port, uint8_t val, void *priv)
                     if (valxor)
                         ali5123_serial_handler(dev, (dev->regs[0x2d] & 0x20) ? 1 : 2);
                     break;
+
+                default:
+                    break;
             }
             break;
+
+        default:
+            break;
     }
 }
 
diff --git a/src/sio/sio_detect.c b/src/sio/sio_detect.c
index 36c12cd54..01c41224a 100644
--- a/src/sio/sio_detect.c
+++ b/src/sio/sio_detect.c
@@ -26,8 +26,9 @@
 #include <86box/fdd.h>
 #include <86box/fdc.h>
 #include <86box/sio.h>
+#include <86box/plat_unused.h>
 
-typedef struct {
+typedef struct sio_detect_t {
     uint8_t regs[2];
 } sio_detect_t;
 
@@ -62,7 +63,7 @@ sio_detect_close(void *priv)
 }
 
 static void *
-sio_detect_init(const device_t *info)
+sio_detect_init(UNUSED(const device_t *info))
 {
     sio_detect_t *dev = (sio_detect_t *) malloc(sizeof(sio_detect_t));
     memset(dev, 0, sizeof(sio_detect_t));
diff --git a/src/sio/sio_f82c710.c b/src/sio/sio_f82c710.c
index ad347fb03..fc0eb2aae 100644
--- a/src/sio/sio_f82c710.c
+++ b/src/sio/sio_f82c710.c
@@ -151,6 +151,9 @@ f82c606_update_ports(upc_t *dev, int set)
         case 0xc0:
             uart2_int = COM2_IRQ;
             break;
+
+        default:
+            break;
     }
 
     switch (dev->regs[8] & 0x30) {
@@ -163,6 +166,9 @@ f82c606_update_ports(upc_t *dev, int set)
         case 0x30:
             uart2_int = COM1_IRQ;
             break;
+
+        default:
+            break;
     }
 
     switch (dev->regs[8] & 0x0c) {
@@ -175,6 +181,9 @@ f82c606_update_ports(upc_t *dev, int set)
         case 0x0c:
             lpt1_int = LPT2_IRQ;
             break;
+
+        default:
+            break;
     }
 
     switch (dev->regs[8] & 0x03) {
@@ -187,6 +196,9 @@ f82c606_update_ports(upc_t *dev, int set)
         case 0x03:
             lpt1_int = LPT1_IRQ;
             break;
+
+        default:
+            break;
     }
 
     if (dev->regs[0] & 1) {
diff --git a/src/sio/sio_fdc37c669.c b/src/sio/sio_fdc37c669.c
index 219213d69..3a1bcf5af 100644
--- a/src/sio/sio_fdc37c669.c
+++ b/src/sio/sio_fdc37c669.c
@@ -32,11 +32,13 @@
 #include <86box/fdc.h>
 #include <86box/sio.h>
 
-typedef struct {
-    uint8_t id, tries,
-        regs[42];
-    int locked, rw_locked,
-        cur_reg;
+typedef struct fdc37c669_t {
+    uint8_t   id;
+    uint8_t   tries;
+    uint8_t   regs[42];
+    int       locked;
+    int       rw_locked;
+    int       cur_reg;
     fdc_t    *fdc;
     serial_t *uart[2];
 } fdc37c669_t;
@@ -62,6 +64,9 @@ make_port(fdc37c669_t *dev, uint8_t reg)
         case 0x25:
             mask = 0xfe;
             break;
+
+        default:
+            break;
     }
 
     p = ((uint16_t) (dev->regs[reg] & mask)) << 2;
@@ -217,6 +222,9 @@ fdc37c669_write(uint16_t port, uint8_t val, void *priv)
                     serial_setup(dev->uart[0], make_port(dev, 0x24), (dev->regs[0x28] & 0xf0) >> 4);
             }
             break;
+
+        default:
+            break;
     }
 }
 
diff --git a/src/sio/sio_fdc37c67x.c b/src/sio/sio_fdc37c67x.c
index 35c7ccc5b..3cfa5a38d 100644
--- a/src/sio/sio_fdc37c67x.c
+++ b/src/sio/sio_fdc37c67x.c
@@ -33,19 +33,23 @@
 #include <86box/fdc.h>
 #include "cpu.h"
 #include <86box/sio.h>
+#include <86box/plat_unused.h>
 
 #define AB_RST 0x80
 
-typedef struct {
-    uint8_t chip_id, is_apm,
-        tries,
-        gpio_regs[2], auxio_reg,
-        regs[48],
-        ld_regs[11][256];
-    uint16_t gpio_base, /* Set to EA */
-        auxio_base, sio_base;
-    int locked,
-        cur_reg;
+typedef struct fdc37c67x_t {
+    uint8_t   chip_id;
+    uint8_t   is_apm;
+    uint8_t   tries;
+    uint8_t   gpio_regs[2];
+    uint8_t   auxio_reg;
+    uint8_t   regs[48];
+    uint8_t   ld_regs[11][256];
+    uint16_t  gpio_base; /* Set to EA */
+    uint16_t  auxio_base;
+    uint16_t  sio_base;
+    int       locked;
+    int       cur_reg;
     fdc_t    *fdc;
     serial_t *uart[2];
 } fdc37c67x_t;
@@ -65,7 +69,7 @@ make_port(fdc37c67x_t *dev, uint8_t ld)
 }
 
 static uint8_t
-fdc37c67x_auxio_read(uint16_t port, void *priv)
+fdc37c67x_auxio_read(UNUSED(uint16_t port), void *priv)
 {
     fdc37c67x_t *dev = (fdc37c67x_t *) priv;
 
@@ -73,7 +77,7 @@ fdc37c67x_auxio_read(uint16_t port, void *priv)
 }
 
 static void
-fdc37c67x_auxio_write(uint16_t port, uint8_t val, void *priv)
+fdc37c67x_auxio_write(UNUSED(uint16_t port), uint8_t val, void *priv)
 {
     fdc37c67x_t *dev = (fdc37c67x_t *) priv;
 
@@ -168,7 +172,7 @@ fdc37c67x_auxio_handler(fdc37c67x_t *dev)
 }
 
 static void
-fdc37c67x_sio_handler(fdc37c67x_t *dev)
+fdc37c67x_sio_handler(UNUSED(fdc37c67x_t *dev))
 {
 #if 0
     if (dev->sio_base) {
@@ -207,6 +211,9 @@ fdc37c67x_gpio_handler(fdc37c67x_t *dev)
             case 3:
                 ld_port = 0xea; /* Default */
                 break;
+
+            default:
+                break;
         }
         dev->gpio_base = ld_port;
         if (ld_port > 0x0000)
@@ -279,6 +286,9 @@ fdc37c67x_write(uint16_t port, uint8_t val, void *priv)
                         case 0x02:
                         case 0x07:
                             return;
+
+                        default:
+                            break;
                     }
                 dev->ld_regs[dev->regs[7]][dev->cur_reg] = val | keep;
             }
@@ -306,6 +316,9 @@ fdc37c67x_write(uint16_t port, uint8_t val, void *priv)
             case 0x26:
             case 0x27:
                 fdc37c67x_sio_handler(dev);
+
+            default:
+                break;
         }
 
         return;
@@ -359,6 +372,9 @@ fdc37c67x_write(uint16_t port, uint8_t val, void *priv)
                     if (valxor & 0x18)
                         fdc_update_drvrate(dev->fdc, 3, (val & 0x18) >> 3);
                     break;
+
+                default:
+                    break;
             }
             break;
         case 3:
@@ -375,6 +391,9 @@ fdc37c67x_write(uint16_t port, uint8_t val, void *priv)
                     if (dev->cur_reg == 0x70)
                         fdc37c67x_smi_handler(dev);
                     break;
+
+                default:
+                    break;
             }
             break;
         case 4:
@@ -391,6 +410,9 @@ fdc37c67x_write(uint16_t port, uint8_t val, void *priv)
                     if (dev->cur_reg == 0x70)
                         fdc37c67x_smi_handler(dev);
                     break;
+
+                default:
+                    break;
             }
             break;
         case 5:
@@ -407,6 +429,9 @@ fdc37c67x_write(uint16_t port, uint8_t val, void *priv)
                     if (dev->cur_reg == 0x70)
                         fdc37c67x_smi_handler(dev);
                     break;
+
+                default:
+                    break;
             }
             break;
         case 8:
@@ -423,8 +448,14 @@ fdc37c67x_write(uint16_t port, uint8_t val, void *priv)
                 case 0xb5:
                     fdc37c67x_smi_handler(dev);
                     break;
+
+                default:
+                    break;
             }
             break;
+
+        default:
+            break;
     }
 }
 
@@ -570,7 +601,9 @@ fdc37c67x_init(const device_t *info)
     dev->chip_id = info->local & 0xff;
 
     dev->gpio_regs[0] = 0xff;
-    // dev->gpio_regs[1] = (info->local == 0x0030) ? 0xff : 0xfd;
+#if 0
+    dev->gpio_regs[1] = (info->local == 0x0030) ? 0xff : 0xfd;
+#endif
     dev->gpio_regs[1] = (dev->chip_id == 0x30) ? 0xff : 0xfd;
 
     fdc37c67x_reset(dev);
diff --git a/src/sio/sio_fdc37c6xx.c b/src/sio/sio_fdc37c6xx.c
index 8400ca1e0..10eccf97c 100644
--- a/src/sio/sio_fdc37c6xx.c
+++ b/src/sio/sio_fdc37c6xx.c
@@ -35,12 +35,15 @@
 #include <86box/fdc.h>
 #include <86box/sio.h>
 
-typedef struct {
-    uint8_t max_reg, chip_id,
-        tries, has_ide,
-        regs[16];
-    int cur_reg,
-        com3_addr, com4_addr;
+typedef struct fdc37c6xx_t {
+    uint8_t   max_reg;
+    uint8_t   chip_id;
+    uint8_t   tries;
+    uint8_t   has_ide;
+    uint8_t   regs[16];
+    int       cur_reg;
+    int       com3_addr;
+    int       com4_addr;
     fdc_t    *fdc;
     serial_t *uart[2];
 } fdc37c6xx_t;
@@ -65,6 +68,9 @@ set_com34_addr(fdc37c6xx_t *dev)
             dev->com3_addr = 0x220;
             dev->com4_addr = 0x228;
             break;
+
+        default:
+            break;
     }
 }
 
@@ -92,6 +98,9 @@ set_serial_addr(fdc37c6xx_t *dev, int port)
             case 3:
                 serial_setup(dev->uart[port], dev->com4_addr, COM4_IRQ);
                 break;
+
+            default:
+                break;
         }
     }
 
@@ -115,6 +124,9 @@ lpt1_handler(fdc37c6xx_t *dev)
             lpt1_init(LPT2_ADDR);
             lpt1_irq(7 /*5*/);
             break;
+
+        default:
+            break;
     }
 }
 
@@ -206,6 +218,9 @@ fdc37c6xx_write(uint16_t port, uint8_t val, void *priv)
                     if (valxor & 0x20)
                         fdc_set_swap(dev->fdc, (dev->regs[5] & 0x20) >> 5);
                     break;
+
+                default:
+                    break;
             }
         }
     } else if ((port == FDC_PRIMARY_ADDR) && (val == 0x55))
diff --git a/src/sio/sio_fdc37c93x.c b/src/sio/sio_fdc37c93x.c
index 80f451e06..0330297fb 100644
--- a/src/sio/sio_fdc37c93x.c
+++ b/src/sio/sio_fdc37c93x.c
@@ -35,10 +35,11 @@
 #include <86box/apm.h>
 #include <86box/acpi.h>
 #include <86box/sio.h>
+#include <86box/plat_unused.h>
 
 #define AB_RST 0x80
 
-typedef struct {
+typedef struct access_bus_t {
     uint8_t  control;
     uint8_t  status;
     uint8_t  own_addr;
@@ -47,16 +48,19 @@ typedef struct {
     uint16_t base;
 } access_bus_t;
 
-typedef struct {
-    uint8_t chip_id, is_apm,
-        tries,
-        gpio_regs[2], auxio_reg,
-        regs[48],
-        ld_regs[11][256];
-    uint16_t gpio_base, /* Set to EA */
-        auxio_base, nvr_sec_base;
-    int locked,
-        cur_reg;
+typedef struct fdc37c93x_t {
+    uint8_t       chip_id;
+    uint8_t       is_apm;
+    uint8_t       tries;
+    uint8_t       gpio_regs[2];
+    uint8_t       auxio_reg;
+    uint8_t       regs[48];
+    uint8_t       ld_regs[11][256];
+    uint16_t      gpio_base; /* Set to EA */
+    uint16_t      auxio_base;
+    uint16_t      nvr_sec_base;
+    int           locked;
+    int           cur_reg;
     fdc_t        *fdc;
     serial_t     *uart[2];
     access_bus_t *access_bus;
@@ -87,7 +91,7 @@ make_port_sec(fdc37c93x_t *dev, uint8_t ld)
 }
 
 static uint8_t
-fdc37c93x_auxio_read(uint16_t port, void *priv)
+fdc37c93x_auxio_read(UNUSED(uint16_t port), void *priv)
 {
     fdc37c93x_t *dev = (fdc37c93x_t *) priv;
 
@@ -95,7 +99,7 @@ fdc37c93x_auxio_read(uint16_t port, void *priv)
 }
 
 static void
-fdc37c93x_auxio_write(uint16_t port, uint8_t val, void *priv)
+fdc37c93x_auxio_write(UNUSED(uint16_t port), uint8_t val, void *priv)
 {
     fdc37c93x_t *dev = (fdc37c93x_t *) priv;
 
@@ -239,6 +243,9 @@ fdc37c93x_gpio_handler(fdc37c93x_t *dev)
             case 3:
                 ld_port = 0xea; /* Default */
                 break;
+
+            default:
+                break;
         }
         dev->gpio_base = ld_port;
         if (ld_port > 0x0000)
@@ -266,6 +273,9 @@ fdc37c93x_access_bus_read(uint16_t port, void *priv)
         case 3:
             ret = (dev->clock & 0x87);
             break;
+
+        default:
+            break;
     }
 
     return ret;
@@ -290,6 +300,9 @@ fdc37c93x_access_bus_write(uint16_t port, uint8_t val, void *priv)
             dev->clock &= 0x80;
             dev->clock |= (val & 0x07);
             break;
+
+        default:
+            break;
     }
 }
 
@@ -410,6 +423,9 @@ fdc37c93x_write(uint16_t port, uint8_t val, void *priv)
                             if (!dev->is_apm)
                                 return;
                             break;
+
+                        default:
+                            break;
                     }
                 dev->ld_regs[dev->regs[7]][dev->cur_reg] = val | keep;
             }
@@ -436,6 +452,9 @@ fdc37c93x_write(uint16_t port, uint8_t val, void *priv)
                 if ((valxor & 0x40) && (dev->chip_id != 0x02))
                     fdc37c93x_access_bus_handler(dev);
                 break;
+
+            default:
+                break;
         }
 
         return;
@@ -489,6 +508,9 @@ fdc37c93x_write(uint16_t port, uint8_t val, void *priv)
                     if (valxor & 0x18)
                         fdc_update_drvrate(dev->fdc, 3, (val & 0x18) >> 3);
                     break;
+
+                default:
+                    break;
             }
             break;
         case 3:
@@ -503,6 +525,9 @@ fdc37c93x_write(uint16_t port, uint8_t val, void *priv)
                     if (valxor)
                         fdc37c93x_lpt_handler(dev);
                     break;
+
+                default:
+                    break;
             }
             break;
         case 4:
@@ -517,6 +542,9 @@ fdc37c93x_write(uint16_t port, uint8_t val, void *priv)
                     if (valxor)
                         fdc37c93x_serial_handler(dev, 0);
                     break;
+
+                default:
+                    break;
             }
             break;
         case 5:
@@ -531,6 +559,9 @@ fdc37c93x_write(uint16_t port, uint8_t val, void *priv)
                     if (valxor)
                         fdc37c93x_serial_handler(dev, 1);
                     break;
+
+                default:
+                    break;
             }
             break;
         case 6:
@@ -554,8 +585,8 @@ fdc37c93x_write(uint16_t port, uint8_t val, void *priv)
                         nvr_lock_set(0xe0, 0x20, !!(dev->ld_regs[6][dev->cur_reg] & 0x08), dev->nvr);
                         if (dev->ld_regs[6][dev->cur_reg] & 0x80)
                             switch ((dev->ld_regs[6][dev->cur_reg] >> 4) & 0x07) {
-                                case 0x00:
                                 default:
+                                case 0x00:
                                     nvr_bank_set(0, 0xff, dev->nvr);
                                     nvr_bank_set(1, 1, dev->nvr);
                                     break;
@@ -588,6 +619,9 @@ fdc37c93x_write(uint16_t port, uint8_t val, void *priv)
                         }
                     }
                     break;
+
+                default:
+                    break;
             }
             break;
         case 8:
@@ -600,6 +634,9 @@ fdc37c93x_write(uint16_t port, uint8_t val, void *priv)
                     if (valxor)
                         fdc37c93x_auxio_handler(dev);
                     break;
+
+                default:
+                    break;
             }
             break;
         case 9:
@@ -614,6 +651,9 @@ fdc37c93x_write(uint16_t port, uint8_t val, void *priv)
                     if (valxor)
                         fdc37c93x_access_bus_handler(dev);
                     break;
+
+                default:
+                    break;
             }
             break;
         case 10:
@@ -628,8 +668,14 @@ fdc37c93x_write(uint16_t port, uint8_t val, void *priv)
                     if (valxor)
                         fdc37c93x_acpi_handler(dev);
                     break;
+
+                default:
+                    break;
             }
             break;
+
+        default:
+            break;
     }
 }
 
@@ -783,7 +829,7 @@ access_bus_close(void *priv)
 }
 
 static void *
-access_bus_init(const device_t *info)
+access_bus_init(UNUSED(const device_t *info))
 {
     access_bus_t *dev = (access_bus_t *) malloc(sizeof(access_bus_t));
     memset(dev, 0, sizeof(access_bus_t));
@@ -830,7 +876,9 @@ fdc37c93x_init(const device_t *info)
     is_compaq    = (info->local >> 8) & 0x02;
 
     dev->gpio_regs[0] = 0xff;
-    // dev->gpio_regs[1] = (info->local == 0x0030) ? 0xff : 0xfd;
+#if 0
+    dev->gpio_regs[1] = (info->local == 0x0030) ? 0xff : 0xfd;
+#endif
     dev->gpio_regs[1] = (dev->chip_id == 0x30) ? 0xff : 0xfd;
 
     if (dev->chip_id == 0x30) {
diff --git a/src/sio/sio_fdc37m60x.c b/src/sio/sio_fdc37m60x.c
index af238c51f..190e65453 100644
--- a/src/sio/sio_fdc37m60x.c
+++ b/src/sio/sio_fdc37m60x.c
@@ -68,9 +68,12 @@ fdc37m60x_log(const char *fmt, ...)
 #    define fdc37m60x_log(fmt, ...)
 #endif
 
-typedef struct
-{
-    uint8_t  index, regs[256], device_regs[10][256], cfg_lock, ide_function;
+typedef struct fdc37m60x_t {
+    uint8_t  index;
+    uint8_t  regs[256];
+    uint8_t  device_regs[10][256];
+    uint8_t  cfg_lock;
+    uint8_t  ide_function;
     uint16_t sio_index_port;
 
     fdc_t    *fdc;
@@ -141,6 +144,9 @@ fdc37m60x_write(uint16_t addr, uint8_t val, void *priv)
                         dev->device_regs[CURRENT_LOGICAL_DEVICE][INDEX] = (INDEX == 0x30) ? (val & 1) : val;
                     fdc37m60x_logical_device_handler(dev);
                     break;
+
+                default:
+                    break;
             }
         }
     } else {
@@ -242,6 +248,9 @@ fdc37m60x_logical_device_handler(fdc37m60x_t *dev)
         case 0x05:
             fdc37m60x_uart_handler(1, dev);
             break;
+
+        default:
+            break;
     }
 }
 
diff --git a/src/sio/sio_it8661f.c b/src/sio/sio_it8661f.c
index 9b3baa188..5e5ccb655 100644
--- a/src/sio/sio_it8661f.c
+++ b/src/sio/sio_it8661f.c
@@ -32,16 +32,19 @@
 #include <86box/fdc.h>
 #include <86box/fdd_common.h>
 #include <86box/sio.h>
+#include <86box/plat_unused.h>
 
 #define LDN dev->regs[7]
 
-typedef struct
-{
+typedef struct it8661f_t {
     fdc_t    *fdc_controller;
     serial_t *uart[2];
 
-    uint8_t index, regs[256], device_regs[6][256];
-    int     unlocked, enumerator;
+    uint8_t index;
+    uint8_t regs[256];
+    uint8_t device_regs[6][256];
+    int     unlocked;
+    int     enumerator;
 } it8661f_t;
 
 static uint8_t mb_pnp_key[32] = { 0x6a, 0xb5, 0xda, 0xed, 0xf6, 0xfb, 0x7d, 0xbe, 0xdf, 0x6f, 0x37, 0x1b, 0x0d, 0x86, 0xc3, 0x61, 0xb0, 0x58, 0x2c, 0x16, 0x8b, 0x45, 0xa2, 0xd1, 0xe8, 0x74, 0x3a, 0x9d, 0xce, 0xe7, 0x73, 0x39 };
@@ -99,6 +102,9 @@ it8661_fdc(uint16_t addr, uint8_t val, it8661f_t *dev)
             case 0xf0:
                 dev->device_regs[0][addr] = val & 0x0f;
                 break;
+
+            default:
+                break;
         }
 
         fdc_set_base(dev->fdc_controller, (dev->device_regs[0][0x60] << 8) | (dev->device_regs[0][0x61]));
@@ -140,6 +146,9 @@ it8661_serial(int uart, uint16_t addr, uint8_t val, it8661f_t *dev)
             case 0xf0:
                 dev->device_regs[1 + uart][addr] = val & 3;
                 break;
+
+            default:
+                break;
         }
 
         serial_setup(dev->uart[uart], (dev->device_regs[1 + uart][0x60] << 8) | (dev->device_regs[1 + uart][0x61]), dev->device_regs[1 + uart][0x70] & 0x0f);
@@ -177,6 +186,9 @@ it8661_lpt(uint16_t addr, uint8_t val, it8661f_t *dev)
             case 0xf0:
                 dev->device_regs[3][addr] = val & 3;
                 break;
+
+            default:
+                break;
         }
 
         lpt1_init((dev->device_regs[3][0x60] << 8) | (dev->device_regs[3][0x61]));
@@ -201,6 +213,9 @@ it8661_ldn(uint16_t addr, uint8_t val, it8661f_t *dev)
         case 3:
             it8661_lpt(addr, val, dev);
             break;
+
+        default:
+            break;
     }
 }
 
@@ -246,6 +261,9 @@ it8661f_write(uint16_t addr, uint8_t val, void *priv)
                 }
             }
             break;
+
+        default:
+            break;
     }
 
     return;
@@ -300,7 +318,7 @@ it8661f_close(void *priv)
 }
 
 static void *
-it8661f_init(const device_t *info)
+it8661f_init(UNUSED(const device_t *info))
 {
     it8661f_t *dev = (it8661f_t *) malloc(sizeof(it8661f_t));
     memset(dev, 0, sizeof(it8661f_t));
diff --git a/src/sio/sio_pc87306.c b/src/sio/sio_pc87306.c
index 088467eeb..656b36d92 100644
--- a/src/sio/sio_pc87306.c
+++ b/src/sio/sio_pc87306.c
@@ -34,10 +34,12 @@
 #include <86box/fdd.h>
 #include <86box/fdc.h>
 #include <86box/sio.h>
+#include <86box/plat_unused.h>
 
-typedef struct {
-    uint8_t tries,
-        regs[29], gpio[2];
+typedef struct pc87306_t {
+    uint8_t   tries;
+    uint8_t   regs[29];
+    uint8_t   gpio[2];
     int       cur_reg;
     fdc_t    *fdc;
     serial_t *uart[2];
@@ -112,6 +114,9 @@ lpt1_handler(pc87306_t *dev)
             lpt_port = 0x000;
             lpt_irq  = 0xff;
             break;
+
+        default:
+            break;
     }
 
     if (dev->regs[0x1b] & 0x10)
@@ -165,6 +170,9 @@ serial_handler(pc87306_t *dev, int uart)
                 case 3:
                     serial_setup(dev->uart[uart], 0x220, irq);
                     break;
+
+                default:
+                    break;
             }
             break;
         case 3:
@@ -181,8 +189,14 @@ serial_handler(pc87306_t *dev, int uart)
                 case 3:
                     serial_setup(dev->uart[uart], 0x228, irq);
                     break;
+
+                default:
+                    break;
             }
             break;
+
+        default:
+            break;
     }
 }
 
@@ -324,6 +338,9 @@ pc87306_write(uint16_t port, uint8_t val, void *priv)
                     serial_handler(dev, 1);
             }
             break;
+
+        default:
+            break;
     }
 }
 
@@ -392,7 +409,7 @@ pc87306_close(void *priv)
 }
 
 static void *
-pc87306_init(const device_t *info)
+pc87306_init(UNUSED(const device_t *info))
 {
     pc87306_t *dev = (pc87306_t *) malloc(sizeof(pc87306_t));
     memset(dev, 0, sizeof(pc87306_t));
diff --git a/src/sio/sio_pc87307.c b/src/sio/sio_pc87307.c
index 78cfdcd99..e66467d62 100644
--- a/src/sio/sio_pc87307.c
+++ b/src/sio/sio_pc87307.c
@@ -35,13 +35,17 @@
 #include <86box/fdc.h>
 #include <86box/sio.h>
 
-typedef struct {
-    uint8_t id, pm_idx,
-        regs[48], ld_regs[256][208],
-        pcregs[16], gpio[2][4],
-        pm[8];
-    uint16_t gpio_base, gpio_base2,
-        pm_base;
+typedef struct pc87307_t {
+    uint8_t   id;
+    uint8_t   pm_idx;
+    uint8_t   regs[48];
+    uint8_t   ld_regs[256][208];
+    uint8_t   pcregs[16];
+    uint8_t   gpio[2][4];
+    uint8_t   pm[8];
+    uint16_t  gpio_base;
+    uint16_t  gpio_base2;
+    uint16_t  pm_base;
     int       cur_reg;
     fdc_t    *fdc;
     serial_t *uart[2];
@@ -74,6 +78,9 @@ pc87307_gpio_read(uint16_t port, void *priv)
             mask = dev->gpio[bank][0x0001];
             ret  = (ret & mask) | (pins & ~mask);
             break;
+
+        default:
+            break;
     }
 
     return ret;
@@ -122,6 +129,9 @@ pc87307_pm_write(uint16_t port, uint8_t val, void *priv)
                 serial_handler(dev, 1);
                 serial_handler(dev, 0);
                 break;
+
+            default:
+                break;
         }
     }
 }
@@ -307,6 +317,9 @@ pc87307_write(uint16_t port, uint8_t val, void *priv)
                 case 0x08:
                     pm_handler(dev);
                     break;
+
+                default:
+                    break;
             }
             break;
         case 0x60:
@@ -333,6 +346,9 @@ pc87307_write(uint16_t port, uint8_t val, void *priv)
                 case 0x08:
                     pm_handler(dev);
                     break;
+
+                default:
+                    break;
             }
             break;
         case 0x61:
@@ -364,6 +380,9 @@ pc87307_write(uint16_t port, uint8_t val, void *priv)
                     dev->ld_regs[dev->regs[0x07]][dev->cur_reg - 0x30] = val & 0xfe;
                     pm_handler(dev);
                     break;
+
+                default:
+                    break;
             }
             break;
         case 0x63:
@@ -396,6 +415,9 @@ pc87307_write(uint16_t port, uint8_t val, void *priv)
                 case 0x08:
                     pm_handler(dev);
                     break;
+
+                default:
+                    break;
             }
             break;
         case 0xf0:
@@ -416,12 +438,18 @@ pc87307_write(uint16_t port, uint8_t val, void *priv)
                 case 0x06:
                     dev->ld_regs[dev->regs[0x07]][dev->cur_reg - 0x30] = val & 0x87;
                     break;
+
+                default:
+                    break;
             }
             break;
         case 0xf1:
             if (dev->regs[0x07] == 0x03)
                 dev->ld_regs[dev->regs[0x07]][dev->cur_reg - 0x30] = val & 0x0f;
             break;
+
+        default:
+            break;
     }
 }
 
@@ -519,8 +547,10 @@ pc87307_reset(pc87307_t *dev)
     dev->ld_regs[0x08][0x44] = 0x04;
     dev->ld_regs[0x08][0x45] = 0x04;
 
-    // dev->gpio[0] = 0xff;
-    // dev->gpio[1] = 0xfb;
+#if 0
+    dev->gpio[0] = 0xff;
+    dev->gpio[1] = 0xfb;
+#endif
     dev->gpio[0][0] = 0xff;
     dev->gpio[0][1] = 0x00;
     dev->gpio[0][2] = 0x00;
diff --git a/src/sio/sio_pc87309.c b/src/sio/sio_pc87309.c
index dfa41f601..edd1fb0b8 100644
--- a/src/sio/sio_pc87309.c
+++ b/src/sio/sio_pc87309.c
@@ -35,10 +35,12 @@
 #include <86box/fdc.h>
 #include <86box/sio.h>
 
-typedef struct {
-    uint8_t id, pm_idx,
-        regs[48], ld_regs[256][208],
-        pm[8];
+typedef struct pc87309_t {
+    uint8_t   id;
+    uint8_t   pm_idx;
+    uint8_t   regs[48];
+    uint8_t   ld_regs[256][208];
+    uint8_t   pm[8];
     uint16_t  pm_base;
     int       cur_reg;
     fdc_t    *fdc;
@@ -64,6 +66,9 @@ pc87309_pm_write(uint16_t port, uint8_t val, void *priv)
                 serial_handler(dev, 1);
                 serial_handler(dev, 0);
                 break;
+
+            default:
+                break;
         }
     } else
         dev->pm_idx = val & 0x07;
@@ -221,6 +226,9 @@ pc87309_write(uint16_t port, uint8_t val, void *priv)
                 case 0x04:
                     pm_handler(dev);
                     break;
+
+                default:
+                    break;
             }
             break;
         case 0x60:
@@ -244,6 +252,9 @@ pc87309_write(uint16_t port, uint8_t val, void *priv)
                 case 0x04:
                     pm_handler(dev);
                     break;
+
+                default:
+                    break;
             }
             break;
         case 0x63:
@@ -275,6 +286,9 @@ pc87309_write(uint16_t port, uint8_t val, void *priv)
                 case 0x06:
                     dev->ld_regs[dev->regs[0x07]][dev->cur_reg - 0x30] = val & 0xf8;
                     break;
+
+                default:
+                    break;
             }
             break;
         case 0x70:
@@ -296,6 +310,9 @@ pc87309_write(uint16_t port, uint8_t val, void *priv)
                 case 0x04:
                     pm_handler(dev);
                     break;
+
+                default:
+                    break;
             }
             break;
         case 0xf0:
@@ -316,12 +333,18 @@ pc87309_write(uint16_t port, uint8_t val, void *priv)
                 case 0x06:
                     dev->ld_regs[dev->regs[0x07]][dev->cur_reg - 0x30] = val & 0xc1;
                     break;
+
+                default:
+                    break;
             }
             break;
         case 0xf1:
             if (dev->regs[0x07] == 0x00)
                 dev->ld_regs[dev->regs[0x07]][dev->cur_reg - 0x30] = val & 0x0f;
             break;
+
+        default:
+            break;
     }
 }
 
diff --git a/src/sio/sio_pc87310.c b/src/sio/sio_pc87310.c
index eacbadf89..d567bd4d5 100644
--- a/src/sio/sio_pc87310.c
+++ b/src/sio/sio_pc87310.c
@@ -40,6 +40,7 @@
 #include <86box/fdd.h>
 #include <86box/fdc.h>
 #include <86box/sio.h>
+#include <86box/plat_unused.h>
 
 #define HAS_IDE_FUNCTIONALITY dev->ide_function
 
@@ -61,9 +62,10 @@ pc87310_log(const char *fmt, ...)
 #    define pc87310_log(fmt, ...)
 #endif
 
-typedef struct {
-    uint8_t tries, ide_function,
-        reg;
+typedef struct pc87310_t {
+    uint8_t   tries;
+    uint8_t   ide_function;
+    uint8_t   reg;
     fdc_t    *fdc;
     serial_t *uart[2];
 } pc87310_t;
@@ -97,6 +99,9 @@ lpt1_handler(pc87310_t *dev)
             lpt_port = 0x000;
             lpt_irq  = 0xff;
             break;
+
+        default:
+            break;
     }
 
     if (lpt_port)
@@ -127,7 +132,7 @@ serial_handler(pc87310_t *dev, int uart)
 }
 
 static void
-pc87310_write(uint16_t port, uint8_t val, void *priv)
+pc87310_write(UNUSED(uint16_t port), uint8_t val, void *priv)
 {
     pc87310_t *dev = (pc87310_t *) priv;
     uint8_t    valxor;
@@ -189,7 +194,7 @@ pc87310_write(uint16_t port, uint8_t val, void *priv)
 }
 
 uint8_t
-pc87310_read(uint16_t port, void *priv)
+pc87310_read(UNUSED(uint16_t port), void *priv)
 {
     pc87310_t *dev = (pc87310_t *) priv;
     uint8_t    ret = 0xff;
@@ -219,7 +224,9 @@ pc87310_reset(pc87310_t *dev)
     serial_handler(dev, 0);
     serial_handler(dev, 1);
     fdc_reset(dev->fdc);
-    // ide_pri_enable();
+#if 0
+    ide_pri_enable();
+#endif
 }
 
 static void
diff --git a/src/sio/sio_pc87311.c b/src/sio/sio_pc87311.c
index c795da8a9..143712692 100644
--- a/src/sio/sio_pc87311.c
+++ b/src/sio/sio_pc87311.c
@@ -32,6 +32,7 @@
 #include <86box/fdd.h>
 #include <86box/fdc.h>
 #include <86box/sio.h>
+#include <86box/plat_unused.h>
 
 #define HAS_IDE_FUNCTIONALITY dev->ide_function
 
@@ -64,10 +65,13 @@ pc87311_log(const char *fmt, ...)
 #    define pc87311_log(fmt, ...)
 #endif
 
-typedef struct
-{
-    uint8_t   index, regs[256], cfg_lock, ide_function;
-    uint16_t  base, irq;
+typedef struct pc87311_t {
+    uint8_t   index;
+    uint8_t   regs[256];
+    uint8_t   cfg_lock;
+    uint8_t   ide_function;
+    uint16_t  base;
+    uint16_t  irq;
     fdc_t    *fdc_controller;
     serial_t *uart[2];
 
@@ -102,15 +106,21 @@ pc87311_write(uint16_t addr, uint8_t val, void *priv)
                 case 0x02:
                     POWER_TEST = val;
                     break;
+
+                default:
+                    break;
             }
             break;
+
+        default:
+            break;
     }
 
     pc87311_enable(dev);
 }
 
 static uint8_t
-pc87311_read(uint16_t addr, void *priv)
+pc87311_read(UNUSED(uint16_t addr), void *priv)
 {
     pc87311_t *dev = (pc87311_t *) priv;
 
@@ -181,6 +191,9 @@ pc87311_uart_handler(uint8_t num, pc87311_t *dev)
             dev->base = com4(dev);
             dev->irq  = COM4_IRQ;
             break;
+
+        default:
+            break;
     }
     serial_setup(dev->uart[num & 1], dev->base, dev->irq);
     pc87311_log("PC87311-UART%01x: BASE %04x IRQ %01x\n", num & 1, dev->base, dev->irq);
@@ -203,6 +216,9 @@ pc87311_lpt_handler(pc87311_t *dev)
             dev->base = LPT2_ADDR;
             dev->irq  = LPT2_IRQ;
             break;
+
+        default:
+            break;
     }
     lpt1_init(dev->base);
     lpt1_irq(dev->irq);
diff --git a/src/sio/sio_pc87332.c b/src/sio/sio_pc87332.c
index 7f48b899b..5cbf9f694 100644
--- a/src/sio/sio_pc87332.c
+++ b/src/sio/sio_pc87332.c
@@ -35,9 +35,11 @@
 #include <86box/fdc.h>
 #include <86box/sio.h>
 
-typedef struct {
-    uint8_t tries, has_ide,
-        fdc_on, regs[15];
+typedef struct pc87332_t {
+    uint8_t   tries;
+    uint8_t   has_ide;
+    uint8_t   fdc_on;
+    uint8_t   regs[15];
     int       cur_reg;
     fdc_t    *fdc;
     serial_t *uart[2];
@@ -69,6 +71,9 @@ lpt1_handler(pc87332_t *dev)
             lpt_port = 0x000;
             lpt_irq  = 0xff;
             break;
+
+        default:
+            break;
     }
 
     if (lpt_port)
@@ -105,6 +110,9 @@ serial_handler(pc87332_t *dev, int uart)
                 case 3:
                     serial_setup(dev->uart[uart], 0x220, COM3_IRQ);
                     break;
+
+                default:
+                    break;
             }
             break;
         case 3:
@@ -121,8 +129,14 @@ serial_handler(pc87332_t *dev, int uart)
                 case 3:
                     serial_setup(dev->uart[uart], 0x228, COM4_IRQ);
                     break;
+
+                default:
+                    break;
             }
             break;
+
+        default:
+            break;
     }
 }
 
@@ -238,6 +252,9 @@ pc87332_write(uint16_t port, uint8_t val, void *priv)
                     lpt1_handler(dev);
             }
             break;
+
+        default:
+            break;
     }
 }
 
diff --git a/src/sio/sio_prime3b.c b/src/sio/sio_prime3b.c
index a204fde5b..32be854c7 100644
--- a/src/sio/sio_prime3b.c
+++ b/src/sio/sio_prime3b.c
@@ -32,6 +32,7 @@
 #include <86box/fdd.h>
 #include <86box/fdc.h>
 #include <86box/sio.h>
+#include <86box/plat_unused.h>
 
 #define FSR                   dev->regs[0xa0]
 #define ASR                   dev->regs[0xa1]
@@ -56,10 +57,13 @@ prime3b_log(const char *fmt, ...)
 #    define prime3b_log(fmt, ...)
 #endif
 
-typedef struct
-{
-    uint8_t  index, regs[256], cfg_lock, ide_function;
-    uint16_t com3_addr, com4_addr;
+typedef struct prime3b_t {
+    uint8_t  index;
+    uint8_t  regs[256];
+    uint8_t  cfg_lock;
+    uint8_t  ide_function;
+    uint16_t com3_addr;
+    uint16_t com4_addr;
 
     fdc_t    *fdc_controller;
     serial_t *uart[2];
@@ -121,17 +125,23 @@ prime3b_write(uint16_t addr, uint8_t val, void *priv)
                         dev->com3_addr = 0x220;
                         dev->com4_addr = 0x228;
                         break;
+
+                    default:
+                        break;
                 }
                 break;
             case 0xa5: /* ECP Register */
                 dev->regs[0xa5] = val;
                 break;
+
+            default:
+                break;
         }
     }
 }
 
 static uint8_t
-prime3b_read(uint16_t addr, void *priv)
+prime3b_read(UNUSED(uint16_t addr), void *priv)
 {
     prime3b_t *dev = (prime3b_t *) priv;
 
diff --git a/src/sio/sio_prime3c.c b/src/sio/sio_prime3c.c
index eae3fe1b2..160f23aab 100644
--- a/src/sio/sio_prime3c.c
+++ b/src/sio/sio_prime3c.c
@@ -32,6 +32,7 @@
 #include <86box/fdd.h>
 #include <86box/fdc.h>
 #include <86box/sio.h>
+#include <86box/plat_unused.h>
 
 #ifdef ENABLE_PRIME3C_LOG
 int prime3c_do_log = ENABLE_PRIME3C_LOG;
@@ -75,9 +76,11 @@ prime3c_log(const char *fmt, ...)
 /* IDE functionality(Note on Init) */
 #define HAS_IDE_FUNCTIONALITY dev->ide_function
 
-typedef struct
-{
-    uint8_t index, regs[256], cfg_lock, ide_function;
+typedef struct prime3c_t {
+    uint8_t index;
+    uint8_t regs[256];
+    uint8_t cfg_lock;
+    uint8_t ide_function;
 
     fdc_t    *fdc_controller;
     serial_t *uart[2];
@@ -189,14 +192,20 @@ prime3c_write(uint16_t addr, uint8_t val, void *priv)
                     case 0xd8:
                         dev->regs[dev->index] = val;
                         break;
+
+                    default:
+                        break;
                 }
             }
             break;
+
+        default:
+            break;
     }
 }
 
 static uint8_t
-prime3c_read(uint16_t addr, void *priv)
+prime3c_read(UNUSED(uint16_t addr), void *priv)
 {
     prime3c_t *dev = (prime3c_t *) priv;
 
diff --git a/src/sio/sio_um8669f.c b/src/sio/sio_um8669f.c
index bd153a9e4..256c614ae 100644
--- a/src/sio/sio_um8669f.c
+++ b/src/sio/sio_um8669f.c
@@ -37,6 +37,7 @@
 #include <86box/gameport.h>
 #include <86box/sio.h>
 #include <86box/isapnp.h>
+#include <86box/plat_unused.h>
 
 /* This ROM was reconstructed out of many assumptions, some of which based on the IT8671F. */
 static uint8_t um8669f_pnp_rom[] = {
@@ -112,7 +113,8 @@ um8669f_log(const char *fmt, ...)
 #endif
 
 typedef struct um8669f_t {
-    int                     locked, cur_reg_108;
+    int                     locked;
+    int                     cur_reg_108;
     void                   *pnp_card;
     isapnp_device_config_t *pnp_config[5];
 
@@ -184,6 +186,9 @@ um8669f_pnp_config_changed(uint8_t ld, isapnp_device_config_t *config, void *pri
                 um8669f_log("UM8669F: Game port disabled\n");
                 gameport_remap(dev->gameport, 0);
             }
+
+        default:
+            break;
     }
 }
 
@@ -263,7 +268,7 @@ um8669f_close(void *priv)
 }
 
 static void *
-um8669f_init(const device_t *info)
+um8669f_init(UNUSED(const device_t *info))
 {
     um8669f_log("UM8669F: init()\n");
 
diff --git a/src/sio/sio_vt82c686.c b/src/sio/sio_vt82c686.c
index b03207041..f22af07df 100644
--- a/src/sio/sio_vt82c686.c
+++ b/src/sio/sio_vt82c686.c
@@ -31,10 +31,17 @@
 #include <86box/fdd.h>
 #include <86box/fdc.h>
 #include <86box/sio.h>
+#include <86box/plat_unused.h>
 
-typedef struct {
-    uint8_t cur_reg, last_val, regs[25],
-        fdc_dma, fdc_irq, uart_irq[2], lpt_dma, lpt_irq;
+typedef struct vt82c686_t {
+    uint8_t   cur_reg;
+    uint8_t   last_val;
+    uint8_t   regs[25];
+    uint8_t   fdc_dma;
+    uint8_t   fdc_irq;
+    uint8_t   uart_irq[2];
+    uint8_t   lpt_dma;
+    uint8_t   lpt_irq;
     fdc_t    *fdc;
     serial_t *uart[2];
 } vt82c686_t;
@@ -189,6 +196,9 @@ vt82c686_write(uint16_t port, uint8_t val, void *priv)
             dev->regs[reg] &= 0xf7;
             vt82c686_fdc_handler(dev);
             break;
+
+        default:
+            break;
     }
 }
 
@@ -242,6 +252,9 @@ vt82c686_sio_write(uint8_t addr, uint8_t val, void *priv)
             if (val & 0x02)
                 io_sethandler(FDC_PRIMARY_ADDR, 2, vt82c686_read, NULL, NULL, vt82c686_write, NULL, NULL, dev);
             break;
+
+        default:
+            break;
     }
 }
 
@@ -272,7 +285,7 @@ vt82c686_close(void *priv)
 }
 
 static void *
-vt82c686_init(const device_t *info)
+vt82c686_init(UNUSED(const device_t *info))
 {
     vt82c686_t *dev = (vt82c686_t *) malloc(sizeof(vt82c686_t));
     memset(dev, 0, sizeof(vt82c686_t));
diff --git a/src/sio/sio_w83787f.c b/src/sio/sio_w83787f.c
index 280bab6e9..1870741c7 100644
--- a/src/sio/sio_w83787f.c
+++ b/src/sio/sio_w83787f.c
@@ -74,13 +74,16 @@ w83787_log(const char *fmt, ...)
 
 #define HAS_IDE_FUNCTIONALITY dev->ide_function
 
-typedef struct {
-    uint8_t  tries, regs[42];
+typedef struct w83787f_t {
+    uint8_t  tries;
+    uint8_t  regs[42];
     uint16_t reg_init;
-    int      locked, rw_locked,
-        cur_reg,
-        key, ide_function,
-        ide_start;
+    int      locked;
+    int      rw_locked;
+    int      cur_reg;
+    int      key;
+    int      ide_function;
+    int      ide_start;
     fdc_t    *fdc;
     serial_t *uart[2];
     void     *gameport;
@@ -342,6 +345,9 @@ w83787f_write(uint16_t port, uint8_t val, void *priv)
             if (valxor & 0x20)
                 w83787f_remap(dev);
             break;
+
+        default:
+            break;
     }
 }
 
diff --git a/src/sio/sio_w83877f.c b/src/sio/sio_w83877f.c
index 4f8b45118..8cbb82876 100644
--- a/src/sio/sio_w83877f.c
+++ b/src/sio/sio_w83877f.c
@@ -56,13 +56,16 @@
 #define PRTIQS     (dev->regs[0x27] & 0x0f)
 #define ECPIRQ     ((dev->regs[0x27] >> 5) & 0x07)
 
-typedef struct {
-    uint8_t  tries, regs[42];
-    uint16_t reg_init;
-    int      locked, rw_locked,
-        cur_reg,
-        base_address, key,
-        key_times;
+typedef struct w83877f_t {
+    uint8_t   tries;
+    uint8_t   regs[42];
+    uint16_t  reg_init;
+    int       locked;
+    int       rw_locked;
+    int       cur_reg;
+    int       base_address;
+    int       key;
+    int       key_times;
     fdc_t    *fdc;
     serial_t *uart[2];
 } w83877f_t;
@@ -140,6 +143,9 @@ make_port(w83877f_t *dev, uint8_t reg)
             if ((p < 0x100) || (p > 0x3F8))
                 p = COM2_ADDR;
             break;
+
+        default:
+            break;
     }
 
     return p;
@@ -358,6 +364,9 @@ w83877f_write(uint16_t port, uint8_t val, void *priv)
                 w83877f_serial_handler(dev, 0);
             }
             break;
+
+        default:
+            break;
     }
 }
 
diff --git a/src/sio/sio_w83977f.c b/src/sio/sio_w83977f.c
index 30a4c8b52..063f0ca69 100644
--- a/src/sio/sio_w83977f.c
+++ b/src/sio/sio_w83977f.c
@@ -36,13 +36,17 @@
 
 #define HEFRAS (dev->regs[0x26] & 0x40)
 
-typedef struct {
-    uint8_t id, tries,
-        regs[48],
-        dev_regs[256][208];
-    int locked, rw_locked,
-        cur_reg, base_address,
-        type, hefras;
+typedef struct w83977f_t {
+    uint8_t   id;
+    uint8_t   tries;
+    uint8_t   regs[48];
+    uint8_t   dev_regs[256][208];
+    int       locked;
+    int       rw_locked;
+    int       cur_reg;
+    int       base_address;
+    int       type;
+    int       hefras;
     fdc_t    *fdc;
     serial_t *uart[2];
 } w83977f_t;
@@ -145,6 +149,9 @@ w83977f_serial_handler(w83977f_t *dev, int uart)
         case 0x03:
             clock_src = 24000000.0 / 1.625;
             break;
+
+        default:
+            break;
     }
 
     serial_set_clock_src(dev->uart[uart], clock_src);
@@ -194,8 +201,10 @@ w83977f_write(uint16_t port, uint8_t val, void *priv)
 
     switch (dev->cur_reg) {
         case 0x02:
-            /* if (valxor & 0x02)
-                    softresetx86(); */
+#if 0
+            if (valxor & 0x02)
+                    softresetx86();
+#endif
             break;
         case 0x22:
             if (valxor & 0x20)
@@ -226,6 +235,9 @@ w83977f_write(uint16_t port, uint8_t val, void *priv)
                     case 0x03:
                         w83977f_serial_handler(dev, ld - 2);
                         break;
+
+                    default:
+                        break;
                 }
             break;
         case 0x60:
@@ -242,6 +254,9 @@ w83977f_write(uint16_t port, uint8_t val, void *priv)
                     case 0x03:
                         w83977f_serial_handler(dev, ld - 2);
                         break;
+
+                    default:
+                        break;
                 }
             break;
         case 0x70:
@@ -257,6 +272,9 @@ w83977f_write(uint16_t port, uint8_t val, void *priv)
                     case 0x03:
                         w83977f_serial_handler(dev, ld - 2);
                         break;
+
+                    default:
+                        break;
                 }
             break;
         case 0xf0:
@@ -281,6 +299,9 @@ w83977f_write(uint16_t port, uint8_t val, void *priv)
                     if (valxor & 0x03)
                         w83977f_serial_handler(dev, ld - 2);
                     break;
+
+                default:
+                    break;
             }
             break;
         case 0xf1:
@@ -298,6 +319,9 @@ w83977f_write(uint16_t port, uint8_t val, void *priv)
                     if (!dev->id && (valxor & 0x01))
                         fdc_set_swwp(dev->fdc, (val & 0x01) ? 1 : 0);
                     break;
+
+                default:
+                    break;
             }
             break;
         case 0xf2:
@@ -315,6 +339,9 @@ w83977f_write(uint16_t port, uint8_t val, void *priv)
                     if (!dev->id && (valxor & 0x03))
                         fdc_update_rwc(dev->fdc, 0, val & 0x03);
                     break;
+
+                default:
+                    break;
             }
             break;
         case 0xf4:
@@ -329,8 +356,14 @@ w83977f_write(uint16_t port, uint8_t val, void *priv)
                     if (!dev->id && (valxor & 0x18))
                         fdc_update_drvrate(dev->fdc, dev->cur_reg & 0x03, (val & 0x18) >> 3);
                     break;
+
+                default:
+                    break;
             }
             break;
+
+        default:
+            break;
     }
 }
 
diff --git a/src/sound/CMakeLists.txt b/src/sound/CMakeLists.txt
index c530d7ac8..ec4f8823b 100644
--- a/src/sound/CMakeLists.txt
+++ b/src/sound/CMakeLists.txt
@@ -85,12 +85,13 @@ if(RTMIDI)
 endif()
 
 if(FLUIDSYNTH)
-    if(APPLE)
-        find_library(FLUIDSYNTH_LIB fluidsynth)
-        if (NOT FLUIDSYNTH_LIB)
-            message(WARNING "Could not find fluid synth. The library will not be bundled and any related features will not work.")
-        endif()
-    endif ()
+    find_package(PkgConfig REQUIRED)
+    pkg_check_modules(FLUIDSYNTH REQUIRED IMPORTED_TARGET fluidsynth)
+    target_link_libraries(86Box PkgConfig::FLUIDSYNTH)
+    if(STATIC_BUILD)
+        target_link_libraries(86Box -static ${FLUIDSYNTH_STATIC_LIBRARIES} -fopenmp)
+    endif()
+
     target_compile_definitions(snd PRIVATE USE_FLUIDSYNTH)
     target_sources(snd PRIVATE midi_fluidsynth.c)
 endif()
diff --git a/src/sound/midi_fluidsynth.c b/src/sound/midi_fluidsynth.c
index 2372203d7..efe8159fc 100644
--- a/src/sound/midi_fluidsynth.c
+++ b/src/sound/midi_fluidsynth.c
@@ -8,104 +8,33 @@
 #    ifdef __unix__
 #        include <unistd.h>
 #    endif
+#    define FLUIDSYNTH_NOT_A_DLL
+#    include <fluidsynth.h>
 
 #    include <86box/86box.h>
 #    include <86box/config.h>
 #    include <86box/device.h>
 #    include <86box/midi.h>
-#    include <86box/plat.h>
-#    include <86box/plat_dynld.h>
 #    include <86box/thread.h>
 #    include <86box/sound.h>
-#    include <86box/ui.h>
-
-#    define FLUID_CHORUS_DEFAULT_N     3
-#    define FLUID_CHORUS_DEFAULT_LEVEL 2.0f
-#    define FLUID_CHORUS_DEFAULT_SPEED 0.3f
-#    define FLUID_CHORUS_DEFAULT_DEPTH 8.0f
-#    define FLUID_CHORUS_DEFAULT_TYPE  FLUID_CHORUS_MOD_SINE
 
 #    define RENDER_RATE                100
 #    define BUFFER_SEGMENTS            10
 
-enum fluid_chorus_mod {
-    FLUID_CHORUS_MOD_SINE     = 0,
-    FLUID_CHORUS_MOD_TRIANGLE = 1
-};
-
-enum fluid_interp {
-    FLUID_INTERP_NONE     = 0,
-    FLUID_INTERP_LINEAR   = 1,
-    FLUID_INTERP_DEFAULT  = 4,
-    FLUID_INTERP_4THORDER = 4,
-    FLUID_INTERP_7THORDER = 7,
-    FLUID_INTERP_HIGHEST  = 7
-};
+/* Check the FluidSynth version to determine wheteher to use the older reverb/chorus
+   control functions that were deprecated in 2.2.0, or their newer replacements */
+#    if (FLUIDSYNTH_VERSION_MAJOR < 2) || ((FLUIDSYNTH_VERSION_MAJOR == 2) && (FLUIDSYNTH_VERSION_MINOR < 2))
+#        define USE_OLD_FLUIDSYNTH_API
+#    endif
 
 extern void givealbuffer_midi(void *buf, uint32_t size);
 extern void al_set_midi(int freq, int buf_size);
 
-static void *fluidsynth_handle; /* handle to FluidSynth DLL */
-
-/* Pointers to the real functions. */
-// clang-format off
-static void *(*f_new_fluid_settings)(void);
-static void  (*f_delete_fluid_settings)(void *settings);
-static int   (*f_fluid_settings_setnum)(void *settings, const char *name, double val);
-static int   (*f_fluid_settings_getnum)(void *settings, const char *name, double *val);
-static void *(*f_new_fluid_synth)(void *settings);
-static int   (*f_delete_fluid_synth)(void *synth);
-static int   (*f_fluid_synth_noteon)(void *synth, int chan, int key, int vel);
-static int   (*f_fluid_synth_noteoff)(void *synth, int chan, int key);
-static int   (*f_fluid_synth_cc)(void *synth, int chan, int ctrl, int val);
-static int   (*f_fluid_synth_channel_pressure)(void *synth, int chan, int val);
-static int   (*f_fluid_synth_sysex)(void *synth, const char *data, int len, char *response, int *response_len, int *handled, int dryrun);
-static int   (*f_fluid_synth_pitch_bend)(void *synth, int chan, int val);
-static int   (*f_fluid_synth_program_change)(void *synth, int chan, int program);
-static int   (*f_fluid_synth_sfload)(void *synth, const char *filename, int reset_presets);
-static int   (*f_fluid_synth_set_interp_method)(void *synth, int chan, int interp_method);
-static void  (*f_fluid_synth_set_reverb)(void *synth, double roomsize, double damping, double width, double level);
-static void  (*f_fluid_synth_set_reverb_on)(void *synth, int on);
-static void  (*f_fluid_synth_set_chorus)(void *synth, int nr, double level, double speed, double depth_ms, int type);
-static void  (*f_fluid_synth_set_chorus_on)(void *synth, int on);
-static int   (*f_fluid_synth_write_s16)(void *synth, int len, void *lout, int loff, int lincr, void *rout, int roff, int rincr);
-static int   (*f_fluid_synth_write_float)(void *synth, int len, void *lout, int loff, int lincr, void *rout, int roff, int rincr);
-static char *(*f_fluid_version_str)(void);
-// clang-format on
-
-static dllimp_t fluidsynth_imports[] = {
-  // clang-format off
-  { "new_fluid_settings",            &f_new_fluid_settings            },
-  { "delete_fluid_settings",         &f_delete_fluid_settings         },
-  { "fluid_settings_setnum",         &f_fluid_settings_setnum         },
-  { "fluid_settings_getnum",         &f_fluid_settings_getnum         },
-  { "new_fluid_synth",               &f_new_fluid_synth               },
-  { "delete_fluid_synth",            &f_delete_fluid_synth            },
-  { "fluid_synth_noteon",            &f_fluid_synth_noteon            },
-  { "fluid_synth_noteoff",           &f_fluid_synth_noteoff           },
-  { "fluid_synth_cc",                &f_fluid_synth_cc                },
-  { "fluid_synth_channel_pressure",  &f_fluid_synth_channel_pressure  },
-  { "fluid_synth_sysex",             &f_fluid_synth_sysex             },
-  { "fluid_synth_pitch_bend",        &f_fluid_synth_pitch_bend        },
-  { "fluid_synth_program_change",    &f_fluid_synth_program_change    },
-  { "fluid_synth_sfload",            &f_fluid_synth_sfload            },
-  { "fluid_synth_set_interp_method", &f_fluid_synth_set_interp_method },
-  { "fluid_synth_set_reverb",        &f_fluid_synth_set_reverb        },
-  { "fluid_synth_set_reverb_on",     &f_fluid_synth_set_reverb_on     },
-  { "fluid_synth_set_chorus",        &f_fluid_synth_set_chorus        },
-  { "fluid_synth_set_chorus_on",     &f_fluid_synth_set_chorus_on     },
-  { "fluid_synth_write_s16",         &f_fluid_synth_write_s16         },
-  { "fluid_synth_write_float",       &f_fluid_synth_write_float       },
-  { "fluid_version_str",             &f_fluid_version_str             },
-  { NULL,                            NULL                             },
-  // clang-format on
-};
-
 typedef struct fluidsynth {
-    void *settings;
-    void *synth;
-    int   samplerate;
-    int   sound_font;
+    fluid_settings_t *settings;
+    fluid_synth_t    *synth;
+    int               samplerate;
+    int               sound_font;
 
     thread_t *thread_h;
     event_t  *event, *start_event;
@@ -153,7 +82,7 @@ fluidsynth_thread(void *param)
             float *buf = (float *) ((uint8_t *) data->buffer + buf_pos);
             memset(buf, 0, buf_size);
             if (data->synth)
-                f_fluid_synth_write_float(data->synth, buf_size / (2 * sizeof(float)), buf, 0, 2, buf, 1, 2);
+                fluid_synth_write_float(data->synth, buf_size / (2 * sizeof(float)), buf, 0, 2, buf, 1, 2);
             buf_pos += buf_size;
             if (buf_pos >= data->buf_size) {
                 givealbuffer_midi(data->buffer, data->buf_size / sizeof(float));
@@ -163,7 +92,7 @@ fluidsynth_thread(void *param)
             int16_t *buf = (int16_t *) ((uint8_t *) data->buffer_int16 + buf_pos);
             memset(buf, 0, buf_size);
             if (data->synth)
-                f_fluid_synth_write_s16(data->synth, buf_size / (2 * sizeof(int16_t)), buf, 0, 2, buf, 1, 2);
+                fluid_synth_write_s16(data->synth, buf_size / (2 * sizeof(int16_t)), buf, 0, 2, buf, 1, 2);
             buf_pos += buf_size;
             if (buf_pos >= data->buf_size) {
                 givealbuffer_midi(data->buffer_int16, data->buf_size / sizeof(int16_t));
@@ -187,24 +116,24 @@ fluidsynth_msg(uint8_t *msg)
 
     switch (cmd) {
         case 0x80: /* Note Off */
-            f_fluid_synth_noteoff(data->synth, chan, param1);
+            fluid_synth_noteoff(data->synth, chan, param1);
             break;
         case 0x90: /* Note On */
-            f_fluid_synth_noteon(data->synth, chan, param1, param2);
+            fluid_synth_noteon(data->synth, chan, param1, param2);
             break;
         case 0xA0: /* Aftertouch */
             break;
         case 0xB0: /* Control Change */
-            f_fluid_synth_cc(data->synth, chan, param1, param2);
+            fluid_synth_cc(data->synth, chan, param1, param2);
             break;
         case 0xC0: /* Program Change */
-            f_fluid_synth_program_change(data->synth, chan, param1);
+            fluid_synth_program_change(data->synth, chan, param1);
             break;
         case 0xD0: /* Channel Pressure */
-            f_fluid_synth_channel_pressure(data->synth, chan, param1);
+            fluid_synth_channel_pressure(data->synth, chan, param1);
             break;
         case 0xE0: /* Pitch Bend */
-            f_fluid_synth_pitch_bend(data->synth, chan, (param2 << 7) | param1);
+            fluid_synth_pitch_bend(data->synth, chan, (param2 << 7) | param1);
             break;
         case 0xF0: /* SysEx */
             break;
@@ -218,7 +147,7 @@ fluidsynth_sysex(uint8_t *data, unsigned int len)
 {
     fluidsynth_t *d = &fsdev;
 
-    f_fluid_synth_sysex(d->synth, (const char *) data, len, 0, 0, 0, 0);
+    fluid_synth_sysex(d->synth, (const char *) data, len, 0, 0, 0, 0);
 }
 
 void *
@@ -229,31 +158,12 @@ fluidsynth_init(const device_t *info)
 
     memset(data, 0, sizeof(fluidsynth_t));
 
-    /* Try loading the DLL. */
-#    ifdef _WIN32
-#        if (!(defined __amd64__ || defined _M_X64 || defined __aarch64__ || defined _M_ARM64))
-    fluidsynth_handle = dynld_module("libfluidsynth.dll", fluidsynth_imports);
-#        else
-    fluidsynth_handle = dynld_module("libfluidsynth64.dll", fluidsynth_imports);
-#        endif
-#    elif defined __APPLE__
-    fluidsynth_handle = dynld_module("libfluidsynth.dylib", fluidsynth_imports);
-#    else
-    fluidsynth_handle = dynld_module("libfluidsynth.so.3", fluidsynth_imports);
-    if (fluidsynth_handle == NULL)
-        fluidsynth_handle = dynld_module("libfluidsynth.so.2", fluidsynth_imports);
-#    endif
-    if (fluidsynth_handle == NULL) {
-        ui_msgbox_header(MBX_ERROR, (wchar_t *) IDS_2080, (wchar_t *) IDS_2134);
-        return NULL;
-    }
+    data->settings = new_fluid_settings();
 
-    data->settings = f_new_fluid_settings();
+    fluid_settings_setnum(data->settings, "synth.sample-rate", 44100);
+    fluid_settings_setnum(data->settings, "synth.gain", device_get_config_int("output_gain") / 100.0f);
 
-    f_fluid_settings_setnum(data->settings, "synth.sample-rate", 44100);
-    f_fluid_settings_setnum(data->settings, "synth.gain", device_get_config_int("output_gain") / 100.0f);
-
-    data->synth = f_new_fluid_synth(data->settings);
+    data->synth = new_fluid_synth(data->settings);
 
     const char *sound_font = (char *) device_get_config_string("sound_font");
 #    ifdef __unix__
@@ -261,10 +171,14 @@ fluidsynth_init(const device_t *info)
         sound_font = (access("/usr/share/sounds/sf2/FluidR3_GM.sf2", F_OK) == 0 ? "/usr/share/sounds/sf2/FluidR3_GM.sf2" :
                       (access("/usr/share/soundfonts/default.sf2", F_OK) == 0 ? "/usr/share/soundfonts/default.sf2" : ""));
 #    endif
-    data->sound_font = f_fluid_synth_sfload(data->synth, sound_font, 1);
+    data->sound_font = fluid_synth_sfload(data->synth, sound_font, 1);
 
     if (device_get_config_int("chorus")) {
-        f_fluid_synth_set_chorus_on(data->synth, 1);
+#    ifndef USE_OLD_FLUIDSYNTH_API
+        fluid_synth_chorus_on(data->synth, -1, 1);
+#    else
+        fluid_synth_set_chorus_on(data->synth, 1);
+#    endif
 
         int    chorus_voices = device_get_config_int("chorus_voices");
         double chorus_level  = device_get_config_int("chorus_level") / 100.0;
@@ -277,21 +191,48 @@ fluidsynth_init(const device_t *info)
         else
             chorus_waveform = FLUID_CHORUS_MOD_TRIANGLE;
 
-        f_fluid_synth_set_chorus(data->synth, chorus_voices, chorus_level, chorus_speed, chorus_depth, chorus_waveform);
+#    ifndef USE_OLD_FLUIDSYNTH_API
+        fluid_synth_set_chorus_group_nr(data->synth, -1, chorus_voices);
+        fluid_synth_set_chorus_group_level(data->synth, -1, chorus_level);
+        fluid_synth_set_chorus_group_speed(data->synth, -1, chorus_speed);
+        fluid_synth_set_chorus_group_depth(data->synth, -1, chorus_depth);
+        fluid_synth_set_chorus_group_type(data->synth, -1, chorus_waveform);
+#    else
+        fluid_synth_set_chorus(data->synth, chorus_voices, chorus_level, chorus_speed, chorus_depth, chorus_waveform);
+#    endif
     } else
-        f_fluid_synth_set_chorus_on(data->synth, 0);
+#    ifndef USE_OLD_FLUIDSYNTH_API
+        fluid_synth_chorus_on(data->synth, -1, 0);
+#    else
+        fluid_synth_set_chorus_on(data->synth, 0);
+#    endif
 
     if (device_get_config_int("reverb")) {
-        f_fluid_synth_set_reverb_on(data->synth, 1);
+#    ifndef USE_OLD_FLUIDSYNTH_API
+        fluid_synth_reverb_on(data->synth, -1, 1);
+#    else
+        fluid_synth_set_reverb_on(data->synth, 1);
+#    endif
 
         double reverb_room_size = device_get_config_int("reverb_room_size") / 100.0;
         double reverb_damping   = device_get_config_int("reverb_damping") / 100.0;
-        int    reverb_width     = device_get_config_int("reverb_width");
+        double reverb_width     = device_get_config_int("reverb_width") / 10.0;
         double reverb_level     = device_get_config_int("reverb_level") / 100.0;
 
-        f_fluid_synth_set_reverb(data->synth, reverb_room_size, reverb_damping, reverb_width, reverb_level);
+#    ifndef USE_OLD_FLUIDSYNTH_API
+        fluid_synth_set_reverb_group_roomsize(data->synth, -1, reverb_room_size);
+        fluid_synth_set_reverb_group_damp(data->synth, -1, reverb_damping);
+        fluid_synth_set_reverb_group_width(data->synth, -1, reverb_width);
+        fluid_synth_set_reverb_group_level(data->synth, -1, reverb_level);
+#    else
+        fluid_synth_set_reverb(data->synth, reverb_room_size, reverb_damping, reverb_width, reverb_level);
+#    endif
     } else
-        f_fluid_synth_set_reverb_on(data->synth, 0);
+#    ifndef USE_OLD_FLUIDSYNTH_API
+        fluid_synth_reverb_on(data->synth, -1, 0);
+#    else
+        fluid_synth_set_reverb_on(data->synth, 0);
+#    endif
 
     int interpolation    = device_get_config_int("interpolation");
     int fs_interpolation = FLUID_INTERP_4THORDER;
@@ -305,10 +246,10 @@ fluidsynth_init(const device_t *info)
     else if (interpolation == 3)
         fs_interpolation = FLUID_INTERP_7THORDER;
 
-    f_fluid_synth_set_interp_method(data->synth, -1, fs_interpolation);
+    fluid_synth_set_interp_method(data->synth, -1, fs_interpolation);
 
     double samplerate;
-    f_fluid_settings_getnum(data->settings, "synth.sample-rate", &samplerate);
+    fluid_settings_getnum(data->settings, "synth.sample-rate", &samplerate);
     data->samplerate = (int) samplerate;
     if (sound_is_float) {
         data->buf_size     = (data->samplerate / RENDER_RATE) * 2 * sizeof(float) * BUFFER_SEGMENTS;
@@ -357,12 +298,12 @@ fluidsynth_close(void *p)
     thread_wait(data->thread_h);
 
     if (data->synth) {
-        f_delete_fluid_synth(data->synth);
+        delete_fluid_synth(data->synth);
         data->synth = NULL;
     }
 
     if (data->settings) {
-        f_delete_fluid_settings(data->settings);
+        delete_fluid_settings(data->settings);
         data->settings = NULL;
     }
 
@@ -375,12 +316,6 @@ fluidsynth_close(void *p)
         free(data->buffer_int16);
         data->buffer_int16 = NULL;
     }
-
-    /* Unload the DLL if possible. */
-    if (fluidsynth_handle != NULL) {
-        dynld_close(fluidsynth_handle);
-        fluidsynth_handle = NULL;
-    }
 }
 
 static const device_config_t fluidsynth_config[] = {
@@ -407,7 +342,7 @@ static const device_config_t fluidsynth_config[] = {
         .name = "chorus",
         .description = "Chorus",
         .type = CONFIG_BINARY,
-        .default_int = 0
+        .default_int = 1
     },
     {
         .name = "chorus_voices",
@@ -429,7 +364,7 @@ static const device_config_t fluidsynth_config[] = {
             .min = 0,
             .max = 100
         },
-        .default_int = 100
+        .default_int = 20
     },
     {
         .name = "chorus_speed",
@@ -437,7 +372,7 @@ static const device_config_t fluidsynth_config[] = {
         .type = CONFIG_SPINNER,
         .spinner =
         {
-            .min = 30,
+            .min = 10,
             .max = 500
         },
         .default_int = 30
@@ -449,7 +384,7 @@ static const device_config_t fluidsynth_config[] = {
         .spinner =
         {
             .min = 0,
-            .max = 210
+            .max = 2560
         },
         .default_int = 80
     },
@@ -474,7 +409,7 @@ static const device_config_t fluidsynth_config[] = {
         .name = "reverb",
         .description = "Reverb",
         .type = CONFIG_BINARY,
-        .default_int = 0
+        .default_int = 1
     },
     {
         .name = "reverb_room_size",
@@ -483,7 +418,7 @@ static const device_config_t fluidsynth_config[] = {
         .spinner =
         {
             .min = 0,
-            .max = 120
+            .max = 100
         },
         .default_int = 20
     },
@@ -505,9 +440,9 @@ static const device_config_t fluidsynth_config[] = {
         .spinner =
         {
             .min = 0,
-            .max = 100
+            .max = 1000
         },
-        .default_int = 1
+        .default_int = 5
     },
     {
         .name = "reverb_level",
diff --git a/src/sound/snd_adlibgold.c b/src/sound/snd_adlibgold.c
index c1b2c3867..e4330cba3 100644
--- a/src/sound/snd_adlibgold.c
+++ b/src/sound/snd_adlibgold.c
@@ -20,7 +20,7 @@
 
 typedef struct adgold_t {
     int adgold_irq_status;
-    int irq, dma, hdma;
+    int irq, dma;
 
     uint8_t adgold_eeprom[0x1a];
 
@@ -547,6 +547,7 @@ adgold_read(uint16_t addr, void *p)
             temp = adgold->adgold_mma_status;
             adgold->adgold_mma_status &= ~0xf3; /*JUKEGOLD expects timer status flags to auto-clear*/
             adgold_update_irq_status(adgold);
+            picintc(1 << adgold->irq);
             break;
         case 5:
             if (adgold->adgold_mma_addr >= 0xf)
@@ -652,7 +653,8 @@ adgold_timer_poll(void *p)
 {
     adgold_t *adgold = (adgold_t *) p;
 
-    timer_advance_u64(&adgold->adgold_mma_timer_count, (uint64_t) ((double) TIMER_USEC * 1.88964));
+	/*A small timer period will result in hangs.*/
+    timer_on_auto(&adgold->adgold_mma_timer_count, 4.88964);
 
     if (adgold->adgold_midi_ctrl & 0x3f) {
         if ((adgold->adgold_midi_ctrl & 0x3f) != 0x3f) {
@@ -925,7 +927,7 @@ adgold_init(const device_t *info)
     adgold->adgold_eeprom[0x10] = 0xff;
     adgold->adgold_eeprom[0x11] = 0x20;
     adgold->adgold_eeprom[0x12] = 0x00;
-    adgold->adgold_eeprom[0x13] = 0xa0;
+    adgold->adgold_eeprom[0x13] = 0x00;
     adgold->adgold_eeprom[0x14] = 0x00;
     adgold->adgold_eeprom[0x15] = 0x388 / 8; /*Present at 388-38f*/
     adgold->adgold_eeprom[0x16] = 0x00;
@@ -957,6 +959,7 @@ adgold_init(const device_t *info)
             break;
     }
     adgold->adgold_eeprom[0x13] |= (adgold->dma << 3);
+    adgold->adgold_eeprom[0x14] |= (adgold->dma << 4);
     memcpy(adgold->adgold_38x_regs, adgold->adgold_eeprom, 0x19);
     adgold->vol_l      = attenuation[adgold->adgold_eeprom[0x04] & 0x3f];
     adgold->vol_r      = attenuation[adgold->adgold_eeprom[0x05] & 0x3f];
diff --git a/src/timer.c b/src/timer.c
index 2d75e532e..dde6fbdbc 100644
--- a/src/timer.c
+++ b/src/timer.c
@@ -159,12 +159,12 @@ timer_init(void)
 }
 
 void
-timer_add(pc_timer_t *timer, void (*callback)(void *p), void *p, int start_timer)
+timer_add(pc_timer_t *timer, void (*callback)(void *p), void *priv, int start_timer)
 {
     memset(timer, 0, sizeof(pc_timer_t));
 
     timer->callback = callback;
-    timer->p        = p;
+    timer->p        = priv;
     timer->flags    = 0;
     timer->prev = timer->next = NULL;
     if (start_timer)
diff --git a/src/unix/unix.c b/src/unix/unix.c
index 4579f21d0..c9be5d80f 100644
--- a/src/unix/unix.c
+++ b/src/unix/unix.c
@@ -36,6 +36,7 @@
 #include <86box/unix_sdl.h>
 #include <86box/timer.h>
 #include <86box/nvr.h>
+#include <86box/version.h>
 #include <86box/video.h>
 #include <86box/ui.h>
 #include <86box/gdbstub.h>
diff --git a/src/upi42.c b/src/upi42.c
index e35a0752f..79f4ec9c2 100644
--- a/src/upi42.c
+++ b/src/upi42.c
@@ -18,6 +18,7 @@
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
+#include <86box/plat_unused.h>
 
 #ifdef UPI42_STANDALONE
 #    define fatal(...)              \
@@ -160,7 +161,7 @@ upi42_op_MOV_A_imm(upi42_t *upi42, uint32_t fetchdat)
 }
 
 static int
-upi42_op_MOV_A_PSW(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_MOV_A_PSW(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->a = upi42->psw;
     upi42_mirror_f0(upi42);
@@ -168,7 +169,7 @@ upi42_op_MOV_A_PSW(upi42_t *upi42, uint32_t fetchdat)
 }
 
 static int
-upi42_op_MOV_PSW_A(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_MOV_PSW_A(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->psw = upi42->a;
     upi42_mirror_f0(upi42);
@@ -176,28 +177,28 @@ upi42_op_MOV_PSW_A(upi42_t *upi42, uint32_t fetchdat)
 }
 
 static int
-upi42_op_MOV_A_T(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_MOV_A_T(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->a = upi42->t;
     return 1;
 }
 
 static int
-upi42_op_MOV_T_A(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_MOV_T_A(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->t = upi42->a;
     return 1;
 }
 
 static int
-upi42_op_MOV_STS_A(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_MOV_STS_A(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->sts = (upi42->a & 0xf0) | (upi42->sts & 0x0f);
     return 1;
 }
 
 static int
-upi42_op_MOVP_A_indA(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_MOVP_A_indA(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->a = upi42->rom[(upi42->pc & 0xff00) | upi42->a];
     upi42->cycs--;
@@ -205,7 +206,7 @@ upi42_op_MOVP_A_indA(upi42_t *upi42, uint32_t fetchdat)
 }
 
 static int
-upi42_op_MOVP3_A_indA(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_MOVP3_A_indA(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->a = upi42->rom[0x300 | upi42->a];
     upi42->cycs--;
@@ -242,7 +243,7 @@ upi42_op_XCHD_A_indRr(upi42_t *upi42, uint32_t fetchdat)
 }
 
 static int
-upi42_op_SWAP_A(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_SWAP_A(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->a = (upi42->a << 4) | (upi42->a >> 4);
     return 1;
@@ -258,7 +259,7 @@ upi42_op_IN_A_Pp(upi42_t *upi42, uint32_t fetchdat)
 }
 
 static int
-upi42_op_IN_A_DBB(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_IN_A_DBB(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->a = upi42->dbb_in;
     upi42->sts &= ~0x02; /* clear IBF */
@@ -266,7 +267,7 @@ upi42_op_IN_A_DBB(upi42_t *upi42, uint32_t fetchdat)
 }
 
 static int
-upi42_op_OUTL_Pp_A(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_OUTL_Pp_A(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->ports_out[fetchdat & 3] = upi42->a;
     upi42->cycs--;
@@ -274,7 +275,7 @@ upi42_op_OUTL_Pp_A(upi42_t *upi42, uint32_t fetchdat)
 }
 
 static int
-upi42_op_OUT_DBB_A(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_OUT_DBB_A(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->dbb_out = upi42->a;
     upi42->sts |= 0x01; /* set OBF */
@@ -397,21 +398,21 @@ upi42_op_ORLD_Pp_A(upi42_t *upi42, uint32_t fetchdat)
 }
 
 static int
-upi42_op_RR_A(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_RR_A(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->a = (upi42->a << 7) | (upi42->a >> 1);
     return 1;
 }
 
 static int
-upi42_op_RL_A(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_RL_A(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->a = (upi42->a >> 7) | (upi42->a << 1);
     return 1;
 }
 
 static int
-upi42_op_RRC_A(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_RRC_A(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     uint8_t temp = upi42->a;
     upi42->a     = (upi42->psw & 0x80) | (temp >> 1);
@@ -420,7 +421,7 @@ upi42_op_RRC_A(upi42_t *upi42, uint32_t fetchdat)
 }
 
 static int
-upi42_op_RLC_A(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_RLC_A(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     uint8_t temp = upi42->a;
     upi42->a     = (temp << 1) | (upi42->psw >> 7);
@@ -429,7 +430,7 @@ upi42_op_RLC_A(upi42_t *upi42, uint32_t fetchdat)
 }
 
 static int
-upi42_op_INC_A(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_INC_A(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->a++;
     return 1;
@@ -450,7 +451,7 @@ upi42_op_INC_indRr(upi42_t *upi42, uint32_t fetchdat)
 }
 
 static int
-upi42_op_DEC_A(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_DEC_A(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->a--;
     return 1;
@@ -533,21 +534,21 @@ upi42_op_ADDC_A_imm(upi42_t *upi42, uint32_t fetchdat)
 }
 
 static int
-upi42_op_CLR_A(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_CLR_A(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->a = 0;
     return 1;
 }
 
 static int
-upi42_op_CPL_A(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_CPL_A(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->a = ~upi42->a;
     return 1;
 }
 
 static int
-upi42_op_DA_A(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_DA_A(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     if (((upi42->a & 0x0f) > 9) || (upi42->psw & 0x40))
         upi42->a += 6;
@@ -560,21 +561,21 @@ upi42_op_DA_A(upi42_t *upi42, uint32_t fetchdat)
 }
 
 static int
-upi42_op_CLR_C(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_CLR_C(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->psw &= ~0x80;
     return 1;
 }
 
 static int
-upi42_op_CPL_C(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_CPL_C(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->psw ^= 0x80;
     return 1;
 }
 
 static int
-upi42_op_CLR_F0(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_CLR_F0(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->psw &= ~0x20;
     upi42_mirror_f0(upi42);
@@ -582,7 +583,7 @@ upi42_op_CLR_F0(upi42_t *upi42, uint32_t fetchdat)
 }
 
 static int
-upi42_op_CPL_F0(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_CPL_F0(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->psw ^= 0x20;
     upi42_mirror_f0(upi42);
@@ -590,21 +591,21 @@ upi42_op_CPL_F0(upi42_t *upi42, uint32_t fetchdat)
 }
 
 static int
-upi42_op_CLR_F1(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_CLR_F1(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->sts &= ~0x08;
     return 1;
 }
 
 static int
-upi42_op_CPL_F1(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_CPL_F1(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->sts ^= 0x08;
     return 1;
 }
 
 static int
-upi42_op_EN_I(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_EN_I(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->i              = 1;
     upi42->skip_timer_inc = 1;
@@ -612,7 +613,7 @@ upi42_op_EN_I(upi42_t *upi42, uint32_t fetchdat)
 }
 
 static int
-upi42_op_DIS_I(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_DIS_I(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->i              = 0;
     upi42->skip_timer_inc = 1;
@@ -620,21 +621,21 @@ upi42_op_DIS_I(upi42_t *upi42, uint32_t fetchdat)
 }
 
 static int
-upi42_op_EN_TCNTI(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_EN_TCNTI(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->tcnti = 1;
     return 1;
 }
 
 static int
-upi42_op_DIS_TCNTI(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_DIS_TCNTI(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->tcnti = upi42->tcnti_raise = 0;
     return 1;
 }
 
 static int
-upi42_op_STRT_T(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_STRT_T(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->run_timer      = 1;
     upi42->prescaler      = 0;
@@ -643,7 +644,7 @@ upi42_op_STRT_T(upi42_t *upi42, uint32_t fetchdat)
 }
 
 static int
-upi42_op_STRT_CNT(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_STRT_CNT(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->run_counter    = 1;
     upi42->skip_timer_inc = 1;
@@ -651,7 +652,7 @@ upi42_op_STRT_CNT(upi42_t *upi42, uint32_t fetchdat)
 }
 
 static int
-upi42_op_STOP_TCNT(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_STOP_TCNT(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->run_timer = upi42->run_counter = 0;
     upi42->skip_timer_inc                 = 1;
@@ -659,35 +660,35 @@ upi42_op_STOP_TCNT(upi42_t *upi42, uint32_t fetchdat)
 }
 
 static int
-upi42_op_SEL_PMB0(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_SEL_PMB0(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->dbf = 0;
     return 1;
 }
 
 static int
-upi42_op_SEL_PMB1(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_SEL_PMB1(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->dbf = 1;
     return 1;
 }
 
 static int
-upi42_op_SEL_RB0(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_SEL_RB0(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->psw &= ~0x10;
     return 1;
 }
 
 static int
-upi42_op_SEL_RB1(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_SEL_RB1(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->psw |= 0x10;
     return 1;
 }
 
 static int
-upi42_op_NOP(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_NOP(UNUSED(upi42_t *upi42), UNUSED(uint32_t fetchdat))
 {
     return 1;
 }
@@ -741,7 +742,7 @@ upi42_op_JMP_imm(upi42_t *upi42, uint32_t fetchdat)
 }
 
 static int
-upi42_op_JMPP_indA(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_JMPP_indA(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->pc = (upi42->pc & 0xff00) | upi42->a;
     upi42->cycs--;
@@ -774,27 +775,27 @@ UPI42_COND_JMP_IMM(JNIBF, !(upi42->sts & 0x02), )
 UPI42_COND_JMP_IMM(JOBF, upi42->sts & 0x01, )
 
 static int
-upi42_op_EN_A20(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_EN_A20(UNUSED(upi42_t *upi42), UNUSED(uint32_t fetchdat))
 {
     /* Enable fast A20 until reset. */
     return 1;
 }
 
 static int
-upi42_op_EN_DMA(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_EN_DMA(UNUSED(upi42_t *upi42), UNUSED(uint32_t fetchdat))
 {
     return 1;
 }
 
 static int
-upi42_op_EN_FLAGS(upi42_t *upi42, uint32_t fetchdat)
+upi42_op_EN_FLAGS(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     upi42->flags = 1;
     return 1;
 }
 
 static int
-upi42_op_SUSPEND(upi42_t *upi42, uint32_t fetchdatr)
+upi42_op_SUSPEND(upi42_t *upi42, UNUSED(uint32_t fetchdat))
 {
     /* Inhibit execution until reset. */
     upi42->suspend = 1;
@@ -923,6 +924,8 @@ upi42_port_read(void *priv, int port)
 
     /* Apply special meanings. */
     switch (port) {
+        default:
+            break;
     }
 
     upi42_log("UPI42: port_read(%d) = %02X\n", port, ret);
@@ -944,7 +947,7 @@ upi42_port_write(void *priv, int port, uint8_t val)
 /* NOTE: The dbb/sts/cmd functions use I/O handler signatures; port is ignored. */
 
 uint8_t
-upi42_dbb_read(uint16_t port, void *priv)
+upi42_dbb_read(UNUSED(uint16_t port), void *priv)
 {
     upi42_t *upi42 = (upi42_t *) priv;
 
@@ -955,7 +958,7 @@ upi42_dbb_read(uint16_t port, void *priv)
 }
 
 void
-upi42_dbb_write(uint16_t port, uint8_t val, void *priv)
+upi42_dbb_write(UNUSED(uint16_t port), uint8_t val, void *priv)
 {
     upi42_t *upi42 = (upi42_t *) priv;
 
@@ -967,7 +970,7 @@ upi42_dbb_write(uint16_t port, uint8_t val, void *priv)
 }
 
 uint8_t
-upi42_sts_read(uint16_t port, void *priv)
+upi42_sts_read(UNUSED(uint16_t port), void *priv)
 {
     upi42_t *upi42 = (upi42_t *) priv;
 
@@ -977,7 +980,7 @@ upi42_sts_read(uint16_t port, void *priv)
 }
 
 void
-upi42_cmd_write(uint16_t port, uint8_t val, void *priv)
+upi42_cmd_write(UNUSED(uint16_t port), uint8_t val, void *priv)
 {
     upi42_t *upi42 = (upi42_t *) priv;
 
diff --git a/src/usb.c b/src/usb.c
index 5f2b95084..e715268af 100644
--- a/src/usb.c
+++ b/src/usb.c
@@ -31,6 +31,7 @@
 #include <86box/timer.h>
 #include <86box/usb.h>
 #include <86box/dma.h>
+#include <86box/plat_unused.h>
 
 #ifdef ENABLE_USB_LOG
 int usb_do_log = ENABLE_USB_LOG;
@@ -51,8 +52,7 @@ usb_log(const char *fmt, ...)
 #endif
 
 /* OHCI registers */
-enum
-{
+enum {
     OHCI_HcRevision         = 0x00 /* 0x00 */,
     OHCI_HcControl          = 0x01 /* 0x04 */,
     OHCI_HcCommandStatus    = 0x02 /* 0x08 */,
@@ -79,8 +79,7 @@ enum
     OHCI_HcRhPortStatus3    = 0x17 /* 0x5c */
 };
 
-enum
-{
+enum {
     OHCI_aHcRevision         = 0x00,
     OHCI_aHcControl          = 0x04,
     OHCI_aHcCommandStatus    = 0x08,
@@ -108,8 +107,7 @@ enum
 };
 
 /* OHCI HcInterruptEnable/Disable bits */
-enum
-{
+enum {
     OHCI_HcInterruptEnable_SO = 1 << 0,
     OHCI_HcInterruptEnable_WDH = 1 << 1,
     OHCI_HcInterruptEnable_SF = 1 << 2,
@@ -120,8 +118,7 @@ enum
 };
 
 /* OHCI HcControl bits */
-enum
-{
+enum {
     OHCI_HcControl_ControlBulkServiceRatio = 1 << 0,
     OHCI_HcControl_PeriodicListEnable = 1 << 1,
     OHCI_HcControl_IsochronousEnable = 1 << 2,
@@ -147,9 +144,9 @@ usb_interrupt_ohci(usb_t *dev, uint32_t level)
 }
 
 static uint8_t
-uhci_reg_read(uint16_t addr, void *p)
+uhci_reg_read(uint16_t addr, void *priv)
 {
-    usb_t   *dev = (usb_t *) p;
+    usb_t   *dev = (usb_t *) priv;
     uint8_t  ret;
     uint8_t *regs = dev->uhci_io;
 
@@ -161,9 +158,9 @@ uhci_reg_read(uint16_t addr, void *p)
 }
 
 static void
-uhci_reg_write(uint16_t addr, uint8_t val, void *p)
+uhci_reg_write(uint16_t addr, uint8_t val, void *priv)
 {
-    usb_t   *dev  = (usb_t *) p;
+    usb_t   *dev  = (usb_t *) priv;
     uint8_t *regs = dev->uhci_io;
 
     addr &= 0x0000001f;
@@ -185,13 +182,16 @@ uhci_reg_write(uint16_t addr, uint8_t val, void *p)
         case 0x0c:
             regs[0x0c] = (val & 0x7f);
             break;
+
+        default:
+            break;
     }
 }
 
 static void
-uhci_reg_writew(uint16_t addr, uint16_t val, void *p)
+uhci_reg_writew(uint16_t addr, uint16_t val, void *priv)
 {
-    usb_t    *dev  = (usb_t *) p;
+    usb_t    *dev  = (usb_t *) priv;
     uint16_t *regs = (uint16_t *) dev->uhci_io;
 
     addr &= 0x0000001f;
@@ -212,8 +212,8 @@ uhci_reg_writew(uint16_t addr, uint16_t val, void *p)
             regs[addr >> 1] = ((regs[addr >> 1] & 0xedbb) | (val & 0x1244)) & ~(val & 0x080a);
             break;
         default:
-            uhci_reg_write(addr, val & 0xff, p);
-            uhci_reg_write(addr + 1, (val >> 8) & 0xff, p);
+            uhci_reg_write(addr, val & 0xff, priv);
+            uhci_reg_write(addr + 1, (val >> 8) & 0xff, priv);
             break;
     }
 }
@@ -260,9 +260,9 @@ typedef struct
 #define ENDPOINT_DESC_LIMIT 32
 
 static uint8_t
-ohci_mmio_read(uint32_t addr, void *p)
+ohci_mmio_read(uint32_t addr, void *priv)
 {
-    usb_t  *dev = (usb_t *) p;
+    usb_t  *dev = (usb_t *) priv;
     uint8_t ret = 0x00;
 #ifdef ENABLE_USB_LOG
     uint32_t old_addr = addr;
@@ -301,15 +301,15 @@ ohci_mmio_read(uint32_t addr, void *p)
 }
 
 static uint16_t
-ohci_mmio_readw(uint32_t addr, void *p)
+ohci_mmio_readw(uint32_t addr, void *priv)
 {
-    return ohci_mmio_read(addr, p) | (ohci_mmio_read(addr + 1, p) << 8);
+    return ohci_mmio_read(addr, priv) | (ohci_mmio_read(addr + 1, priv) << 8);
 }
 
 static uint32_t
-ohci_mmio_readl(uint32_t addr, void *p)
+ohci_mmio_readl(uint32_t addr, void *priv)
 {
-    return ohci_mmio_readw(addr, p) | (ohci_mmio_readw(addr + 2, p) << 16);
+    return ohci_mmio_readw(addr, priv) | (ohci_mmio_readw(addr + 2, priv) << 16);
 }
 
 static void
@@ -367,7 +367,7 @@ static int ohci_copy_td_input(usb_t* dev, usb_td_t *td,
 }
 #endif
 
-static int ohci_copy_td_output(usb_t* dev, usb_td_t *td,
+static int ohci_copy_td_output(UNUSED(usb_t* dev), usb_td_t *td,
                         uint8_t *buf, int len)
 {
     uint32_t ptr;
@@ -504,6 +504,9 @@ ohci_service_transfer_desc(usb_t* dev, usb_ed_t* endpoint_desc)
             switch (device_result) {
                 case USB_ERROR_NAK:
                     return 1;
+
+                default:
+                    break;
             }
             dev->ohci_interrupt_counter = 0;
         }
@@ -676,9 +679,9 @@ ohci_soft_reset(usb_t* dev)
 }
 
 static void
-ohci_mmio_write(uint32_t addr, uint8_t val, void *p)
+ohci_mmio_write(uint32_t addr, uint8_t val, void *priv)
 {
-    usb_t  *dev = (usb_t *) p;
+    usb_t  *dev = (usb_t *) priv;
     uint8_t old;
 
 #ifdef ENABLE_USB_LOG
@@ -888,8 +891,10 @@ ohci_mmio_write(uint32_t addr, uint8_t val, void *p)
 
             if (!(dev->ohci_mmio[addr >> 2].b[addr & 3] & 0x04) && (old & 0x04))
                 dev->ohci_mmio[(addr + 2) >> 2].b[(addr + 2) & 3] |= 0x04;
-            /* if (!(dev->ohci_mmio[addr >> 2].b[addr & 3] & 0x02))
-                    dev->ohci_mmio[(addr + 2) >> 2].b[(addr + 2) & 3] |= 0x02; */
+#if 0
+            if (!(dev->ohci_mmio[addr >> 2].b[addr & 3] & 0x02))
+                dev->ohci_mmio[(addr + 2) >> 2].b[(addr + 2) & 3] |= 0x02;
+#endif
             return;
         case OHCI_aHcRhPortStatus1 + 1:
             if ((val & 0x02) && ((dev->ohci_mmio[OHCI_HcRhDescriptorA].b[1] & 0x03) == 0x00) && (dev->ohci_mmio[OHCI_HcRhDescriptorB].b[2] & 0x02)) {
@@ -924,23 +929,26 @@ ohci_mmio_write(uint32_t addr, uint8_t val, void *p)
         case OHCI_aHcPeriodCurrentED:
             dev->ohci_mmio[addr >> 2].b[addr & 3] = (val & 0xf0);
             return;
+
+        default:
+            break;
     }
 
     dev->ohci_mmio[addr >> 2].b[addr & 3] = val;
 }
 
 static void
-ohci_mmio_writew(uint32_t addr, uint16_t val, void *p)
+ohci_mmio_writew(uint32_t addr, uint16_t val, void *priv)
 {
-    ohci_mmio_write(addr, val & 0xff, p);
-    ohci_mmio_write(addr + 1, val >> 8, p);
+    ohci_mmio_write(addr, val & 0xff, priv);
+    ohci_mmio_write(addr + 1, val >> 8, priv);
 }
 
 static void
-ohci_mmio_writel(uint32_t addr, uint32_t val, void *p)
+ohci_mmio_writel(uint32_t addr, uint32_t val, void *priv)
 {
-    ohci_mmio_writew(addr, val & 0xffff, p);
-    ohci_mmio_writew(addr + 2, val >> 16, p);
+    ohci_mmio_writew(addr, val & 0xffff, priv);
+    ohci_mmio_writew(addr + 2, val >> 16, priv);
 }
 
 void
@@ -964,7 +972,7 @@ usb_attach_device(usb_t *dev, usb_device_t* device, uint8_t bus_type)
     switch (bus_type) {
         case USB_BUS_OHCI:
             {
-                for (int i = 0; i < 2; i++) {
+                for (uint8_t i = 0; i < 2; i++) {
                     if (!dev->ohci_devices[i]) {
                         uint32_t old = dev->ohci_mmio[OHCI_HcRhPortStatus1 + (4 * i)].l;
                         dev->ohci_devices[i] = device;
@@ -981,8 +989,11 @@ usb_attach_device(usb_t *dev, usb_device_t* device, uint8_t bus_type)
                 }
             }
             break;
+
+        default:
+            break;
     }
-    return 255;
+    return 0xff;
 }
 
 void
@@ -1011,6 +1022,9 @@ usb_detach_device(usb_t *dev, uint8_t port, uint8_t bus_type)
 
             }
             break;
+
+        default:
+            break;
     }
     return;
 }
@@ -1046,7 +1060,7 @@ usb_close(void *priv)
 }
 
 static void *
-usb_init_ext(const device_t *info, void *params)
+usb_init_ext(UNUSED(const device_t *info), void *params)
 {
     usb_t *dev;
 
diff --git a/src/video/CMakeLists.txt b/src/video/CMakeLists.txt
index 70ef72b7a..2395040ab 100644
--- a/src/video/CMakeLists.txt
+++ b/src/video/CMakeLists.txt
@@ -18,7 +18,7 @@ add_library(vid OBJECT agpgart.c video.c vid_table.c vid_cga.c vid_cga_comp.c
     vid_incolor.c vid_colorplus.c vid_genius.c vid_pgc.c vid_im1024.c
     vid_sigma.c vid_wy700.c vid_ega.c vid_ega_render.c vid_svga.c vid_8514a.c
     vid_svga_render.c vid_ddc.c vid_vga.c vid_ati_eeprom.c vid_ati18800.c
-    vid_ati28800.c vid_ati_mach64.c vid_ati68860_ramdac.c vid_bt48x_ramdac.c
+    vid_ati28800.c vid_ati_mach8.c vid_ati_mach64.c vid_ati68860_ramdac.c vid_bt48x_ramdac.c
     vid_av9194.c vid_icd2061.c vid_ics2494.c vid_ics2595.c vid_cl54xx.c
     vid_et3000.c vid_et4000.c vid_sc1148x_ramdac.c vid_sc1502x_ramdac.c
     vid_et4000w32.c vid_stg_ramdac.c vid_ht216.c vid_oak_oti.c vid_paradise.c
diff --git a/src/video/vid_8514a.c b/src/video/vid_8514a.c
index 9ec196d86..115105d7a 100644
--- a/src/video/vid_8514a.c
+++ b/src/video/vid_8514a.c
@@ -7,13 +7,13 @@
  *          This file is part of the 86Box distribution.
  *
  *          Emulation of the 8514/A card from IBM for the MCA bus and
- *          generic ISA bus clones without vendor extensions.
+ *          ISA bus clones.
  *
  *
  *
  * Authors: TheCollector1995.
  *
- *          Copyright 2022 TheCollector1995.
+ *          Copyright 2022-2023 TheCollector1995.
  */
 #include <stdarg.h>
 #include <stdint.h>
@@ -21,6 +21,7 @@
 #include <string.h>
 #include <stdlib.h>
 #include <wchar.h>
+#include <stdatomic.h>
 #define HAVE_STDARG_H
 #include <86box/86box.h>
 #include <86box/device.h>
@@ -42,17 +43,53 @@ static void     ibm8514_accel_outw(uint16_t port, uint16_t val, void *p);
 static uint8_t  ibm8514_accel_inb(uint16_t port, void *p);
 static uint16_t ibm8514_accel_inw(uint16_t port, void *p);
 
+#ifdef ENABLE_IBM8514_LOG
+int ibm8514_do_log = ENABLE_IBM8514_LOG;
+
+static void
+ibm8514_log(const char *fmt, ...)
+{
+    va_list ap;
+
+    if (ibm8514_do_log) {
+        va_start(ap, fmt);
+        pclog_ex(fmt, ap);
+        va_end(ap);
+    }
+}
+#else
+#    define ibm8514_log(fmt, ...)
+#endif
+
 #define READ_PIXTRANS_WORD(cx, n)                                                                    \
     if ((cmd <= 1) || (cmd == 5)) {                                                                    \
-        temp = dev->vram[((dev->accel.cy * dev->pitch) + (cx) + (n)) & dev->vram_mask];             \
-        temp |= (dev->vram[((dev->accel.cy * dev->pitch) + (cx) + (n + 1)) & dev->vram_mask] << 8); \
+        if (dev->local) { \
+            temp = svga->vram[((dev->accel.cy * dev->pitch) + (cx) + (n)) & svga->vram_mask];             \
+            temp |= (svga->vram[((dev->accel.cy * dev->pitch) + (cx) + (n + 1)) & svga->vram_mask] << 8); \
+        } else { \
+            temp = dev->vram[((dev->accel.cy * dev->pitch) + (cx) + (n)) & dev->vram_mask];             \
+            temp |= (dev->vram[((dev->accel.cy * dev->pitch) + (cx) + (n + 1)) & dev->vram_mask] << 8); \
+        } \
     } else {                                                                                         \
-        temp = dev->vram[(dev->accel.dest + (cx) + (n)) & dev->vram_mask];                           \
-        temp |= (dev->vram[(dev->accel.dest + (cx) + (n + 1)) & dev->vram_mask] << 8);               \
+        if (dev->local) { \
+            temp = svga->vram[(dev->accel.dest + (cx) + (n)) & svga->vram_mask];                           \
+            temp |= (svga->vram[(dev->accel.dest + (cx) + (n + 1)) & svga->vram_mask] << 8);               \
+        } else { \
+            temp = dev->vram[(dev->accel.dest + (cx) + (n)) & dev->vram_mask];                           \
+            temp |= (dev->vram[(dev->accel.dest + (cx) + (n + 1)) & dev->vram_mask] << 8);               \
+        } \
     }
 
 #define READ(addr, dat) \
-    dat = dev->vram[(addr) & (dev->vram_mask)];
+    if (dev->local) { \
+        if ((svga->bpp == 15) || (svga->bpp == 16)) { \
+            dat = vram_w[(addr) & (svga->vram_mask >> 1)]; \
+        } else { \
+            dat = (svga->vram[(addr) & (svga->vram_mask)]); \
+        } \
+    } else { \
+        dat = (dev->vram[(addr) & (dev->vram_mask)]); \
+    }
 
 #define MIX(mixmode, dest_dat, src_dat)                                                       \
     {                                                                                         \
@@ -157,8 +194,20 @@ static uint16_t ibm8514_accel_inw(uint16_t port, void *p);
     }
 
 #define WRITE(addr, dat)                                         \
-    dev->vram[((addr)) & (dev->vram_mask)]                = dat; \
-    dev->changedvram[(((addr)) & (dev->vram_mask)) >> 12] = changeframecount;
+    if (dev->local) { \
+        if ((svga->bpp == 15) || (svga->bpp == 16)) { \
+            vram_w[((addr)) & (svga->vram_mask >> 1)]                    = dat; \
+            svga->changedvram[(((addr)) & (svga->vram_mask >> 1)) >> 11] = changeframecount; \
+        } else { \
+            svga->vram[((addr)) & (svga->vram_mask)]                = dat; \
+            svga->changedvram[(((addr)) & (svga->vram_mask)) >> 12] = changeframecount; \
+        } \
+    } else { \
+        dev->vram[((addr)) & (dev->vram_mask)]                = dat; \
+        dev->changedvram[(((addr)) & (dev->vram_mask)) >> 12] = changeframecount; \
+    }
+
+int ibm8514_has_vga = 0;
 
 int
 ibm8514_cpu_src(svga_t *svga)
@@ -557,10 +606,11 @@ ibm8514_accel_out_fifo(svga_t *svga, uint16_t port, uint32_t val, int len)
                 dev->accel.short_stroke = val;
                 dev->accel.cx           = dev->accel.cur_x;
                 dev->accel.cy           = dev->accel.cur_y;
-                if (dev->accel.cur_x & 0x400)
-                    dev->accel.cx |= ~0x3ff;
-                if (dev->accel.cur_y & 0x400)
-                    dev->accel.cy |= ~0x3ff;
+                if (dev->accel.cur_x >= 0x600)
+                    dev->accel.cx |= ~0x5ff;
+
+                if (dev->accel.cur_y >= 0x600)
+                    dev->accel.cy |= ~0x5ff;
 
                 if (dev->accel.cmd & 0x1000) {
                     ibm8514_short_stroke_start(-1, 0, -1, 0, svga, dev->accel.short_stroke & 0xff, len);
@@ -573,15 +623,16 @@ ibm8514_accel_out_fifo(svga_t *svga, uint16_t port, uint32_t val, int len)
             break;
         case 0x9ee9:
         case 0xdee9:
+            dev->accel.ssv_state = 1;
             if (len == 1) {
                 dev->accel.short_stroke = (dev->accel.short_stroke & 0xff) | (val << 8);
                 dev->accel.cx           = dev->accel.cur_x;
                 dev->accel.cy           = dev->accel.cur_y;
-                if (dev->accel.cur_x & 0x400)
-                    dev->accel.cx |= ~0x3ff;
-                if (dev->accel.cur_y & 0x400)
-                    dev->accel.cy |= ~0x3ff;
+                if (dev->accel.cur_x >= 0x600)
+                    dev->accel.cx |= ~0x5ff;
 
+                if (dev->accel.cur_y >= 0x600)
+                    dev->accel.cy |= ~0x5ff;
                 if (dev->accel.cmd & 0x1000) {
                     ibm8514_short_stroke_start(-1, 0, -1, 0, svga, dev->accel.short_stroke & 0xff, len);
                     ibm8514_short_stroke_start(-1, 0, -1, 0, svga, dev->accel.short_stroke >> 8, len);
@@ -701,10 +752,14 @@ ibm8514_accel_out_fifo(svga_t *svga, uint16_t port, uint32_t val, int len)
                 dev->accel.multifunc_cntl                             = val;
                 dev->accel.multifunc[dev->accel.multifunc_cntl >> 12] = dev->accel.multifunc_cntl & 0xfff;
                 if ((dev->accel.multifunc_cntl >> 12) == 1) {
-                    dev->accel.clip_top = val & 0x3ff;
+                    dev->accel.clip_top = val & 0x7ff;
+                    if (val & 0x400)
+                        dev->accel.clip_top |= ~0x3ff;
                 }
                 if ((dev->accel.multifunc_cntl >> 12) == 2) {
-                    dev->accel.clip_left = val & 0x3ff;
+                    dev->accel.clip_left = val & 0x7ff;
+                    if (val & 0x400)
+                        dev->accel.clip_left |= ~0x3ff;
                 }
                 if (port == 0xfee8)
                     dev->accel.cmd_back = 1;
@@ -820,24 +875,24 @@ ibm8514_accel_out(uint16_t port, uint32_t val, svga_t *svga, int len)
             case 0x2e9:
                 if (len != 1) {
                     dev->htotal = (dev->htotal & 0xff) | (val << 8);
-                    // pclog("IBM 8514/A: H_TOTAL write 02E8 = %d\n", dev->htotal + 1);
+                    ibm8514_log("IBM 8514/A: H_TOTAL write 02E8 = %d\n", dev->htotal + 1);
                     svga_recalctimings(svga);
                 }
                 break;
 
             case 0x6e8:
                 dev->hdisp = val;
-                // pclog("IBM 8514/A: H_DISP write 06E8 = %d\n", dev->hdisp + 1);
+                ibm8514_log("IBM 8514/A: H_DISP write 06E8 = %d\n", dev->hdisp + 1);
                 svga_recalctimings(svga);
                 break;
 
             case 0xae8:
-                // pclog("IBM 8514/A: H_SYNC_STRT write 0AE8 = %d\n", val + 1);
+                ibm8514_log("IBM 8514/A: H_SYNC_STRT write 0AE8 = %d\n", val + 1);
                 svga_recalctimings(svga);
                 break;
 
             case 0xee8:
-                // pclog("IBM 8514/A: H_SYNC_WID write 0EE8 = %d\n", val + 1);
+                ibm8514_log("IBM 8514/A: H_SYNC_WID write 0EE8 = %d\n", val + 1);
                 svga_recalctimings(svga);
                 break;
 
@@ -852,7 +907,7 @@ ibm8514_accel_out(uint16_t port, uint32_t val, svga_t *svga, int len)
             case 0x12e9:
                 if (len == 1) {
                     dev->vtotal = (dev->vtotal & 0xff) | ((val & 0x1f) << 8);
-                    // pclog("IBM 8514/A: V_TOTAL write 12E8 = %d\n", dev->vtotal);
+                    ibm8514_log("IBM 8514/A: V_TOTAL write 12E8 = %d\n", dev->vtotal);
                     svga_recalctimings(svga);
                 }
                 break;
@@ -868,7 +923,7 @@ ibm8514_accel_out(uint16_t port, uint32_t val, svga_t *svga, int len)
             case 0x16e9:
                 if (len == 1) {
                     dev->vdisp = (dev->vdisp & 0xff) | ((val & 0x1f) << 8);
-                    // pclog("IBM 8514/A: V_DISP write 16E8 = %d\n", dev->vdisp);
+                    ibm8514_log("IBM 8514/A: V_DISP write 16E8 = %d\n", dev->vdisp);
                     svga_recalctimings(svga);
                 }
                 break;
@@ -884,21 +939,21 @@ ibm8514_accel_out(uint16_t port, uint32_t val, svga_t *svga, int len)
             case 0x1ae9:
                 if (len == 1) {
                     dev->vsyncstart = (dev->vsyncstart & 0xff) | ((val & 0x1f) << 8);
-                    // pclog("IBM 8514/A: V_SYNC_STRT write 1AE8 = %d\n", dev->vsyncstart);
+                    ibm8514_log("IBM 8514/A: V_SYNC_STRT write 1AE8 = %d\n", dev->vsyncstart);
                     svga_recalctimings(svga);
                 }
                 break;
 
             case 0x1ee8:
                 dev->vsyncwidth = val;
-                // pclog("IBM 8514/A: V_SYNC_WID write 1EE8 = %02x\n", val);
+                ibm8514_log("IBM 8514/A: V_SYNC_WID write 1EE8 = %02x\n", val);
                 svga_recalctimings(svga);
                 break;
 
             case 0x22e8:
                 dev->disp_cntl = val & 0x7e;
                 dev->interlace = !!(val & 0x10);
-                // pclog("IBM 8514/A: DISP_CNTL write 22E8 = %02x, SCANMODULOS = %d\n", dev->disp_cntl, dev->scanmodulos);
+                ibm8514_log("IBM 8514/A: DISP_CNTL write 22E8 = %02x, SCANMODULOS = %d\n", dev->disp_cntl, dev->scanmodulos);
                 svga_recalctimings(svga);
                 break;
 
@@ -922,7 +977,7 @@ ibm8514_accel_out(uint16_t port, uint32_t val, svga_t *svga, int len)
                 dev->accel.advfunc_cntl = val & 7;
                 ibm8514_on = (dev->accel.advfunc_cntl & 1);
                 vga_on = !ibm8514_on;
-                //pclog("IBM 8514/A: VGA ON = %i, val = %02x\n", vga_on, val);
+                ibm8514_log("IBM 8514/A: VGA ON = %i, val = %02x\n", vga_on, val);
                 svga_recalctimings(svga);
                 break;
         }
@@ -1105,26 +1160,33 @@ void
 ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat, svga_t *svga, int len)
 {
     ibm8514_t *dev = &svga->dev8514;
-    uint8_t  src_dat = 0;
-    uint8_t dest_dat;
-    uint8_t old_dest_dat;
-    int      frgd_mix;
-    int      bkgd_mix;
-    uint16_t clip_b          = dev->accel.multifunc[3] & 0x7ff;
-    uint16_t clip_r          = dev->accel.multifunc[4] & 0x7ff;
+    uint16_t *vram_w = (uint16_t *)svga->vram;
+    uint16_t  src_dat = 0, dest_dat, old_dest_dat;
+    int      frgd_mix, bkgd_mix;
+    uint16_t clip_b          = dev->accel.multifunc[3];
+    uint16_t clip_r          = dev->accel.multifunc[4];
     int      pixcntl         = (dev->accel.multifunc[0x0a] >> 6) & 3;
-    uint8_t  mix_mask        = 0x80;
-    uint8_t  compare         = dev->accel.color_cmp & 0xff;
+    uint16_t  mix_mask       = ((svga->bpp == 8) || (svga->bpp == 24)) ? 0x80 : 0x8000;
+    uint16_t compare         = dev->accel.color_cmp;
     int      compare_mode    = dev->accel.multifunc[0x0a] & 0x38;
     int      cmd             = dev->accel.cmd >> 13;
-    uint8_t  wrt_mask        = dev->accel.wrt_mask & 0xff;
-    uint8_t  rd_mask         = ((dev->accel.rd_mask & 0x01) << 7) | ((dev->accel.rd_mask & 0xfe) >> 1);
-    uint8_t  rd_mask_polygon = dev->accel.rd_mask & 0xff;
-    uint8_t  frgd_color      = dev->accel.frgd_color;
-    uint8_t  bkgd_color      = dev->accel.bkgd_color;
+    uint16_t  wrt_mask        = dev->accel.wrt_mask;
+    uint16_t  rd_mask         = dev->accel.rd_mask;
+    uint16_t  rd_mask_polygon = dev->accel.rd_mask;
+    uint16_t  frgd_color      = dev->accel.frgd_color;
+    uint16_t  bkgd_color      = dev->accel.bkgd_color;
     uint32_t old_mix_dat;
     int      and3     = dev->accel.cur_x & 3;
-    uint8_t  poly_src = 0;
+    uint16_t  poly_src = 0;
+
+    if ((svga->bpp == 8) || (svga->bpp == 24)) {
+        compare &= 0xff;
+        frgd_color &= 0xff;
+        bkgd_color &= 0xff;
+        rd_mask = ((dev->accel.rd_mask & 0x01) << 7) | ((dev->accel.rd_mask & 0xfe) >> 1);
+        rd_mask &= 0xff;
+        rd_mask_polygon &= 0xff;
+    }
 
     if (dev->accel.cmd & 0x100) {
         dev->force_busy  = 1;
@@ -1147,6 +1209,11 @@ ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat
         } else {
             count >>= 3;
         }
+
+        if (((svga->bpp == 15) || (svga->bpp == 16))) {
+            if ((dev->accel.cmd & 0x200) && (count == 2))
+                count >>= 1;
+        }
     }
 
     if (pixcntl == 1) {
@@ -1246,7 +1313,7 @@ ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat
                                 src_dat = frgd_color;
                                 break;
                             case 2:
-                                src_dat = cpu_dat & 0xff;
+                                src_dat = cpu_dat;
                                 break;
                             case 3:
                                 src_dat = 0;
@@ -1255,17 +1322,26 @@ ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat
                         READ((dev->accel.cy * dev->pitch) + dev->accel.cx, dest_dat);
 
                         if ((compare_mode == 0) || ((compare_mode == 0x10) && (dest_dat >= compare)) || ((compare_mode == 0x18) && (dest_dat < compare)) || ((compare_mode == 0x20) && (dest_dat != compare)) || ((compare_mode == 0x28) && (dest_dat == compare)) || ((compare_mode == 0x30) && (dest_dat <= compare)) || ((compare_mode == 0x38) && (dest_dat > compare))) {
+							old_dest_dat = dest_dat;
                             MIX(mix_dat & mix_mask, dest_dat, src_dat);
+							dest_dat = (dest_dat & wrt_mask) | (old_dest_dat & ~wrt_mask);
 
                             if (dev->accel.ssv_draw) {
-                                WRITE((dev->accel.cy * dev->pitch) + dev->accel.cx, dest_dat);
+                                if ((dev->accel.cmd & 4) && dev->accel.ssv_len) {
+                                    WRITE((dev->accel.cy * dev->pitch) + dev->accel.cx, dest_dat);
+                                } else if (!(dev->accel.cmd & 4)) {
+                                    WRITE((dev->accel.cy * dev->pitch) + dev->accel.cx, dest_dat);
+                                }
                             }
                         }
                     }
 
                     mix_dat <<= 1;
                     mix_dat |= 1;
-                    cpu_dat >>= 8;
+                    if ((svga->bpp == 15) || (svga->bpp == 16))
+                        cpu_dat >>= 16;
+                    else
+                        cpu_dat >>= 8;
 
                     if (!dev->accel.ssv_len)
                         break;
@@ -1303,10 +1379,116 @@ ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat
 
                     dev->accel.ssv_len--;
                 }
+            } else {
+                while (count-- && (dev->accel.ssv_len >= 0)) {
+                    if ((dev->accel.cx >= dev->accel.clip_left && dev->accel.cx <= clip_r && dev->accel.cy >= dev->accel.clip_top && dev->accel.cy <= clip_b)) {
+                        switch ((mix_dat & mix_mask) ? frgd_mix : bkgd_mix) {
+                            case 0:
+                                src_dat = bkgd_color;
+                                break;
+                            case 1:
+                                src_dat = frgd_color;
+                                break;
+                            case 2:
+                                src_dat = cpu_dat;
+                                break;
+                            case 3:
+                                src_dat = 0;
+                                break;
+                        }
 
-                dev->accel.cur_x = dev->accel.cx;
-                dev->accel.cur_y = dev->accel.cy;
+                        READ((dev->accel.cy * dev->pitch) + dev->accel.cx, dest_dat);
+
+                        if ((compare_mode == 0) || ((compare_mode == 0x10) && (dest_dat >= compare)) || ((compare_mode == 0x18) && (dest_dat < compare)) || ((compare_mode == 0x20) && (dest_dat != compare)) || ((compare_mode == 0x28) && (dest_dat == compare)) || ((compare_mode == 0x30) && (dest_dat <= compare)) || ((compare_mode == 0x38) && (dest_dat > compare))) {
+                            old_dest_dat = dest_dat;
+                            MIX(mix_dat & mix_mask, dest_dat, src_dat);
+                            dest_dat = (dest_dat & wrt_mask) | (old_dest_dat & ~wrt_mask);
+
+                            if (dev->accel.ssv_draw) {
+                                if ((dev->accel.cmd & 4) && dev->accel.ssv_len) {
+                                    WRITE((dev->accel.cy * dev->pitch) + dev->accel.cx, dest_dat);
+                                } else if (!(dev->accel.cmd & 4)) {
+                                    WRITE((dev->accel.cy * dev->pitch) + dev->accel.cx, dest_dat);
+                                }
+                            }
+                        }
+                    }
+
+                    mix_dat <<= 1;
+                    mix_dat |= 1;
+                    if ((svga->bpp == 15) || (svga->bpp == 16))
+                        cpu_dat >>= 16;
+                    else
+                        cpu_dat >>= 8;
+
+                    if (!dev->accel.ssv_len)
+                        break;
+
+                    if (dev->accel.err_term >= dev->accel.maj_axis_pcnt) {
+                        dev->accel.err_term += dev->accel.destx_distp;
+                        /*Step minor axis*/
+                        switch (dev->accel.cmd & 0xe0) {
+                            case 0x00:
+                                dev->accel.cy--;
+                                break;
+                            case 0x20:
+                                dev->accel.cy--;
+                                break;
+                            case 0x40:
+                                dev->accel.cx--;
+                                break;
+                            case 0x60:
+                                dev->accel.cx++;
+                                break;
+                            case 0x80:
+                                dev->accel.cy++;
+                                break;
+                            case 0xa0:
+                                dev->accel.cy++;
+                                break;
+                            case 0xc0:
+                                dev->accel.cx--;
+                                break;
+                            case 0xe0:
+                                dev->accel.cx++;
+                                break;
+                        }
+                    } else
+                        dev->accel.err_term += dev->accel.desty_axstp;
+
+                    /*Step major axis*/
+                    switch (dev->accel.cmd & 0xe0) {
+                        case 0x00:
+                            dev->accel.cx--;
+                            break;
+                        case 0x20:
+                            dev->accel.cx++;
+                            break;
+                        case 0x40:
+                            dev->accel.cy--;
+                            break;
+                        case 0x60:
+                            dev->accel.cy--;
+                            break;
+                        case 0x80:
+                            dev->accel.cx--;
+                            break;
+                        case 0xa0:
+                            dev->accel.cx++;
+                            break;
+                        case 0xc0:
+                            dev->accel.cy++;
+                            break;
+                        case 0xe0:
+                            dev->accel.cy++;
+                            break;
+                    }
+
+                    dev->accel.ssv_len--;
+                }
             }
+            dev->accel.cur_x = dev->accel.cx;
+            dev->accel.cur_y = dev->accel.cy;
             break;
 
         case 1: /*Draw line*/
@@ -1315,11 +1497,11 @@ ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat
                 dev->accel.cx       = dev->accel.cur_x;
                 dev->accel.cy       = dev->accel.cur_y;
 
-                if (dev->accel.cur_x & 0x400) {
-                    dev->accel.cx |= ~0x3ff;
+                if (dev->accel.cur_x >= 0x600) {
+                    dev->accel.cx |= ~0x5ff;
                 }
-                if (dev->accel.cur_y & 0x400) {
-                    dev->accel.cy |= ~0x3ff;
+                if (dev->accel.cur_y >= 0x600) {
+                    dev->accel.cy |= ~0x5ff;
                 }
 
                 dev->accel.sy = dev->accel.maj_axis_pcnt;
@@ -1392,7 +1574,7 @@ ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat
                                     src_dat = frgd_color;
                                     break;
                                 case 2:
-                                    src_dat = cpu_dat & 0xff;
+                                    src_dat = cpu_dat;
                                     break;
                                 case 3:
                                     src_dat = 0;
@@ -1477,7 +1659,10 @@ ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat
 
                     mix_dat <<= 1;
                     mix_dat |= 1;
-                    cpu_dat >>= 8;
+                    if ((svga->bpp == 15) || (svga->bpp == 16))
+                        cpu_dat >>= 16;
+                    else
+                        cpu_dat >>= 8;
 
                     if (dev->accel.sy == 0) {
                         break;
@@ -1538,7 +1723,7 @@ ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat
                                         src_dat = frgd_color;
                                         break;
                                     case 2:
-                                        src_dat = cpu_dat & 0xff;
+                                        src_dat = cpu_dat;
                                         break;
                                     case 3:
                                         src_dat = 0;
@@ -1561,7 +1746,10 @@ ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat
 
                         dev->accel.temp_cnt--;
                         mix_dat >>= 1;
-                        cpu_dat >>= 8;
+                        if ((svga->bpp == 15) || (svga->bpp == 16))
+                            cpu_dat >>= 16;
+                        else
+                            cpu_dat >>= 8;
 
                         if (dev->accel.sy == 0) {
                             break;
@@ -1654,7 +1842,7 @@ ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat
                                         src_dat = frgd_color;
                                         break;
                                     case 2:
-                                        src_dat = cpu_dat & 0xff;
+                                        src_dat = cpu_dat;
                                         break;
                                     case 3:
                                         src_dat = 0;
@@ -1677,7 +1865,10 @@ ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat
 
                         mix_dat <<= 1;
                         mix_dat |= 1;
-                        cpu_dat >>= 8;
+                        if ((svga->bpp == 15) || (svga->bpp == 16))
+                            cpu_dat >>= 16;
+                        else
+                            cpu_dat >>= 8;
 
                         if (dev->accel.sy == 0) {
                             break;
@@ -1767,14 +1958,18 @@ ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat
                 dev->accel.sx = dev->accel.maj_axis_pcnt & 0x7ff;
                 dev->accel.sy = dev->accel.multifunc[0] & 0x7ff;
 
-                dev->accel.cx = dev->accel.cur_x & 0x3ff;
-                if (dev->accel.cur_x & 0x400)
-                    dev->accel.cx |= ~0x3ff;
-                dev->accel.cy = dev->accel.cur_y & 0x3ff;
-                if (dev->accel.cur_y & 0x400)
-                    dev->accel.cy |= ~0x3ff;
+                dev->accel.cx = dev->accel.cur_x;
+                if (dev->accel.cur_x >= 0x600)
+                    dev->accel.cx |= ~0x5ff;
+                dev->accel.cy = dev->accel.cur_y;
+                if (dev->accel.cur_y >= 0x600)
+                    dev->accel.cy |= ~0x5ff;
 
-                dev->accel.dest       = dev->accel.cy * dev->pitch;
+                if (dev->local && dev->accel.ge_offset && (svga->bpp == 24))
+                    dev->accel.dest       = (dev->accel.ge_offset << 2) + (dev->accel.cy * dev->pitch);
+                else
+                    dev->accel.dest       = dev->accel.cy * dev->pitch;
+                   
                 dev->accel.fill_state = 0;
 
                 if (cmd == 4)
@@ -1822,7 +2017,10 @@ ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat
                         if (!(dev->accel.cmd & 0x40) && (frgd_mix == 2) && (bkgd_mix == 2) && (pixcntl == 0) && (cmd == 2)) {
                             if (!(dev->accel.sx & 1)) {
                                 dev->accel.output      = 1;
-                                dev->accel.newdest_out = (dev->accel.cy + 1) * dev->pitch;
+                                if (dev->local && dev->accel.ge_offset && (svga->bpp == 24))
+                                    dev->accel.newdest_out = (dev->accel.ge_offset << 2) + ((dev->accel.cy + 1) * dev->pitch);
+                                else
+                                    dev->accel.newdest_out = (dev->accel.cy + 1) * dev->pitch;
                             }
                         }
                     }
@@ -1864,7 +2062,7 @@ rect_fill_pix:
                                         src_dat = frgd_color;
                                         break;
                                     case 2:
-                                        src_dat = cpu_dat & 0xff;
+                                        src_dat = cpu_dat;
                                         break;
                                     case 3:
                                         src_dat = 0;
@@ -1939,7 +2137,10 @@ rect_fill_pix:
 
                             mix_dat <<= 1;
                             mix_dat |= 1;
-                            cpu_dat >>= 8;
+                            if ((svga->bpp == 15) || (svga->bpp == 16))
+                                cpu_dat >>= 16;
+                            else
+                                cpu_dat >>= 8;
 
                             switch (dev->accel.cmd & 0xe0) {
                                 case 0x00:
@@ -2033,7 +2234,7 @@ rect_fill_pix:
                                             src_dat = frgd_color;
                                             break;
                                         case 2:
-                                            src_dat = cpu_dat & 0xff;
+                                            src_dat = cpu_dat;
                                             break;
                                         case 3:
                                             src_dat = 0;
@@ -2052,7 +2253,10 @@ rect_fill_pix:
 
                             mix_dat <<= 1;
                             mix_dat |= 1;
-                            cpu_dat >>= 8;
+                            if ((svga->bpp == 15) || (svga->bpp == 16))
+                                cpu_dat >>= 16;
+                            else
+                                cpu_dat >>= 8;
 
                             if (dev->accel.cmd & 0x20)
                                 dev->accel.cx++;
@@ -2077,7 +2281,11 @@ rect_fill_pix:
                                 else
                                     dev->accel.cy--;
 
-                                dev->accel.dest = dev->accel.cy * dev->pitch;
+                                if (dev->local && dev->accel.ge_offset && (svga->bpp == 24))
+                                    dev->accel.dest       = (dev->accel.ge_offset << 2) + (dev->accel.cy * dev->pitch);
+                                else
+                                    dev->accel.dest       = dev->accel.cy * dev->pitch;
+
                                 dev->accel.sy--;
                                 return;
                             }
@@ -2107,7 +2315,7 @@ rect_fill_pix:
                                             src_dat = frgd_color;
                                             break;
                                         case 2:
-                                            src_dat = cpu_dat & 0xff;
+                                            src_dat = cpu_dat;
                                             break;
                                         case 3:
                                             src_dat = 0;
@@ -2125,7 +2333,10 @@ rect_fill_pix:
                                 }
                             }
                             mix_dat >>= 1;
-                            cpu_dat >>= 8;
+                            if ((svga->bpp == 15) || (svga->bpp == 16))
+                                cpu_dat >>= 16;
+                            else
+                                cpu_dat >>= 8;
 
                             if (dev->accel.cmd & 0x20)
                                 dev->accel.cx++;
@@ -2148,9 +2359,9 @@ rect_fill_pix:
 
                                 if (dev->accel.cmd & 2) {
                                     if (dev->accel.cmd & 0x1000) {
-                                        dev->accel.cx = dev->accel.cur_x & 0x3ff;
-                                        if (dev->accel.cur_x & 0x400)
-                                            dev->accel.cx |= ~0x3ff;
+                                        dev->accel.cx = dev->accel.cur_x;
+                                        if (dev->accel.cur_x >= 0x600)
+                                            dev->accel.cx |= ~0x5ff;
                                     }
                                 }
 
@@ -2159,7 +2370,11 @@ rect_fill_pix:
                                 else
                                     dev->accel.cy--;
 
-                                dev->accel.dest = dev->accel.cy * dev->pitch;
+                                if (dev->local && dev->accel.ge_offset && (svga->bpp == 24))
+                                    dev->accel.dest       = (dev->accel.ge_offset << 2) + (dev->accel.cy * dev->pitch);
+                                else
+                                    dev->accel.dest       = dev->accel.cy * dev->pitch;
+
                                 dev->accel.sy--;
                                 return;
                             }
@@ -2262,7 +2477,10 @@ rect_fill_pix:
                                 }
                                 mix_dat <<= 1;
                                 mix_dat |= 1;
-                                cpu_dat >>= 8;
+                                if ((svga->bpp == 15) || (svga->bpp == 16))
+                                    cpu_dat >>= 16;
+                                else
+                                    cpu_dat >>= 8;
 
                                 if (dev->accel.cmd & 0x20)
                                     dev->accel.cx++;
@@ -2333,7 +2551,7 @@ rect_fill_pix:
                                                 src_dat = frgd_color;
                                                 break;
                                             case 2:
-                                                src_dat = cpu_dat & 0xff;
+                                                src_dat = cpu_dat;
                                                 break;
                                             case 3:
                                                 src_dat = 0;
@@ -2358,7 +2576,10 @@ rect_fill_pix:
 
                                 mix_dat <<= 1;
                                 mix_dat |= 1;
-                                cpu_dat >>= 8;
+                                if ((svga->bpp == 15) || (svga->bpp == 16))
+                                    cpu_dat >>= 16;
+                                else
+                                    cpu_dat >>= 8;
 
                                 if (dev->accel.cmd & 0x20)
                                     dev->accel.cx++;
@@ -2671,8 +2892,12 @@ rect_fill:
             if (!cpu_input) {
                 dev->accel.cx    = dev->accel.cur_x;
                 dev->accel.cy    = dev->accel.cur_y;
-                dev->accel.oldcy = dev->accel.cy;
+                if (dev->accel.cur_x >= 0x600)
+                    dev->accel.cx |= ~0x5ff;
 
+                if (dev->accel.cur_y >= 0x600)
+                    dev->accel.cy |= ~0x5ff;
+                dev->accel.oldcy = dev->accel.cy;
                 dev->accel.sy = 0;
 
                 if (ibm8514_cpu_src(svga)) {
@@ -2687,9 +2912,6 @@ rect_fill:
             }
 
             while (count-- && (dev->accel.sy >= 0)) {
-                if (dev->accel.cur_x > 1023)
-                    dev->accel.cx = 0;
-
                 if ((dev->accel.cx) >= dev->accel.clip_left && ((dev->accel.cx) <= clip_r) && (dev->accel.cy) >= dev->accel.clip_top && (dev->accel.cy) <= clip_b) {
                     switch ((mix_dat & mix_mask) ? frgd_mix : bkgd_mix) {
                         case 0:
@@ -2699,7 +2921,7 @@ rect_fill:
                             src_dat = frgd_color;
                             break;
                         case 2:
-                            src_dat = cpu_dat & 0xff;
+                            src_dat = cpu_dat;
                             break;
                         case 3:
                             src_dat = 0;
@@ -2726,7 +2948,10 @@ rect_fill:
 
                 mix_dat <<= 1;
                 mix_dat |= 1;
-                cpu_dat >>= 8;
+                if ((svga->bpp == 15) || (svga->bpp == 16))
+                    cpu_dat >>= 16;
+                else
+                    cpu_dat >>= 8;
 
                 if (dev->accel.sy == dev->accel.maj_axis_pcnt) {
                     break;
@@ -2813,21 +3038,21 @@ rect_fill:
                 dev->accel.sx = dev->accel.maj_axis_pcnt & 0x7ff;
                 dev->accel.sy = dev->accel.multifunc[0] & 0x7ff;
 
-                dev->accel.dx = dev->accel.destx_distp & 0x3ff;
-                dev->accel.dy = dev->accel.desty_axstp & 0x3ff;
+                dev->accel.dx = dev->accel.destx_distp;
+                dev->accel.dy = dev->accel.desty_axstp;
 
-                if (dev->accel.destx_distp & 0x400)
-                    dev->accel.dx |= ~0x3ff;
-                if (dev->accel.desty_axstp & 0x400)
-                    dev->accel.dy |= ~0x3ff;
+                if (dev->accel.destx_distp >= 0x600)
+                    dev->accel.dx |= ~0x5ff;
+                if (dev->accel.desty_axstp >= 0x600)
+                    dev->accel.dy |= ~0x5ff;
 
-                dev->accel.cx = dev->accel.cur_x & 0x3ff;
-                dev->accel.cy = dev->accel.cur_y & 0x3ff;
+                dev->accel.cx = dev->accel.cur_x;
+                dev->accel.cy = dev->accel.cur_y;
 
-                if (dev->accel.cur_x & 0x400)
-                    dev->accel.cx |= ~0x3ff;
-                if (dev->accel.cur_y & 0x400)
-                    dev->accel.cy |= ~0x3ff;
+                if (dev->accel.cur_x >= 0x600)
+                    dev->accel.cx |= ~0x5ff;
+                if (dev->accel.cur_y >= 0x600)
+                    dev->accel.cy |= ~0x5ff;
 
                 dev->accel.src  = dev->accel.cy * dev->pitch;
                 dev->accel.dest = dev->accel.dy * dev->pitch;
@@ -2876,7 +3101,7 @@ bitblt_pix:
                                         src_dat = frgd_color;
                                         break;
                                     case 2:
-                                        src_dat = cpu_dat & 0xff;
+                                        src_dat = cpu_dat;
                                         break;
                                     case 3:
                                         READ(dev->accel.src + dev->accel.cx, src_dat);
@@ -2900,12 +3125,18 @@ bitblt_pix:
 
                             mix_dat <<= 1;
                             mix_dat |= 1;
-                            cpu_dat >>= 8;
-
-                            if (dev->accel.cmd & 0x20)
-                                dev->accel.cx++;
+                            if ((svga->bpp == 15) || (svga->bpp == 16))
+                                cpu_dat >>= 16;
                             else
+                                cpu_dat >>= 8;
+
+                            if (dev->accel.cmd & 0x20) {
+                                dev->accel.dx++;
+                                dev->accel.cx++;
+                            } else {
+                                dev->accel.dx--;
                                 dev->accel.cx--;
+                            }
 
                             dev->accel.sx--;
                             if (dev->accel.sx < 0) {
@@ -2916,16 +3147,23 @@ bitblt_pix:
                                 }
 
                                 if (dev->accel.cmd & 0x20) {
+                                    dev->accel.dx -= (dev->accel.sx) + 1;
                                     dev->accel.cx -= (dev->accel.sx) + 1;
-                                } else
+                                } else {
+                                    dev->accel.dx += (dev->accel.sx) + 1;
                                     dev->accel.cx += (dev->accel.sx) + 1;
+                                }
 
-                                if (dev->accel.cmd & 0x80)
+                                if (dev->accel.cmd & 0x80) {
+                                    dev->accel.dy++;
                                     dev->accel.cy++;
-                                else
+                                } else {
+                                    dev->accel.dy--;
                                     dev->accel.cy--;
+                                }
 
-                                dev->accel.dest = dev->accel.cy * dev->pitch;
+                                dev->accel.src = dev->accel.cy * dev->pitch;
+                                dev->accel.dest = dev->accel.dy * dev->pitch;
                                 dev->accel.sy--;
                                 return;
                             }
@@ -2952,7 +3190,7 @@ bitblt_pix:
                                         src_dat = frgd_color;
                                         break;
                                     case 2:
-                                        src_dat = cpu_dat & 0xff;
+                                        src_dat = cpu_dat;
                                         break;
                                     case 3:
                                         READ(dev->accel.src + dev->accel.cx, src_dat);
@@ -2974,7 +3212,10 @@ bitblt_pix:
                                 }
                             }
                             mix_dat >>= 1;
-                            cpu_dat >>= 8;
+                            if ((svga->bpp == 15) || (svga->bpp == 16))
+                                cpu_dat >>= 16;
+                            else
+                                cpu_dat >>= 8;
 
                             if (dev->accel.cmd & 0x20) {
                                 dev->accel.dx++;
@@ -3003,12 +3244,12 @@ bitblt_pix:
 
                                 if (dev->accel.cmd & 2) {
                                     if (dev->accel.cmd & 0x1000) {
-                                        dev->accel.cx = dev->accel.cur_x & 0x3ff;
-                                        if (dev->accel.cur_x & 0x400)
-                                            dev->accel.cx |= ~0x3ff;
-                                        dev->accel.dx = dev->accel.destx_distp & 0x3ff;
-                                        if (dev->accel.destx_distp & 0x400)
-                                            dev->accel.dx |= ~0x3ff;
+                                        dev->accel.cx = dev->accel.cur_x;
+                                        if (dev->accel.cur_x >= 0x600)
+                                            dev->accel.cx |= ~0x5ff;
+                                        dev->accel.dx = dev->accel.destx_distp;
+                                        if (dev->accel.destx_distp >= 0x600)
+                                            dev->accel.dx |= ~0x5ff;
                                     }
                                 }
 
@@ -3056,7 +3297,7 @@ bitblt_pix:
                                         src_dat = frgd_color;
                                         break;
                                     case 2:
-                                        src_dat = cpu_dat & 0xff;
+                                        src_dat = cpu_dat;
                                         break;
                                     case 3:
                                         READ(dev->accel.src + dev->accel.cx, src_dat);
@@ -3079,7 +3320,10 @@ bitblt_pix:
                             }
                             mix_dat <<= 1;
                             mix_dat |= 1;
-                            cpu_dat >>= 8;
+                            if ((svga->bpp == 15) || (svga->bpp == 16))
+                                cpu_dat >>= 16;
+                            else
+                                cpu_dat >>= 8;
 
                             if (dev->accel.cmd & 0x20) {
                                 dev->accel.dx++;
@@ -3123,7 +3367,7 @@ bitblt:
                             count               = dev->accel.maj_axis_pcnt + 1;
                             dev->accel.temp_cnt = 8;
                             while (count-- && dev->accel.sy >= 0) {
-                                if (dev->accel.temp_cnt == 0) {
+                                if (!dev->accel.temp_cnt) {
                                     mix_dat >>= 8;
                                     dev->accel.temp_cnt = 8;
                                 }
@@ -3197,7 +3441,7 @@ bitblt:
                         } else {
                             dev->accel.temp_cnt = 8;
                             while (count-- && dev->accel.sy >= 0) {
-                                if (dev->accel.temp_cnt == 0) {
+                                if (!dev->accel.temp_cnt) {
                                     dev->accel.temp_cnt = 8;
                                     mix_dat             = old_mix_dat;
                                 }
@@ -3269,6 +3513,32 @@ bitblt:
                             }
                         }
                     } else {
+                        if ((svga->bpp == 24) && dev->local && (dev->accel.cmd == 0xc2b5)) {
+                            int64_t x, cx, dx;
+
+                            cx = (int64_t)dev->accel.cx;
+                            dx = (int64_t)dev->accel.dx;
+
+                            while (1) {
+                                if (((dx) >= (((int64_t)dev->accel.clip_left) * 3) && (dx) <= (((uint64_t)clip_r) * 3) && dev->accel.dy >= (dev->accel.clip_top << 1) && dev->accel.dy <= (clip_b << 1))) {
+                                    READ(dev->accel.src + (dev->accel.ge_offset << 2) + cx, src_dat);
+                                    READ(dev->accel.dest + (dev->accel.ge_offset << 2) + dx, dest_dat);
+
+                                    dest_dat = (src_dat & wrt_mask) | (dest_dat & ~wrt_mask);
+
+                                    WRITE(dev->accel.dest + (dev->accel.ge_offset << 2) + dx, dest_dat);
+                                }
+
+                                cx++;
+                                dx++;
+
+                                dev->accel.sx--;
+                                if (dev->accel.sx < 0)
+                                    return;
+                            }
+                            return;
+                        }
+
                         while (count-- && dev->accel.sy >= 0) {
                             if (dev->accel.dx >= dev->accel.clip_left && dev->accel.dx <= clip_r && dev->accel.dy >= dev->accel.clip_top && dev->accel.dy <= clip_b) {
                                 if (pixcntl == 3) {
@@ -3309,7 +3579,13 @@ bitblt:
                                     MIX(mix_dat & mix_mask, dest_dat, src_dat);
                                     dest_dat = (dest_dat & wrt_mask) | (old_dest_dat & ~wrt_mask);
 
-                                    WRITE(dev->accel.dest + dev->accel.dx, dest_dat);
+                                    if (dev->accel.cmd & 4) {
+                                        if (dev->accel.sx > 0) {
+                                            WRITE(dev->accel.dest + dev->accel.dx, dest_dat);
+                                        }
+                                    } else {
+                                        WRITE(dev->accel.dest + dev->accel.dx, dest_dat);
+                                    }
                                 }
                             }
                             mix_dat <<= 1;
@@ -3547,7 +3823,7 @@ ibm8514_recalctimings(svga_t *svga)
 
     if (ibm8514_on) {
         dev->h_disp                     = (dev->hdisp + 1) << 3;
-        dev->pitch                      = dev->h_disp;
+        dev->pitch                      = (dev->accel.advfunc_cntl & 4) ? 1024 : 640;
         dev->h_total                    = (dev->htotal + 1);
         dev->v_total                    = (dev->vtotal + 1);
         dev->v_syncstart                = (dev->vsyncstart + 1);
@@ -3555,46 +3831,9 @@ ibm8514_recalctimings(svga_t *svga)
         dev->dispend                   = ((dev->vdisp >> 1) + 1);
         if (dev->dispend == 766)
             dev->dispend = 768;
-        //pclog("HDISP = %d, VTOTAL = %d, DISPEND = %d, hires = %02x\n", dev->h_disp, dev->v_total, dev->dispend, dev->accel.advfunc_cntl & 4);
-        {
-#if 0
-            if (dev->dispend == 480) {
-                dev->h_disp = 640;
-                dev->rowoffset = 128;
-                dev->pitch = 1024;
-            } else if ((dev->dispend == 600) || (dev->dispend == 598)) {
-                dev->h_disp = 800;
-                dev->rowoffset = 128;
-                dev->dispend = 600;
-                dev->pitch = 1024;
-                if (!dev->vtotal)
-                    dev->v_total = 816;
-                if (!dev->vsyncstart)
-                    dev->v_syncstart = dev->dispend;
-            } else {
-                if (dev->accel.advfunc_cntl & 4) {
-                    if (!dev->hdisp) {
-                        dev->rowoffset = 128;
-                        dev->h_disp    = 1024;
-                        dev->pitch = dev->h_disp;
-                    }
-                    if (!dev->vtotal)
-                        dev->v_total = 816;
-                    if (!dev->vsyncstart)
-                        dev->v_syncstart = dev->dispend;
-                } else {
-                    dev->rowoffset = 128;
-                    dev->h_disp    = 640;
-                    dev->dispend = 480;
-                    dev->pitch = 1024;
-                    if (!dev->vtotal)
-                        dev->v_total = 816;
-                    if (!dev->vsyncstart)
-                        dev->v_syncstart = dev->dispend;
-                }
-            }
-#endif
-        }
+
+        if (dev->dispend == 598)
+            dev->dispend = 600;
 
         if (dev->accel.advfunc_cntl & 4) {
             if (!vga_on && dev->ibm_mode) {
@@ -3617,15 +3856,21 @@ ibm8514_recalctimings(svga_t *svga)
                 dev->v_syncstart >>= 1;
                 dev->v_total >>= 1;
             }
-            dev->rowoffset = 0x80;
-            dev->pitch = 1024;
 
-            // pclog("1024x768 clock mode, hdisp = %d, htotal = %d, vtotal = %d, vsyncstart = %d, interlace = %02x\n", dev->h_disp, dev->h_total, dev->v_total, dev->v_syncstart, dev->interlace);
+            if (ibm8514_has_vga) {
+                dev->pitch = dev->ext_pitch;
+                dev->rowoffset = dev->ext_crt_pitch;
+            } else
+                dev->rowoffset = 128;
+
+            ibm8514_log("1024x768 clock mode, hdisp = %d, htotal = %d, vtotal = %d, vsyncstart = %d, interlace = %02x\n", dev->h_disp, dev->h_total, dev->v_total, dev->v_syncstart, dev->interlace);
             svga->clock = (cpuclock * (double) (1ULL << 32)) / 44900000.0;
         } else {
             if (!vga_on && dev->ibm_mode) {
-                dev->h_disp = 640;
-                dev->dispend = 480;
+                if (dev->h_disp == 1024) {
+                    dev->h_disp = 640;
+                    dev->dispend = 480;
+                }
             }
 
             if (dev->interlace) {
@@ -3636,15 +3881,19 @@ ibm8514_recalctimings(svga_t *svga)
                 dev->v_syncstart >>= 1;
                 dev->v_total >>= 1;
             }
-            dev->rowoffset = 0x80;
-            dev->pitch = 1024;
+
+            if (ibm8514_has_vga) {
+                dev->pitch = dev->ext_pitch;
+                dev->rowoffset = dev->ext_crt_pitch;
+            } else
+                dev->rowoffset = 128;
 
             svga->clock = (cpuclock * (double) (1ULL << 32)) / 25175000.0;
         }
         svga->render = ibm8514_render_8bpp;
-        //pclog("Pitch = %d, mode = %d.\n", dev->pitch, dev->ibm_mode);
+        ibm8514_log("BPP=%d, Pitch = %d, rowoffset = %d, crtc13 = %02x, mode = %d, highres bit = %02x, has_vga? = %d.\n", dev->bpp, dev->pitch, dev->rowoffset, svga->crtc[0x13], dev->ibm_mode, dev->accel.advfunc_cntl & 4, ibm8514_has_vga);
     }
-    // pclog("8514 enabled, hdisp=%d, vtotal=%d, htotal=%d, dispend=%d, rowoffset=%d, split=%d, vsyncstart=%d, split=%08x\n", dev->hdisp, dev->vtotal, dev->htotal, dev->dispend, dev->rowoffset, dev->split, dev->vsyncstart, dev->split);
+    ibm8514_log("8514 enabled, hdisp=%d, vtotal=%d, htotal=%d, dispend=%d, rowoffset=%d, split=%d, vsyncstart=%d, split=%08x\n", dev->hdisp, dev->vtotal, dev->htotal, dev->dispend, dev->rowoffset, dev->split, dev->vsyncstart, dev->split);
 }
 
 static uint8_t
@@ -3697,6 +3946,7 @@ static void
 
     dev->type = info->flags;
     dev->ibm_mode = 1;
+    dev->bpp = 8;
 
     ibm8514_io_set(svga);
 
@@ -3770,7 +4020,7 @@ const device_t ibm8514_mca_device = {
 void
 ibm8514_device_add(void)
 {
-    if (!ibm8514_enabled)
+    if (!ibm8514_enabled || (ibm8514_enabled && ibm8514_has_vga))
         return;
 
     if (machine_has_bus(machine, MACHINE_BUS_MCA))
diff --git a/src/video/vid_ati_eeprom.c b/src/video/vid_ati_eeprom.c
index 2e10216e4..10b37212d 100644
--- a/src/video/vid_ati_eeprom.c
+++ b/src/video/vid_ati_eeprom.c
@@ -45,6 +45,26 @@ ati_eeprom_load(ati_eeprom_t *eeprom, char *fn, int type)
     fclose(f);
 }
 
+void
+ati_eeprom_load_mach8(ati_eeprom_t *eeprom, char *fn)
+{
+    FILE *f;
+    int   size;
+    eeprom->type = 0;
+    strncpy(eeprom->fn, fn, sizeof(eeprom->fn) - 1);
+    f    = nvr_fopen(eeprom->fn, "rb");
+    size = 128;
+    if (!f) { /*The ATI Graphics Ultra bios expects an immediate write to nvram if none is present at boot time otherwise
+            it would hang the machine.*/
+        memset(eeprom->data, 0, size);
+        f = nvr_fopen(eeprom->fn, "wb");
+        fwrite(eeprom->data, 1, size, f);
+    }
+    if (fread(eeprom->data, 1, size, f) != size)
+        memset(eeprom->data, 0, size);
+    fclose(f);
+}
+
 void
 ati_eeprom_save(ati_eeprom_t *eeprom)
 {
diff --git a/src/video/vid_ati_mach64.c b/src/video/vid_ati_mach64.c
index 273dbdcf6..8c32929ce 100644
--- a/src/video/vid_ati_mach64.c
+++ b/src/video/vid_ati_mach64.c
@@ -172,7 +172,7 @@ typedef struct mach64_t {
     uint32_t write_mask;
     uint32_t chain_mask;
 
-    uint32_t linear_base, old_linear_base;
+    uint32_t linear_base;
     uint32_t io_base;
 
     struct
@@ -185,6 +185,7 @@ typedef struct mach64_t {
         int      src_x_start, src_y_start;
         int      xinc, yinc;
         int      x_count, y_count;
+        int      xx_count;
         int      src_x_count, src_y_count;
         int      src_width1, src_height1;
         int      src_width2, src_height2;
@@ -487,35 +488,35 @@ mach64_recalctimings(svga_t *svga)
             case BPP_4:
                 if (mach64->type != MACH64_GX)
                     svga->render = svga_render_4bpp_highres;
-                svga->hdisp *= 8;
+                svga->hdisp <<= 3;
                 break;
             case BPP_8:
                 if (mach64->type != MACH64_GX)
                     svga->render = svga_render_8bpp_highres;
-                svga->hdisp *= 8;
-                svga->rowoffset /= 2;
+                svga->hdisp <<= 3;
+                svga->rowoffset >>= 1;
                 break;
             case BPP_15:
                 if (mach64->type != MACH64_GX)
                     svga->render = svga_render_15bpp_highres;
-                svga->hdisp *= 8;
+                svga->hdisp <<= 3;
                 break;
             case BPP_16:
                 if (mach64->type != MACH64_GX)
                     svga->render = svga_render_16bpp_highres;
-                svga->hdisp *= 8;
+                svga->hdisp <<= 3;
                 break;
             case BPP_24:
                 if (mach64->type != MACH64_GX)
                     svga->render = svga_render_24bpp_highres;
-                svga->hdisp *= 8;
+                svga->hdisp <<= 3;
                 svga->rowoffset = (svga->rowoffset * 3) / 2;
                 break;
             case BPP_32:
                 if (mach64->type != MACH64_GX)
                     svga->render = svga_render_32bpp_highres;
-                svga->hdisp *= 8;
-                svga->rowoffset *= 2;
+                svga->hdisp <<= 3;
+                svga->rowoffset <<= 1;
                 break;
         }
 
@@ -530,7 +531,7 @@ mach64_updatemapping(mach64_t *mach64)
 {
     svga_t *svga = &mach64->svga;
 
-    if (!(mach64->pci_regs[PCI_REG_COMMAND] & PCI_COMMAND_MEM)) {
+    if (mach64->pci && !(mach64->pci_regs[PCI_REG_COMMAND] & PCI_COMMAND_MEM)) {
         mach64_log("Update mapping - PCI disabled\n");
         mem_mapping_disable(&svga->mapping);
         mem_mapping_disable(&mach64->linear_mapping);
@@ -543,47 +544,60 @@ mach64_updatemapping(mach64_t *mach64)
     mem_mapping_disable(&mach64->mmio_mapping);
     switch (svga->gdcreg[6] & 0xc) {
         case 0x0: /*128k at A0000*/
-            mem_mapping_set_handler(&mach64->svga.mapping, mach64_read, mach64_readw, mach64_readl, mach64_write, mach64_writew, mach64_writel);
-            mem_mapping_set_p(&mach64->svga.mapping, mach64);
+            mem_mapping_set_handler(&svga->mapping, mach64_read, mach64_readw, mach64_readl, mach64_write, mach64_writew, mach64_writel);
+            mem_mapping_set_p(&svga->mapping, mach64);
             mem_mapping_set_addr(&svga->mapping, 0xa0000, 0x20000);
             mem_mapping_enable(&mach64->mmio_mapping);
             svga->banked_mask = 0xffff;
             break;
         case 0x4: /*64k at A0000*/
-            mem_mapping_set_handler(&mach64->svga.mapping, mach64_read, mach64_readw, mach64_readl, mach64_write, mach64_writew, mach64_writel);
-            mem_mapping_set_p(&mach64->svga.mapping, mach64);
+            mem_mapping_set_handler(&svga->mapping, mach64_read, mach64_readw, mach64_readl, mach64_write, mach64_writew, mach64_writel);
+            mem_mapping_set_p(&svga->mapping, mach64);
             mem_mapping_set_addr(&svga->mapping, 0xa0000, 0x10000);
             svga->banked_mask = 0xffff;
             break;
         case 0x8: /*32k at B0000*/
-            mem_mapping_set_handler(&mach64->svga.mapping, svga_read, svga_readw, svga_readl, svga_write, svga_writew, svga_writel);
-            mem_mapping_set_p(&mach64->svga.mapping, svga);
+            mem_mapping_set_handler(&svga->mapping, svga_read, svga_readw, svga_readl, svga_write, svga_writew, svga_writel);
+            mem_mapping_set_p(&svga->mapping, svga);
             mem_mapping_set_addr(&svga->mapping, 0xb0000, 0x08000);
             svga->banked_mask = 0x7fff;
             break;
         case 0xC: /*32k at B8000*/
-            mem_mapping_set_handler(&mach64->svga.mapping, svga_read, svga_readw, svga_readl, svga_write, svga_writew, svga_writel);
-            mem_mapping_set_p(&mach64->svga.mapping, svga);
+            mem_mapping_set_handler(&svga->mapping, svga_read, svga_readw, svga_readl, svga_write, svga_writew, svga_writel);
+            mem_mapping_set_p(&svga->mapping, svga);
             mem_mapping_set_addr(&svga->mapping, 0xb8000, 0x08000);
             svga->banked_mask = 0x7fff;
             break;
     }
+
+    mach64_log("Mach64 linear aperture = %08x.\n", mach64->linear_base);
     if (mach64->linear_base) {
         if (mach64->type == MACH64_GX) {
             if ((mach64->config_cntl & 3) == 2) {
                 /*8 MB aperture*/
                 mem_mapping_set_addr(&mach64->linear_mapping, mach64->linear_base, (8 << 20) - 0x4000);
                 mem_mapping_set_addr(&mach64->mmio_linear_mapping, mach64->linear_base + ((8 << 20) - 0x4000), 0x4000);
-            } else {
+            } else if ((mach64->config_cntl & 3) == 1) {
                 /*4 MB aperture*/
                 mem_mapping_set_addr(&mach64->linear_mapping, mach64->linear_base, (4 << 20) - 0x4000);
                 mem_mapping_set_addr(&mach64->mmio_linear_mapping, mach64->linear_base + ((4 << 20) - 0x4000), 0x4000);
+            } else {
+                /*Disable aperture on reserved values*/
+                mem_mapping_disable(&mach64->linear_mapping);
+                mem_mapping_disable(&mach64->mmio_linear_mapping);
             }
         } else {
-            /*2*8 MB aperture*/
-            mem_mapping_set_addr(&mach64->linear_mapping, mach64->linear_base, (8 << 20) - 0x4000);
-            mem_mapping_set_addr(&mach64->mmio_linear_mapping, mach64->linear_base + ((8 << 20) - 0x4000), 0x4000);
-            mem_mapping_set_addr(&mach64->mmio_linear_mapping_2, mach64->linear_base + ((16 << 20) - 0x4000), 0x4000);
+            if ((mach64->config_cntl & 3) == 2) {
+                /*2*8 MB aperture*/
+                mem_mapping_set_addr(&mach64->linear_mapping, mach64->linear_base, (8 << 20) - 0x4000);
+                mem_mapping_set_addr(&mach64->mmio_linear_mapping, mach64->linear_base + ((8 << 20) - 0x4000), 0x4000);
+                mem_mapping_set_addr(&mach64->mmio_linear_mapping_2, mach64->linear_base + ((16 << 20) - 0x4000), 0x4000);
+            } else {
+                /*Disable aperture on reserved values*/
+                mem_mapping_disable(&mach64->linear_mapping);
+                mem_mapping_disable(&mach64->mmio_linear_mapping);
+                mem_mapping_disable(&mach64->mmio_linear_mapping_2);
+            }
         }
     } else {
         mem_mapping_disable(&mach64->linear_mapping);
@@ -621,7 +635,7 @@ mach64_wait_fifo_idle(mach64_t *mach64)
 }
 
 #define READ8(addr, var)                \
-    switch ((addr) &3) {                \
+    switch ((addr) & 3) {               \
         case 0:                         \
             ret = (var) &0xff;          \
             break;                      \
@@ -637,7 +651,7 @@ mach64_wait_fifo_idle(mach64_t *mach64)
     }
 
 #define WRITE8(addr, var, val)                        \
-    switch ((addr) &3) {                              \
+    switch ((addr) & 3) {                             \
         case 0:                                       \
             var = (var & 0xffffff00) | (val);         \
             break;                                    \
@@ -1181,8 +1195,13 @@ mach64_start_fill(mach64_t *mach64)
 {
     mach64->accel.dst_x       = 0;
     mach64->accel.dst_y       = 0;
+
     mach64->accel.dst_x_start = (mach64->dst_y_x >> 16) & 0xfff;
-    mach64->accel.dst_y_start = mach64->dst_y_x & 0xfff;
+    if (((mach64->dst_y_x >> 16) & 0x1000))
+        mach64->accel.dst_x_start |= ~0xfff;
+    mach64->accel.dst_y_start = mach64->dst_y_x & 0x3fff;
+    if (mach64->dst_y_x & 0x4000)
+        mach64->accel.dst_y_start |= ~0x3fff;
 
     mach64->accel.dst_width  = (mach64->dst_height_width >> 16) & 0x1fff;
     mach64->accel.dst_height = mach64->dst_height_width & 0x1fff;
@@ -1193,11 +1212,18 @@ mach64_start_fill(mach64_t *mach64)
     }
 
     mach64->accel.x_count = mach64->accel.dst_width;
+    mach64->accel.xx_count = 0;
 
     mach64->accel.src_x       = 0;
     mach64->accel.src_y       = 0;
+
     mach64->accel.src_x_start = (mach64->src_y_x >> 16) & 0xfff;
-    mach64->accel.src_y_start = mach64->src_y_x & 0xfff;
+    if (((mach64->src_y_x >> 16) & 0x1000))
+        mach64->accel.src_x_start |= ~0xfff;
+    mach64->accel.src_y_start = mach64->src_y_x & 0x3fff;
+    if (mach64->src_y_x & 0x4000)
+        mach64->accel.src_y_start |= ~0x3fff;
+
     if (mach64->src_cntl & SRC_LINEAR_EN)
         mach64->accel.src_x_count = 0x7ffffff; /*Essentially infinite*/
     else
@@ -1219,11 +1245,11 @@ mach64_start_fill(mach64_t *mach64)
                mach64->src_height1_width1,
                mach64->src_height2_width2);
 
-    mach64->accel.src_pitch  = (mach64->src_off_pitch >> 22) * 8;
-    mach64->accel.src_offset = (mach64->src_off_pitch & 0xfffff) * 8;
+    mach64->accel.src_pitch  = (mach64->src_off_pitch >> 22) << 3;
+    mach64->accel.src_offset = (mach64->src_off_pitch & 0xfffff) << 3;
 
-    mach64->accel.dst_pitch  = (mach64->dst_off_pitch >> 22) * 8;
-    mach64->accel.dst_offset = (mach64->dst_off_pitch & 0xfffff) * 8;
+    mach64->accel.dst_pitch  = (mach64->dst_off_pitch >> 22) << 3;
+    mach64->accel.dst_offset = (mach64->dst_off_pitch & 0xfffff) << 3;
 
     mach64->accel.mix_fg = (mach64->dp_mix >> 16) & 0x1f;
     mach64->accel.mix_bg = mach64->dp_mix & 0x1f;
@@ -1308,16 +1334,24 @@ void
 mach64_start_line(mach64_t *mach64)
 {
     mach64->accel.dst_x = (mach64->dst_y_x >> 16) & 0xfff;
-    mach64->accel.dst_y = mach64->dst_y_x & 0xfff;
+    if (((mach64->dst_y_x >> 16) & 0x1000))
+        mach64->accel.dst_x |= ~0xfff;
+    mach64->accel.dst_y = mach64->dst_y_x & 0x3fff;
+    if (mach64->dst_y_x & 0x4000)
+        mach64->accel.dst_y |= ~0x3fff;
 
     mach64->accel.src_x = (mach64->src_y_x >> 16) & 0xfff;
-    mach64->accel.src_y = mach64->src_y_x & 0xfff;
+    if (((mach64->src_y_x >> 16) & 0x1000))
+        mach64->accel.src_x |= ~0xfff;
+    mach64->accel.src_y = mach64->src_y_x & 0x3fff;
+    if (mach64->src_y_x & 0x4000)
+        mach64->accel.src_y |= ~0x3fff;
 
-    mach64->accel.src_pitch  = (mach64->src_off_pitch >> 22) * 8;
-    mach64->accel.src_offset = (mach64->src_off_pitch & 0xfffff) * 8;
+    mach64->accel.src_pitch  = (mach64->src_off_pitch >> 22) << 3;
+    mach64->accel.src_offset = (mach64->src_off_pitch & 0xfffff) << 3;
 
-    mach64->accel.dst_pitch  = (mach64->dst_off_pitch >> 22) * 8;
-    mach64->accel.dst_offset = (mach64->dst_off_pitch & 0xfffff) * 8;
+    mach64->accel.dst_pitch  = (mach64->dst_off_pitch >> 22) << 3;
+    mach64->accel.dst_offset = (mach64->dst_off_pitch & 0xfffff) << 3;
 
     mach64->accel.mix_fg = (mach64->dp_mix >> 16) & 0x1f;
     mach64->accel.mix_bg = mach64->dp_mix & 0x1f;
@@ -1344,9 +1378,6 @@ mach64_start_line(mach64_t *mach64)
     else
         mach64->accel.dst_offset >>= mach64->accel.dst_size;
 
-    /*        mach64->accel.src_pitch *= mach64_inc[mach64->accel.src_pix_width];
-            mach64->accel.dst_pitch *= mach64_inc[mach64->accel.dst_pix_width];*/
-
     mach64->accel.source_host = ((mach64->dp_src & 7) == SRC_HOST) || (((mach64->dp_src >> 8) & 7) == SRC_HOST);
 
     for (uint8_t y = 0; y < 8; y++) {
@@ -1446,6 +1477,7 @@ mach64_start_line(mach64_t *mach64)
             break;                                               \
         case 0x17:                                               \
             dest_dat = (dest_dat + src_dat) >> 1;                \
+            break;                                               \
     }
 
 #define WRITE(addr, width)                                                                  \
@@ -1483,24 +1515,31 @@ mach64_blit(uint32_t cpu_dat, int count, mach64_t *mach64)
         mach64_log("mach64_blit : return as not busy\n");
         return;
     }
+
     switch (mach64->accel.op) {
         case OP_RECT:
             while (count) {
+                uint8_t write_mask = 0;
                 uint32_t src_dat  = 0;
                 uint32_t dest_dat;
                 uint32_t host_dat = 0;
                 uint32_t old_dest_dat;
                 int      mix   = 0;
-                int      dst_x = (mach64->accel.dst_x + mach64->accel.dst_x_start) & 0xfff;
-                int      dst_y = (mach64->accel.dst_y + mach64->accel.dst_y_start) & 0xfff;
+                int      dst_x;
+                int      dst_y;
                 int      src_x;
-                int      src_y = (mach64->accel.src_y + mach64->accel.src_y_start) & 0xfff;
+                int      src_y;
+
+                dst_x = (mach64->accel.dst_x + mach64->accel.dst_x_start) & 0xfff;
+                dst_y = (mach64->accel.dst_y + mach64->accel.dst_y_start) & 0x3fff;
 
                 if (mach64->src_cntl & SRC_LINEAR_EN)
                     src_x = mach64->accel.src_x;
                 else
                     src_x = (mach64->accel.src_x + mach64->accel.src_x_start) & 0xfff;
 
+                src_y = (mach64->accel.src_y + mach64->accel.src_y_start) & 0x3fff;
+
                 if (mach64->accel.source_host) {
                     host_dat = cpu_dat;
                     switch (mach64->accel.host_size) {
@@ -1525,7 +1564,7 @@ mach64_blit(uint32_t cpu_dat, int count, mach64_t *mach64)
                             mix = cpu_dat & 1;
                             cpu_dat >>= 1;
                         } else {
-                            mix = cpu_dat >> 31;
+                            mix = cpu_dat >> 0x1f;
                             cpu_dat <<= 1;
                         }
                         break;
@@ -1544,7 +1583,7 @@ mach64_blit(uint32_t cpu_dat, int count, mach64_t *mach64)
                         break;
                 }
 
-                if (dst_x >= mach64->accel.sc_left && dst_x <= mach64->accel.sc_right && dst_y >= mach64->accel.sc_top && dst_y <= mach64->accel.sc_bottom) {
+                if ((dst_x) >= mach64->accel.sc_left && (dst_x) <= mach64->accel.sc_right && (dst_y) >= mach64->accel.sc_top && (dst_y) <= mach64->accel.sc_bottom) {
                     switch (mix ? mach64->accel.source_fg : mach64->accel.source_bg) {
                         case SRC_HOST:
                             src_dat = host_dat;
@@ -1553,24 +1592,42 @@ mach64_blit(uint32_t cpu_dat, int count, mach64_t *mach64)
                             READ(mach64->accel.src_offset + (src_y * mach64->accel.src_pitch) + src_x, src_dat, mach64->accel.src_size);
                             break;
                         case SRC_FG:
-                            if ((mach64->dst_cntl & (DST_LAST_PEL | DST_X_DIR | DST_Y_DIR | DST_24_ROT_EN)) == (DST_LAST_PEL | DST_X_DIR | DST_Y_DIR | DST_24_ROT_EN)) {
-                                if ((mach64->accel.x_count % 3) == 2)
-                                    src_dat = mach64->accel.dp_frgd_clr & 0xff;
-                                else if ((mach64->accel.x_count % 3) == 1)
-                                    src_dat = (mach64->accel.dp_frgd_clr >> 8) & 0xff;
-                                else if ((mach64->accel.x_count % 3) == 0)
-                                    src_dat = (mach64->accel.dp_frgd_clr >> 16) & 0xff;
+                            if (((mach64->crtc_gen_cntl >> 8) & 7) == BPP_24) {
+                                if (mach64->accel.xinc == -1) {
+                                    if ((mach64->accel.xx_count % 3) == 2)
+                                        src_dat = mach64->accel.dp_frgd_clr & 0xff;
+                                    else if ((mach64->accel.xx_count % 3) == 1)
+                                        src_dat = (mach64->accel.dp_frgd_clr >> 8) & 0xff;
+                                    else
+                                        src_dat = (mach64->accel.dp_frgd_clr >> 16) & 0xff;
+                                } else {
+                                    if ((mach64->accel.xx_count % 3) == 2)
+                                        src_dat = (mach64->accel.dp_frgd_clr >> 16) & 0xff;
+                                    else if ((mach64->accel.xx_count % 3) == 1)
+                                        src_dat = (mach64->accel.dp_frgd_clr >> 8) & 0xff;
+                                    else
+                                        src_dat = mach64->accel.dp_frgd_clr & 0xff;
+                                }
                             } else
                                 src_dat = mach64->accel.dp_frgd_clr;
                             break;
                         case SRC_BG:
-                            if ((mach64->dst_cntl & (DST_LAST_PEL | DST_X_DIR | DST_Y_DIR | DST_24_ROT_EN)) == (DST_LAST_PEL | DST_X_DIR | DST_Y_DIR | DST_24_ROT_EN)) {
-                                if ((mach64->accel.x_count % 3) == 2)
-                                    src_dat = mach64->accel.dp_bkgd_clr & 0xff;
-                                else if ((mach64->accel.x_count % 3) == 1)
-                                    src_dat = (mach64->accel.dp_bkgd_clr >> 8) & 0xff;
-                                else if ((mach64->accel.x_count % 3) == 0)
-                                    src_dat = (mach64->accel.dp_bkgd_clr >> 16) & 0xff;
+                            if (((mach64->crtc_gen_cntl >> 8) & 7) == BPP_24) {
+                                if (mach64->accel.xinc == -1) {
+                                    if ((mach64->accel.xx_count % 3) == 2)
+                                        src_dat = mach64->accel.dp_bkgd_clr & 0xff;
+                                    else if ((mach64->accel.xx_count % 3) == 1)
+                                        src_dat = (mach64->accel.dp_bkgd_clr >> 8) & 0xff;
+                                    else
+                                        src_dat = (mach64->accel.dp_bkgd_clr >> 16) & 0xff;
+                                } else {
+                                    if ((mach64->accel.xx_count % 3) == 2)
+                                        src_dat = (mach64->accel.dp_bkgd_clr >> 16) & 0xff;
+                                    else if ((mach64->accel.xx_count % 3) == 1)
+                                        src_dat = (mach64->accel.dp_bkgd_clr >> 8) & 0xff;
+                                    else
+                                        src_dat = mach64->accel.dp_bkgd_clr & 0xff;
+                                }
                             } else
                                 src_dat = mach64->accel.dp_bkgd_clr;
                             break;
@@ -1595,7 +1652,7 @@ mach64_blit(uint32_t cpu_dat, int count, mach64_t *mach64)
                     }
 
                     if (!(mach64->dst_cntl & DST_POLYGON_EN) || mach64->accel.poly_draw) {
-                        READ(mach64->accel.dst_offset + (dst_y * mach64->accel.dst_pitch) + dst_x, dest_dat, mach64->accel.dst_size);
+                        READ(mach64->accel.dst_offset + ((dst_y) * mach64->accel.dst_pitch) + (dst_x), dest_dat, mach64->accel.dst_size);
 
                         switch (mach64->accel.clr_cmp_fn) {
                             case 1: /*TRUE*/
@@ -1612,19 +1669,30 @@ mach64_blit(uint32_t cpu_dat, int count, mach64_t *mach64)
                         if (!cmp_clr) {
                             old_dest_dat = dest_dat;
                             MIX
-                                dest_dat
-                                = (dest_dat & mach64->accel.write_mask) | (old_dest_dat & ~mach64->accel.write_mask);
+
+                            if (((mach64->crtc_gen_cntl >> 8) & 7) == BPP_24) {
+                                if (mach64->accel.xinc == -1) {
+                                    if ((mach64->accel.xx_count % 3) == 2)
+                                        write_mask = mach64->accel.write_mask & 0xff;
+                                    else if ((mach64->accel.xx_count % 3) == 1)
+                                        write_mask = (mach64->accel.write_mask >> 8) & 0xff;
+                                    else
+                                        write_mask = (mach64->accel.write_mask >> 16) & 0xff;
+                                } else {
+                                    if ((mach64->accel.xx_count % 3) == 2)
+                                        write_mask = (mach64->accel.write_mask >> 16) & 0xff;
+                                    else if ((mach64->accel.xx_count % 3) == 1)
+                                        write_mask = (mach64->accel.write_mask >> 8) & 0xff;
+                                    else
+                                        write_mask = mach64->accel.write_mask & 0xff;
+                                }
+                                dest_dat = (dest_dat & write_mask) | (old_dest_dat & ~write_mask);
+                            } else {
+                                dest_dat = (dest_dat & mach64->accel.write_mask) | (old_dest_dat & ~mach64->accel.write_mask);
+                            }
                         }
 
-                        WRITE(mach64->accel.dst_offset + (dst_y * mach64->accel.dst_pitch) + dst_x, mach64->accel.dst_size);
-                    }
-                }
-
-                if (((mach64->crtc_gen_cntl >> 8) & 7) == BPP_24) {
-                    if ((mach64->dst_cntl & (DST_LAST_PEL | DST_X_DIR | DST_Y_DIR | DST_24_ROT_EN)) != (DST_LAST_PEL | DST_X_DIR | DST_Y_DIR | DST_24_ROT_EN)) {
-                        mach64->accel.dp_frgd_clr = ((mach64->accel.dp_frgd_clr >> 8) & 0xffff) | (mach64->accel.dp_frgd_clr << 16);
-                        mach64->accel.dp_bkgd_clr = ((mach64->accel.dp_bkgd_clr >> 8) & 0xffff) | (mach64->accel.dp_bkgd_clr << 16);
-                        mach64->accel.write_mask  = ((mach64->accel.write_mask >> 8) & 0xffff) | (mach64->accel.write_mask << 16);
+                        WRITE(mach64->accel.dst_offset + ((dst_y) * mach64->accel.dst_pitch) + (dst_x), mach64->accel.dst_size);
                     }
                 }
 
@@ -1636,14 +1704,18 @@ mach64_blit(uint32_t cpu_dat, int count, mach64_t *mach64)
                         mach64->accel.src_x = 0;
                         if ((mach64->src_cntl & (SRC_PATT_ROT_EN | SRC_PATT_EN)) == (SRC_PATT_ROT_EN | SRC_PATT_EN)) {
                             mach64->accel.src_x_start = (mach64->src_y_x_start >> 16) & 0xfff;
+                            if ((mach64->src_y_x_start >> 16) & 0x1000)
+                                mach64->accel.src_x_start |= ~0xfff;
                             mach64->accel.src_x_count = mach64->accel.src_width2;
                         } else
                             mach64->accel.src_x_count = mach64->accel.src_width1;
                     }
                 }
 
+                mach64->accel.xx_count++;
                 mach64->accel.x_count--;
                 if (mach64->accel.x_count <= 0) {
+                    mach64->accel.xx_count = 0;
                     mach64->accel.x_count = mach64->accel.dst_width;
                     mach64->accel.dst_x   = 0;
                     mach64->accel.dst_y += mach64->accel.yinc;
@@ -1657,7 +1729,9 @@ mach64_blit(uint32_t cpu_dat, int count, mach64_t *mach64)
                         if (mach64->accel.src_y_count <= 0) {
                             mach64->accel.src_y = 0;
                             if ((mach64->src_cntl & (SRC_PATT_ROT_EN | SRC_PATT_EN)) == (SRC_PATT_ROT_EN | SRC_PATT_EN)) {
-                                mach64->accel.src_y_start = mach64->src_y_x_start & 0xfff;
+                                mach64->accel.src_y_start = mach64->src_y_x_start & 0x3fff;
+                                if (mach64->src_y_x_start & 0x4000)
+                                    mach64->accel.src_y_start |= ~0x3fff;
                                 mach64->accel.src_y_count = mach64->accel.src_height2;
                             } else
                                 mach64->accel.src_y_count = mach64->accel.src_height1;
@@ -1702,7 +1776,7 @@ mach64_blit(uint32_t cpu_dat, int count, mach64_t *mach64)
 
                     if (mach64->accel.source_host) {
                         host_dat = cpu_dat;
-                        switch (mach64->accel.src_size) {
+                        switch (mach64->accel.host_size) {
                             case 0:
                                 cpu_dat >>= 8;
                                 count -= 8;
@@ -1783,12 +1857,10 @@ mach64_blit(uint32_t cpu_dat, int count, mach64_t *mach64)
                         if (!cmp_clr)
                             MIX
 
-                                if (!(mach64->dst_cntl & DST_Y_MAJOR))
-                            {
-                                if (x == 0)
-                                    dest_dat &= ~1;
-                            }
-                        else {
+                        if (!(mach64->dst_cntl & DST_Y_MAJOR)) {
+                            if (!x)
+                                dest_dat &= ~1;
+                        } else {
                             if (x == (mach64->accel.x_count - 1))
                                 dest_dat &= ~1;
                         }
@@ -1831,7 +1903,7 @@ mach64_blit(uint32_t cpu_dat, int count, mach64_t *mach64)
 
                     if (mach64->accel.source_host) {
                         host_dat = cpu_dat;
-                        switch (mach64->accel.src_size) {
+                        switch (mach64->accel.host_size) {
                             case 0:
                                 cpu_dat >>= 8;
                                 count -= 8;
@@ -1909,7 +1981,7 @@ mach64_blit(uint32_t cpu_dat, int count, mach64_t *mach64)
                         if (!cmp_clr)
                             MIX
 
-                                WRITE(mach64->accel.dst_offset + (mach64->accel.dst_y * mach64->accel.dst_pitch) + mach64->accel.dst_x, mach64->accel.dst_size);
+                        WRITE(mach64->accel.dst_offset + (mach64->accel.dst_y * mach64->accel.dst_pitch) + mach64->accel.dst_x, mach64->accel.dst_size);
                     }
 
                     mach64->accel.x_count--;
@@ -2347,6 +2419,7 @@ mach64_ext_readb(uint32_t addr, void *p)
                     mach64->config_cntl = (mach64->config_cntl & ~0x3ff0) | ((mach64->linear_base >> 22) << 4);
                 else
                     mach64->config_cntl = (mach64->config_cntl & ~0x3ff0) | ((mach64->linear_base >> 24) << 4);
+
                 READ8(addr, mach64->config_cntl);
                 break;
             case 0xe0:
@@ -2667,11 +2740,15 @@ mach64_ext_readb(uint32_t addr, void *p)
 
             case 0x310:
             case 0x311:
-                if (!FIFO_EMPTY)
-                    wake_fifo_thread(mach64);
-                ret = 0;
-                if (FIFO_FULL)
-                    ret = 0xff;
+                if (((mach64->crtc_gen_cntl >> 8) & 7) == BPP_24) {
+                    ret = 0;
+                } else {
+                    if (!FIFO_EMPTY)
+                        wake_fifo_thread(mach64);
+                    ret = 0;
+                    if (FIFO_FULL)
+                        ret = 0xff;
+                }
                 break;
 
             case 0x320:
@@ -2697,7 +2774,10 @@ mach64_ext_readb(uint32_t addr, void *p)
                 break;
 
             case 0x338:
-                ret = FIFO_EMPTY ? 0 : 1;
+                if (((mach64->crtc_gen_cntl >> 8) & 7) == BPP_24)
+                    ret = 0;
+                else
+                    ret = FIFO_EMPTY ? 0 : 1;
                 break;
 
             default:
@@ -2876,7 +2956,10 @@ mach64_ext_writeb(uint32_t addr, uint8_t val, void *p)
 
         mach64_log("nmach64_ext_writeb: addr=%04x val=%02x\n", addr, val);
     } else if (addr & 0x300) {
-        mach64_queue(mach64, addr & 0x3ff, val, FIFO_WRITE_BYTE);
+        if (((mach64->crtc_gen_cntl >> 8) & 7) == BPP_24)
+            mach64_accel_write_fifo(mach64, addr & 0x3ff, val);
+        else
+            mach64_queue(mach64, addr & 0x3ff, val, FIFO_WRITE_BYTE);
     } else
         switch (addr & 0x3ff) {
             case 0x00:
@@ -3099,7 +3182,10 @@ mach64_ext_writew(uint32_t addr, uint16_t val, void *p)
         mach64_ext_writeb(addr, val, p);
         mach64_ext_writeb(addr + 1, val >> 8, p);
     } else if (addr & 0x300) {
-        mach64_queue(mach64, addr & 0x3fe, val, FIFO_WRITE_WORD);
+        if (((mach64->crtc_gen_cntl >> 8) & 7) == BPP_24)
+            mach64_accel_write_fifo_w(mach64, addr & 0x3fe, val);
+        else
+            mach64_queue(mach64, addr & 0x3fe, val, FIFO_WRITE_WORD);
     } else
         switch (addr & 0x3fe) {
             default:
@@ -3120,7 +3206,10 @@ mach64_ext_writel(uint32_t addr, uint32_t val, void *p)
         mach64_ext_writew(addr, val, p);
         mach64_ext_writew(addr + 2, val >> 16, p);
     } else if (addr & 0x300) {
-        mach64_queue(mach64, addr & 0x3fc, val, FIFO_WRITE_DWORD);
+        if (((mach64->crtc_gen_cntl >> 8) & 7) == BPP_24)
+            mach64_accel_write_fifo_l(mach64, addr & 0x3fc, val);
+        else
+            mach64_queue(mach64, addr & 0x3fc, val, FIFO_WRITE_DWORD);
     } else
         switch (addr & 0x3fc) {
             default:
@@ -4210,9 +4299,8 @@ mach64_common_init(const device_t *info)
 
     mach64_io_set(mach64);
 
-    if (info->flags & DEVICE_PCI) {
+    if (info->flags & DEVICE_PCI)
         mach64->card = pci_add_card(PCI_ADD_VIDEO, mach64_pci_read, mach64_pci_write, mach64);
-    }
 
     mach64->pci_regs[PCI_REG_COMMAND] = 3;
     mach64->pci_regs[0x30]            = 0x00;
@@ -4252,7 +4340,7 @@ mach64gx_init(const device_t *info)
     mach64->type           = MACH64_GX;
     mach64->pci            = !!(info->flags & DEVICE_PCI);
     mach64->pci_id         = (int) 'X' | ((int) 'G' << 8);
-    mach64->config_chip_id = 0x020000d7;
+    mach64->config_chip_id = 0x000000d7;
     mach64->dac_cntl       = 5 << 16;             /*ATI 68860 RAMDAC*/
     mach64->config_stat0   = (5 << 9) | (3 << 3); /*ATI-68860, 256Kx16 DRAM*/
     if (info->flags & DEVICE_PCI)
@@ -4282,10 +4370,7 @@ mach64vt2_init(const device_t *info)
     mach64_t *mach64 = mach64_common_init(info);
     svga_t   *svga   = &mach64->svga;
 
-    if (info->flags & DEVICE_PCI)
-        video_inform(VIDEO_FLAG_TYPE_SPECIAL, &timing_mach64_pci);
-    else
-        video_inform(VIDEO_FLAG_TYPE_SPECIAL, &timing_mach64_vlb);
+    video_inform(VIDEO_FLAG_TYPE_SPECIAL, &timing_mach64_pci);
 
     mach64->type                 = MACH64_VT2;
     mach64->pci                  = 1;
@@ -4299,8 +4384,7 @@ mach64vt2_init(const device_t *info)
 
     rom_init(&mach64->bios_rom, BIOS_ROMVT2_PATH, 0xc0000, 0x8000, 0x7fff, 0, MEM_MAPPING_EXTERNAL);
 
-    if (info->flags & DEVICE_PCI)
-        mem_mapping_disable(&mach64->bios_rom.mapping);
+    mem_mapping_disable(&mach64->bios_rom.mapping);
 
     svga->vblank_start = mach64_vblank_start;
 
diff --git a/src/video/vid_ati_mach8.c b/src/video/vid_ati_mach8.c
new file mode 100644
index 000000000..9c0cbb616
--- /dev/null
+++ b/src/video/vid_ati_mach8.c
@@ -0,0 +1,5566 @@
+/*
+ * 86Box    A hypervisor and IBM PC system emulator that specializes in
+ *          running old operating systems and software designed for IBM
+ *          PC systems and compatibles from 1981 through fairly recent
+ *          system designs based on the PCI bus.
+ *
+ *          This file is part of the 86Box distribution.
+ *
+ *          Emulation of the 8514/A-compatible Mach8 and Mach32 graphics
+ *          chips from ATI for the ISA/VLB/MCA/PCI buses.
+ *
+ *
+ *
+ * Authors: TheCollector1995.
+ *
+ *          Copyright 2022-2023 TheCollector1995.
+ */
+#include <stdarg.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <string.h>
+#include <stdlib.h>
+#include <wchar.h>
+#include <stdatomic.h>
+#define HAVE_STDARG_H
+#include <86box/86box.h>
+#include <86box/device.h>
+#include <86box/io.h>
+#include <86box/mem.h>
+#include <86box/timer.h>
+#include <86box/mca.h>
+#include <86box/pci.h>
+#include <86box/rom.h>
+#include <86box/plat.h>
+#include <86box/thread.h>
+#include <86box/video.h>
+#include <86box/i2c.h>
+#include <86box/vid_ddc.h>
+#include <86box/vid_svga.h>
+#include <86box/vid_svga_render.h>
+#include <86box/vid_ati_eeprom.h>
+
+#define BIOS_MACH8_ROM_PATH             "roms/video/mach8/BIOS.BIN"
+#define BIOS_MACH32_ISA_ROM_PATH        "roms/video/mach32/MACH32ISA.VBI"
+#define BIOS_MACH32_VLB_ROM_PATH        "roms/video/mach32/MACH32VLB.VBI"
+#define BIOS_MACH32_PCI_ROM_PATH        "roms/video/mach32/MACH32PCI.BIN"
+
+typedef struct mach_t {
+    ati_eeprom_t eeprom;
+    svga_t       svga;
+
+    rom_t bios_rom;
+    mem_mapping_t mmio_linear_mapping;
+
+    int mca_bus;
+    int pci_bus;
+    int vlb_bus;
+    uint8_t regs[256];
+    uint8_t pci_regs[256];
+    uint8_t int_line;
+    int     card;
+    int     index;
+
+    uint32_t memory;
+
+    uint16_t config1;
+    uint16_t config2;
+
+    uint8_t pos_regs[8];
+    uint8_t cursor_col_0, cursor_col_1;
+    uint8_t ext_cur_col_0_r, ext_cur_col_1_r;
+    uint8_t ext_cur_col_0_g, ext_cur_col_1_g;
+    uint16_t cursor_offset_lo, cursor_offset_hi;
+    uint16_t cursor_x, cursor_y;
+    uint16_t misc;
+    uint16_t memory_aperture;
+    uint16_t local_cntl;
+    uint32_t linear_base;
+    uint8_t ap_size;
+    uint8_t bank_w, bank_r;
+
+    struct {
+        uint8_t line_idx;
+        int16_t line_array[6];
+        uint8_t patt_idx;
+        uint8_t patt_len;
+        uint8_t pix_trans[2];
+        uint8_t eeprom_control;
+        uint16_t dest_x_end;
+        uint16_t dest_x_start;
+        uint16_t dest_y_end;
+        uint16_t src_x_end;
+        uint16_t src_x_start;
+        uint16_t src_x, src_y;
+        int16_t bres_count;
+        uint16_t clock_sel;
+        uint16_t crt_offset_lo;
+        uint16_t crt_offset_hi;
+        uint16_t dest_cmp_fn;
+        uint16_t dp_config;
+        uint16_t ext_ge_config;
+        uint16_t ge_offset_lo;
+        uint16_t ge_offset_hi;
+        uint16_t linedraw_opt;
+        uint16_t max_waitstates;
+        uint8_t patt_data_idx;
+        uint8_t patt_data[0x18];
+        uint16_t scan_to_x;
+        uint16_t scratch0;
+        uint16_t scratch1;
+        uint16_t test;
+        uint16_t pattern;
+        uint8_t test2[2], test3[2];
+        int src_y_dir;
+        int cmd_type;
+        int block_write_mono_pattern_enable;
+        int mono_pattern_enable;
+        int16_t cx_end_line, cy_end_line;
+        int16_t cx, cx_end, cy_end, dx, dx_end, dy_end;
+        int16_t dx_start, dy_start;
+        int16_t cy, sx_start, sx_end;
+        int16_t sx, x_count, xx_count, xxx_count;
+        int16_t sy, y_count;
+        int16_t err;
+        int16_t width, src_width;
+        int16_t height;
+        int poly_src, temp_cnt;
+        int stepx, stepy, src_stepx;
+        uint8_t color_pattern[16];
+        uint8_t color_pattern_full[32];
+        uint16_t color_pattern_word[8];
+        int mono_pattern[8][8];
+        uint32_t ge_offset;
+        uint32_t crt_offset;
+        uint32_t patt_len_reg;
+        int poly_fill;
+        uint16_t dst_clr_cmp_mask;
+        int clip_overrun;
+        int color_pattern_idx;
+    } accel;
+
+    atomic_int force_busy, force_busy2;
+} mach_t;
+
+static video_timings_t timing_gfxultra_isa = { .type = VIDEO_ISA, .write_b = 3, .write_w = 3, .write_l = 6, .read_b = 5, .read_w = 5, .read_l = 10 };
+static video_timings_t timing_mach32_vlb = { .type = VIDEO_BUS, .write_b = 2, .write_w = 2, .write_l = 1, .read_b = 20, .read_w = 20, .read_l = 21 };
+static video_timings_t timing_mach32_pci = { .type = VIDEO_PCI, .write_b = 2, .write_w = 2, .write_l = 1, .read_b = 20, .read_w = 20, .read_l = 21 };
+
+
+static void     mach_accel_outb(uint16_t port, uint8_t val, void *p);
+static void     mach_accel_outw(uint16_t port, uint16_t val, void *p);
+static uint8_t  mach_accel_inb(uint16_t port, void *p);
+static uint16_t mach_accel_inw(uint16_t port, void *p);
+
+static void mach32_updatemapping(mach_t *mach);
+
+#ifdef ENABLE_MACH_LOG
+int mach_do_log = ENABLE_MACH_LOG;
+
+static void
+mach_log(const char *fmt, ...)
+{
+    va_list ap;
+
+    if (mach_do_log) {
+        va_start(ap, fmt);
+        pclog_ex(fmt, ap);
+        va_end(ap);
+    }
+}
+#else
+#    define mach_log(fmt, ...)
+#endif
+
+#define READ_PIXTRANS_BYTE_IO(cx, n, vgacore) \
+    if ((mach->accel.cmd_type == 2) || (mach->accel.cmd_type == 5)) { \
+        if (vgacore) { \
+            if ((svga->bpp == 15) || (svga->bpp == 16)) \
+                if (n == 0) \
+                    mach->accel.pix_trans[(n)] = vram_w[(dev->accel.dest + (cx) + (n)) & (svga->vram_mask >> 1)] & 0xff; \
+                else \
+                    mach->accel.pix_trans[(n)] = vram_w[(dev->accel.dest + (cx) + (n)) & (svga->vram_mask >> 1)] >> 8; \
+            else \
+                mach->accel.pix_trans[(n)] = svga->vram[(dev->accel.dest + (cx) + (n)) & svga->vram_mask]; \
+        } else \
+            mach->accel.pix_trans[(n)] = dev->vram[(dev->accel.dest + (cx) + (n)) & dev->vram_mask]; \
+    }
+
+#define READ_PIXTRANS_WORD(cx, n, vgacore)                                                                    \
+    if ((cmd == 0) || (cmd == 1) || (cmd == 5) || (mach->accel.cmd_type == -1)) { \
+        if (vgacore) { \
+            if ((svga->bpp == 15) || (svga->bpp == 16)) { \
+                temp = vram_w[((dev->accel.cy * dev->pitch) + (cx) + (n)) & (svga->vram_mask >> 1)];             \
+            } else { \
+                temp = svga->vram[((dev->accel.cy * dev->pitch) + (cx) + (n)) & svga->vram_mask];             \
+                temp |= (svga->vram[((dev->accel.cy * dev->pitch) + (cx) + (n + 1)) & svga->vram_mask] << 8); \
+            } \
+        } else { \
+            temp = dev->vram[((dev->accel.cy * dev->pitch) + (cx) + (n)) & dev->vram_mask];             \
+            temp |= (dev->vram[((dev->accel.cy * dev->pitch) + (cx) + (n + 1)) & dev->vram_mask] << 8); \
+        } \
+    } else if ((mach->accel.cmd_type == 2) || (mach->accel.cmd_type == 5)) { \
+        if ((svga->bpp == 8) || (svga->bpp == 24)) { \
+            if (vgacore) { \
+                temp = svga->vram[((dev->accel.dest) + (cx) + (n)) & svga->vram_mask];                           \
+                temp |= (svga->vram[((dev->accel.dest) + (cx) + (n + 1)) & svga->vram_mask] << 8);               \
+            } else { \
+                temp = dev->vram[((dev->accel.dest) + (cx) + (n)) & dev->vram_mask];                           \
+                temp |= (dev->vram[((dev->accel.dest) + (cx) + (n + 1)) & dev->vram_mask] << 8);               \
+            } \
+        } else if ((svga->bpp == 15) || (svga->bpp == 16)) { \
+            temp = vram_w[((dev->accel.dest) + (cx) + (n)) & (svga->vram_mask >> 1)];                           \
+        } \
+    } else if ((mach->accel.cmd_type == 3) || (mach->accel.cmd_type == 4)) { \
+        if ((svga->bpp == 8) || (svga->bpp == 24)) { \
+            if (vgacore) { \
+                temp = svga->vram[((mach->accel.ge_offset << 2) + ((dev->accel.cy) * (dev->pitch)) + (cx) + (n)) & svga->vram_mask]; \
+                temp |= (svga->vram[((mach->accel.ge_offset << 2) + ((dev->accel.cy) * (dev->pitch)) + (cx) + (n + 1)) & svga->vram_mask] << 8); \
+            } else { \
+                temp = dev->vram[((mach->accel.ge_offset << 2) + ((dev->accel.cy) * (dev->pitch)) + (cx) + (n)) & dev->vram_mask]; \
+                temp |= (dev->vram[((mach->accel.ge_offset << 2) + ((dev->accel.cy) * (dev->pitch)) + (cx) + (n + 1)) & dev->vram_mask] << 8); \
+            } \
+        } else if ((svga->bpp == 15) || (svga->bpp == 16)) { \
+            temp = vram_w[((mach->accel.ge_offset << 1) + ((dev->accel.cy) * (dev->pitch)) + (cx) + (n)) & (svga->vram_mask >> 1)]; \
+        } \
+    }
+
+#define READ(addr, dat, vgacore) \
+        if ((svga->bpp == 8) || (svga->bpp == 24)) \
+            dat = vgacore ? (svga->vram[(addr) & (svga->vram_mask)]) : (dev->vram[(addr) & (dev->vram_mask)]); \
+        else if ((svga->bpp == 15) || (svga->bpp == 16)) \
+            dat = vram_w[(addr) & (svga->vram_mask >> 1)];
+
+#define MIX(mixmode, dest_dat, src_dat)                                                       \
+    {                                                                                         \
+        switch ((mixmode) ? (dev->accel.frgd_mix & 0x1f) : (dev->accel.bkgd_mix & 0x1f)) {    \
+            case 0x00:                                                                        \
+                dest_dat = ~dest_dat;                                                         \
+                break;                                                                        \
+            case 0x01:                                                                        \
+                dest_dat = 0;                                                                 \
+                break;                                                                        \
+            case 0x02:                                                                        \
+                dest_dat = ~0;                                                                \
+                break;                                                                        \
+            case 0x03:                                                                        \
+                dest_dat = dest_dat;                                                          \
+                break;                                                                        \
+            case 0x04:                                                                        \
+                dest_dat = ~src_dat;                                                          \
+                break;                                                                        \
+            case 0x05:                                                                        \
+                dest_dat = src_dat ^ dest_dat;                                                \
+                break;                                                                        \
+            case 0x06:                                                                        \
+                dest_dat = ~(src_dat ^ dest_dat);                                             \
+                break;                                                                        \
+            case 0x07:                                                                        \
+                dest_dat = src_dat;                                                           \
+                break;                                                                        \
+            case 0x08:                                                                        \
+                dest_dat = ~(src_dat & dest_dat);                                             \
+                break;                                                                        \
+            case 0x09:                                                                        \
+                dest_dat = ~src_dat | dest_dat;                                               \
+                break;                                                                        \
+            case 0x0a:                                                                        \
+                dest_dat = src_dat | ~dest_dat;                                               \
+                break;                                                                        \
+            case 0x0b:                                                                        \
+                dest_dat = src_dat | dest_dat;                                                \
+                break;                                                                        \
+            case 0x0c:                                                                        \
+                dest_dat = src_dat & dest_dat;                                                \
+                break;                                                                        \
+            case 0x0d:                                                                        \
+                dest_dat = src_dat & ~dest_dat;                                               \
+                break;                                                                        \
+            case 0x0e:                                                                        \
+                dest_dat = ~src_dat & dest_dat;                                               \
+                break;                                                                        \
+            case 0x0f:                                                                        \
+                dest_dat = ~(src_dat | dest_dat);                                             \
+                break;                                                                        \
+            case 0x10:                                                                        \
+                dest_dat = MIN(src_dat, dest_dat);                                            \
+                break;                                                                        \
+            case 0x11:                                                                        \
+                dest_dat = dest_dat - src_dat;                                                \
+                break;                                                                        \
+            case 0x12:                                                                        \
+                dest_dat = src_dat - dest_dat;                                                \
+                break;                                                                        \
+            case 0x13:                                                                        \
+                dest_dat = src_dat + dest_dat;                                                \
+                break;                                                                        \
+            case 0x14:                                                                        \
+                dest_dat = MAX(src_dat, dest_dat);                                            \
+                break;                                                                        \
+            case 0x15:                                                                        \
+                dest_dat = (dest_dat - src_dat) / 2;                                          \
+                break;                                                                        \
+            case 0x16:                                                                        \
+                dest_dat = (src_dat - dest_dat) / 2;                                          \
+                break;                                                                        \
+            case 0x17:                                                                        \
+                dest_dat = (dest_dat + src_dat) / 2;                                          \
+                break;                                                                        \
+            case 0x18:                                                                        \
+                dest_dat = MAX(0, (dest_dat - src_dat));                                      \
+                break;                                                                        \
+            case 0x19:                                                                        \
+                dest_dat = MAX(0, (dest_dat - src_dat));                                      \
+                break;                                                                        \
+            case 0x1a:                                                                        \
+                dest_dat = MAX(0, (src_dat - dest_dat));                                      \
+                break;                                                                        \
+            case 0x1b:                                                                        \
+                dest_dat = MIN(0xff, (dest_dat + src_dat));                                   \
+                break;                                                                        \
+            case 0x1c:                                                                        \
+                dest_dat = MAX(0, (dest_dat - src_dat)) / 2;                                  \
+                break;                                                                        \
+            case 0x1d:                                                                        \
+                dest_dat = MAX(0, (dest_dat - src_dat)) / 2;                                  \
+                break;                                                                        \
+            case 0x1e:                                                                        \
+                dest_dat = MAX(0, (src_dat - dest_dat)) / 2;                                  \
+                break;                                                                        \
+            case 0x1f:                                                                        \
+                dest_dat = (0xff < (src_dat + dest_dat)) ? 0xff : ((src_dat + dest_dat) / 2); \
+                break;                                                                        \
+        }                                                                                     \
+    }
+
+
+#define WRITE(addr, dat, vgacore)                                         \
+    if ((svga->bpp == 8) || (svga->bpp == 24)) { \
+        if (vgacore) { \
+            svga->vram[((addr)) & (svga->vram_mask)]                = dat; \
+            svga->changedvram[(((addr)) & (svga->vram_mask)) >> 12] = changeframecount; \
+        } else { \
+            dev->vram[((addr)) & (dev->vram_mask)]                = dat; \
+            dev->changedvram[(((addr)) & (dev->vram_mask)) >> 12] = changeframecount; \
+        } \
+    } else if ((svga->bpp == 15) || (svga->bpp == 16)) { \
+        vram_w[((addr)) & (svga->vram_mask >> 1)]                = dat; \
+        svga->changedvram[(((addr)) & (svga->vram_mask >> 1)) >> 11] = changeframecount; \
+    }
+
+
+static int
+mach_pixel_write(mach_t *mach)
+{
+    if (mach->accel.dp_config & 1)
+        return 1;
+
+    return 0;
+}
+
+static int
+mach_pixel_read(mach_t *mach)
+{
+    if (mach->accel.dp_config & 1)
+        return 0;
+
+    return 1;
+}
+
+static void
+mach_accel_start(int cmd_type, int cpu_input, int count, uint32_t mix_dat, uint32_t cpu_dat, mach_t *mach, ibm8514_t *dev, int len)
+{
+    svga_t *svga = &mach->svga;
+    int compare_mode;
+    int poly_src = 0;
+    uint16_t rd_mask = dev->accel.rd_mask;
+    uint16_t wrt_mask = dev->accel.wrt_mask;
+    uint16_t dest_cmp_clr = dev->accel.color_cmp;
+    int frgd_sel, bkgd_sel, mono_src;
+    int compare = 0;
+    uint16_t src_dat = 0, dest_dat = 0;
+    uint16_t old_dest_dat;
+    uint16_t *vram_w = (uint16_t *) svga->vram;
+    uint16_t mix = 0;
+    int16_t clip_l = dev->accel.clip_left & 0x7ff;
+    int16_t clip_t = dev->accel.clip_top & 0x7ff;
+    int16_t clip_r = dev->accel.multifunc[4] & 0x7ff;
+    int16_t clip_b = dev->accel.multifunc[3] & 0x7ff;
+    uint32_t mono_dat0 = 0, mono_dat1 = 0;
+
+    if ((svga->bpp == 8) || (svga->bpp == 24)) {
+        rd_mask &= 0xff;
+        dest_cmp_clr &= 0xff;
+    }
+
+    compare_mode = (mach->accel.dest_cmp_fn >> 3) & 7;
+    frgd_sel = (mach->accel.dp_config >> 13) & 7;
+    bkgd_sel = (mach->accel.dp_config >> 7) & 3;
+    mono_src = (mach->accel.dp_config >> 5) & 3;
+
+    mach->accel.ge_offset = (mach->accel.ge_offset_lo | (mach->accel.ge_offset_hi << 16));
+
+    if ((mono_src == 2) || (bkgd_sel == 2) || (frgd_sel == 2) || mach_pixel_read(mach)) {
+        mach->force_busy = 1;
+        mach->force_busy2 = 1;
+        dev->force_busy = 1;
+        dev->force_busy2 = 1;
+    }
+
+    if (cpu_input) {
+        if ((svga->bpp == 15) || (svga->bpp == 16)) {
+            if ((mach->accel.dp_config & 0x200) && (count == 2)) {
+                count >>= 1;
+            }
+        }
+    }
+
+    if ((svga->bpp == 8) || (svga->bpp == 15) || (svga->bpp == 16) || (svga->bpp == 24)) {
+        if (svga->bpp == 24)
+            mach_log("24BPP: CMDType=%d, cwh(%d,%d,%d,%d), dpconfig=%04x\n", cmd_type, clip_l, clip_r, clip_t, clip_b, mach->accel.dp_config);
+        else
+            mach_log("BPP=%d, CMDType = %d, offs=%08x, DPCONFIG = %04x, cnt = %d, input = %d, mono_src = %d, frgdsel = %d, dstx = %d, dstxend = %d, pitch = %d, extcrt = %d, rw = %x, monpattern = %x.\n", svga->bpp, cmd_type, mach->accel.ge_offset, mach->accel.dp_config, count, cpu_input, mono_src, frgd_sel, dev->accel.cur_x, mach->accel.dest_x_end, dev->ext_pitch, dev->ext_crt_pitch, mach->accel.dp_config & 1, mach->accel.mono_pattern_enable);
+    }
+
+    switch (cmd_type) {
+        case 1: /*Extended Raw Linedraw from bres_count register (0x96ee)*/
+            if (!cpu_input) {
+                dev->accel.dx = dev->accel.cur_x;
+                if (dev->accel.cur_x >= 0x600)
+                    dev->accel.dx |= ~0x5ff;
+                dev->accel.dy = dev->accel.cur_y;
+                if (dev->accel.cur_y >= 0x600)
+                    dev->accel.dy |= ~0x5ff;
+
+                dev->accel.cx = dev->accel.destx_distp;
+                if (dev->accel.destx_distp >= 0x600)
+                    dev->accel.cx |= ~0x5ff;
+                dev->accel.cy = dev->accel.desty_axstp;
+                if (dev->accel.desty_axstp >= 0x600)
+                    dev->accel.cy |= ~0x5ff;
+
+                mach->accel.width = mach->accel.bres_count;
+                dev->accel.sx = 0;
+                mach->accel.poly_fill = 0;
+
+                mach->accel.color_pattern_idx = ((dev->accel.cx + (dev->accel.cy << 3)) & mach->accel.patt_len);
+
+                mach->accel.stepx = (mach->accel.linedraw_opt & 0x20) ? 1 : -1;
+                mach->accel.stepy = (mach->accel.linedraw_opt & 0x80) ? 1 : -1;
+
+                mach_log("Extended bresenham, CUR(%d,%d), DEST(%d,%d), width = %d, options = %04x, dpconfig = %04x, opt_ena = %03x.\n", dev->accel.dx, dev->accel.dy, dev->accel.cx, dev->accel.cy, mach->accel.width, mach->accel.linedraw_opt, mach->accel.dp_config, mach->accel.max_waitstates & 0x100);
+
+                if ((mono_src == 2) || (bkgd_sel == 2) || (frgd_sel == 2) || mach_pixel_read(mach)) {
+                    if (mach_pixel_write(mach)) {
+                        dev->data_available = 0;
+                        dev->data_available2 = 0;
+                        return;
+                    } else if (mach_pixel_read(mach)) {
+                        dev->data_available = 1;
+                        dev->data_available2 = 1;
+                        return;
+                    }
+                }
+            }
+
+            if (frgd_sel == 5) {
+                for (int x = 0; x <= mach->accel.patt_len; x++) {
+                    mach->accel.color_pattern[x] = mach->accel.patt_data[x & mach->accel.patt_len];
+                }
+
+                /*The destination coordinates should match the pattern index.*/
+                if (mach->accel.color_pattern_idx != mach->accel.patt_idx)
+                    mach->accel.color_pattern_idx = mach->accel.patt_idx;
+            }
+
+            if (mono_src == 1) {
+                count = mach->accel.width;
+                mix_dat = mach->accel.patt_data[0x10];
+                dev->accel.temp_cnt = 8;
+            }
+
+            if (mach->accel.linedraw_opt & 0x08) { /*Vector Line*/
+                while (count--) {
+                    switch (mono_src) {
+                        case 0:
+                            mix = 1;
+                            break;
+                        case 1:
+                            if (dev->accel.temp_cnt == 0) {
+                                dev->accel.temp_cnt = 8;
+                                mix_dat >>= 8;
+                            }
+                            mix = (mix_dat & 0x80);
+                            dev->accel.temp_cnt--;
+                            mix_dat <<= 1;
+                            mix_dat |= 1;
+                            break;
+                        case 2:
+                            if (mach->accel.dp_config & 0x1000) {
+                                mix = mix_dat >> 0x1f;
+                                mix_dat <<= 1;
+                            } else {
+                                if (mach->accel.dp_config & 0x200) {
+                                    mix = mix_dat & 1;
+                                    mix_dat >>= 1;
+                                } else {
+                                    mix = mix_dat & 0x80;
+                                    mix_dat <<= 1;
+                                    mix_dat |= 1;
+                                }
+                            }
+                            break;
+                        case 3:
+                            if ((svga->bpp == 15) || (svga->bpp == 16)) {
+                                READ((mach->accel.ge_offset << 1) + ((dev->accel.cy) * (dev->pitch)) + (dev->accel.cx), mix, dev->local);
+                            } else {
+                                READ((mach->accel.ge_offset << 2) + ((dev->accel.cy) * (dev->pitch)) + (dev->accel.cx), mix, dev->local);
+                            }
+                            mix = (mix & rd_mask) == rd_mask;
+                            break;
+                    }
+
+                    if ((((dev->accel.dx) >= clip_l) && ((dev->accel.dx) <= clip_r) && ((dev->accel.dy) >= clip_t) && ((dev->accel.dy) <= clip_b))) {
+                        if (mach->accel.linedraw_opt & 0x02) {
+                            if ((svga->bpp == 15) || (svga->bpp == 16)) {
+                                READ((mach->accel.ge_offset << 1) + ((dev->accel.cy) * (dev->pitch)) + (dev->accel.cx), poly_src, dev->local);
+                            } else {
+                                READ((mach->accel.ge_offset << 2) + ((dev->accel.cy) * (dev->pitch)) + (dev->accel.cx), poly_src, dev->local);
+                            }
+                            poly_src = ((poly_src & rd_mask) == rd_mask);
+                            if (poly_src)
+                                mach->accel.poly_fill = !mach->accel.poly_fill;
+                        }
+
+                        if (!mach->accel.poly_fill || !(mach->accel.linedraw_opt & 0x02)) {
+                            switch (mix ? frgd_sel : bkgd_sel) {
+                                case 0:
+                                    src_dat = dev->accel.bkgd_color;
+                                    break;
+                                case 1:
+                                    src_dat = dev->accel.frgd_color;
+                                    break;
+                                case 2:
+                                    src_dat = cpu_dat;
+                                    break;
+                                case 3:
+                                    if (mach_pixel_read(mach))
+                                        src_dat = cpu_dat;
+                                    else {
+                                        if ((svga->bpp == 15) || (svga->bpp == 16)) {
+                                            READ((mach->accel.ge_offset << 1) + ((dev->accel.cy) * (dev->pitch)) + (dev->accel.cx), mix, dev->local);
+                                        } else {
+                                            READ((mach->accel.ge_offset << 2) + ((dev->accel.cy) * (dev->pitch)) + (dev->accel.cx), mix, dev->local);
+                                        }
+                                        if (mono_src == 3) {
+                                            src_dat = (src_dat & rd_mask) == rd_mask;
+                                        }
+                                    }
+                                    break;
+                                case 5:
+                                    if (mix) {
+                                        src_dat = mach->accel.color_pattern[((dev->accel.dx) + ((dev->accel.dy) << 3)) & mach->accel.patt_len];
+                                    } else
+                                        src_dat = 0;
+                                    break;
+                            }
+
+                            if ((svga->bpp == 15) || (svga->bpp == 16)) {
+                                READ((mach->accel.ge_offset << 1) + ((dev->accel.dy) * (dev->pitch)) + (dev->accel.dx), dest_dat, dev->local);
+                            } else {
+                                READ((mach->accel.ge_offset << 2) + ((dev->accel.dy) * (dev->pitch)) + (dev->accel.dx), dest_dat, dev->local);
+                            }
+                        }
+
+                        switch (compare_mode) {
+                            case 1:
+                                compare = 1;
+                                break;
+                            case 2:
+                                compare = ((dest_dat) >= dest_cmp_clr) ? 0 : 1;
+                                break;
+                            case 3:
+                                compare = ((dest_dat) < dest_cmp_clr) ? 0 : 1;
+                                break;
+                            case 4:
+                                compare = ((dest_dat) != dest_cmp_clr) ? 0 : 1;
+                                break;
+                            case 5:
+                                compare = ((dest_dat) == dest_cmp_clr) ? 0 : 1;
+                                break;
+                            case 6:
+                                compare = ((dest_dat) <= dest_cmp_clr) ? 0 : 1;
+                                break;
+                            case 7:
+                                compare = ((dest_dat) > dest_cmp_clr) ? 0 : 1;
+                                break;
+                        }
+
+                        if (!compare) {
+                            if (mach_pixel_write(mach)) {
+                                old_dest_dat = dest_dat;
+                                MIX(mix, dest_dat, src_dat);
+                                dest_dat = (dest_dat & wrt_mask) | (old_dest_dat & ~wrt_mask);
+                            }
+                        }
+
+                        if (mach->accel.dp_config & 0x10) {
+                            if (mach->accel.linedraw_opt & 0x04) {
+                                if (dev->accel.sx < mach->accel.width) {
+                                    if ((svga->bpp == 15) || (svga->bpp == 16)) {
+                                        WRITE((mach->accel.ge_offset << 1) + ((dev->accel.dy) * (dev->pitch)) + (dev->accel.dx), dest_dat, dev->local);
+                                    } else {
+                                        WRITE((mach->accel.ge_offset << 2) + ((dev->accel.dy) * (dev->pitch)) + (dev->accel.dx), dest_dat, dev->local);
+                                    }
+                                }
+                            } else {
+                                if ((svga->bpp == 15) || (svga->bpp == 16)) {
+                                    WRITE((mach->accel.ge_offset << 1) + ((dev->accel.dy) * (dev->pitch)) + (dev->accel.dx), dest_dat, dev->local);
+                                } else {
+                                    WRITE((mach->accel.ge_offset << 2) + ((dev->accel.dy) * (dev->pitch)) + (dev->accel.dx), dest_dat, dev->local);
+                                }
+                            }
+                        }
+                    }
+
+                    if ((mono_src == 1) && !count)
+                        break;
+                    else if ((mono_src != 1) && (dev->accel.sx >= mach->accel.width))
+                        break;
+
+                    if (svga->bpp == 8)
+                        cpu_dat >>= 8;
+                    else
+                        cpu_dat >>= 16;
+
+                    switch (mach->accel.linedraw_opt & 0xe0) {
+                        case 0x00:
+                            dev->accel.cx++;
+                            dev->accel.dx++;
+                            break;
+                        case 0x20:
+                            dev->accel.cx++;
+                            dev->accel.dx++;
+                            dev->accel.cy--;
+                            dev->accel.dy--;
+                            break;
+                        case 0x40:
+                            dev->accel.cy--;
+                            dev->accel.dy--;
+                            break;
+                        case 0x60:
+                            dev->accel.cx--;
+                            dev->accel.dx--;
+                            dev->accel.cy--;
+                            dev->accel.dy--;
+                            break;
+                        case 0x80:
+                            dev->accel.cx--;
+                            dev->accel.dx--;
+                            break;
+                        case 0xa0:
+                            dev->accel.cx--;
+                            dev->accel.dx--;
+                            dev->accel.cy++;
+                            dev->accel.dy++;
+                            break;
+                        case 0xc0:
+                            dev->accel.cy++;
+                            dev->accel.dy++;
+                            break;
+                        case 0xe0:
+                            dev->accel.cx++;
+                            dev->accel.dx++;
+                            dev->accel.cy++;
+                            dev->accel.dy++;
+                            break;
+                    }
+
+                    dev->accel.sx++;
+                }
+            } else { /*Bresenham*/
+                while (count--) {
+                    switch (mono_src) {
+                        case 0:
+                            mix = 1;
+                            break;
+                        case 1:
+                            if (dev->accel.temp_cnt == 0) {
+                                dev->accel.temp_cnt = 8;
+                                mix_dat >>= 8;
+                            }
+                            mix = (mix_dat & 0x80);
+                            dev->accel.temp_cnt--;
+                            mix_dat <<= 1;
+                            mix_dat |= 1;
+                            break;
+                        case 2:
+                            if (mach->accel.dp_config & 0x1000) {
+                                mix = mix_dat >> 0x1f;
+                                mix_dat <<= 1;
+                            } else {
+                                if (mach->accel.dp_config & 0x200) {
+                                    mix = mix_dat & 1;
+                                    mix_dat >>= 1;
+                                } else {
+                                    mix = mix_dat & 0x80;
+                                    mix_dat <<= 1;
+                                    mix_dat |= 1;
+                                }
+                            }
+                            break;
+                        case 3:
+                            if ((svga->bpp == 15) || (svga->bpp == 16)) {
+                                READ((mach->accel.ge_offset << 1) + ((dev->accel.cy) * (dev->pitch)) + (dev->accel.cx), mix, dev->local);
+                            } else {
+                                READ((mach->accel.ge_offset << 2) + ((dev->accel.cy) * (dev->pitch)) + (dev->accel.cx), mix, dev->local);
+                            }
+                            mix = (mix & rd_mask) == rd_mask;
+                            break;
+                    }
+
+                    if ((((dev->accel.dx) >= clip_l) && ((dev->accel.dx) <= clip_r) && ((dev->accel.dy) >= clip_t) && ((dev->accel.dy) <= clip_b))) {
+                        if (mach->accel.linedraw_opt & 0x02) {
+                            if ((svga->bpp == 15) || (svga->bpp == 16)) {
+                                READ((mach->accel.ge_offset << 1) + ((dev->accel.cy) * (dev->pitch)) + (dev->accel.cx), poly_src, dev->local);
+                            } else {
+                                READ((mach->accel.ge_offset << 2) + ((dev->accel.cy) * (dev->pitch)) + (dev->accel.cx), poly_src, dev->local);
+                            }
+                            poly_src = ((poly_src & rd_mask) == rd_mask);
+                            if (poly_src)
+                                mach->accel.poly_fill = !mach->accel.poly_fill;
+                        }
+
+                        if (!mach->accel.poly_fill || !(mach->accel.linedraw_opt & 0x02)) {
+                            switch (mix ? frgd_sel : bkgd_sel) {
+                                case 0:
+                                    src_dat = dev->accel.bkgd_color;
+                                    break;
+                                case 1:
+                                    src_dat = dev->accel.frgd_color;
+                                    break;
+                                case 2:
+                                    src_dat = cpu_dat;
+                                    break;
+                                case 3:
+                                    if (mach_pixel_read(mach))
+                                        src_dat = cpu_dat;
+                                    else {
+                                        if ((svga->bpp == 15) || (svga->bpp == 16)) {
+                                            READ((mach->accel.ge_offset << 1) + ((dev->accel.cy) * (dev->pitch)) + (dev->accel.cx), mix, dev->local);
+                                        } else {
+                                            READ((mach->accel.ge_offset << 2) + ((dev->accel.cy) * (dev->pitch)) + (dev->accel.cx), mix, dev->local);
+                                        }
+                                        if (mono_src == 3) {
+                                            src_dat = (src_dat & rd_mask) == rd_mask;
+                                        }
+                                    }
+                                    break;
+                                case 5:
+                                    if (mix) {
+                                        src_dat = mach->accel.color_pattern[((dev->accel.dx) + ((dev->accel.dy) << 3)) & mach->accel.patt_len];
+                                    } else
+                                        src_dat = 0;
+                                    break;
+                            }
+
+                            if ((svga->bpp == 15) || (svga->bpp == 16)) {
+                                READ((mach->accel.ge_offset << 1) + ((dev->accel.dy) * (dev->pitch)) + (dev->accel.dx), dest_dat, dev->local);
+                            } else {
+                                READ((mach->accel.ge_offset << 2) + ((dev->accel.dy) * (dev->pitch)) + (dev->accel.dx), dest_dat, dev->local);
+                            }
+                        }
+
+                        switch (compare_mode) {
+                            case 1:
+                                compare = 1;
+                                break;
+                            case 2:
+                                compare = ((dest_dat) >= dest_cmp_clr) ? 0 : 1;
+                                break;
+                            case 3:
+                                compare = ((dest_dat) < dest_cmp_clr) ? 0 : 1;
+                                break;
+                            case 4:
+                                compare = ((dest_dat) != dest_cmp_clr) ? 0 : 1;
+                                break;
+                            case 5:
+                                compare = ((dest_dat) == dest_cmp_clr) ? 0 : 1;
+                                break;
+                            case 6:
+                                compare = ((dest_dat) <= dest_cmp_clr) ? 0 : 1;
+                                break;
+                            case 7:
+                                compare = ((dest_dat) > dest_cmp_clr) ? 0 : 1;
+                                break;
+                        }
+
+                        if (!compare) {
+                            if (mach_pixel_write(mach)) {
+                                old_dest_dat = dest_dat;
+                                MIX(mix, dest_dat, src_dat);
+                                dest_dat = (dest_dat & wrt_mask) | (old_dest_dat & ~wrt_mask);
+                            }
+                        }
+
+                        if (mach->accel.dp_config & 0x10) {
+                            if (mach->accel.linedraw_opt & 0x04) {
+                                if (dev->accel.sx < mach->accel.width) {
+                                    if ((svga->bpp == 15) || (svga->bpp == 16)) {
+                                        WRITE((mach->accel.ge_offset << 1) + ((dev->accel.dy) * (dev->pitch)) + (dev->accel.dx), dest_dat, dev->local);
+                                    } else {
+                                        WRITE((mach->accel.ge_offset << 2) + ((dev->accel.dy) * (dev->pitch)) + (dev->accel.dx), dest_dat, dev->local);
+                                    }
+                                }
+                            } else {
+                                if ((svga->bpp == 15) || (svga->bpp == 16)) {
+                                    WRITE((mach->accel.ge_offset << 1) + ((dev->accel.dy) * (dev->pitch)) + (dev->accel.dx), dest_dat, dev->local);
+                                } else {
+                                    WRITE((mach->accel.ge_offset << 2) + ((dev->accel.dy) * (dev->pitch)) + (dev->accel.dx), dest_dat, dev->local);
+                                }
+                            }
+                        }
+                    }
+
+                    if ((mono_src == 1) && !count)
+                        break;
+                    else if ((mono_src != 1) && (dev->accel.sx >= mach->accel.width))
+                        break;
+
+                    if (svga->bpp == 8)
+                        cpu_dat >>= 8;
+                    else
+                        cpu_dat >>= 16;
+
+                    if (mach->accel.linedraw_opt & 0x40) {
+                        dev->accel.dy += mach->accel.stepy;
+                        if ((frgd_sel == 3) || (bkgd_sel == 3))
+                            dev->accel.cy += mach->accel.stepy;
+
+                        if (dev->accel.err_term >= 0) {
+                            dev->accel.err_term += dev->accel.destx_distp;
+                            dev->accel.dx += mach->accel.stepx;
+                            if ((frgd_sel == 3) || (bkgd_sel == 3))
+                                dev->accel.cx += mach->accel.stepx;
+                        } else {
+                            dev->accel.err_term += dev->accel.desty_axstp;
+                        }
+                    } else {
+                        dev->accel.dx += mach->accel.stepx;
+                        if ((frgd_sel == 3) || (bkgd_sel == 3))
+                            dev->accel.cx += mach->accel.stepx;
+
+                        if (dev->accel.err_term >= 0) {
+                            dev->accel.err_term += dev->accel.destx_distp;
+                            dev->accel.dy += mach->accel.stepy;
+                            if ((frgd_sel == 3) || (bkgd_sel == 3))
+                                dev->accel.cy += mach->accel.stepy;
+                        } else {
+                            dev->accel.err_term += dev->accel.desty_axstp;
+                        }
+                    }
+
+                    dev->accel.sx++;
+                }
+            }
+            dev->accel.cur_x = dev->accel.dx;
+            dev->accel.cur_y = dev->accel.dy;
+            break;
+
+        case 2: /*Non-conforming BitBLT from dest_y_end register (0xaeee)*/
+            if (!cpu_input) {
+                mach->accel.stepx = 0;
+                mach->accel.stepy = 0;
+
+                dev->accel.dx = dev->accel.cur_x;
+                if (dev->accel.cur_x >= 0x600)
+                    dev->accel.dx |= ~0x5ff;
+
+                dev->accel.dy = dev->accel.cur_y;
+                if (dev->accel.cur_y >= 0x600)
+                    dev->accel.dy |= ~0x5ff;
+
+                /*Destination Width*/
+                if (mach->accel.dest_x_start != dev->accel.dx)
+                    mach->accel.dest_x_start = dev->accel.dx;
+
+                mach->accel.dx_start = mach->accel.dest_x_start;
+                if (mach->accel.dest_x_start >= 0x600)
+                    mach->accel.dx_start |= ~0x5ff;
+
+                mach->accel.dx_end = mach->accel.dest_x_end;
+                if (mach->accel.dest_x_end >= 0x600)
+                    mach->accel.dx_end |= ~0x5ff;
+
+                if (mach->accel.dx_end > mach->accel.dx_start) {
+                    mach->accel.width = (mach->accel.dx_end - mach->accel.dx_start);
+                    mach->accel.stepx = 1;
+                } else if (mach->accel.dx_end < mach->accel.dx_start) {
+                    mach->accel.width = (mach->accel.dx_start - mach->accel.dx_end);
+                    mach->accel.stepx = -1;
+                    if (dev->accel.dx > 0)
+                        dev->accel.dx--;
+                    mach_log("BitBLT: Dst Negative X, dxstart = %d, end = %d, width = %d, dx = %d, dpconfig = %04x.\n", mach->accel.dest_x_start, mach->accel.dest_x_end, mach->accel.width, dev->accel.dx, mach->accel.dp_config);
+                } else {
+                    mach->accel.stepx = 1;
+                    mach->accel.width = 0;
+                    mach_log("BitBLT: Dst Indeterminate X, dpconfig = %04x, destxend = %d, destxstart = %d.\n", mach->accel.dp_config, mach->accel.dest_x_end, mach->accel.dest_x_start);
+                }
+
+                dev->accel.sx = 0;
+                mach->accel.poly_fill = 0;
+                mach->accel.color_pattern_idx = ((dev->accel.dx + (dev->accel.dy << 3)) & mach->accel.patt_len);
+                if ((svga->bpp == 24) && (mono_src != 1)) {
+                    if (mach->accel.color_pattern_idx == mach->accel.patt_len)
+                        mach->accel.color_pattern_idx = mach->accel.patt_data_idx;
+                } else if ((svga->bpp == 24) && (frgd_sel == 5) && (mono_src == 1) && (mach->accel.patt_len_reg & 0x4000))
+                    mach->accel.color_pattern_idx = 0;
+
+                /*Height*/
+                mach->accel.dy_start = dev->accel.cur_y;
+                if (dev->accel.cur_y >= 0x600)
+                    mach->accel.dy_start |= ~0x5ff;
+                mach->accel.dy_end = mach->accel.dest_y_end;
+                if (mach->accel.dest_y_end >= 0x600)
+                    mach->accel.dy_end |= ~0x5ff;
+
+                if (mach->accel.dy_end > mach->accel.dy_start) {
+                    mach->accel.height = (mach->accel.dy_end - mach->accel.dy_start);
+                    mach->accel.stepy = 1;
+                } else if (mach->accel.dy_end < mach->accel.dy_start) {
+                    mach->accel.height = (mach->accel.dy_start - mach->accel.dy_end);
+                    mach->accel.stepy = -1;
+                } else {
+                    mach->accel.height = 0;
+                    mach->accel.stepy = 1;
+                }
+
+                dev->accel.sy = 0;
+                if ((svga->bpp == 15) || (svga->bpp == 16))
+                    dev->accel.dest = (mach->accel.ge_offset << 1) + (dev->accel.dy * (dev->pitch));
+                else
+                    dev->accel.dest = (mach->accel.ge_offset << 2) + (dev->accel.dy * (dev->pitch));
+
+                mach->accel.src_stepx = 0;
+
+                /*Source Width*/
+                dev->accel.cx = mach->accel.src_x;
+                if (mach->accel.src_x >= 0x600)
+                    dev->accel.cx |= ~0x5ff;
+
+                dev->accel.cy = mach->accel.src_y;
+                if (mach->accel.src_y >= 0x600)
+                    dev->accel.cy |= ~0x5ff;
+
+                mach->accel.sx_start = mach->accel.src_x_start;
+                if (mach->accel.src_x_start >= 0x600)
+                    mach->accel.sx_start |= ~0x5ff;
+
+                mach->accel.sx_end = mach->accel.src_x_end;
+                if (mach->accel.src_x_end >= 0x600)
+                    mach->accel.sx_end |= ~0x5ff;
+
+                if (mach->accel.sx_end > mach->accel.sx_start) {
+                    mach->accel.src_width = (mach->accel.sx_end - mach->accel.sx_start);
+                    mach->accel.src_stepx = 1;
+                    mach_log("BitBLT: Src Positive X: wh(%d,%d), srcwidth = %d, coordinates: %d,%d px, start: %d, end: %d px, stepx = %d, dpconfig = %04x, oddwidth = %d.\n", mach->accel.width, mach->accel.height, mach->accel.src_width, dev->accel.cx, dev->accel.cy, mach->accel.src_x_start, mach->accel.src_x_end, mach->accel.src_stepx, mach->accel.dp_config, mach->accel.src_width & 1);
+                } else if (mach->accel.sx_end < mach->accel.sx_start) {
+                    mach->accel.src_width = (mach->accel.sx_start - mach->accel.sx_end);
+                    mach->accel.src_stepx = -1;
+                    if (dev->accel.cx > 0)
+                        dev->accel.cx--;
+                    mach_log("BitBLT: Src Negative X: width = %d, coordinates: %d,%d px, end: %d px, stepx = %d, dpconfig = %04x, oddwidth = %d.\n", mach->accel.src_width, dev->accel.cx, dev->accel.cy, mach->accel.src_x_end, mach->accel.src_stepx, mach->accel.dp_config, mach->accel.src_width & 1);
+                } else {
+                    mach->accel.src_stepx = 1;
+                    mach->accel.src_width = 0;
+                    mach_log("BitBLT: Src Indeterminate X: width = %d, coordinates: %d,%d px, end: %d px, stepx = %d, dpconfig = %04x, oddwidth = %d.\n", mach->accel.src_width, dev->accel.cx, dev->accel.cy, mach->accel.src_x_end, mach->accel.src_stepx, mach->accel.dp_config, mach->accel.src_width & 1);
+                }
+                mach->accel.sx = 0;
+                if ((svga->bpp == 15) || (svga->bpp == 16))
+                    dev->accel.src = (mach->accel.ge_offset << 1) + (dev->accel.cy * (dev->pitch));
+                else
+                    dev->accel.src = (mach->accel.ge_offset << 2) + (dev->accel.cy * (dev->pitch));
+
+                if ((svga->bpp == 24) && (frgd_sel == 5)) {
+                    mach_log("BitBLT=%04x, WH(%d,%d), SRCWidth=%d, c(%d,%d), s(%d,%d).\n", mach->accel.dp_config, mach->accel.width, mach->accel.height, mach->accel.src_width, dev->accel.dx, dev->accel.dy, dev->accel.cx, dev->accel.cy);
+                } else
+                    mach_log("BitBLT=%04x, Pitch=%d, C(%d,%d), SRCWidth=%d, WH(%d,%d), geoffset=%08x.\n", mach->accel.dp_config, dev->ext_pitch, dev->accel.cx, dev->accel.cy, mach->accel.src_width, mach->accel.width, mach->accel.height, (mach->accel.ge_offset << 2));
+
+                if (mono_src == 1) {
+                    if ((mach->accel.mono_pattern_enable) && !(mach->accel.patt_len_reg & 0x4000)) {
+                        mono_dat0 = mach->accel.patt_data[0x10];
+                        mono_dat0 |= (mach->accel.patt_data[0x11] << 8);
+                        mono_dat0 |= (mach->accel.patt_data[0x12] << 16);
+                        mono_dat0 |= (mach->accel.patt_data[0x13] << 24);
+                        mono_dat1 = mach->accel.patt_data[0x14];
+                        mono_dat1 |= (mach->accel.patt_data[0x15] << 8);
+                        mono_dat1 |= (mach->accel.patt_data[0x16] << 16);
+                        mono_dat1 |= (mach->accel.patt_data[0x17] << 24);
+
+                        for (uint8_t y = 0; y < 8; y++) {
+                            for (uint8_t x = 0; x < 8; x++) {
+                                uint32_t temp = (y & 4) ? mono_dat1 : mono_dat0;
+                                mach->accel.mono_pattern[y][7 - x] = (temp >> (x + ((y & 3) << 3))) & 1;
+                            }
+                        }
+                    }
+                }
+
+                if ((mono_src == 2) || (bkgd_sel == 2) || (frgd_sel == 2) || mach_pixel_read(mach)) {
+                    if (mach_pixel_write(mach)) {
+                        dev->data_available = 0;
+                        dev->data_available2 = 0;
+                        return;
+                    } else if (mach_pixel_read(mach)) {
+                        dev->data_available = 1;
+                        dev->data_available2 = 1;
+                        return;
+                    }
+                }
+            }
+
+            if (mono_src == 1) {
+                if (!mach->accel.mono_pattern_enable && !(mach->accel.patt_len_reg & 0x4000)) {
+                    count = mach->accel.width;
+                    mix_dat = mach->accel.patt_data[0x10] ^ ((mach->accel.patt_idx & 1) ? 0xff : 0);
+                    dev->accel.temp_cnt = 8;
+                }
+            }
+
+            if (frgd_sel == 5) {
+                if ((svga->bpp == 15) || (svga->bpp == 16)) {
+                    for (int x = 0; x <= mach->accel.patt_len; x += 2) {
+                        mach->accel.color_pattern_word[x + (mach->accel.color_pattern_idx & 1)] = (mach->accel.patt_data[x & mach->accel.patt_len] & 0xff);
+                        mach->accel.color_pattern_word[x + (mach->accel.color_pattern_idx & 1)] |= (mach->accel.patt_data[(x + 1) & mach->accel.patt_len] << 8);
+                    }
+                } else {
+                    if ((svga->bpp == 24) && (mach->accel.patt_len < 3)) {
+                        for (int x = 0; x <= mach->accel.patt_len; x++) {
+                            mach->accel.color_pattern[x] = mach->accel.patt_data[x];
+                            mach_log("BITBLT: Color Pattern 24bpp[%d]=%02x, dataidx=%d, pattlen=%d.\n", x, mach->accel.color_pattern[x], mach->accel.patt_data_idx, mach->accel.patt_len);
+                        }
+                    } else {
+                        for (int x = 0; x <= mach->accel.patt_len; x++) {
+                            mach->accel.color_pattern[x] = mach->accel.patt_data[x & mach->accel.patt_len];
+                        }
+                    }
+                }
+
+                /*The destination coordinates should match the pattern index.*/
+                if (mach->accel.color_pattern_idx != mach->accel.patt_idx)
+                    mach->accel.color_pattern_idx = mach->accel.patt_idx;
+            }
+
+            if ((mach->accel.dy_end == mach->accel.dy_start)) {
+                mach_log("No DEST.\n");
+                return;
+            }
+
+            if ((mono_src == 3) || (bkgd_sel == 3) || (frgd_sel == 3)) {
+                if (mach->accel.sx_end == mach->accel.sx_start) {
+                    mach_log("No SRC.\n");
+                    return;
+                }
+            }
+
+            if (cpu_input) {
+                if (mach->accel.dp_config == 0x3251) {
+                    if (dev->accel.sy == mach->accel.height)
+                        return;
+                }
+            }
+
+            while (count--) {
+                switch (mono_src) {
+                    case 0:
+                        mix = 1;
+                        break;
+                    case 1:
+                        if (mach->accel.mono_pattern_enable) {
+                            mix = mach->accel.mono_pattern[dev->accel.dy & 7][dev->accel.dx & 7];
+                        } else {
+                            if ((svga->bpp == 24) && (frgd_sel == 5) && (mach->accel.patt_len_reg & 0x4000))
+                                mix = 1;
+                            else {
+                                if (!dev->accel.temp_cnt) {
+                                    dev->accel.temp_cnt = 8;
+                                    mix_dat >>= 8;
+                                }
+                                mix = (mix_dat & 0x80);
+                                dev->accel.temp_cnt--;
+                                mix_dat <<= 1;
+                                mix_dat |= 1;
+                            }
+                        }
+                        break;
+                    case 2:
+                        if (mach->accel.dp_config & 0x1000) {
+                            mix = mix_dat >> 0x1f;
+                            mix_dat <<= 1;
+                        } else {
+                            if (mach->accel.dp_config & 0x200) {
+                                mix = mix_dat & 1;
+                                mix_dat >>= 1;
+                            } else {
+                                mix = mix_dat & 0x80;
+                                mix_dat <<= 1;
+                                mix_dat |= 1;
+                            }
+                        }
+                        break;
+                    case 3:
+                        READ(dev->accel.src + ((dev->accel.cx)), mix, dev->local);
+                        mix = (mix & rd_mask) == rd_mask;
+                        break;
+                }
+
+                if (((dev->accel.dx) >= (clip_l) && (dev->accel.dx) <= (clip_r) &&
+                    (dev->accel.dy) >= (clip_t) && (dev->accel.dy) <= (clip_b))) {
+                    if (mach->accel.dp_config & 0x02) {
+                        READ(dev->accel.src + (dev->accel.cx), poly_src, dev->local);
+                        poly_src = ((poly_src & rd_mask) == rd_mask);
+                        if (poly_src)
+                            mach->accel.poly_fill = !mach->accel.poly_fill;
+                    }
+
+                    if (!mach->accel.poly_fill || !(mach->accel.dp_config & 0x02)) {
+                        switch (mix ? frgd_sel : bkgd_sel) {
+                            case 0:
+                                src_dat = dev->accel.bkgd_color;
+                                break;
+                            case 1:
+                                src_dat = dev->accel.frgd_color;
+                                break;
+                            case 2:
+                                src_dat = cpu_dat;
+                                break;
+                            case 3:
+                                if (mach_pixel_read(mach))
+                                    src_dat = cpu_dat;
+                                else {
+                                    READ(dev->accel.src + (dev->accel.cx), src_dat, dev->local);
+                                    if (mono_src == 3) {
+                                        src_dat = (src_dat & rd_mask) == rd_mask;
+                                    }
+                                }
+                                break;
+                            case 5:
+                                if (mix) {
+                                    if ((svga->bpp == 15) || (svga->bpp == 16)) {
+                                        src_dat = mach->accel.color_pattern_word[mach->accel.color_pattern_idx];
+                                    } else {
+                                        src_dat = mach->accel.color_pattern[mach->accel.color_pattern_idx];
+                                    }
+                                } else
+                                    src_dat = 0;
+                                break;
+                        }
+                    }
+
+                    if ((svga->bpp == 24) && (mono_src == 1) && (frgd_sel == 5) && (mach->accel.patt_len_reg & 0x4000)) {
+                        if (dev->accel.sy & 1) {
+                            READ(dev->accel.dest + dev->accel.dx - dev->ext_pitch, dest_dat, dev->local);
+                        } else {
+                            READ(dev->accel.dest + dev->accel.dx, dest_dat, dev->local);
+                        }
+                    } else {
+                        READ(dev->accel.dest + dev->accel.dx, dest_dat, dev->local);
+                    }
+
+                    switch (compare_mode) {
+                        case 1:
+                            compare = 1;
+                            break;
+                        case 2:
+                            compare = ((dest_dat) >= dest_cmp_clr) ? 0 : 1;
+                            break;
+                        case 3:
+                            compare = ((dest_dat) < dest_cmp_clr) ? 0 : 1;
+                            break;
+                        case 4:
+                            compare = ((dest_dat) != dest_cmp_clr) ? 0 : 1;
+                            break;
+                        case 5:
+                            compare = ((dest_dat) == dest_cmp_clr) ? 0 : 1;
+                            break;
+                        case 6:
+                            compare = ((dest_dat) <= dest_cmp_clr) ? 0 : 1;
+                            break;
+                        case 7:
+                            compare = ((dest_dat) > dest_cmp_clr) ? 0 : 1;
+                            break;
+                    }
+
+                    if (!compare) {
+                        if (mach_pixel_write(mach)) {
+                            old_dest_dat = dest_dat;
+                            MIX(mix, dest_dat, src_dat);
+                            dest_dat = (dest_dat & wrt_mask) | (old_dest_dat & ~wrt_mask);
+                        }
+                    }
+
+                    if (mach->accel.dp_config & 0x10) {
+                        if ((svga->bpp == 24) && (mono_src == 1) && (frgd_sel == 5) && (mach->accel.patt_len_reg & 0x4000)) {
+                            if (dev->accel.sy & 1) {
+                                WRITE(dev->accel.dest + dev->accel.dx - dev->ext_pitch, dest_dat, dev->local);
+                            } else {
+                                WRITE(dev->accel.dest + dev->accel.dx, dest_dat, dev->local);
+                            }
+                        } else {
+                            WRITE(dev->accel.dest + dev->accel.dx, dest_dat, dev->local);
+                        }
+                    }
+                }
+
+                if ((svga->bpp == 8) || (svga->bpp == 24))
+                    cpu_dat >>= 8;
+                else
+                    cpu_dat >>= 16;
+
+                if ((mono_src == 3) || (frgd_sel == 3) || (bkgd_sel == 3)) {
+                    dev->accel.cx += mach->accel.src_stepx;
+                    mach->accel.sx++;
+                    if (mach->accel.sx >= mach->accel.src_width) {
+                        mach->accel.sx = 0;
+                        if (mach->accel.src_stepx == -1)
+                            dev->accel.cx += mach->accel.src_width;
+                        else
+                            dev->accel.cx -= mach->accel.src_width;
+                        dev->accel.cy += (mach->accel.src_y_dir ? 1 : -1);
+                        if ((svga->bpp == 15) || (svga->bpp == 16))
+                            dev->accel.src = (mach->accel.ge_offset << 1) + (dev->accel.cy * (dev->pitch));
+                        else
+                            dev->accel.src = (mach->accel.ge_offset << 2) + (dev->accel.cy * (dev->pitch));
+                    }
+                }
+
+                dev->accel.dx += mach->accel.stepx;
+
+                if ((svga->bpp == 8) || ((svga->bpp == 24) && (mach->accel.patt_len >= 3) && (mono_src != 1)))
+                    mach->accel.color_pattern_idx = (mach->accel.color_pattern_idx + mach->accel.stepx) & mach->accel.patt_len;
+
+                if ((svga->bpp == 24) && (mach->accel.color_pattern_idx == mach->accel.patt_len) && (mach->accel.patt_len >= 3) && (mono_src != 1)) {
+                    mach->accel.color_pattern_idx = mach->accel.patt_data_idx;
+                } else if ((svga->bpp == 24) && (mach->accel.patt_len < 3)) {
+                    if (mach->accel.patt_len == 2) {
+                        mach->accel.color_pattern_idx++;
+                        if (mach->accel.color_pattern_idx == 3)
+                            mach->accel.color_pattern_idx = 0;
+                    } else {
+                        mach->accel.color_pattern_idx = (mach->accel.color_pattern_idx + mach->accel.stepx) & mach->accel.patt_len;
+                    }
+                } else if ((svga->bpp == 24) && (mach->accel.patt_len_reg & 0x4000) && (frgd_sel == 5)) {
+                    mach->accel.color_pattern_idx++;
+                    if (mach->accel.color_pattern_idx == 3)
+                        mach->accel.color_pattern_idx = 0;
+                }
+
+                if ((svga->bpp == 15) || (svga->bpp == 16)) {
+                    mach->accel.color_pattern_idx = (mach->accel.color_pattern_idx + mach->accel.stepx) & mach->accel.patt_len;
+                    mach->accel.color_pattern_idx = (mach->accel.color_pattern_idx + mach->accel.stepx) & mach->accel.patt_len;
+                }
+
+                dev->accel.sx++;
+                if (dev->accel.sx >= mach->accel.width) {
+                    mach->accel.poly_fill = 0;
+                    dev->accel.sx = 0;
+                    if (mach->accel.stepx == -1)
+                        dev->accel.dx += mach->accel.width;
+                    else
+                        dev->accel.dx -= mach->accel.width;
+
+                    dev->accel.dy += mach->accel.stepy;
+                    dev->accel.sy++;
+
+                    if ((svga->bpp == 15) || (svga->bpp == 16))
+                        dev->accel.dest = (mach->accel.ge_offset << 1) + (dev->accel.dy * (dev->pitch));
+                    else {
+                        dev->accel.dest = (mach->accel.ge_offset << 2) + (dev->accel.dy * (dev->pitch));
+                    }
+                    if ((mono_src == 1) && (svga->bpp == 24) && (frgd_sel == 5))
+                        mach->accel.color_pattern_idx = 0;
+                    else
+                        mach->accel.color_pattern_idx = ((dev->accel.dx + (dev->accel.dy << 3)) & mach->accel.patt_len);
+
+                    if ((svga->bpp == 24) && (mach->accel.color_pattern_idx == mach->accel.patt_len) && (mono_src != 1))
+                        mach->accel.color_pattern_idx = 0;
+                    if ((mono_src == 1) && !mach->accel.mono_pattern_enable && !(mach->accel.patt_len_reg & 0x4000)) {
+                        dev->accel.cur_x = dev->accel.dx;
+                        dev->accel.cur_y = dev->accel.dy;
+                        return;
+                    }
+                    if (dev->accel.sy >= mach->accel.height) {
+                        if ((mono_src == 2) || (mono_src == 3) || (frgd_sel == 2) || (frgd_sel == 3) || (bkgd_sel == 2) || (bkgd_sel == 3))
+                            return;
+                        if ((mono_src == 1) && (frgd_sel == 5) && (svga->bpp == 24) && (mach->accel.patt_len_reg & 0x4000))
+                            return;
+                        dev->accel.cur_x = dev->accel.dx;
+                        dev->accel.cur_y = dev->accel.dy;
+                        return;
+                    }
+                }
+            }
+            break;
+
+        case 3: /*Direct Linedraw (Polyline) from linedraw indexes (0xfeee)*/
+        case 4:
+            if (!cpu_input) {
+                dev->accel.cx = dev->accel.cur_x;
+                dev->accel.cy = dev->accel.cur_y;
+
+                if (dev->accel.cur_x >= 0x600) {
+                    mach_log("Linedraw XOver = %d.\n", dev->accel.cur_x);
+                    dev->accel.cx |= ~0x5ff;
+                }
+                if (dev->accel.cur_y >= 0x600) {
+                    mach_log("Linedraw YOver = %d.\n", dev->accel.cur_y);
+                    dev->accel.cy |= ~0x5ff;
+                }
+
+                dev->accel.dx = ABS(mach->accel.cx_end_line - dev->accel.cx) << 1;
+                dev->accel.dy = ABS(mach->accel.cy_end_line - dev->accel.cy) << 1;
+
+                mach->accel.stepx = (mach->accel.cx_end_line < dev->accel.cx) ? -1 : 1;
+                mach->accel.stepy = (mach->accel.cy_end_line < dev->accel.cy) ? -1 : 1;
+
+                dev->accel.sx = 0;
+
+                mach_log("Linedraw: c(%d,%d), d(%d,%d), cend(%d,%d).\n", dev->accel.cur_x, dev->accel.cur_y, dev->accel.dx, dev->accel.dy, mach->accel.cx_end_line, mach->accel.cy_end_line);
+
+                if ((mono_src == 2) || (bkgd_sel == 2) || (frgd_sel == 2) || mach_pixel_read(mach)) {
+                    if (mach_pixel_write(mach)) {
+                        dev->data_available = 0;
+                        dev->data_available2 = 0;
+                        return;
+                    } else if (mach_pixel_read(mach)) {
+                        dev->data_available = 1;
+                        dev->data_available2 = 1;
+                        return;
+                    }
+                }
+            }
+
+            if (frgd_sel == 5) {
+                for (int x = 0; x <= mach->accel.patt_len; x++) {
+                    mach->accel.color_pattern[x] = mach->accel.patt_data[x & mach->accel.patt_len];
+                }
+            }
+
+            if (mono_src == 1) {
+                mix_dat = mach->accel.patt_data[0x10];
+                dev->accel.temp_cnt = 8;
+            }
+
+            count = (dev->accel.dx > dev->accel.dy) ? (dev->accel.dx >> 1) : (dev->accel.dy >> 1);
+            mach->accel.width = count;
+
+            if (dev->accel.dx > dev->accel.dy) {
+                mach->accel.err = (dev->accel.dy - dev->accel.dx) >> 1;
+                if (mono_src == 1) {
+                    while (count--) {
+                        if (!dev->accel.temp_cnt) {
+                            dev->accel.temp_cnt = 8;
+                            mix_dat >>= 8;
+                        }
+                        mix = (mix_dat & 0x80);
+                        dev->accel.temp_cnt--;
+                        mix_dat <<= 1;
+                        mix_dat |= 1;
+
+                        if ((((dev->accel.cx) >= clip_l) && ((dev->accel.cx) <= clip_r) && ((dev->accel.cy) >= clip_t) && ((dev->accel.cy) <= clip_b))) {
+                            mach->accel.clip_overrun = 0;
+                            switch (mix ? frgd_sel : bkgd_sel) {
+                                case 0:
+                                    src_dat = dev->accel.bkgd_color;
+                                    break;
+                                case 1:
+                                    src_dat = dev->accel.frgd_color;
+                                    break;
+                                case 2:
+                                    src_dat = cpu_dat;
+                                    break;
+                                case 3:
+                                    if (mach_pixel_read(mach))
+                                        src_dat = cpu_dat;
+                                    else {
+                                        src_dat = 0;
+                                    }
+                                    break;
+                                case 5:
+                                    if (mix) {
+                                        src_dat = mach->accel.color_pattern[((dev->accel.cx) + ((dev->accel.cy) << 3)) & mach->accel.patt_len];
+                                    } else
+                                        src_dat = 0;
+                                    break;
+                            }
+
+                            if ((svga->bpp == 15) || (svga->bpp == 16)) {
+                                READ((mach->accel.ge_offset << 1) + ((dev->accel.cy) * (dev->pitch)) + (dev->accel.cx), dest_dat, dev->local);
+                            } else {
+                                READ((mach->accel.ge_offset << 2) + ((dev->accel.cy) * (dev->pitch)) + (dev->accel.cx), dest_dat, dev->local);
+                            }
+
+                            switch (compare_mode) {
+                                case 1:
+                                    compare = 1;
+                                    break;
+                                case 2:
+                                    compare = ((dest_dat) >= dest_cmp_clr) ? 0 : 1;
+                                    break;
+                                case 3:
+                                    compare = ((dest_dat) < dest_cmp_clr) ? 0 : 1;
+                                    break;
+                                case 4:
+                                    compare = ((dest_dat) != dest_cmp_clr) ? 0 : 1;
+                                    break;
+                                case 5:
+                                    compare = ((dest_dat) == dest_cmp_clr) ? 0 : 1;
+                                    break;
+                                case 6:
+                                    compare = ((dest_dat) <= dest_cmp_clr) ? 0 : 1;
+                                    break;
+                                case 7:
+                                    compare = ((dest_dat) > dest_cmp_clr) ? 0 : 1;
+                                    break;
+                            }
+
+                            if (!compare) {
+                                if (mach_pixel_write(mach)) {
+                                    old_dest_dat = dest_dat;
+                                    MIX(mix, dest_dat, src_dat);
+                                    dest_dat = (dest_dat & wrt_mask) | (old_dest_dat & ~wrt_mask);
+                                }
+                            }
+                            if ((mach->accel.dp_config & 0x10) && (cmd_type == 3)) {
+                                if ((svga->bpp == 15) || (svga->bpp == 16)) {
+                                    WRITE((mach->accel.ge_offset << 1) + ((dev->accel.cy) * (dev->pitch)) + (dev->accel.cx), dest_dat, dev->local);
+                                } else {
+                                    WRITE((mach->accel.ge_offset << 2) + ((dev->accel.cy) * (dev->pitch)) + (dev->accel.cx), dest_dat, dev->local);
+                                }
+                            }
+                        } else
+                            mach->accel.clip_overrun = ((mach->accel.clip_overrun + 1) & 0x0f);
+
+                        if (!count)
+                            break;
+
+                        if (svga->bpp == 8)
+                            cpu_dat >>= 8;
+                        else
+                            cpu_dat >>= 16;
+
+                        if (mach->accel.err >= 0) {
+                            dev->accel.cy += mach->accel.stepy;
+                            mach->accel.err -= dev->accel.dx;
+                        }
+                        dev->accel.cx += mach->accel.stepx;
+                        mach->accel.err += dev->accel.dy;
+                    }
+                } else {
+                    while (count--) {
+                        switch (mono_src) {
+                            case 0:
+                            case 3:
+                                mix = 1;
+                                break;
+                            case 2:
+                                if (mach->accel.dp_config & 0x1000) {
+                                    mix = mix_dat >> 0x1f;
+                                    mix_dat <<= 1;
+                                } else {
+                                    if (mach->accel.dp_config & 0x200) {
+                                        mix = mix_dat & 1;
+                                        mix_dat >>= 1;
+                                    } else {
+                                        mix = mix_dat & 0x80;
+                                        mix_dat <<= 1;
+                                        mix_dat |= 1;
+                                    }
+                                }
+                                break;
+                        }
+
+                        if ((((dev->accel.cx) >= clip_l) && ((dev->accel.cx) <= clip_r) && ((dev->accel.cy) >= clip_t) && ((dev->accel.cy) <= clip_b))) {
+                            mach->accel.clip_overrun = 0;
+                            switch (mix ? frgd_sel : bkgd_sel) {
+                                case 0:
+                                    src_dat = dev->accel.bkgd_color;
+                                    break;
+                                case 1:
+                                    src_dat = dev->accel.frgd_color;
+                                    break;
+                                case 2:
+                                    src_dat = cpu_dat;
+                                    break;
+                                case 3:
+                                    if (mach_pixel_read(mach))
+                                        src_dat = cpu_dat;
+                                    else {
+                                        src_dat = 0;
+                                    }
+                                    break;
+                                case 5:
+                                    if (mix) {
+                                        src_dat = mach->accel.color_pattern[((dev->accel.cx) + ((dev->accel.cy) << 3)) & mach->accel.patt_len];
+                                    } else
+                                        src_dat = 0;
+                                    break;
+                            }
+
+                            if ((svga->bpp == 15) || (svga->bpp == 16)) {
+                                READ((mach->accel.ge_offset << 1) + ((dev->accel.cy) * (dev->pitch)) + (dev->accel.cx), dest_dat, dev->local);
+                            } else {
+                                READ((mach->accel.ge_offset << 2) + ((dev->accel.cy) * (dev->pitch)) + (dev->accel.cx), dest_dat, dev->local);
+                            }
+
+                            switch (compare_mode) {
+                                case 1:
+                                    compare = 1;
+                                    break;
+                                case 2:
+                                    compare = ((dest_dat) >= dest_cmp_clr) ? 0 : 1;
+                                    break;
+                                case 3:
+                                    compare = ((dest_dat) < dest_cmp_clr) ? 0 : 1;
+                                    break;
+                                case 4:
+                                    compare = ((dest_dat) != dest_cmp_clr) ? 0 : 1;
+                                    break;
+                                case 5:
+                                    compare = ((dest_dat) == dest_cmp_clr) ? 0 : 1;
+                                    break;
+                                case 6:
+                                    compare = ((dest_dat) <= dest_cmp_clr) ? 0 : 1;
+                                    break;
+                                case 7:
+                                    compare = ((dest_dat) > dest_cmp_clr) ? 0 : 1;
+                                    break;
+                            }
+
+                            if (!compare) {
+                                if (mach_pixel_write(mach)) {
+                                    old_dest_dat = dest_dat;
+                                    MIX(mix, dest_dat, src_dat);
+                                    dest_dat = (dest_dat & wrt_mask) | (old_dest_dat & ~wrt_mask);
+                                }
+                            }
+
+                            if ((mach->accel.dp_config & 0x10) && (cmd_type == 3)) {
+                                if (mach->accel.linedraw_opt & 0x04) {
+                                    if (dev->accel.sx < mach->accel.width) {
+                                        if ((svga->bpp == 15) || (svga->bpp == 16)) {
+                                            WRITE((mach->accel.ge_offset << 1) + ((dev->accel.cy) * (dev->pitch)) + (dev->accel.cx), dest_dat, dev->local);
+                                        } else {
+                                            WRITE((mach->accel.ge_offset << 2) + ((dev->accel.cy) * (dev->pitch)) + (dev->accel.cx), dest_dat, dev->local);
+                                        }
+                                    }
+                                } else {
+                                    if ((svga->bpp == 15) || (svga->bpp == 16)) {
+                                        WRITE((mach->accel.ge_offset << 1) + ((dev->accel.cy) * (dev->pitch)) + (dev->accel.cx), dest_dat, dev->local);
+                                    } else {
+                                        WRITE((mach->accel.ge_offset << 2) + ((dev->accel.cy) * (dev->pitch)) + (dev->accel.cx), dest_dat, dev->local);
+                                    }
+                                }
+                            }
+                        } else
+                            mach->accel.clip_overrun = ((mach->accel.clip_overrun + 1) & 0x0f);
+
+                        if (dev->accel.sx >= mach->accel.width)
+                            break;
+
+                        if (svga->bpp == 8)
+                            cpu_dat >>= 8;
+                        else
+                            cpu_dat >>= 16;
+
+                        if (mach->accel.err >= 0) {
+                            dev->accel.cy += mach->accel.stepy;
+                            mach->accel.err -= dev->accel.dx;
+                        }
+                        dev->accel.cx += mach->accel.stepx;
+                        mach->accel.err += dev->accel.dy;
+
+                        dev->accel.sx++;
+                    }
+                }
+            } else {
+                mach->accel.err = (dev->accel.dx - dev->accel.dy) >> 1;
+                if (mono_src == 1) {
+                    while (count--) {
+                        if (dev->accel.temp_cnt == 0) {
+                            dev->accel.temp_cnt = 8;
+                            mix_dat >>= 8;
+                        }
+                        mix = (mix_dat & 0x80);
+                        dev->accel.temp_cnt--;
+                        mix_dat <<= 1;
+                        mix_dat |= 1;
+
+                        if ((((dev->accel.cx) >= clip_l) && ((dev->accel.cx) <= clip_r) && ((dev->accel.cy) >= clip_t) && ((dev->accel.cy) <= clip_b))) {
+                            mach->accel.clip_overrun = 0;
+                            switch (mix ? frgd_sel : bkgd_sel) {
+                                case 0:
+                                    src_dat = dev->accel.bkgd_color;
+                                    break;
+                                case 1:
+                                    src_dat = dev->accel.frgd_color;
+                                    break;
+                                case 2:
+                                    src_dat = cpu_dat;
+                                    break;
+                                case 3:
+                                    if (mach_pixel_read(mach))
+                                        src_dat = cpu_dat;
+                                    else {
+                                        src_dat = 0;
+                                    }
+                                    break;
+                                case 5:
+                                    if (mix) {
+                                        src_dat = mach->accel.color_pattern[((dev->accel.cx) + ((dev->accel.cy) << 3)) & mach->accel.patt_len];
+                                    } else
+                                        src_dat = 0;
+                                    break;
+                            }
+
+                            if ((svga->bpp == 15) || (svga->bpp == 16)) {
+                                READ((mach->accel.ge_offset << 1) + ((dev->accel.cy) * (dev->pitch)) + (dev->accel.cx), dest_dat, dev->local);
+                            } else {
+                                READ((mach->accel.ge_offset << 2) + ((dev->accel.cy) * (dev->pitch)) + (dev->accel.cx), dest_dat, dev->local);
+                            }
+                            switch (compare_mode) {
+                                case 1:
+                                    compare = 1;
+                                    break;
+                                case 2:
+                                    compare = ((dest_dat) >= dest_cmp_clr) ? 0 : 1;
+                                    break;
+                                case 3:
+                                    compare = ((dest_dat) < dest_cmp_clr) ? 0 : 1;
+                                    break;
+                                case 4:
+                                    compare = ((dest_dat) != dest_cmp_clr) ? 0 : 1;
+                                    break;
+                                case 5:
+                                    compare = ((dest_dat) == dest_cmp_clr) ? 0 : 1;
+                                    break;
+                                case 6:
+                                    compare = ((dest_dat) <= dest_cmp_clr) ? 0 : 1;
+                                    break;
+                                case 7:
+                                    compare = ((dest_dat) > dest_cmp_clr) ? 0 : 1;
+                                    break;
+                            }
+
+                            if (!compare) {
+                                if (mach_pixel_write(mach)) {
+                                    old_dest_dat = dest_dat;
+                                    MIX(mix, dest_dat, src_dat);
+                                    dest_dat = (dest_dat & wrt_mask) | (old_dest_dat & ~wrt_mask);
+                                }
+                            }
+
+                            if ((mach->accel.dp_config & 0x10) && (cmd_type == 3)) {
+                                if ((svga->bpp == 15) || (svga->bpp == 16)) {
+                                    WRITE((mach->accel.ge_offset << 1) + ((dev->accel.cy) * (dev->pitch)) + (dev->accel.cx), dest_dat, dev->local);
+                                } else {
+                                    WRITE((mach->accel.ge_offset << 2) + ((dev->accel.cy) * (dev->pitch)) + (dev->accel.cx), dest_dat, dev->local);
+                                }
+                            }
+                        } else
+                            mach->accel.clip_overrun = ((mach->accel.clip_overrun + 1) & 0x0f);
+
+                        if (!count)
+                            break;
+
+                        if (svga->bpp == 8)
+                            cpu_dat >>= 8;
+                        else
+                            cpu_dat >>= 16;
+
+                        if (mach->accel.err >= 0) {
+                            dev->accel.cx += mach->accel.stepx;
+                            mach->accel.err -= dev->accel.dy;
+                        }
+                        dev->accel.cy += mach->accel.stepy;
+                        mach->accel.err += dev->accel.dx;
+                    }
+                } else {
+                    while (count--) {
+                        switch (mono_src) {
+                            case 0:
+                            case 3:
+                                mix = 1;
+                                break;
+                            case 2:
+                                if (mach->accel.dp_config & 0x1000) {
+                                    mix = mix_dat >> 0x1f;
+                                    mix_dat <<= 1;
+                                } else {
+                                    if (mach->accel.dp_config & 0x200) {
+                                        mix = mix_dat & 1;
+                                        mix_dat >>= 1;
+                                    } else {
+                                        mix = mix_dat & 0x80;
+                                        mix_dat <<= 1;
+                                        mix_dat |= 1;
+                                    }
+                                }
+                                break;
+                        }
+
+                        if ((((dev->accel.cx) >= clip_l) && ((dev->accel.cx) <= clip_r) && ((dev->accel.cy) >= clip_t) && ((dev->accel.cy) <= clip_b))) {
+                            mach->accel.clip_overrun = 0;
+                            switch (mix ? frgd_sel : bkgd_sel) {
+                                case 0:
+                                    src_dat = dev->accel.bkgd_color;
+                                    break;
+                                case 1:
+                                    src_dat = dev->accel.frgd_color;
+                                    break;
+                                case 2:
+                                    src_dat = cpu_dat;
+                                    break;
+                                case 3:
+                                    if (mach_pixel_read(mach))
+                                        src_dat = cpu_dat;
+                                    else {
+                                        src_dat = 0;
+                                    }
+                                    break;
+                                case 5:
+                                    if (mix) {
+                                        src_dat = mach->accel.color_pattern[((dev->accel.cx) + ((dev->accel.cy) << 3)) & mach->accel.patt_len];
+                                    } else
+                                        src_dat = 0;
+                                    break;
+                            }
+
+                            if ((svga->bpp == 15) || (svga->bpp == 16)) {
+                                READ((mach->accel.ge_offset << 1) + ((dev->accel.cy) * (dev->pitch)) + (dev->accel.cx), dest_dat, dev->local);
+                            } else {
+                                READ((mach->accel.ge_offset << 2) + ((dev->accel.cy) * (dev->pitch)) + (dev->accel.cx), dest_dat, dev->local);
+                            }
+
+                            switch (compare_mode) {
+                                case 1:
+                                    compare = 1;
+                                    break;
+                                case 2:
+                                    compare = ((dest_dat) >= dest_cmp_clr) ? 0 : 1;
+                                    break;
+                                case 3:
+                                    compare = ((dest_dat) < dest_cmp_clr) ? 0 : 1;
+                                    break;
+                                case 4:
+                                    compare = ((dest_dat) != dest_cmp_clr) ? 0 : 1;
+                                    break;
+                                case 5:
+                                    compare = ((dest_dat) == dest_cmp_clr) ? 0 : 1;
+                                    break;
+                                case 6:
+                                    compare = ((dest_dat) <= dest_cmp_clr) ? 0 : 1;
+                                    break;
+                                case 7:
+                                    compare = ((dest_dat) > dest_cmp_clr) ? 0 : 1;
+                                    break;
+                            }
+
+                            if (!compare) {
+                                if (mach_pixel_write(mach)) {
+                                    old_dest_dat = dest_dat;
+                                    MIX(mix, dest_dat, src_dat);
+                                    dest_dat = (dest_dat & wrt_mask) | (old_dest_dat & ~wrt_mask);
+                                }
+                            }
+
+                            if ((mach->accel.dp_config & 0x10) && (cmd_type == 3)) {
+                                if (mach->accel.linedraw_opt & 0x04) {
+                                    if (dev->accel.sx < mach->accel.width) {
+                                        if ((svga->bpp == 15) || (svga->bpp == 16)) {
+                                            WRITE((mach->accel.ge_offset << 1) + ((dev->accel.cy) * (dev->pitch)) + (dev->accel.cx), dest_dat, dev->local);
+                                        } else {
+                                            WRITE((mach->accel.ge_offset << 2) + ((dev->accel.cy) * (dev->pitch)) + (dev->accel.cx), dest_dat, dev->local);
+                                        }
+                                    }
+                                } else {
+                                    if ((svga->bpp == 15) || (svga->bpp == 16)) {
+                                        WRITE((mach->accel.ge_offset << 1) + ((dev->accel.cy) * (dev->pitch)) + (dev->accel.cx), dest_dat, dev->local);
+                                    } else {
+                                        WRITE((mach->accel.ge_offset << 2) + ((dev->accel.cy) * (dev->pitch)) + (dev->accel.cx), dest_dat, dev->local);
+                                    }
+                                }
+                            }
+                        } else
+                            mach->accel.clip_overrun = ((mach->accel.clip_overrun + 1) & 0x0f);
+
+                        if (dev->accel.sx >= mach->accel.width)
+                            break;
+
+                        if (svga->bpp == 8)
+                            cpu_dat >>= 8;
+                        else
+                            cpu_dat >>= 16;
+
+                        if (mach->accel.err >= 0) {
+                            dev->accel.cx += mach->accel.stepx;
+                            mach->accel.err -= dev->accel.dy;
+                        }
+                        dev->accel.cy += mach->accel.stepy;
+                        mach->accel.err += dev->accel.dx;
+
+                        dev->accel.sx++;
+                    }
+                }
+            }
+            mach->accel.line_array[(cmd_type == 4) ? 4 : 0] = dev->accel.cx;
+            mach->accel.line_array[(cmd_type == 4) ? 5 : 1] = dev->accel.cy;
+            dev->accel.cur_x = mach->accel.line_array[(cmd_type == 4) ? 4 : 0];
+            dev->accel.cur_y = mach->accel.line_array[(cmd_type == 4) ? 5 : 1];
+            break;
+
+        case 5: /*Horizontal Raster Draw from scan_to_x register (0xcaee)*/
+            if (!cpu_input) {
+                mach->accel.stepx = 0;
+                mach->accel.stepy = 0;
+
+                dev->accel.dx = dev->accel.cur_x;
+                if (dev->accel.cur_x >= 0x600)
+                    dev->accel.dx |= ~0x5ff;
+                dev->accel.dy = dev->accel.cur_y;
+                if (dev->accel.cur_y >= 0x600)
+                    dev->accel.dy |= ~0x5ff;
+
+                /*Destination Width*/
+                mach->accel.dx_start = dev->accel.cur_x;
+                if (dev->accel.cur_x >= 0x600)
+                    mach->accel.dx_start |= ~0x5ff;
+                mach->accel.dx_end = mach->accel.scan_to_x;
+                if (mach->accel.scan_to_x >= 0x600)
+                    mach->accel.dx_end |= ~0x5ff;
+
+                if (mach->accel.dx_end > mach->accel.dx_start) {
+                    mach->accel.width = (mach->accel.dx_end - mach->accel.dx_start);
+                    mach->accel.stepx = 1;
+                } else if (mach->accel.dx_end < mach->accel.dx_start) {
+                    mach->accel.width = (mach->accel.dx_start - mach->accel.dx_end);
+                    mach->accel.stepx = -1;
+                    if (dev->accel.dx > 0)
+                        dev->accel.dx--;
+                } else {
+                    mach->accel.stepx = 1;
+                    mach->accel.width = 0;
+                }
+
+                dev->accel.sx = 0;
+                if ((svga->bpp == 24) && (mach->accel.patt_len < 0x17))
+                    mach->accel.color_pattern_idx = 0;
+
+                /*Step Y*/
+                mach->accel.dy_start = dev->accel.cur_y;
+                if (dev->accel.cur_y >= 0x600)
+                    mach->accel.dy_start |= ~0x5ff;
+                mach->accel.dy_end = mach->accel.dest_y_end;
+                if (mach->accel.dest_y_end >= 0x600)
+                    mach->accel.dy_end |= ~0x5ff;
+
+                if (mach->accel.dy_end > mach->accel.dy_start) {
+                    mach->accel.stepy = 1;
+                } else if (mach->accel.dy_end < mach->accel.dy_start) {
+                    mach->accel.stepy = -1;
+                } else {
+                    mach->accel.stepy = 0;
+                }
+
+                if ((svga->bpp == 15) || (svga->bpp == 16))
+                    dev->accel.dest = (mach->accel.ge_offset << 1) + (dev->accel.dy * (dev->pitch));
+                else
+                    dev->accel.dest = (mach->accel.ge_offset << 2) + (dev->accel.dy * (dev->pitch));
+
+                mach->accel.src_stepx = 0;
+
+                /*Source Width*/
+                dev->accel.cx = mach->accel.src_x;
+                if (mach->accel.src_x >= 0x600)
+                    dev->accel.cx |= ~0x5ff;
+                dev->accel.cy = mach->accel.src_y;
+                if (mach->accel.src_y >= 0x600)
+                    dev->accel.cy |= ~0x5ff;
+
+                mach->accel.sx_start = mach->accel.src_x_start;
+                if (mach->accel.src_x_start >= 0x600)
+                    mach->accel.sx_start |= ~0x5ff;
+
+                mach->accel.sx_end = mach->accel.src_x_end;
+                if (mach->accel.src_x_end >= 0x600)
+                    mach->accel.sx_end |= ~0x5ff;
+
+                if (mach->accel.sx_end > mach->accel.sx_start) {
+                    mach->accel.src_width = (mach->accel.sx_end - mach->accel.sx_start);
+                    mach->accel.src_stepx = 1;
+                } else if (mach->accel.sx_end < mach->accel.sx_start) {
+                    mach->accel.src_width = (mach->accel.sx_start - mach->accel.sx_end);
+                    mach->accel.src_stepx = -1;
+                    if (dev->accel.cx > 0)
+                        dev->accel.cx--;
+                } else {
+                    mach->accel.src_stepx = 1;
+                    mach->accel.src_width = 0;
+                }
+
+                mach->accel.sx = 0;
+                if ((svga->bpp == 15) || (svga->bpp == 16))
+                    dev->accel.src = (mach->accel.ge_offset << 1) + (dev->accel.cy * (dev->pitch));
+                else
+                    dev->accel.src = (mach->accel.ge_offset << 2) + (dev->accel.cy * (dev->pitch));
+
+                if ((svga->bpp == 24) && (frgd_sel == 5)) {
+                    if (mach->accel.patt_len == 0x17)
+                        mach->accel.color_pattern_idx = 0;
+                    dev->accel.x1 = dev->accel.dx + mach->accel.width;
+                    if (dev->accel.x1 == dev->pitch) {
+                        dev->accel.x2 = mach->accel.width & 1;
+                    } else if ((dev->accel.x1 == mach->accel.width) && (dev->accel.dy & 1) && !dev->accel.y1 && dev->accel.x2) {
+                        if (mach->accel.patt_len == 0x17)
+                            mach->accel.color_pattern_idx = 3;
+                        dev->accel.x3 = 1;
+                    } else
+                        dev->accel.x3 = 0;
+                } else
+                    mach_log("ScanToX=%04x, Pitch=%d, C(%d,%d), SRCWidth=%d, WH(%d,%d), geoffset=%08x.\n", mach->accel.dp_config, dev->ext_pitch, dev->accel.cx, dev->accel.cy, mach->accel.src_width, mach->accel.width, mach->accel.height, (mach->accel.ge_offset << 1));
+
+                dev->accel.y1 = 0;
+
+                if ((mono_src == 2) || (bkgd_sel == 2) || (frgd_sel == 2) || mach_pixel_read(mach)) {
+                    if (mach_pixel_write(mach)) {
+                        dev->data_available = 0;
+                        dev->data_available2 = 0;
+                        return;
+                    } else if (mach_pixel_read(mach)) {
+                        dev->data_available = 1;
+                        dev->data_available2 = 1;
+                        return;
+                    }
+                }
+            }
+
+            if (mono_src == 1) {
+                count = mach->accel.width;
+                mix_dat = mach->accel.patt_data[0x10];
+                dev->accel.temp_cnt = 8;
+            }
+
+            if (frgd_sel == 5) {
+                if (svga->bpp != 24) {
+                    for (int x = 0; x <= mach->accel.patt_len; x++) {
+                        mach->accel.color_pattern[x] = mach->accel.patt_data[x & mach->accel.patt_len];
+                    }
+                } else {
+                    if (mach->accel.patt_len == 0x17) {
+                        for (int x = 0; x <= mach->accel.patt_len; x++) {
+                            mach->accel.color_pattern_full[x] = mach->accel.patt_data[x];
+                            mach_log("ScanToX: Color Pattern 24bpp[%d]=%02x, dataidx=%d, pattlen=%d.\n", x, mach->accel.color_pattern_full[x], mach->accel.patt_data_idx, mach->accel.patt_len);
+                        }
+                    } else {
+                        for (int x = 0; x <= mach->accel.patt_len; x++) {
+                            mach->accel.color_pattern[x] = mach->accel.patt_data[x];
+                            mach_log("ScanToX: Color Pattern 24bpp[%d]=%02x, dataidx=%d, pattlen=%d.\n", x, mach->accel.color_pattern[x], mach->accel.patt_data_idx, mach->accel.patt_len);
+                        }
+                    }
+                }
+            }
+
+            while (count--) {
+                switch (mono_src) {
+                    case 0:
+                        mix = 1;
+                        break;
+                    case 1:
+                        if (dev->accel.temp_cnt == 0) {
+                            dev->accel.temp_cnt = 8;
+                            mix_dat >>= 8;
+                        }
+                        mix = (mix_dat & 0x80);
+                        dev->accel.temp_cnt--;
+                        mix_dat <<= 1;
+                        mix_dat |= 1;
+                        break;
+                    case 2:
+                        if (mach->accel.dp_config & 0x1000) {
+                            mix = mix_dat >> 0x1f;
+                            mix_dat <<= 1;
+                        } else {
+                            if (mach->accel.dp_config & 0x200) {
+                                mix = mix_dat & 1;
+                                mix_dat >>= 1;
+                            } else {
+                                mix = mix_dat & 0x80;
+                                mix_dat <<= 1;
+                                mix_dat |= 1;
+                            }
+                        }
+                        break;
+                    case 3:
+                        READ(dev->accel.src + (dev->accel.cx), mix, dev->local);
+                        mix = (mix & rd_mask) == rd_mask;
+                        break;
+                }
+
+                if ((dev->accel.dx) >= (clip_l) && (dev->accel.dx) <= (clip_r) &&
+                    (dev->accel.dy) >= (clip_t) && (dev->accel.dy) <= (clip_b)) {
+                    switch (mix ? frgd_sel : bkgd_sel) {
+                        case 0:
+                            src_dat = dev->accel.bkgd_color;
+                            break;
+                        case 1:
+                            src_dat = dev->accel.frgd_color;
+                            break;
+                        case 2:
+                            src_dat = cpu_dat;
+                            break;
+                        case 3:
+                            if (mach_pixel_read(mach))
+                                src_dat = cpu_dat;
+                            else {
+                                READ(dev->accel.src + (dev->accel.cx), src_dat, dev->local);
+                                if (mono_src == 3) {
+                                    src_dat = (src_dat & rd_mask) == rd_mask;
+                                }
+                            }
+                            break;
+                        case 5:
+                            if (mix) {
+                                if (svga->bpp == 24) {
+                                    if (mach->accel.patt_len == 0x17)
+                                        src_dat = mach->accel.color_pattern_full[mach->accel.color_pattern_idx];
+                                    else
+                                        src_dat = mach->accel.color_pattern[mach->accel.color_pattern_idx];
+                                } else
+                                    src_dat = mach->accel.color_pattern[(dev->accel.dx + (dev->accel.dy << 3)) & mach->accel.patt_len];
+                            } else
+                                src_dat = 0;
+                            break;
+                    }
+
+                    READ(dev->accel.dest + (dev->accel.dx), dest_dat, dev->local);
+
+                    switch (compare_mode) {
+                        case 1:
+                            compare = 1;
+                            break;
+                        case 2:
+                            compare = ((dest_dat) >= dest_cmp_clr) ? 0 : 1;
+                            break;
+                        case 3:
+                            compare = ((dest_dat) < dest_cmp_clr) ? 0 : 1;
+                            break;
+                        case 4:
+                            compare = ((dest_dat) != dest_cmp_clr) ? 0 : 1;
+                            break;
+                        case 5:
+                            compare = ((dest_dat) == dest_cmp_clr) ? 0 : 1;
+                            break;
+                        case 6:
+                            compare = ((dest_dat) <= dest_cmp_clr) ? 0 : 1;
+                            break;
+                        case 7:
+                            compare = ((dest_dat) > dest_cmp_clr) ? 0 : 1;
+                            break;
+                    }
+
+                    if (!compare) {
+                        if (mach_pixel_write(mach)) {
+                            old_dest_dat = dest_dat;
+                            MIX(mix, dest_dat, src_dat);
+                            dest_dat = (dest_dat & wrt_mask) | (old_dest_dat & ~wrt_mask);
+                        }
+                    }
+
+                    if (mach->accel.dp_config & 0x10) {
+                        WRITE(dev->accel.dest + (dev->accel.dx), dest_dat, dev->local);
+                    }
+                }
+
+                if ((svga->bpp == 8) || (svga->bpp == 24))
+                    cpu_dat >>= 8;
+                else
+                    cpu_dat >>= 16;
+
+                dev->accel.cx += mach->accel.src_stepx;
+                mach->accel.sx++;
+                if (mach->accel.sx >= mach->accel.src_width) {
+                    mach->accel.sx = 0;
+                    if (mach->accel.src_stepx == -1) {
+                        dev->accel.cx += mach->accel.src_width;
+                    } else
+                        dev->accel.cx -= mach->accel.src_width;
+                    dev->accel.cy += (mach->accel.src_y_dir ? 1 : -1);
+                    if ((svga->bpp == 15) || (svga->bpp == 16))
+                        dev->accel.src = (mach->accel.ge_offset << 1) + (dev->accel.cy * (dev->pitch));
+                    else
+                        dev->accel.src = (mach->accel.ge_offset << 2) + (dev->accel.cy * (dev->pitch));
+                }
+
+                dev->accel.dx += mach->accel.stepx;
+                if ((svga->bpp == 24) && (mach->accel.patt_len == 0x17)) {
+                    mach->accel.color_pattern_idx++;
+                    if (dev->accel.x3) {
+                        if (mach->accel.color_pattern_idx == 9)
+                            mach->accel.color_pattern_idx = 3;
+                    } else {
+                        if (mach->accel.color_pattern_idx == 6)
+                            mach->accel.color_pattern_idx = 0;
+                    }
+                } else if ((svga->bpp == 24) && (mach->accel.patt_len < 3)) {
+                    mach->accel.color_pattern_idx++;
+                    if (mach->accel.color_pattern_idx == 3)
+                        mach->accel.color_pattern_idx = 0;
+                } else
+                    mach->accel.color_pattern_idx = (mach->accel.color_pattern_idx + mach->accel.stepx) & mach->accel.patt_len;
+
+                dev->accel.sx++;
+                if (dev->accel.sx >= mach->accel.width) {
+                    dev->accel.sx = 0;
+                    dev->accel.dy += mach->accel.stepy;
+                    if ((svga->bpp == 15) || (svga->bpp == 16))
+                        dev->accel.dest = (mach->accel.ge_offset << 1) + (dev->accel.dy * (dev->pitch));
+                    else
+                        dev->accel.dest = (mach->accel.ge_offset << 2) + (dev->accel.dy * (dev->pitch));
+                    if (mach->accel.line_idx == 2) {
+                        mach->accel.line_array[0] = dev->accel.dx;
+                        mach->accel.line_array[4] = dev->accel.dx;
+                    }
+                    return;
+                }
+            }
+            break;
+    }
+}
+
+static void
+mach_accel_out_pixtrans(mach_t *mach, ibm8514_t *dev, uint16_t port, uint16_t val, uint16_t len)
+{
+    int frgd_sel, bkgd_sel, mono_src;
+
+    frgd_sel = (mach->accel.dp_config >> 13) & 7;
+    bkgd_sel = (mach->accel.dp_config >> 7) & 3;
+    mono_src = (mach->accel.dp_config >> 5) & 3;
+
+    if ((mach->accel.dp_config & 4) && (mach->accel.cmd_type != 5)) {
+        val = (val >> 8) | (val << 8);
+    }
+
+    switch (mach->accel.dp_config & 0x200) {
+        case 0x000: /*8-bit size*/
+            if ((mono_src == 2)) {
+                if ((frgd_sel != 2) && (bkgd_sel != 2)) {
+                    if ((mach->accel.dp_config & 0x1000) && dev->local)
+                        val = (val >> 8) | (val << 8);
+                    mach_accel_start(mach->accel.cmd_type, 1, 8, val | (val << 16), 0, mach, dev, len);
+                } else
+                    mach_accel_start(mach->accel.cmd_type, 1, 1, -1, val | (val << 16), mach, dev, len);
+            } else
+                mach_accel_start(mach->accel.cmd_type, 1, 1, -1, val | (val << 16), mach, dev, len);
+            break;
+        case 0x200: /*16-bit size*/
+            if ((mono_src == 2)) {
+                if ((frgd_sel != 2) && (bkgd_sel != 2)) {
+                    if (mach->accel.dp_config & 0x1000)
+                        val = (val >> 8) | (val << 8);
+                    mach_accel_start(mach->accel.cmd_type, 1, 16, val | (val << 16), 0, mach, dev, len);
+                } else {
+                    mach_accel_start(mach->accel.cmd_type, 1, 2, -1, val | (val << 16), mach, dev, len);
+                }
+            } else {
+                mach_accel_start(mach->accel.cmd_type, 1, 2, -1, val | (val << 16), mach, dev, len);
+            }
+            break;
+    }
+}
+
+static void
+mach_out(uint16_t addr, uint8_t val, void *p)
+{
+    mach_t     *mach     = (mach_t *) p;
+    svga_t     *svga     = &mach->svga;
+    ibm8514_t  *dev      = &svga->dev8514;
+    uint8_t     old;
+    uint8_t     rs2;
+
+    if (((addr & 0xFFF0) == 0x3D0 || (addr & 0xFFF0) == 0x3B0) && !(svga->miscout & 1))
+        addr ^= 0x60;
+
+    switch (addr) {
+        case 0x1ce:
+            mach->index = val;
+            break;
+        case 0x1cf:
+            old = mach->regs[mach->index];
+            mach->regs[mach->index] = val;
+            mach_log("ATI VGA write reg=0x%02X, val=0x%02X\n", mach->index, val);
+            switch (mach->index) {
+                case 0xa3:
+                    if ((old ^ val) & 0x10)
+                        svga_recalctimings(svga);
+                    break;
+                case 0xa7:
+                    if ((old ^ val) & 0x80)
+                        svga_recalctimings(svga);
+                    break;
+                case 0xad:
+                    if (dev->local) {
+                        if ((old ^ val) & 0x0c)
+                            svga_recalctimings(svga);
+                    }
+                    break;
+                case 0xb0:
+                    if ((old ^ val) & 0x60)
+                        svga_recalctimings(svga);
+                    break;
+                case 0xae:
+                case 0xb2:
+                case 0xbe:
+                    mach_log("ATI VGA write reg=0x%02X, val=0x%02X\n", mach->index, val);
+                    if (mach->regs[0xbe] & 0x08) { /* Read/write bank mode */
+                        mach->bank_r = (((mach->regs[0xb2] & 1) << 3) | ((mach->regs[0xb2] & 0xe0) >> 5));
+                        mach->bank_w = ((mach->regs[0xb2] & 0x1e) >> 1);
+                        if (dev->local) {
+                            mach->bank_r |= (((mach->regs[0xae] & 0x0c) << 2));
+                            mach->bank_w |= (((mach->regs[0xae] & 3) << 4));
+                        }
+                        if (ibm8514_on)
+                            mach_log("Separate B2Bank = %02x, AEbank = %02x.\n", mach->regs[0xb2], mach->regs[0xae]);
+                    } else { /* Single bank mode */
+                        mach->bank_w = ((mach->regs[0xb2] & 0x1e) >> 1);
+                        if (dev->local) {
+                            mach->bank_w |= (((mach->regs[0xae] & 3) << 4));
+                        }
+                        mach->bank_r = mach->bank_w;
+                        if (ibm8514_on)
+                            mach_log("Single B2Bank = %02x, AEbank = %02x.\n", mach->regs[0xb2], mach->regs[0xae]);
+                    }
+                    svga->read_bank = mach->bank_r << 16;
+                    svga->write_bank = mach->bank_w << 16;
+
+                    if (mach->index == 0xbe) {
+                        if ((old ^ val) & 0x10)
+                            svga_recalctimings(svga);
+                    }
+                    break;
+                case 0xbd:
+                    if ((old ^ val) & 4) {
+                        mach32_updatemapping(mach);
+                    }
+                    break;
+                case 0xb3:
+                    ati_eeprom_write(&mach->eeprom, val & 8, val & 2, val & 1);
+                    break;
+                case 0xb6:
+                    if ((old ^ val) & 0x10)
+                        svga_recalctimings(svga);
+                    break;
+                case 0xb8:
+                    if (dev->local) {
+                        if ((old ^ val) & 0x40)
+                            svga_recalctimings(svga);
+                    } else {
+                        if ((old ^ val) & 0xc0)
+                            svga_recalctimings(svga);
+                    }
+                    break;
+                case 0xb9:
+                    if ((old ^ val) & 2)
+                        svga_recalctimings(svga);
+                    break;
+            }
+            break;
+
+        case 0x2ea:
+        case 0x2eb:
+        case 0x2ec:
+        case 0x2ed:
+            rs2 = !!(mach->accel.ext_ge_config & 0x1000);
+            if (dev->local) {
+                if (mach->pci_bus)
+                    ati68860_ramdac_out((addr & 3) | (rs2 << 2), val, svga->ramdac, svga);
+                else
+                    svga_out(addr, val, svga);
+            } else
+                svga_out(addr, val, svga);
+            return;
+
+        case 0x3C6:
+        case 0x3C7:
+        case 0x3C8:
+        case 0x3C9:
+            rs2 = !!(mach->accel.ext_ge_config & 0x1000);
+            if (dev->local) {
+                if (mach->pci_bus)
+                    ati68860_ramdac_out((addr & 3) | (rs2 << 2), val, svga->ramdac, svga);
+                else
+                    svga_out(addr, val, svga);
+            } else
+                svga_out(addr, val, svga);
+            return;
+
+        case 0x3CF:
+            if (svga->gdcaddr == 6) {
+                uint8_t old_val = svga->gdcreg[6];
+                svga->gdcreg[6] = val;
+                if ((svga->gdcreg[6] & 0xc) != (old_val & 0xc))
+                    mach32_updatemapping(mach);
+                return;
+            }
+            break;
+
+        case 0x3D4:
+            svga->crtcreg = val & 0x3f;
+            return;
+        case 0x3D5:
+            if ((svga->crtcreg < 7) && (svga->crtc[0x11] & 0x80))
+                return;
+            if ((svga->crtcreg == 7) && (svga->crtc[0x11] & 0x80))
+                val = (svga->crtc[7] & ~0x10) | (val & 0x10);
+
+            old                       = svga->crtc[svga->crtcreg];
+            svga->crtc[svga->crtcreg] = val;
+            if (old != val) {
+                if (svga->crtcreg < 0xe || svga->crtcreg > 0x10) {
+                    if ((svga->crtcreg == 0xc) || (svga->crtcreg == 0xd)) {
+                        svga->fullchange = 3;
+                        svga->ma_latch   = ((svga->crtc[0xc] << 8) | svga->crtc[0xd]) + ((svga->crtc[8] & 0x60) >> 5);
+                    } else {
+                        svga->fullchange = changeframecount;
+                        svga_recalctimings(svga);
+                    }
+                }
+            }
+            break;
+    }
+    svga_out(addr, val, svga);
+}
+
+static uint8_t
+mach_in(uint16_t addr, void *p)
+{
+    mach_t     *mach     = (mach_t *) p;
+    svga_t     *svga     = &mach->svga;
+    ibm8514_t  *dev      = &svga->dev8514;
+    uint8_t     temp;
+    uint8_t     rs2;
+
+    if (((addr & 0xFFF0) == 0x3D0 || (addr & 0xFFF0) == 0x3B0) && !(svga->miscout & 1))
+        addr ^= 0x60;
+
+    switch (addr) {
+        case 0x1ce:
+            temp = mach->index;
+            break;
+        case 0x1cf:
+            switch (mach->index) {
+                case 0xa8:
+                    temp = (svga->vc >> 8) & 3;
+                    break;
+                case 0xa9:
+                    temp = svga->vc & 0xff;
+                    break;
+                case 0xb0:
+                    temp = mach->regs[0xb0] | 0x80;
+                    if (dev->local) { /*Mach32 VGA 1MB memory*/
+                        temp |= 0x08;
+                        temp &= ~0x10;
+                    } else { /*ATI 28800 VGA 512kB memory*/
+                        temp &= ~0x08;
+                        temp |= 0x10;
+                    }
+                    break;
+                case 0xb7:
+                    temp = mach->regs[0xb7] & ~8;
+                    if (ati_eeprom_read(&mach->eeprom))
+                        temp |= 8;
+                    break;
+
+                default:
+                    temp = mach->regs[mach->index];
+                    break;
+            }
+            break;
+
+        case 0x2ea:
+        case 0x2eb:
+        case 0x2ec:
+        case 0x2ed:
+            rs2 = !!(mach->accel.ext_ge_config & 0x1000);
+            if (dev->local) {
+                if (mach->pci_bus)
+                    return ati68860_ramdac_in((addr & 3) | (rs2 << 2), svga->ramdac, svga);
+                else
+                    return svga_in(addr, svga);
+            }
+            return svga_in(addr, svga);
+
+        case 0x3C6:
+        case 0x3C7:
+        case 0x3C8:
+        case 0x3C9:
+            rs2 = !!(mach->accel.ext_ge_config & 0x1000);
+            if (dev->local) {
+                if (mach->pci_bus)
+                    return ati68860_ramdac_in((addr & 3) | (rs2 << 2), svga->ramdac, svga);
+                else
+                    return svga_in(addr, svga);
+            }
+            return svga_in(addr, svga);
+
+        case 0x3D4:
+            temp = svga->crtcreg;
+            break;
+        case 0x3D5:
+            temp = svga->crtc[svga->crtcreg];
+            break;
+        case 0x3DA:
+            svga->attrff = 0;
+            if (svga->cgastat & 0x01)
+                svga->cgastat &= ~0x38;
+            else
+                svga->cgastat ^= 0x38;
+            return svga->cgastat;
+
+        default:
+            temp = svga_in(addr, svga);
+            break;
+    }
+    return temp;
+}
+
+static void
+mach_recalctimings(svga_t *svga)
+{
+    mach_t *mach = (mach_t *) svga->p;
+    ibm8514_t *dev = &svga->dev8514;
+
+    if (vga_on && !ibm8514_on) {
+        switch (((mach->regs[0xbe] & 0x10) >> 1) | ((mach->regs[0xb9] & 2) << 1) | ((svga->miscout & 0x0c) >> 2)) {
+            case 0x00:
+                svga->clock = (cpuclock * (double) (1ull << 32)) / 42954000.0;
+                break;
+            case 0x01:
+                svga->clock = (cpuclock * (double) (1ull << 32)) / 48771000.0;
+                break;
+            case 0x02:
+                mach_log("clock 2\n");
+                break;
+            case 0x03:
+                svga->clock = (cpuclock * (double) (1ull << 32)) / 36000000.0;
+                break;
+            case 0x04:
+                svga->clock = (cpuclock * (double) (1ull << 32)) / 50350000.0;
+                break;
+            case 0x05:
+                svga->clock = (cpuclock * (double) (1ull << 32)) / 56640000.0;
+                break;
+            case 0x06:
+                mach_log("clock 2\n");
+                break;
+            case 0x07:
+                svga->clock = (cpuclock * (double) (1ull << 32)) / 44900000.0;
+                break;
+            case 0x08:
+                svga->clock = (cpuclock * (double) (1ull << 32)) / 30240000.0;
+                break;
+            case 0x09:
+                svga->clock = (cpuclock * (double) (1ull << 32)) / 32000000.0;
+                break;
+            case 0x0A:
+                svga->clock = (cpuclock * (double) (1ull << 32)) / 37500000.0;
+                break;
+            case 0x0B:
+                svga->clock = (cpuclock * (double) (1ull << 32)) / 39000000.0;
+                break;
+            case 0x0C:
+                svga->clock = (cpuclock * (double) (1ull << 32)) / 50350000.0;
+                break;
+            case 0x0D:
+                svga->clock = (cpuclock * (double) (1ull << 32)) / 56644000.0;
+                break;
+            case 0x0E:
+                svga->clock = (cpuclock * (double) (1ull << 32)) / 75000000.0;
+                break;
+            case 0x0F:
+                svga->clock = (cpuclock * (double) (1ull << 32)) / 65000000.0;
+                break;
+            default:
+                break;
+        }
+    }
+
+    if (mach->regs[0xa3] & 0x10)
+        svga->ma_latch |= 0x10000;
+
+    if (mach->regs[0xb0] & 0x40)
+        svga->ma_latch |= 0x20000;
+
+    if (dev->local) {
+        if (mach->regs[0xad] & 0x04)
+            svga->ma_latch |= 0x40000;
+
+        if (mach->regs[0xad] & 0x08)
+            svga->ma_latch |= 0x80000;
+
+        if (mach->regs[0xb8] & 0x40)
+            svga->clock *= 2;
+    } else {
+        switch (mach->regs[0xb8] & 0xc0) {
+            case 0x40:
+                svga->clock *= 2;
+                break;
+            case 0x80:
+                svga->clock *= 3;
+                break;
+            case 0xc0:
+                svga->clock *= 4;
+                break;
+        }
+    }
+
+    if (mach->regs[0xa7] & 0x80)
+        svga->clock *= 3;
+
+    if (mach->regs[0xb6] & 0x10) {
+        svga->hdisp <<= 1;
+        svga->htotal <<= 1;
+        svga->rowoffset <<= 1;
+        svga->gdcreg[5] &= ~0x40;
+    }
+
+    if (mach->regs[0xb0] & 0x20) {
+        svga->gdcreg[5] |= 0x40;
+    }
+
+    if (vga_on && !ibm8514_on) {
+        if (!svga->scrblank && (svga->crtc[0x17] & 0x80) && svga->attr_palette_enable) {
+            if ((svga->gdcreg[6] & 1) || (svga->attrregs[0x10] & 1)) {
+                switch (svga->gdcreg[5] & 0x60) {
+                    case 0x00:
+                        if (svga->seqregs[1] & 8) /*Low res (320)*/
+                            svga->render = svga_render_4bpp_lowres;
+                        else
+                            svga->render = svga_render_4bpp_highres;
+                        break;
+                    case 0x20:                    /*4 colours*/
+                        if (svga->seqregs[1] & 8) /*Low res (320)*/
+                            svga->render = svga_render_2bpp_lowres;
+                        else
+                            svga->render = svga_render_2bpp_highres;
+                        break;
+                    case 0x40:
+                    case 0x60: /*256+ colours*/
+                        switch (svga->bpp) {
+                            case 8:
+                                svga->map8 = svga->pallook;
+                                if (svga->lowres)
+                                    svga->render = svga_render_8bpp_lowres;
+                                else {
+                                    svga->render = svga_render_8bpp_highres;
+                                    svga->ma_latch <<= 1;
+                                    svga->rowoffset <<= 1;
+                                }
+                                break;
+
+                        }
+                        break;
+                }
+            }
+        }
+    } else if (dev->local) {
+        if (ibm8514_on) {
+            svga->hdisp_time = svga->hdisp  = (dev->hdisp + 1) << 3;
+            dev->pitch                      = (dev->accel.advfunc_cntl & 4) ? 1024 : 640;
+            svga->htotal                    = (dev->htotal + 1);
+            svga->vtotal                    = (dev->vtotal + 1);
+            svga->vsyncstart                = (dev->vsyncstart + 1);
+            svga->rowcount                  = !!(dev->disp_cntl & 0x08);
+            svga->dispend                   = ((dev->vdisp >> 1) + 1);
+            svga->interlace                 = dev->interlace;
+            svga->split                    = 0xffffff;
+            svga->vblankstart              = svga->dispend;
+
+            if (svga->dispend == 766) {
+                svga->dispend = 768;
+                svga->vblankstart = svga->dispend;
+            }
+
+            if (svga->dispend == 598) {
+                svga->dispend = 600;
+                svga->vblankstart = svga->dispend;
+            }
+
+            if (dev->accel.advfunc_cntl & 4) {
+                if (dev->ibm_mode) {
+                    if (svga->hdisp == 8) {
+                        svga->hdisp = 1024;
+                        svga->dispend = 768;
+                        svga->vtotal = 1536;
+                        svga->vsyncstart = 1536;
+                    }
+                }
+
+                if (svga->interlace) {
+                    svga->dispend >>= 1;
+                    svga->vsyncstart >>= 2;
+                    svga->vtotal >>= 2;
+                } else {
+                    svga->vsyncstart >>= 1;
+                    svga->vtotal >>= 1;
+                }
+
+                dev->pitch = dev->ext_pitch;
+                svga->rowoffset = dev->ext_crt_pitch;
+
+                svga->clock = (cpuclock * (double) (1ull << 32)) / 44900000.0;
+            } else {
+                if (dev->ibm_mode) {
+                    if ((svga->hdisp == 1024) && !dev->internal_pitch) {
+                        svga->hdisp = 640;
+                        svga->dispend = 480;
+                    }
+                }
+
+                if (svga->interlace) {
+                    svga->dispend >>= 1;
+                    svga->vsyncstart >>= 2;
+                    svga->vtotal >>= 2;
+                } else {
+                    svga->vsyncstart >>= 1;
+                    svga->vtotal >>= 1;
+                }
+
+                dev->pitch = dev->ext_pitch;
+                svga->rowoffset = dev->ext_crt_pitch;
+
+                svga->clock = (cpuclock * (double) (1ull << 32)) / 25175000.0;
+            }
+            switch (svga->bpp) {
+                case 8:
+                default:
+                    svga->render = svga_render_8bpp_highres;
+                    break;
+                case 15:
+                    svga->render = svga_render_15bpp_highres;
+                    break;
+                case 16:
+                    svga->render = svga_render_16bpp_highres;
+                    break;
+                case 24:
+                    svga->render = svga_render_24bpp_highres;
+                    break;
+            }
+            mach_log("BPP=%d, VRAM Mask=%08x, NormalPitch=%d, CRTPitch=%d, VSYNCSTART=%d, VTOTAL=%d, ROWCOUNT=%d, mode=%d, highres bit=%x, has_vga?=%d, override=%d.\n", svga->bpp, svga->vram_mask, dev->pitch, dev->ext_crt_pitch, svga->vsyncstart, svga->vtotal, svga->rowcount, dev->ibm_mode, dev->accel.advfunc_cntl & 4, ibm8514_has_vga, svga->override);
+        }
+        mach_log("8514 enabled, hdisp=%d, vtotal=%d, htotal=%d, dispend=%d, rowoffset=%d, split=%d, vsyncstart=%d, split=%08x\n", svga->hdisp, svga->vtotal, svga->htotal, svga->dispend, svga->rowoffset, svga->split, svga->vsyncstart, svga->split);
+    }
+}
+
+static void
+mach_accel_out_fifo(mach_t *mach, svga_t *svga, ibm8514_t *dev, uint16_t port, uint32_t val, int len)
+{
+    int frgd_sel, bkgd_sel, mono_src;
+    switch (port) {
+        case 0x82e8:
+        case 0xc2e8:
+            if (len == 1) {
+                dev->accel.cur_y = (dev->accel.cur_y & 0x700) | val;
+            } else {
+                dev->accel.cur_y = val & 0x7ff;
+            }
+            break;
+        case 0x82e9:
+        case 0xc2e9:
+            if (len == 1) {
+                dev->accel.cur_y = (dev->accel.cur_y & 0xff) | ((val & 0x07) << 8);
+            }
+            break;
+
+        case 0x86e8:
+        case 0xc6e8:
+            if (len == 1) {
+                dev->accel.cur_x = (dev->accel.cur_x & 0x700) | val;
+            } else {
+                dev->accel.cur_x = val & 0x7ff;
+            }
+            break;
+        case 0x86e9:
+        case 0xc6e9:
+            if (len == 1) {
+                dev->accel.cur_x = (dev->accel.cur_x & 0xff) | ((val & 0x07) << 8);
+            }
+            break;
+
+        case 0x8ae8:
+        case 0xcae8:
+            if (len == 1)
+                dev->accel.desty_axstp = (dev->accel.desty_axstp & 0x3f00) | val;
+            else {
+                mach->accel.src_y = val;
+                dev->accel.desty_axstp = val & 0x3fff;
+                if (val & 0x2000)
+                    dev->accel.desty_axstp |= ~0x1fff;
+            }
+            break;
+        case 0x8ae9:
+        case 0xcae9:
+            if (len == 1) {
+                dev->accel.desty_axstp = (dev->accel.desty_axstp & 0xff) | ((val & 0x3f) << 8);
+                if (val & 0x20)
+                    dev->accel.desty_axstp |= ~0x1fff;
+            }
+            break;
+
+        case 0x8ee8:
+        case 0xcee8:
+            if (len == 1)
+                dev->accel.destx_distp = (dev->accel.destx_distp & 0x3f00) | val;
+            else {
+                mach->accel.src_x = val;
+                dev->accel.destx_distp = val & 0x3fff;
+                if (val & 0x2000)
+                    dev->accel.destx_distp |= ~0x1fff;
+            }
+            break;
+        case 0x8ee9:
+        case 0xcee9:
+            if (len == 1) {
+                dev->accel.destx_distp = (dev->accel.destx_distp & 0xff) | ((val & 0x3f) << 8);
+                if (val & 0x20)
+                    dev->accel.destx_distp |= ~0x1fff;
+            }
+            break;
+
+        case 0x92e8:
+            if (len != 1)
+                dev->test = val;
+        case 0xd2e8:
+            mach_log("92E8 = %04x\n", val);
+            if (len == 1)
+                dev->accel.err_term = (dev->accel.err_term & 0x3f00) | val;
+            else {
+                dev->accel.err_term = val & 0x3fff;
+                if (val & 0x2000)
+                    dev->accel.err_term |= ~0x1fff;
+            }
+            break;
+        case 0x92e9:
+        case 0xd2e9:
+            if (len == 1) {
+                dev->accel.err_term = (dev->accel.err_term & 0xff) | ((val & 0x3f) << 8);
+                if (val & 0x20)
+                    dev->accel.err_term |= ~0x1fff;
+            }
+            break;
+
+        case 0x96e8:
+        case 0xd6e8:
+            if (len == 1)
+                dev->accel.maj_axis_pcnt = (dev->accel.maj_axis_pcnt & 0x0700) | val;
+            else {
+                mach->accel.test = val & 0x1fff;
+                dev->accel.maj_axis_pcnt = val & 0x07ff;
+            }
+            break;
+        case 0x96e9:
+        case 0xd6e9:
+            if (len == 1) {
+                dev->accel.maj_axis_pcnt = (dev->accel.maj_axis_pcnt & 0xff) | ((val & 0x07) << 8);
+            }
+            break;
+
+        case 0x9ae8:
+        case 0xdae8:
+            dev->accel.ssv_state = 0;
+            if (len == 1)
+                dev->accel.cmd = (dev->accel.cmd & 0xff00) | val;
+            else {
+                dev->data_available  = 0;
+                dev->data_available2 = 0;
+                dev->accel.cmd       = val;
+                mach_log("CMD8514 = %04x.\n", val);
+                mach->accel.cmd_type = -1;
+                if (port == 0xdae8) {
+                    if (dev->accel.cmd & 0x100)
+                        dev->accel.cmd_back = 0;
+                }
+                ibm8514_accel_start(-1, 0, -1, 0, svga, len);
+            }
+            break;
+        case 0x9ae9:
+        case 0xdae9:
+            if (len == 1) {
+                dev->data_available  = 0;
+                dev->data_available2 = 0;
+                dev->accel.cmd       = (dev->accel.cmd & 0xff) | (val << 8);
+                mach->accel.cmd_type = -1;
+                if (port == 0xdae9) {
+                    if (dev->accel.cmd & 0x100)
+                        dev->accel.cmd_back = 0;
+                }
+                ibm8514_accel_start(-1, 0, -1, 0, svga, len);
+            }
+            break;
+
+        case 0x9ee8:
+        case 0xdee8:
+            dev->accel.ssv_state = 1;
+            if (len == 1)
+                dev->accel.short_stroke = (dev->accel.short_stroke & 0xff00) | val;
+            else {
+                dev->accel.short_stroke = val;
+                dev->accel.cx           = dev->accel.cur_x;
+                dev->accel.cy           = dev->accel.cur_y;
+
+                if (dev->accel.cur_x >= 0x600) {
+                    dev->accel.cx |= ~0x5ff;
+                }
+                if (dev->accel.cur_y >= 0x600) {
+                    dev->accel.cy |= ~0x5ff;
+                }
+
+                if (dev->accel.cmd & 0x1000) {
+                    ibm8514_short_stroke_start(-1, 0, -1, 0, svga, dev->accel.short_stroke & 0xff, len);
+                    ibm8514_short_stroke_start(-1, 0, -1, 0, svga, dev->accel.short_stroke >> 8, len);
+                } else {
+                    ibm8514_short_stroke_start(-1, 0, -1, 0, svga, dev->accel.short_stroke >> 8, len);
+                    ibm8514_short_stroke_start(-1, 0, -1, 0, svga, dev->accel.short_stroke & 0xff, len);
+                }
+            }
+            break;
+        case 0x9ee9:
+        case 0xdee9:
+            if (len == 1) {
+                dev->accel.short_stroke = (dev->accel.short_stroke & 0xff) | (val << 8);
+                dev->accel.cx           = dev->accel.cur_x;
+                dev->accel.cy           = dev->accel.cur_y;
+
+                if (dev->accel.cur_x >= 0x600) {
+                    dev->accel.cx |= ~0x5ff;
+                }
+                if (dev->accel.cur_y >= 0x600) {
+                    dev->accel.cy |= ~0x5ff;
+                }
+
+                if (dev->accel.cmd & 0x1000) {
+                    ibm8514_short_stroke_start(-1, 0, -1, 0, svga, dev->accel.short_stroke & 0xff, len);
+                    ibm8514_short_stroke_start(-1, 0, -1, 0, svga, dev->accel.short_stroke >> 8, len);
+                } else {
+                    ibm8514_short_stroke_start(-1, 0, -1, 0, svga, dev->accel.short_stroke >> 8, len);
+                    ibm8514_short_stroke_start(-1, 0, -1, 0, svga, dev->accel.short_stroke & 0xff, len);
+                }
+            }
+            break;
+
+        case 0xa2e8:
+        case 0xe2e8:
+            if (port == 0xe2e8) {
+                if (dev->accel.cmd_back) {
+                    if (len == 1)
+                        dev->accel.bkgd_color = (dev->accel.bkgd_color & 0x00ff) | val;
+                    else
+                        dev->accel.bkgd_color = val;
+                } else {
+                    if (len == 1) {
+                        if (mach->accel.cmd_type >= 0) {
+                            if (mach_pixel_read(mach))
+                                break;
+                            mach->accel.pix_trans[1] = val;
+                        }
+                    } else {
+                        if (mach->accel.cmd_type >= 0) {
+                            if (mach_pixel_read(mach))
+                                break;
+                            mach_accel_out_pixtrans(mach, dev, port, val, len);
+                        } else {
+                            if (ibm8514_cpu_dest(svga))
+                                break;
+                            ibm8514_accel_out_pixtrans(svga, port, val, len);
+                        }
+                    }
+                }
+            } else {
+                if (len == 1)
+                    dev->accel.bkgd_color = (dev->accel.bkgd_color & 0x00ff) | val;
+                else
+                    dev->accel.bkgd_color = val;
+            }
+            break;
+        case 0xa2e9:
+        case 0xe2e9:
+            if (port == 0xe2e9) {
+                if (dev->accel.cmd_back) {
+                    if (len == 1)
+                        dev->accel.bkgd_color = (dev->accel.bkgd_color & 0xff00) | (val << 8);
+                } else {
+                    if (len == 1) {
+                        if (mach->accel.cmd_type >= 0) {
+                            if (mach_pixel_read(mach))
+                                break;
+                            mach->accel.pix_trans[0] = val;
+                            frgd_sel = (mach->accel.dp_config >> 13) & 7;
+                            bkgd_sel = (mach->accel.dp_config >> 7) & 3;
+                            mono_src = (mach->accel.dp_config >> 5) & 3;
+
+                            switch (mach->accel.dp_config & 0x200) {
+                                case 0x000: /*8-bit size*/
+                                    if (mono_src == 2) {
+                                        if ((frgd_sel != 2) && (bkgd_sel != 2)) {
+                                            mach_accel_start(mach->accel.cmd_type, 1, 8, mach->accel.pix_trans[0] | (mach->accel.pix_trans[1] << 8), 0, mach, dev, len);
+                                        } else
+                                            mach_accel_start(mach->accel.cmd_type, 1, 1, -1, mach->accel.pix_trans[0] | (mach->accel.pix_trans[1] << 8), mach, dev, len);
+                                    } else
+                                        mach_accel_start(mach->accel.cmd_type, 1, 1, -1, mach->accel.pix_trans[0] | (mach->accel.pix_trans[1] << 8), mach, dev, len);
+                                    break;
+                                case 0x200: /*16-bit size*/
+                                    if (mono_src == 2) {
+                                        if ((frgd_sel != 2) && (bkgd_sel != 2)) {
+                                            if (mach->accel.dp_config & 0x1000)
+                                                mach_accel_start(mach->accel.cmd_type, 1, 16, mach->accel.pix_trans[1] | (mach->accel.pix_trans[0] << 8), 0, mach, dev, len);
+                                            else
+                                                mach_accel_start(mach->accel.cmd_type, 1, 16, mach->accel.pix_trans[0] | (mach->accel.pix_trans[1] << 8), 0, mach, dev, len);
+                                        } else
+                                            mach_accel_start(mach->accel.cmd_type, 1, 2, -1, mach->accel.pix_trans[0] | (mach->accel.pix_trans[1] << 8), mach, dev, len);
+                                    } else
+                                        mach_accel_start(mach->accel.cmd_type, 1, 2, -1, mach->accel.pix_trans[0] | (mach->accel.pix_trans[1] << 8), mach, dev, len);
+                                    break;
+                            }
+                        }
+                    }
+                }
+            } else {
+                if (len == 1)
+                    dev->accel.bkgd_color = (dev->accel.bkgd_color & 0xff00) | (val << 8);
+            }
+            break;
+
+        case 0xa6e8:
+        case 0xe6e8:
+            if (port == 0xe6e8) {
+                if (dev->accel.cmd_back) {
+                    if (len == 1)
+                        dev->accel.frgd_color = (dev->accel.frgd_color & 0x00ff) | val;
+                    else
+                        dev->accel.frgd_color = val;
+                } else {
+                    if (len == 1) {
+                        if (mach->accel.cmd_type >= 0) {
+                            if (mach_pixel_read(mach))
+                                break;
+                            mach->accel.pix_trans[1] = val;
+                        }
+                    } else {
+                        if (mach->accel.cmd_type >= 0) {
+                            if (mach_pixel_read(mach))
+                                break;
+                            mach_accel_out_pixtrans(mach, dev, port, val, len);
+                        } else {
+                            if (ibm8514_cpu_dest(svga))
+                                break;
+                            ibm8514_accel_out_pixtrans(svga, port, val, len);
+                        }
+                    }
+                }
+            } else {
+                if (len == 1)
+                    dev->accel.frgd_color = (dev->accel.frgd_color & 0x00ff) | val;
+                else
+                    dev->accel.frgd_color = val;
+            }
+            break;
+        case 0xa6e9:
+        case 0xe6e9:
+            if (port == 0xe6e9) {
+                if (dev->accel.cmd_back) {
+                    if (len == 1)
+                        dev->accel.frgd_color = (dev->accel.frgd_color & 0xff00) | (val << 8);
+                } else {
+                    if (len == 1) {
+                        if (mach->accel.cmd_type >= 0) {
+                            if (mach_pixel_read(mach))
+                                break;
+                            mach->accel.pix_trans[0] = val;
+                            frgd_sel = (mach->accel.dp_config >> 13) & 7;
+                            bkgd_sel = (mach->accel.dp_config >> 7) & 3;
+                            mono_src = (mach->accel.dp_config >> 5) & 3;
+
+                            switch (mach->accel.dp_config & 0x200) {
+                                case 0x000: /*8-bit size*/
+                                    if (mono_src == 2) {
+                                        if ((frgd_sel != 2) && (bkgd_sel != 2)) {
+                                            mach_accel_start(mach->accel.cmd_type, 1, 8, mach->accel.pix_trans[0] | (mach->accel.pix_trans[1] << 8), 0, mach, dev, len);
+                                        } else
+                                            mach_accel_start(mach->accel.cmd_type, 1, 1, -1, mach->accel.pix_trans[0] | (mach->accel.pix_trans[1] << 8), mach, dev, len);
+                                    } else
+                                        mach_accel_start(mach->accel.cmd_type, 1, 1, -1, mach->accel.pix_trans[0] | (mach->accel.pix_trans[1] << 8), mach, dev, len);
+                                    break;
+                                case 0x200: /*16-bit size*/
+                                    if (mono_src == 2) {
+                                        if ((frgd_sel != 2) && (bkgd_sel != 2)) {
+                                            if (mach->accel.dp_config & 0x1000)
+                                                mach_accel_start(mach->accel.cmd_type, 1, 16, mach->accel.pix_trans[1] | (mach->accel.pix_trans[0] << 8), 0, mach, dev, len);
+                                            else
+                                                mach_accel_start(mach->accel.cmd_type, 1, 16, mach->accel.pix_trans[0] | (mach->accel.pix_trans[1] << 8), 0, mach, dev, len);
+                                        } else
+                                            mach_accel_start(mach->accel.cmd_type, 1, 2, -1, mach->accel.pix_trans[0] | (mach->accel.pix_trans[1] << 8), mach, dev, len);
+                                    } else
+                                        mach_accel_start(mach->accel.cmd_type, 1, 2, -1, mach->accel.pix_trans[0] | (mach->accel.pix_trans[1] << 8), mach, dev, len);
+                                    break;
+                            }
+                        }
+                    }
+                }
+            } else {
+                if (len == 1)
+                    dev->accel.frgd_color = (dev->accel.frgd_color & 0xff00) | (val << 8);
+            }
+            break;
+
+        case 0xaae8:
+        case 0xeae8:
+            if (len == 1)
+                dev->accel.wrt_mask = (dev->accel.wrt_mask & 0x00ff) | val;
+            else
+                dev->accel.wrt_mask = val;
+            break;
+        case 0xaae9:
+        case 0xeae9:
+            if (len == 1)
+                dev->accel.wrt_mask = (dev->accel.wrt_mask & 0xff00) | (val << 8);
+            break;
+
+        case 0xaee8:
+        case 0xeee8:
+            if (len == 1)
+                dev->accel.rd_mask = (dev->accel.rd_mask & 0x00ff) | val;
+            else
+                dev->accel.rd_mask = val;
+            break;
+        case 0xaee9:
+        case 0xeee9:
+            if (len == 1)
+                dev->accel.rd_mask = (dev->accel.rd_mask & 0xff00) | (val << 8);
+            break;
+
+        case 0xb2e8:
+        case 0xf2e8:
+            if (len == 1)
+                dev->accel.color_cmp = (dev->accel.color_cmp & 0x00ff) | val;
+            else
+                dev->accel.color_cmp = val;
+            break;
+        case 0xb2e9:
+        case 0xf2e9:
+            if (len == 1)
+                dev->accel.color_cmp = (dev->accel.color_cmp & 0xff00) | (val << 8);
+            break;
+
+        case 0xb6e8:
+        case 0xf6e8:
+            dev->accel.bkgd_mix = val & 0xff;
+            break;
+
+        case 0xbae8:
+        case 0xfae8:
+            dev->accel.frgd_mix = val & 0xff;
+            break;
+
+        case 0xbee8:
+        case 0xfee8:
+            if (len == 1)
+                dev->accel.multifunc_cntl = (dev->accel.multifunc_cntl & 0xff00) | val;
+            else {
+                dev->accel.multifunc_cntl                             = val;
+                dev->accel.multifunc[dev->accel.multifunc_cntl >> 12] = dev->accel.multifunc_cntl & 0xfff;
+                if ((dev->accel.multifunc_cntl >> 12) == 1) {
+                    dev->accel.clip_top = val & 0x7ff;
+                }
+                if ((dev->accel.multifunc_cntl >> 12) == 2) {
+                    dev->accel.clip_left = val & 0x7ff;
+                }
+                if ((dev->accel.multifunc_cntl >> 12) == 3) {
+                    dev->accel.multifunc[3] = val & 0x7ff;
+                }
+                if ((dev->accel.multifunc_cntl >> 12) == 4) {
+                    dev->accel.multifunc[4] = val & 0x7ff;
+                }
+                mach_log("CLIPBOTTOM=%d, CLIPRIGHT=%d, bpp=%d, pitch=%d.\n", dev->accel.multifunc[3], dev->accel.multifunc[4], svga->bpp, dev->pitch);
+                if ((dev->accel.multifunc_cntl >> 12) == 5) {
+                    if (!dev->local || !dev->ext_crt_pitch)
+                        dev->ext_crt_pitch = 128;
+                    svga_recalctimings(svga);
+                }
+                if (port == 0xfee8)
+                    dev->accel.cmd_back = 1;
+                else
+                    dev->accel.cmd_back = 0;
+            }
+            break;
+        case 0xbee9:
+        case 0xfee9:
+            if (len == 1) {
+                dev->accel.multifunc_cntl                             = (dev->accel.multifunc_cntl & 0xff) | (val << 8);
+                dev->accel.multifunc[dev->accel.multifunc_cntl >> 12] = dev->accel.multifunc_cntl & 0xfff;
+                if ((dev->accel.multifunc_cntl >> 12) == 1) {
+                    dev->accel.clip_top = dev->accel.multifunc_cntl & 0x7ff;
+                }
+                if ((dev->accel.multifunc_cntl >> 12) == 2) {
+                    dev->accel.clip_left = dev->accel.multifunc_cntl & 0x7ff;
+                }
+                if ((dev->accel.multifunc_cntl >> 12) == 5) {
+                    if (!dev->local || !dev->ext_crt_pitch)
+                        dev->ext_crt_pitch = 128;
+                    svga_recalctimings(svga);
+                }
+                if (port == 0xfee9)
+                    dev->accel.cmd_back = 1;
+                else
+                    dev->accel.cmd_back = 0;
+            }
+            break;
+
+/*ATI Mach8/32 specific registers*/
+        case 0x82ee:
+            mach->accel.patt_data_idx = val & 0x1f;
+            mach_log("Pattern Data Index = %d.\n", val & 0x1f);
+            break;
+
+        case 0x8eee:
+            if (len == 1) {
+                mach->accel.patt_data[mach->accel.patt_data_idx] = val;
+            } else {
+                mach->accel.patt_data[mach->accel.patt_data_idx] = val & 0xff;
+                mach->accel.patt_data[mach->accel.patt_data_idx + 1] = (val >> 8) & 0xff;
+                if (mach->accel.mono_pattern_enable)
+                    mach->accel.patt_data_idx = (mach->accel.patt_data_idx + 2) & 0x17;
+                else {
+                    frgd_sel = (mach->accel.dp_config >> 13) & 7;
+                    mono_src = (mach->accel.dp_config >> 5) & 3;
+                    if ((svga->bpp == 24) && (mach->accel.patt_len == 0x17) && (frgd_sel == 5)) {
+                        mach->accel.patt_data_idx += 2;
+                        dev->accel.y1 = 1;
+                    } else {
+                        if (svga->bpp == 24)
+                            mach->accel.patt_data_idx += 2;
+                        else
+                            mach->accel.patt_data_idx = (mach->accel.patt_data_idx + 2) & mach->accel.patt_len;
+
+                    }
+                    mach_log("ExtCONFIG = %04x, Pattern Mono = %04x, selidx = %d, dataidx = %d, bit 0 = %02x len = %d.\n", mach->accel.ext_ge_config, val, mach->accel.patt_idx, mach->accel.patt_data_idx, val & 1, mach->accel.patt_len);
+                }
+            }
+            break;
+        case 0x8eef:
+            if (len == 1) {
+                mach->accel.patt_data[mach->accel.patt_data_idx + 1] = val;
+                if (mach->accel.mono_pattern_enable)
+                    mach->accel.patt_data_idx = (mach->accel.patt_data_idx + 2) & 7;
+                else {
+                    frgd_sel = (mach->accel.dp_config >> 13) & 7;
+                    if ((svga->bpp == 24) && (mach->accel.patt_len == 0x17) && (frgd_sel == 5)) {
+                        mach->accel.patt_data_idx += 2;
+                        dev->accel.y1 = 1;
+                    } else
+                        mach->accel.patt_data_idx = (mach->accel.patt_data_idx + 2) & mach->accel.patt_len;
+                }
+            }
+            break;
+
+        case 0x96ee:
+            if (len == 1)
+                mach->accel.bres_count = (mach->accel.bres_count & 0x700) | val;
+            else {
+                mach->accel.bres_count = val & 0x7ff;
+                mach_log("96EE line draw.\n");
+                dev->data_available  = 0;
+                dev->data_available2 = 0;
+                mach->accel.cmd_type = 1;
+                mach_accel_start(mach->accel.cmd_type, 0, -1, -1, 0, mach, dev, len);
+            }
+            break;
+        case 0x96ef:
+            if (len == 1) {
+                mach->accel.bres_count = (mach->accel.bres_count & 0xff) | ((val & 0x07) << 8);
+                mach_log("96EE (2) line draw.\n");
+                dev->data_available  = 0;
+                dev->data_available2 = 0;
+                mach->accel.cmd_type = 1;
+                mach_accel_start(mach->accel.cmd_type, 0, -1, -1, 0, mach, dev, len);
+            }
+            break;
+
+        case 0x9aee:
+            mach->accel.line_idx = val & 0x07;
+            break;
+
+        case 0xa2ee:
+            mach_log("Line OPT = %04x\n", val);
+            if (len == 1)
+                mach->accel.linedraw_opt = (mach->accel.linedraw_opt & 0xff00) | val;
+            else {
+                mach->accel.linedraw_opt = val;
+            }
+            break;
+        case 0xa2ef:
+            if (len == 1) {
+                mach->accel.linedraw_opt = (mach->accel.linedraw_opt & 0x00ff) | (val << 8);
+            }
+            break;
+
+        case 0xa6ee:
+            if (len == 1)
+                mach->accel.dest_x_start = (mach->accel.dest_x_start & 0x700) | val;
+            else
+                mach->accel.dest_x_start = val & 0x7ff;
+            break;
+        case 0xa6ef:
+            if (len == 1)
+                mach->accel.dest_x_start = (mach->accel.dest_x_start & 0x0ff) | ((val & 0x07) << 8);
+            break;
+
+        case 0xaaee:
+            if (len == 1)
+                mach->accel.dest_x_end = (mach->accel.dest_x_end & 0x700) | val;
+            else {
+                mach->accel.dest_x_end = val & 0x7ff;
+            }
+            break;
+        case 0xaaef:
+            if (len == 1)
+                mach->accel.dest_x_end = (mach->accel.dest_x_end & 0x0ff) | ((val & 0x07) << 8);
+            break;
+
+        case 0xaeee:
+            mach_log("AEEE write val = %04x.\n", val);
+            if (len == 1)
+                mach->accel.dest_y_end = (mach->accel.dest_y_end & 0x700) | val;
+            else {
+                mach->accel.dest_y_end = val & 0x7ff;
+                if ((val + 1) == 0x10000) {
+                    mach_log("Dest_Y_end overflow val = %04x\n", val);
+                    mach->accel.dest_y_end = 0;
+                }
+                dev->data_available  = 0;
+                dev->data_available2 = 0;
+                mach_log("BitBLT = %04x.\n", mach->accel.dp_config);
+                mach->accel.cmd_type = 2; /*Non-conforming BitBLT from dest_y_end register (0xaeee)*/
+                mach_accel_start(mach->accel.cmd_type, 0, -1, -1, 0, mach, dev, len);
+            }
+            break;
+        case 0xaeef:
+            if (len == 1) {
+                mach->accel.dest_y_end = (mach->accel.dest_y_end & 0x0ff) | ((val & 0x07) << 8);
+                dev->data_available  = 0;
+                dev->data_available2 = 0;
+                mach->accel.cmd_type = 2; /*Non-conforming BitBLT from dest_y_end register (0xaeee)*/
+                mach_accel_start(mach->accel.cmd_type, 0, -1, -1, 0, mach, dev, len);
+            }
+            break;
+
+        case 0xb2ee:
+            if (len == 1)
+                mach->accel.src_x_start = (mach->accel.src_x_start & 0x700) | val;
+            else
+                mach->accel.src_x_start = val & 0x7ff;
+            break;
+        case 0xb2ef:
+            if (len == 1)
+                mach->accel.src_x_start = (mach->accel.src_x_start & 0x0ff) | ((val & 0x07) << 8);
+            break;
+
+        case 0xb6ee:
+            dev->accel.bkgd_mix = val & 0xff;
+            break;
+
+        case 0xbaee:
+            dev->accel.frgd_mix = val & 0xff;
+            break;
+
+        case 0xbeee:
+            if (len == 1)
+                mach->accel.src_x_end = (mach->accel.src_x_end & 0x700) | val;
+            else {
+                mach->accel.src_x_end = val & 0x7ff;
+            }
+            break;
+        case 0xbeef:
+            if (len == 1)
+                mach->accel.src_x_end = (mach->accel.src_x_end & 0x0ff) | ((val & 0x07) << 8);
+            break;
+
+        case 0xc2ee:
+            mach->accel.src_y_dir = val & 1;
+            break;
+
+        case 0xc6ee:
+            mach->accel.cmd_type = 0;
+            mach_log("TODO: Short Stroke.\n");
+            break;
+
+        case 0xcaee:
+            mach_log("CAEE write val = %04x.\n", val);
+            if (len == 1)
+                mach->accel.scan_to_x = (mach->accel.scan_to_x & 0x700) | val;
+            else {
+                mach->accel.scan_to_x = (val & 0x7ff);
+                if ((val + 1) == 0x10000) {
+                    mach_log("Scan_to_X overflow val = %04x\n", val);
+                    mach->accel.scan_to_x = 0;
+                }
+                dev->data_available  = 0;
+                dev->data_available2 = 0;
+                mach->accel.cmd_type = 5; /*Horizontal Raster Draw from scan_to_x register (0xcaee)*/
+                mach_log("ScanToX = %04x.\n", mach->accel.dp_config);
+                mach_accel_start(mach->accel.cmd_type, 0, -1, -1, 0, mach, dev, len);
+            }
+            break;
+        case 0xcaef:
+            if (len == 1) {
+                mach->accel.scan_to_x = (mach->accel.scan_to_x & 0x0ff) | ((val & 0x07) << 8);
+                dev->data_available  = 0;
+                dev->data_available2 = 0;
+                mach->accel.cmd_type = 5; /*Horizontal Raster Draw from scan_to_x register (0xcaee)*/
+                mach_accel_start(mach->accel.cmd_type, 0, -1, -1, 0, mach, dev, len);
+            }
+            break;
+
+        case 0xceee:
+            mach_log("CEEE write val = %04x.\n", val);
+            if (len == 1)
+                mach->accel.dp_config = (mach->accel.dp_config & 0xff00) | val;
+            else {
+                mach->accel.dp_config = val;
+            }
+            break;
+        case 0xceef:
+            if (len == 1) {
+                mach->accel.dp_config = (mach->accel.dp_config & 0x00ff) | (val << 8);
+            }
+            break;
+
+        case 0xd2ee:
+            mach->accel.patt_len = val & 0x1f;
+            mach_log("Pattern Length = %d, val = %04x.\n", val & 0x1f, val);
+            mach->accel.mono_pattern_enable = !!(val & 0x80);
+            if (len != 1) {
+                mach->accel.patt_len_reg = val;
+            } else {
+                mach->accel.patt_len_reg = (mach->accel.patt_len_reg & 0xff00) | val;
+            }
+            break;
+        case 0xd2ef:
+            if (len == 1)
+                mach->accel.patt_len_reg = (mach->accel.patt_len_reg & 0x00ff) | (val << 8);
+            break;
+
+        case 0xd6ee:
+            mach->accel.patt_idx = val & 0x1f;
+            mach_log("Pattern Index = %d, val = %02x.\n", val & 0x1f, val);
+            break;
+
+        case 0xdaee:
+            mach_log("DAEE (extclipl) write val = %d\n", val);
+            if (len == 1)
+                dev->accel.clip_left = (dev->accel.clip_left & 0x700) | val;
+            else {
+                dev->accel.clip_left = val & 0x7ff;
+            }
+            break;
+        case 0xdaef:
+            if (len == 1)
+                dev->accel.clip_left = (dev->accel.clip_left & 0x0ff) | ((val & 0x07) << 8);
+            break;
+
+        case 0xdeee:
+            mach_log("DEEE (extclipt) write val = %d\n", val);
+            if (len == 1)
+                dev->accel.clip_top = (dev->accel.clip_top & 0x700) | val;
+            else {
+                dev->accel.clip_top = val & 0x7ff;
+            }
+            break;
+        case 0xdeef:
+            if (len == 1)
+                dev->accel.clip_top = (dev->accel.clip_top & 0x0ff) | ((val & 0x07) << 8);
+            break;
+
+        case 0xe2ee:
+            mach_log("E2EE (extclipr) write val = %d\n", val);
+            if (len == 1)
+                dev->accel.multifunc[4] = (dev->accel.multifunc[4] & 0x700) | val;
+            else {
+                dev->accel.multifunc[4] = val & 0x7ff;
+            }
+            break;
+        case 0xe2ef:
+            if (len == 1)
+                dev->accel.multifunc[4] = (dev->accel.multifunc[4] & 0x0ff) | ((val & 0x07) << 8);
+            break;
+
+        case 0xe6ee:
+            mach_log("E6EE (extclipb) write val = %d\n", val);
+            if (len == 1)
+                dev->accel.multifunc[3] = (dev->accel.multifunc[3] & 0x700) | val;
+            else {
+                dev->accel.multifunc[3] = val & 0x7ff;
+            }
+            break;
+        case 0xe6ef:
+            if (len == 1)
+                dev->accel.multifunc[3] = (dev->accel.multifunc[3] & 0x0ff) | ((val & 0x07) << 8);
+            break;
+
+        case 0xeeee:
+            if (len == 1)
+                mach->accel.dest_cmp_fn = (mach->accel.dest_cmp_fn & 0xff00) | val;
+            else
+                mach->accel.dest_cmp_fn = val;
+            break;
+        case 0xeeef:
+            if (len == 1)
+                mach->accel.dest_cmp_fn = (mach->accel.dest_cmp_fn & 0x00ff) | (val << 8);
+            break;
+
+        case 0xf2ee:
+            mach_log("F2EE.\n");
+            if (len == 1)
+                mach->accel.dst_clr_cmp_mask = (mach->accel.dst_clr_cmp_mask & 0xff00) | val;
+            else
+                mach->accel.dst_clr_cmp_mask = val;
+            break;
+        case 0xf2ef:
+            if (len == 1)
+                mach->accel.dst_clr_cmp_mask = (mach->accel.dst_clr_cmp_mask & 0x00ff) | (val << 8);
+            break;
+
+        case 0xfeee:
+            if (mach->accel.dp_config == 0x2231 || mach->accel.dp_config == 0x2211)
+                mach_log("FEEE val = %d, lineidx = %d, DPCONFIG = %04x, CPUCX = %04x.\n", val, mach->accel.line_idx, mach->accel.dp_config, CX);
+            if (len != 1) {
+                mach->accel.line_array[mach->accel.line_idx] = val;
+                dev->accel.cur_x = mach->accel.line_array[(mach->accel.line_idx == 4) ? 4 : 0];
+                dev->accel.cur_y = mach->accel.line_array[(mach->accel.line_idx == 5) ? 5 : 1];
+                mach->accel.cx_end_line = mach->accel.line_array[2];
+                mach->accel.cy_end_line = mach->accel.line_array[3];
+                if ((mach->accel.line_idx == 3) || (mach->accel.line_idx == 5)) {
+                    mach->accel.cmd_type = (mach->accel.line_idx == 5) ? 4 : 3;
+                    mach_accel_start(mach->accel.cmd_type, 0, -1, -1, 0, mach, dev, len);
+                    mach->accel.line_idx = (mach->accel.line_idx == 5) ? 4 : 2;
+                    break;
+                }
+                mach->accel.line_idx++;
+            }
+            break;
+    }
+}
+
+static void
+mach_accel_out(uint16_t port, uint32_t val, mach_t *mach, int len)
+{
+    svga_t *svga = &mach->svga;
+    ibm8514_t *dev = &svga->dev8514;
+
+    mach_log("Port accel out = %04x, val = %04x, len = %d.\n", port, val, len);
+
+    if (port & 0x8000) {
+        mach_accel_out_fifo(mach, svga, dev, port, val, len);
+    } else {
+        switch (port) {
+            case 0x2e8:
+                if (len == 1)
+                    dev->htotal = (dev->htotal & 0xff00) | val;
+                else {
+                    dev->htotal = val;
+                    svga_recalctimings(svga);
+                }
+                break;
+            case 0x2e9:
+                if (len != 1) {
+                    dev->htotal = (dev->htotal & 0xff) | (val << 8);
+                    mach_log("ATI 8514/A: H_TOTAL write 02E8 = %d\n", dev->htotal + 1);
+                    svga_recalctimings(svga);
+                }
+                break;
+
+            case 0x6e8:
+                dev->hdisp = val;
+                mach_log("ATI 8514/A: H_DISP write 06E8 = %d\n", dev->hdisp + 1);
+                svga_recalctimings(svga);
+                break;
+
+            case 0xae8:
+                mach_log("ATI 8514/A: H_SYNC_STRT write 0AE8 = %d\n", val + 1);
+                svga_recalctimings(svga);
+                break;
+
+            case 0xee8:
+                mach_log("ATI 8514/A: H_SYNC_WID write 0EE8 = %d\n", val + 1);
+                svga_recalctimings(svga);
+                break;
+
+            case 0x12e8:
+                if (len == 1)
+                    dev->vtotal = (dev->vtotal & 0x1f00) | val;
+                else {
+                    dev->vtotal = val & 0x1fff;
+                    svga_recalctimings(svga);
+                }
+                break;
+            case 0x12e9:
+                if (len == 1) {
+                    dev->vtotal = (dev->vtotal & 0xff) | ((val & 0x1f) << 8);
+                    mach_log("ATI 8514/A: V_TOTAL write 12E8 = %d\n", dev->vtotal);
+                    svga_recalctimings(svga);
+                }
+                break;
+
+            case 0x16e8:
+                if (len == 1)
+                    dev->vdisp = (dev->vdisp & 0x1f00) | val;
+                else {
+                    dev->vdisp = val & 0x1fff;
+                    svga_recalctimings(svga);
+                }
+                break;
+            case 0x16e9:
+                if (len == 1) {
+                    dev->vdisp = (dev->vdisp & 0xff) | ((val & 0x1f) << 8);
+                    mach_log("ATI 8514/A: V_DISP write 16E8 = %d\n", dev->vdisp);
+                    svga_recalctimings(svga);
+                }
+                break;
+
+            case 0x1ae8:
+                if (len == 1)
+                    dev->vsyncstart = (dev->vsyncstart & 0x1f00) | val;
+                else {
+                    dev->vsyncstart = val & 0x1fff;
+                    svga_recalctimings(svga);
+                }
+                break;
+            case 0x1ae9:
+                if (len == 1) {
+                    dev->vsyncstart = (dev->vsyncstart & 0xff) | ((val & 0x1f) << 8);
+                    mach_log("ATI 8514/A: V_SYNC_STRT write 1AE8 = %d\n", dev->vsyncstart);
+                    svga_recalctimings(svga);
+                }
+                break;
+
+            case 0x1ee8:
+                dev->vsyncwidth = val;
+                mach_log("ATI 8514/A: V_SYNC_WID write 1EE8 = %02x\n", val);
+                svga_recalctimings(svga);
+                break;
+
+            case 0x22e8:
+                dev->disp_cntl = val & 0x7e;
+                dev->interlace = !!(val & 0x10);
+                mach_log("ATI 8514/A: DISP_CNTL write 22E8 = %02x, SCANMODULOS = %d\n", dev->disp_cntl, dev->scanmodulos);
+                svga_recalctimings(svga);
+                break;
+
+            case 0x42e8:
+                if (len == 1) {
+                    dev->subsys_stat &= ~val;
+                } else {
+                    dev->subsys_stat &= ~(val & 0xff);
+                    dev->subsys_cntl = (val >> 8);
+                    mach_log("CNTL = %02x.\n", val >> 8);
+                }
+                break;
+            case 0x42e9:
+                if (len == 1) {
+                    dev->subsys_cntl = val;
+                    mach_log("CNTL = %02x.\n", val);
+                }
+                break;
+
+            case 0x4ae8:
+                mach_log("ATI 8514/A: VGA ON (0x4ae8) = %i, val = %02x\n", vga_on, val);
+                if (!val)
+                    break;
+                if (!dev->local || !dev->ext_crt_pitch)
+                    dev->ext_crt_pitch = 128;
+                dev->accel.advfunc_cntl = val & 7;
+                ibm8514_on = (dev->accel.advfunc_cntl & 1);
+                vga_on = !ibm8514_on;
+                dev->ibm_mode = 1;
+                if (ibm8514_on)
+                    svga->adv_flags |= FLAG_ATI;
+                else
+                    svga->adv_flags &= ~FLAG_ATI;
+                svga_recalctimings(svga);
+                break;
+
+            /*ATI Mach8/32 specific registers*/
+            case 0x6ee:
+                mach_log("6EE write val = %02x, len = %d.\n", val, len);
+                break;
+
+            case 0x6ef:
+                mach_log("6EF write val = %02x, len = %d.\n", val, len);
+                break;
+
+            case 0xaee:
+                if (len == 1)
+                    mach->cursor_offset_lo = (mach->cursor_offset_lo & 0xff00) | val;
+                else {
+                    mach_log("AEE val=%02x.\n", val);
+                    mach->cursor_offset_lo = val;
+                    svga->hwcursor.addr = mach->cursor_offset_lo << 2;
+                }
+                break;
+            case 0xaef:
+                if (len == 1) {
+                    mach->cursor_offset_lo = (mach->cursor_offset_lo & 0x00ff) | (val << 8);
+                    svga->hwcursor.addr = mach->cursor_offset_lo << 2;
+                }
+                break;
+
+            case 0xeee:
+                mach->cursor_offset_hi = val & 0x0f;
+                if (len != 1) {
+                    svga->hwcursor.addr = ((mach->cursor_offset_lo | (mach->cursor_offset_hi << 16))) << 2;
+                    svga->hwcursor.ena = !!(val & 0x8000);
+                }
+                mach_log("EEE val=%08x.\n", svga->hwcursor.addr);
+                break;
+            case 0xeef:
+                if (len == 1) {
+                    svga->hwcursor.addr = ((mach->cursor_offset_lo | (mach->cursor_offset_hi << 16))) << 2;
+                    svga->hwcursor.ena = !!(val & 0x80);
+                }
+                break;
+
+            case 0x12ee:
+                if (len == 1) {
+                    svga->hwcursor.x = (svga->hwcursor.x & 0x700) | val;
+                } else {
+                    svga->hwcursor.x = val & 0x7ff;
+                    mach_log("X = %03x.\n", val);
+                }
+                break;
+            case 0x12ef:
+                if (len == 1) {
+                    svga->hwcursor.x = (svga->hwcursor.x & 0x0ff) | ((val & 0x07) << 8);
+                }
+                break;
+
+            case 0x16ee:
+                if (len == 1) {
+                    svga->hwcursor.y = (svga->hwcursor.y & 0xf00) | val;
+                } else {
+                    svga->hwcursor.y = val & 0xfff;
+                }
+                break;
+            case 0x16ef:
+                if (len == 1) {
+                    svga->hwcursor.y = (svga->hwcursor.y & 0x0ff) | ((val & 0x0f) << 8);
+                }
+                break;
+
+            case 0x1aee:
+                if (len != 1) {
+                    mach->cursor_col_0 = val & 0xff;
+                    mach->cursor_col_1 = (val >> 8) & 0xff;
+                } else
+                    mach->cursor_col_0 = val;
+                break;
+            case 0x1aef:
+                if (len == 1)
+                    mach->cursor_col_1 = val;
+                break;
+
+            case 0x1eee:
+                if (len != 1) {
+                    svga->hwcursor.xoff = val & 0x3f;
+                    svga->hwcursor.yoff = (val >> 8) & 0x3f;
+                } else
+                    svga->hwcursor.xoff = val & 0x3f;
+                break;
+            case 0x1eef:
+                if (len == 1)
+                    svga->hwcursor.yoff = val & 0x3f;
+                break;
+
+            case 0x2aee:
+                mach_log("2AEE write val = %04x\n", val);
+                if (len == 1)
+                    mach->accel.crt_offset_lo = (mach->accel.crt_offset_lo & 0xff00) | val;
+                else
+                    mach->accel.crt_offset_lo = val;
+                break;
+            case 0x2aef:
+                if (len == 1)
+                    mach->accel.crt_offset_lo = (mach->accel.crt_offset_lo & 0x00ff) | (val << 8);
+                break;
+
+            case 0x2eee:
+                mach_log("2EEE write val = %04x\n", val);
+                if (len == 1)
+                    mach->accel.crt_offset_hi = (mach->accel.crt_offset_hi & 0xff00) | val;
+                else
+                    mach->accel.crt_offset_hi = val;
+                break;
+            case 0x2eef:
+                if (len == 1)
+                    mach->accel.crt_offset_hi = (mach->accel.crt_offset_hi & 0x00ff) | (val << 8);
+                break;
+
+            case 0x26ee:
+                mach_log("CRT Pitch = %d, original val = %d.\n", val << 3, val);
+                dev->ext_crt_pitch = val;
+                dev->internal_pitch = val;
+                if (svga->bpp > 8) {
+                    if (svga->bpp == 24)
+                        dev->ext_crt_pitch *= 3;
+                    else
+                        dev->ext_crt_pitch <<= 1;
+                }
+                if (dev->local) {
+                    if (!ibm8514_on) {
+                        ibm8514_on ^= 1;
+                        svga->adv_flags |= FLAG_ATI;
+                    }
+                }
+                svga_recalctimings(svga);
+                break;
+
+            case 0x32ee:
+                if (len == 1) {
+                    mach->local_cntl = (mach->local_cntl & 0xff00) | val;
+                } else {
+                    mach->local_cntl = val;
+                    mach32_updatemapping(mach);
+                }
+                break;
+
+            case 0x32ef:
+                if (len == 1) {
+                    mach->local_cntl = (mach->local_cntl & 0x00ff) | (val << 8);
+                    mach32_updatemapping(mach);
+                }
+                break;
+
+            case 0x36ee:
+                if (len == 1) {
+                    mach->misc = (mach->misc & 0xff00) | (val);
+                } else {
+                    mach->misc = val;
+                }
+                break;
+            case 0x36ef:
+                if (len == 1) {
+                    mach->misc = (mach->misc & 0x00ff) | (val << 8);
+                }
+                break;
+
+            case 0x3aee:
+                if (len == 1) {
+                    mach->ext_cur_col_0_g = val;
+                } else {
+                    mach->ext_cur_col_0_g = val & 0xff;
+                    mach->ext_cur_col_0_r = (val >> 8) & 0xff;
+                }
+                break;
+            case 0x3aef:
+                if (len == 1) {
+                    mach->ext_cur_col_0_r = val;
+                }
+                break;
+
+            case 0x3eee:
+                if (len == 1) {
+                    mach->ext_cur_col_1_g = val;
+                } else {
+                    mach->ext_cur_col_1_g = val & 0xff;
+                    mach->ext_cur_col_1_r = (val >> 8) & 0xff;
+                }
+                break;
+            case 0x3eef:
+                if (len == 1) {
+                    mach->ext_cur_col_1_r = val;
+                }
+                break;
+
+            case 0x42ee:
+                mach->accel.test2[0] = val;
+                break;
+            case 0x42ef:
+                mach->accel.test2[1] = val;
+                break;
+
+            case 0x46ee:
+                mach->accel.test3[0] = val;
+                break;
+            case 0x46ef:
+                mach->accel.test3[1] = val;
+                break;
+
+            case 0x4aee:
+                if (len == 1)
+                    mach->accel.clock_sel = (mach->accel.clock_sel & 0xff00) | val;
+                else {
+                    mach->accel.clock_sel = val;
+                    ibm8514_on = (mach->accel.clock_sel & 1);
+                    vga_on = !ibm8514_on;
+                    dev->ibm_mode = 0;
+                    if (ibm8514_on)
+                        svga->adv_flags |= FLAG_ATI;
+                    else
+                        svga->adv_flags &= ~FLAG_ATI;
+                    mach_log("ATI 8514/A: VGA ON (0x4aee) = %i, val = %04x\n", vga_on, val);
+                    svga_recalctimings(svga);
+                }
+                break;
+            case 0x4aef:
+                if (len == 1) {
+                    mach->accel.clock_sel = (mach->accel.clock_sel & 0x00ff) | (val << 8);
+                    ibm8514_on = (mach->accel.clock_sel & 1);
+                    vga_on = !ibm8514_on;
+                    dev->ibm_mode = 0;
+                    if (ibm8514_on)
+                        svga->adv_flags |= FLAG_ATI;
+                    else
+                        svga->adv_flags &= ~FLAG_ATI;
+                    mach_log("ATI 8514/A: VGA ON (0x4aef) = %i, val = %04x\n", vga_on, mach->accel.clock_sel);
+                    svga_recalctimings(svga);
+                }
+                break;
+
+            case 0x52ee:
+                if (len == 1)
+                    mach->accel.scratch0 = (mach->accel.scratch0 & 0xff00) | val;
+                else
+                    mach->accel.scratch0 = val;
+                break;
+            case 0x52ef:
+                if (len == 1)
+                    mach->accel.scratch0 = (mach->accel.scratch0 & 0x00ff) | (val << 8);
+                break;
+
+            case 0x56ee:
+                if (len == 1)
+                    mach->accel.scratch1 = (mach->accel.scratch1 & 0xff00) | val;
+                else
+                    mach->accel.scratch1 = val;
+                break;
+            case 0x56ef:
+                if (len == 1)
+                    mach->accel.scratch1 = (mach->accel.scratch1 & 0x00ff) | (val << 8);
+                break;
+
+            case 0x5aee:
+                mach_log("Shadow set = %04x\n", val);
+                break;
+            case 0x5aef:
+                mach_log("Shadow + 1 set = %02x\n", val);
+                break;
+
+            case 0x5eee:
+                mach_log("Memory Aperture = %04x, len = %d.\n", val, len);
+                if (len == 1) {
+                    mach->memory_aperture = (mach->memory_aperture & 0xff00) | val;
+                } else {
+                    mach->memory_aperture = val;
+                    if (!mach->pci_bus)
+                        mach->linear_base = (mach->memory_aperture & 0xff00) << 12;
+
+                    mach32_updatemapping(mach);
+                }
+                break;
+
+            case 0x5eef:
+                if (len == 1) {
+                    mach->memory_aperture = (mach->memory_aperture & 0x00ff) | (val << 8);
+                    if (!mach->pci_bus)
+                        mach->linear_base = (mach->memory_aperture & 0xff00) << 12;
+
+                    mach32_updatemapping(mach);
+                }
+                break;
+
+            case 0x62ee:
+                mach_log("62EE write val = %04x, len = %d.\n", val, len);
+                break;
+
+            case 0x66ee:
+                mach_log("66EE write val = %04x, len = %d.\n", val, len);
+                break;
+
+            case 0x6aee:
+                mach_log("6AEE write val = %04x.\n", val & 0x400);
+                if (len == 1)
+                    mach->accel.max_waitstates = (mach->accel.max_waitstates & 0xff00) | val;
+                else {
+                    mach->accel.max_waitstates = val;
+                }
+                break;
+            case 0x6aef:
+                if (len == 1)
+                    mach->accel.max_waitstates = (mach->accel.max_waitstates & 0x00ff) | (val << 8);
+                break;
+
+            case 0x6eee:
+                mach_log("6EEE write val = %04x\n", val);
+                if (len == 1)
+                    mach->accel.ge_offset_lo = (mach->accel.ge_offset_lo & 0xff00) | val;
+                else {
+                    mach->accel.ge_offset_lo = val;
+                    dev->accel.ge_offset = mach->accel.ge_offset_lo;
+                }
+                break;
+            case 0x6eef:
+                if (len == 1) {
+                    mach->accel.ge_offset_lo = (mach->accel.ge_offset_lo & 0x00ff) | (val << 8);
+                    dev->accel.ge_offset = mach->accel.ge_offset_lo;
+                }
+                break;
+
+            case 0x72ee:
+                mach_log("72EE write val = %04x\n", val);
+                if (len == 1)
+                    mach->accel.ge_offset_hi = (mach->accel.ge_offset_hi & 0xff00) | val;
+                else {
+                    mach->accel.ge_offset_hi = val;
+                    dev->accel.ge_offset = mach->accel.ge_offset_lo | (mach->accel.ge_offset_hi << 16);
+                }
+                break;
+            case 0x72ef:
+                if (len == 1) {
+                    mach->accel.ge_offset_hi = (mach->accel.ge_offset_hi & 0x00ff) | (val << 8);
+                    dev->accel.ge_offset = mach->accel.ge_offset_lo | (mach->accel.ge_offset_hi << 16);
+                }
+                break;
+
+            case 0x76ee:
+                mach_log("76EE write val=%d shifted, normal=%d.\n", val << 3, val);
+                dev->ext_pitch = val << 3;
+                svga_recalctimings(svga);
+                break;
+
+            case 0x7aee:
+                mach_log("7AEE write val = %04x, len = %d.\n", val, len);
+                if (len == 1)
+                    mach->accel.ext_ge_config = (mach->accel.ext_ge_config & 0xff00) | val;
+                else {
+                    mach->accel.ext_ge_config = val;
+                    dev->ext_crt_pitch = dev->internal_pitch;
+                    switch (mach->accel.ext_ge_config & 0x30) {
+                        case 0:
+                        case 0x10:
+                            svga->bpp = 8;
+                            break;
+                        case 0x20:
+                            if ((mach->accel.ext_ge_config & 0xc0) == 0x40)
+                                svga->bpp = 16;
+                            else
+                                svga->bpp = 15;
+
+                            dev->ext_crt_pitch <<= 1;
+                            break;
+                        case 0x30:
+                            svga->bpp = 24;
+                            dev->ext_crt_pitch *= 3;
+                            break;
+                    }
+                    if (mach->accel.ext_ge_config & 0x800) {
+                        svga_recalctimings(svga);
+                    }
+                    if (!(mach->accel.ext_ge_config & 0x8000) && !(mach->accel.ext_ge_config & 0x800))
+                        svga_recalctimings(svga);
+                }
+                break;
+            case 0x7aef:
+                mach_log("7AEF write val = %02x.\n", val);
+                if (len == 1) {
+                    mach->accel.ext_ge_config = (mach->accel.ext_ge_config & 0x00ff) | (val << 8);
+                    dev->ext_crt_pitch = dev->internal_pitch;
+                    switch (mach->accel.ext_ge_config & 0x30) {
+                        case 0:
+                        case 0x10:
+                            svga->bpp = 8;
+                            break;
+                        case 0x20:
+                            if ((mach->accel.ext_ge_config & 0xc0) == 0x40)
+                                svga->bpp = 16;
+                            else
+                                svga->bpp = 15;
+
+                            dev->ext_crt_pitch <<= 1;
+                            break;
+                        case 0x30:
+                            svga->bpp = 24;
+                            dev->ext_crt_pitch *= 3;
+                            break;
+                    }
+                    if (mach->accel.ext_ge_config & 0x800) {
+                        svga_recalctimings(svga);
+                    }
+                    if (!(mach->accel.ext_ge_config & 0x8000) && !(mach->accel.ext_ge_config & 0x800))
+                        svga_recalctimings(svga);
+                }
+                break;
+
+            case 0x7eee:
+                mach->accel.eeprom_control = val;
+                break;
+        }
+    }
+}
+
+static uint32_t
+mach_accel_in(uint16_t port, mach_t *mach, int len)
+{
+    svga_t *svga = &mach->svga;
+    ibm8514_t *dev = &svga->dev8514;
+    uint16_t *vram_w = (uint16_t *) svga->vram;
+    uint16_t   temp = 0;
+    int        cmd;
+    int vpos = dev->displine + svga->y_add;
+    int vblankend = svga->vblankstart + svga->crtc[0x16];
+    int frgd_sel, bkgd_sel, mono_src;
+
+    switch (port) {
+        case 0x2e8:
+            if (dev->local) {
+                vpos = svga->displine + svga->y_add;
+                if (vblankend > svga->vtotal) {
+                    vblankend -= svga->vtotal;
+                    if (vpos >= svga->vblankstart || vpos <= vblankend)
+                        temp |= 2;
+                } else {
+                     if (vpos >= svga->vblankstart && vpos <= vblankend)
+                        temp |= 2;
+                }
+            } else {
+                vpos = dev->displine + svga->y_add;
+                if (vblankend > dev->vtotal) {
+                    vblankend -= dev->vtotal;
+                    if (vpos >= svga->vblankstart || vpos <= vblankend)
+                        temp |= 2;
+                } else {
+                     if (vpos >= svga->vblankstart && vpos <= vblankend)
+                        temp |= 2;
+                }
+            }
+            break;
+
+        case 0x6e8:
+            temp = dev->hdisp;
+            break;
+
+        case 0x22e8:
+            temp = dev->disp_cntl;
+            break;
+
+        case 0x26e8:
+            if (len == 1)
+                temp = dev->htotal & 0xff;
+            else
+                temp = dev->htotal;
+            break;
+        case 0x26e9:
+            if (len == 1)
+                temp = dev->htotal >> 8;
+            break;
+
+        case 0x2ee8:
+            temp = dev->subsys_cntl;
+            break;
+
+        case 0x42e8:
+            if (dev->local) {
+                vpos = svga->displine + svga->y_add;
+                if (vblankend > svga->vtotal) {
+                    vblankend -= svga->vtotal;
+                    if (vpos >= svga->vblankstart || vpos <= vblankend)
+                        dev->subsys_stat |= 1;
+                } else {
+                     if (vpos >= svga->vblankstart && vpos <= vblankend)
+                        dev->subsys_stat |= 1;
+                }
+            } else {
+                vpos = dev->displine + svga->y_add;
+                if (vblankend > dev->vtotal) {
+                    vblankend -= dev->vtotal;
+                    if (vpos >= svga->vblankstart || vpos <= vblankend)
+                        dev->subsys_stat |= 1;
+                } else {
+                     if (vpos >= svga->vblankstart && vpos <= vblankend)
+                        dev->subsys_stat |= 1;
+                }
+            }
+
+            if (len != 1) {
+                temp = dev->subsys_stat | 0xa0 | 0x8000;
+            } else {
+                temp = dev->subsys_stat | 0xa0;
+            }
+            break;
+
+        case 0x4ae8:
+            temp = dev->accel.advfunc_cntl;
+            break;
+
+        case 0x42e9:
+            if (len == 1) {
+                temp = dev->subsys_stat >> 8;
+                temp |= 0x80;
+            }
+            break;
+
+        case 0x82e8:
+        case 0xc2e8:
+            if (len != 1) {
+                temp = dev->accel.cur_y;
+            }
+            break;
+
+        case 0x86e8:
+        case 0xc6e8:
+            if (len != 1) {
+                temp = dev->accel.cur_x;
+            }
+            break;
+
+        case 0x92e8:
+            if (len != 1) {
+                temp = dev->test;
+            }
+            break;
+
+        case 0x96e8:
+            if (len != 1) {
+                temp = dev->accel.maj_axis_pcnt;
+            }
+            break;
+
+        case 0x9ae8:
+        case 0xdae8:
+            if (len != 1) {
+                if (dev->force_busy)
+                    temp |= 0x200; /*Hardware busy*/
+                dev->force_busy = 0;
+                if (dev->data_available) {
+                    temp |= 0x100; /*Read Data available*/
+                    if (mach->accel.cmd_type >= 0) {
+                        switch (mach->accel.cmd_type) {
+                            case 2:
+                                if (dev->accel.sy >= mach->accel.height)
+                                    dev->data_available = 0;
+                                break;
+                            case 5:
+                                if (dev->accel.sx >= mach->accel.width)
+                                    dev->data_available = 0;
+                                break;
+                            default:
+                                if (dev->accel.sy < 0)
+                                    dev->data_available = 0;
+                                break;
+                        }
+                    } else {
+                        if (dev->accel.sy < 0)
+                            dev->data_available = 0;
+                    }
+                }
+            }
+            mach_log("[%04X:%08X]: 9AE8: Temp = %04x, len = %d\n\n", CS, cpu_state.pc, temp, len);
+            break;
+        case 0x9ae9:
+        case 0xdae9:
+            if (len == 1) {
+                if (dev->force_busy2)
+                    temp |= 2; /*Hardware busy*/
+                dev->force_busy2 = 0;
+                if (dev->data_available2) {
+                    temp |= 1; /*Read Data available*/
+                    if (mach->accel.cmd_type >= 0) {
+                        switch (mach->accel.cmd_type) {
+                            case 2:
+                                if (dev->accel.sy >= mach->accel.height)
+                                    dev->data_available2 = 0;
+                                break;
+                            case 5:
+                                if (dev->accel.sx >= mach->accel.width)
+                                    dev->data_available2 = 0;
+                                break;
+                            default:
+                                if (dev->accel.sy < 0)
+                                    dev->data_available2 = 0;
+                                break;
+                        }
+                    } else {
+                        if (dev->accel.sy < 0)
+                            dev->data_available2 = 0;
+                    }
+                }
+            }
+            mach_log("[%04X:%08X]: 9AE9: Temp = %04x, len = %d\n\n", CS, cpu_state.pc, temp, len);
+            break;
+
+        case 0xe2e8:
+        case 0xe6e8:
+            if (mach->accel.cmd_type >= 0) {
+                if (mach_pixel_read(mach)) {
+                    cmd = -1;
+                    if (len == 1) {
+                        READ_PIXTRANS_BYTE_IO(dev->accel.dx, 1, dev->local)
+
+                        temp = mach->accel.pix_trans[1];
+                    } else {
+                        if (mach->accel.cmd_type == 3) {
+                            READ_PIXTRANS_WORD(dev->accel.cx, 0, dev->local)
+                        } else {
+                            READ_PIXTRANS_WORD(dev->accel.dx, 0, dev->local)
+                        }
+                        mach_accel_out_pixtrans(mach, dev, port, temp, len);
+                    }
+                }
+            } else {
+                if (ibm8514_cpu_dest(svga)) {
+                    cmd = (dev->accel.cmd >> 13);
+                    if (len == 1) {
+                        ; // READ_PIXTRANS_BYTE_IO(0)
+                    } else {
+                        READ_PIXTRANS_WORD(dev->accel.cx, 0, dev->local)
+                        if (dev->accel.input && !dev->accel.odd_in && !dev->accel.sx) {
+                            temp &= ~0xff00;
+                            if (dev->local)
+                                temp |= (svga->vram[(dev->accel.newdest_in + dev->accel.cur_x) & svga->vram_mask] << 8);
+                            else
+                                temp |= (dev->vram[(dev->accel.newdest_in + dev->accel.cur_x) & dev->vram_mask] << 8);
+                        }
+                        if (dev->subsys_stat & 1) {
+                            dev->force_busy = 1;
+                            dev->data_available = 1;
+                        }
+                    }
+                    ibm8514_accel_out_pixtrans(svga, port, temp, len);
+                }
+            }
+            break;
+        case 0xe2e9:
+        case 0xe6e9:
+            if (mach->accel.cmd_type >= 0) {
+                mach_log("%04x pixtrans read, len=%d.\n", port, len);
+                if (mach_pixel_read(mach)) {
+                    if (len == 1) {
+                        cmd = -1;
+                        READ_PIXTRANS_BYTE_IO(dev->accel.dx, 0, dev->local)
+
+                        temp = mach->accel.pix_trans[0];
+                        frgd_sel = (mach->accel.dp_config >> 13) & 7;
+                        bkgd_sel = (mach->accel.dp_config >> 7) & 3;
+                        mono_src = (mach->accel.dp_config >> 5) & 3;
+
+                        switch (mach->accel.dp_config & 0x200) {
+                            case 0x000: /*8-bit size*/
+                                if (mono_src == 2) {
+                                    if ((frgd_sel != 2) && (bkgd_sel != 2)) {
+                                        mach_accel_start(mach->accel.cmd_type, 1, 8, mach->accel.pix_trans[0] | (mach->accel.pix_trans[1] << 8), 0, mach, dev, len);
+                                    } else
+                                        mach_accel_start(mach->accel.cmd_type, 1, 1, -1, mach->accel.pix_trans[0] | (mach->accel.pix_trans[1] << 8), mach, dev, len);
+                                } else
+                                    mach_accel_start(mach->accel.cmd_type, 1, 1, -1, mach->accel.pix_trans[0] | (mach->accel.pix_trans[1] << 8), mach, dev, len);
+                                break;
+                            case 0x200: /*16-bit size*/
+                                if (mono_src == 2) {
+                                    if ((frgd_sel != 2) && (bkgd_sel != 2)) {
+                                        if (mach->accel.dp_config & 0x1000)
+                                            mach_accel_start(mach->accel.cmd_type, 1, 16, mach->accel.pix_trans[1] | (mach->accel.pix_trans[0] << 8), 0, mach, dev, len);
+                                        else
+                                            mach_accel_start(mach->accel.cmd_type, 1, 16, mach->accel.pix_trans[0] | (mach->accel.pix_trans[1] << 8), 0, mach, dev, len);
+                                    } else
+                                        mach_accel_start(mach->accel.cmd_type, 1, 2, -1, mach->accel.pix_trans[0] | (mach->accel.pix_trans[1] << 8), mach, dev, len);
+                                } else
+                                    mach_accel_start(mach->accel.cmd_type, 1, 2, -1, mach->accel.pix_trans[0] | (mach->accel.pix_trans[1] << 8), mach, dev, len);
+                                break;
+                        }
+                    }
+                }
+            }
+            break;
+
+        case 0xbee8:
+        case 0xfee8:
+            if (len != 1) {
+                mach_log("Multifunc_cntl = %d.\n", dev->accel.multifunc_cntl >> 12);
+                switch ((dev->accel.multifunc_cntl >> 12) & 0x0f) {
+                    case 0:
+                        temp = dev->accel.multifunc[0];
+                        break;
+                    case 1:
+                        temp = dev->accel.clip_top;
+                        break;
+                    case 2:
+                        temp = dev->accel.clip_left;
+                        break;
+                    case 3:
+                        temp = dev->accel.multifunc[3];
+                        break;
+                    case 4:
+                        temp = dev->accel.multifunc[4];
+                        break;
+                    case 5:
+                        temp = dev->accel.multifunc[5];
+                        break;
+                    case 8:
+                        temp = dev->accel.multifunc[8];
+                        break;
+                    case 9:
+                        temp = dev->accel.multifunc[9];
+                        break;
+                    case 0x0a:
+                        temp = dev->accel.multifunc[0x0a];
+                        break;
+                }
+            }
+            break;
+
+/*ATI Mach8/32 specific registers*/
+        case 0x12ee:
+            if (len == 1)
+                temp = mach->config1 & 0xff;
+            else
+                temp = mach->config1;
+            break;
+        case 0x12ef:
+            if (len == 1)
+                temp = mach->config1 >> 8;
+            break;
+
+        case 0x16ee:
+            if (len == 1)
+                temp = mach->config2 & 0xff;
+            else
+                temp = mach->config2;
+            break;
+        case 0x16ef:
+            if (len == 1)
+                temp = mach->config2 >> 8;
+            break;
+
+        case 0x32ee:
+            if (len == 1)
+                temp = mach->local_cntl & 0xff;
+            else
+                temp = mach->local_cntl;
+            break;
+        case 0x32ef:
+            if (len == 1)
+                temp = mach->local_cntl >> 8;
+            break;
+
+        case 0x36ee:
+            if (len == 1)
+                temp = mach->misc & 0xff;
+            else
+                temp = mach->misc;
+            break;
+        case 0x36ef:
+            if (len == 1)
+                temp = mach->misc >> 8;
+            break;
+
+        case 0x42ee:
+            temp = mach->accel.test2[0];
+            break;
+        case 0x42ef:
+            temp = mach->accel.test2[1];
+            break;
+
+        case 0x46ee:
+            temp = mach->accel.test3[0];
+            break;
+        case 0x46ef:
+            temp = mach->accel.test3[1];
+            break;
+
+        case 0x4aee:
+            if (len == 1)
+                temp = mach->accel.clock_sel & 0xff;
+            else
+                temp = mach->accel.clock_sel;
+            break;
+        case 0x4aef:
+            if (len == 1)
+                temp = mach->accel.clock_sel >> 8;
+            break;
+
+        case 0x52ee:
+            if (len == 1)
+                temp = mach->accel.scratch0 & 0xff;
+            else
+                temp = mach->accel.scratch0;
+            break;
+        case 0x52ef:
+            if (len == 1)
+                temp = mach->accel.scratch0 >> 8;
+            break;
+
+        case 0x56ee:
+            if (len == 1)
+                temp = mach->accel.scratch1 & 0xff;
+            else
+                temp = mach->accel.scratch1;
+            break;
+        case 0x56ef:
+            if (len == 1)
+                temp = mach->accel.scratch1 >> 8;
+            break;
+
+        case 0x5eee:
+            if (mach->pci_bus)
+                mach->memory_aperture = (mach->memory_aperture & ~0xfff0) | ((mach->linear_base >> 20) << 4);
+
+            if (len == 1)
+                temp = mach->memory_aperture & 0xff;
+            else
+                temp = mach->memory_aperture;
+            break;
+        case 0x5eef:
+            if (len == 1)
+                temp = mach->memory_aperture >> 8;
+            break;
+
+        case 0x62ee:
+            temp = mach->accel.clip_overrun;
+            if (len != 1) {
+                if (mach->force_busy)
+                    temp |= 0x2000;
+                mach->force_busy = 0;
+                if (ati_eeprom_read(&mach->eeprom))
+                    temp |= 0x4000;
+            }
+            mach_log("[%04X:%08X]: 62EE: Temp = %04x, len = %d\n\n", CS, cpu_state.pc, temp, len);
+            break;
+        case 0x62ef:
+            if (len == 1) {
+                if (mach->force_busy2)
+                    temp |= 0x20;
+                mach->force_busy2 = 0;
+                if (ati_eeprom_read(&mach->eeprom))
+                    temp |= 0x40;
+            }
+            mach_log("[%04X:%08X]: 62EF: Temp = %04x, len = %d\n\n", CS, cpu_state.pc, temp, len);
+            break;
+
+        case 0x6aee:
+            if (len == 1)
+                temp = mach->accel.max_waitstates & 0xff;
+            else
+                temp = mach->accel.max_waitstates;
+            break;
+        case 0x6aef:
+            if (len == 1)
+                temp = mach->accel.max_waitstates >> 8;
+            break;
+
+        case 0x72ee:
+            if (len == 1)
+                temp = dev->accel.clip_left & 0xff;
+            else
+                temp = dev->accel.clip_left;
+            break;
+        case 0x72ef:
+            if (len == 1)
+                temp = dev->accel.clip_left >> 8;
+            break;
+
+        case 0x76ee:
+            if (len == 1)
+                temp = dev->accel.clip_top & 0xff;
+            else
+                temp = dev->accel.clip_top;
+            break;
+        case 0x76ef:
+            if (len == 1)
+                temp = dev->accel.clip_top >> 8;
+            break;
+
+        case 0x7aee:
+            if (len == 1)
+                temp = dev->accel.multifunc[4] & 0xff;
+            else
+                temp = dev->accel.multifunc[4];
+            break;
+        case 0x7aef:
+            if (len == 1)
+                temp = dev->accel.multifunc[4] >> 8;
+            break;
+
+        case 0x7eee:
+            if (len == 1)
+                temp = dev->accel.multifunc[3] & 0xff;
+            else
+                temp = dev->accel.multifunc[3];
+            break;
+        case 0x7eef:
+            if (len == 1)
+                temp = dev->accel.multifunc[3] >> 8;
+            break;
+
+        case 0x82ee:
+            temp = mach->accel.patt_data_idx;
+            break;
+
+        case 0x8eee:
+            if (len == 1)
+                temp = mach->accel.ext_ge_config & 0xff;
+            else
+                temp = mach->accel.ext_ge_config;
+            break;
+        case 0x8eef:
+            if (len == 1)
+                temp = mach->accel.ext_ge_config >> 8;
+            break;
+
+        case 0x92ee:
+            temp = mach->accel.eeprom_control;
+            break;
+
+        case 0x96ee:
+            if (len == 1) {
+                temp = dev->accel.maj_axis_pcnt & 0xff;
+            } else {
+                temp = dev->accel.maj_axis_pcnt;
+                if ((mach->accel.test == 0x1555) || (mach->accel.test == 0x0aaa))
+                    temp = mach->accel.test;
+            }
+            break;
+        case 0x96ef:
+            if (len == 1)
+                temp = dev->accel.maj_axis_pcnt >> 8;
+            break;
+
+        case 0xa2ee:
+            if (len == 1)
+                temp = mach->accel.linedraw_opt & 0xff;
+            else {
+                temp = mach->accel.linedraw_opt;
+            }
+            break;
+        case 0xa2ef:
+            if (len == 1)
+                temp = mach->accel.linedraw_opt >> 8;
+            break;
+
+        case 0xb2ee:
+            if (len == 1)
+                temp = dev->hdisp;
+            else {
+                temp = dev->hdisp & 0xff;
+                temp |= (dev->htotal << 8);
+                mach_log("HDISP read=%d, HTOTAL read=%d.\n", temp & 0xff, temp >> 8);
+            }
+            break;
+        case 0xb2ef:
+            if (len == 1) {
+                temp = dev->htotal;
+            }
+            break;
+
+        case 0xc2ee:
+            if (len == 1)
+                temp = dev->vtotal & 0xff;
+            else {
+                temp = dev->vtotal;
+                mach_log("VTOTAL read=%d.\n", temp);
+            }
+            break;
+        case 0xc2ef:
+            if (len == 1)
+                temp = dev->vtotal >> 8;
+            break;
+
+        case 0xc6ee:
+            if (len == 1)
+                temp = dev->vdisp & 0xff;
+            else {
+                temp = dev->vdisp;
+                mach_log("VDISP read=%d.\n", temp);
+            }
+            break;
+        case 0xc6ef:
+            if (len == 1)
+                temp = dev->vdisp >> 8;
+            break;
+
+        case 0xcaee:
+            if (len == 1)
+                temp = dev->vsyncstart & 0xff;
+            else
+                temp = dev->vsyncstart;
+            break;
+        case 0xcaef:
+            if (len == 1)
+                temp = dev->vsyncstart >> 8;
+            break;
+
+        case 0xceee:
+            if (len == 1)
+                temp = svga->vc & 0xff;
+            else
+                temp = svga->vc & 0x7ff;
+            break;
+        case 0xceef:
+            if (len == 1)
+                temp = (svga->vc >> 8) & 7;
+            break;
+
+        case 0xdaee:
+            if (len != 1) {
+                temp = mach->accel.src_x;
+                if (dev->local) {
+                    temp &= 0x7ff;
+                }
+            } else
+                temp = dev->accel.destx_distp & 0xff;
+            break;
+        case 0xdaef:
+            if (len == 1)
+                temp = dev->accel.destx_distp >> 8;
+            break;
+
+        case 0xdeee:
+            if (len != 1) {
+                temp = mach->accel.src_y;
+                if (dev->local)
+                    temp &= 0x7ff;
+            } else
+                temp = dev->accel.desty_axstp & 0xff;
+            break;
+        case 0xdeef:
+            if (len == 1)
+                temp = dev->accel.desty_axstp >> 8;
+            break;
+
+        case 0xfaee:
+            if (len != 1) {
+                if (mach->pci_bus)
+                    temp = 0x0017;
+                else
+                    temp = 0x22f7;
+            } else {
+                if (mach->pci_bus)
+                    temp = 0x17;
+                else
+                    temp = 0xf7;
+            }
+            break;
+        case 0xfaef:
+            if (len == 1) {
+                if (mach->pci_bus)
+                    temp = 0x00;
+                else
+                    temp = 0x22;
+            }
+            break;
+    }
+    if (port != 0x9ae8 && port != 0x9ae9 && port != 0x62ee && port != 0x9aee) {
+        mach_log("Port accel in = %04x, temp = %04x, len = %d, mode = %d.\n", port, temp, len, dev->ibm_mode);
+    }
+    return temp;
+}
+
+static void
+mach_accel_outb(uint16_t port, uint8_t val, void *p)
+{
+    mach_t *mach = (mach_t *) p;
+    mach_accel_out(port, val, mach, 1);
+}
+
+static void
+mach_accel_outw(uint16_t port, uint16_t val, void *p)
+{
+    mach_t *mach = (mach_t *) p;
+    mach_accel_out(port, val, mach, 2);
+}
+
+static uint8_t
+mach_accel_inb(uint16_t port, void *p)
+{
+    mach_t *mach = (mach_t *) p;
+    return mach_accel_in(port, mach, 1);
+}
+
+static uint16_t
+mach_accel_inw(uint16_t port, void *p)
+{
+    mach_t *mach = (mach_t *) p;
+    return mach_accel_in(port, mach, 2);
+}
+
+static void
+mach32_ap_writeb(uint32_t addr, uint8_t val, void *p)
+{
+    mach_t *mach = (mach_t *) p;
+    uint8_t port_dword = addr & 0xfc;
+
+    if (((addr >= ((mach->ap_size << 20) - 0x200)) && (addr < (mach->ap_size << 20)))) {
+        if (addr & 0x100) {
+            mach_log("Port WORDB Write=%04x.\n", 0x02ee + (port_dword << 8));
+            mach_accel_outb(0x02ee + (addr & 1) + (port_dword << 8), val, mach);
+        } else {
+            mach_log("Port WORDB Write=%04x.\n", 0x02e8 + (port_dword << 8));
+            mach_accel_outb(0x02e8 + (addr & 1) + (port_dword << 8), val, mach);
+        }
+    } else {
+        mach_log("Linear WORDB Write=%08x.\n", addr);
+        svga_write_linear(addr, val, &mach->svga);
+    }
+}
+
+static void
+mach32_ap_writew(uint32_t addr, uint16_t val, void *p)
+{
+    mach_t *mach = (mach_t *) p;
+    uint8_t port_dword = addr & 0xfc;
+
+    if (((addr >= ((mach->ap_size << 20) - 0x200)) && (addr < (mach->ap_size << 20)))) {
+        if (addr & 0x100) {
+            mach_log("Port WORDW Write=%04x.\n", 0x02ee + (port_dword << 8));
+            mach_accel_outw(0x02ee + (port_dword << 8), val, mach);
+        } else {
+            mach_log("Port WORDW Write=%04x.\n", 0x02e8 + (port_dword << 8));
+            mach_accel_outw(0x02e8 + (port_dword << 8), val, mach);
+        }
+    } else {
+        mach_log("Linear WORDW Write=%08x.\n", addr);
+        svga_writew_linear(addr, val, &mach->svga);
+    }
+}
+
+static void
+mach32_ap_writel(uint32_t addr, uint32_t val, void *p)
+{
+    mach_t *mach = (mach_t *) p;
+    uint8_t port_dword = addr & 0xfc;
+
+    if (((addr >= ((mach->ap_size << 20) - 0x200)) && (addr < (mach->ap_size << 20)))) {
+        if (addr & 0x100) {
+            mach_log("Port WORDL Write=%04x.\n", 0x02ee + (port_dword << 8));
+            mach_accel_outw(0x02ee + (port_dword << 8), val & 0xffff, mach);
+            mach_accel_outw(0x02ee + (port_dword << 8) + 4, val >> 16, mach);
+        } else {
+            mach_log("Port WORDL Write=%04x.\n", 0x02e8 + (port_dword << 8));
+            mach_accel_outw(0x02e8 + (port_dword << 8), val & 0xffff, mach);
+            mach_accel_outw(0x02e8 + (port_dword << 8) + 4, val >> 16, mach);
+        }
+    } else {
+        mach_log("Linear WORDL Write=%08x, val=%08x, mode=%d, rop=%02x.\n", addr, val, mach->svga.writemode, mach->svga.gdcreg[3] & 0x18);
+        svga_writel_linear(addr, val, &mach->svga);
+    }
+}
+
+static uint8_t
+mach32_ap_readb(uint32_t addr, void *p)
+{
+    mach_t *mach = (mach_t *) p;
+    uint8_t temp;
+    uint8_t port_dword = addr & 0xfc;
+
+    if (((addr >= ((mach->ap_size << 20) - 0x200)) && (addr < (mach->ap_size << 20)))) {
+        if (addr & 0x100) {
+            temp = mach_accel_inb(0x02ee + (addr & 1) + (port_dword << 8), mach);
+        } else {
+            temp = mach_accel_inb(0x02e8 + (addr & 1) + (port_dword << 8), mach);
+        }
+    } else
+        temp = svga_read_linear(addr, &mach->svga);
+
+    return temp;
+}
+
+static uint16_t
+mach32_ap_readw(uint32_t addr, void *p)
+{
+    mach_t *mach = (mach_t *) p;
+    uint16_t temp;
+    uint8_t port_dword = addr & 0xfc;
+
+    if (((addr >= ((mach->ap_size << 20) - 0x200)) && (addr < (mach->ap_size << 20)))) {
+        if (addr & 0x100) {
+            temp = mach_accel_inw(0x02ee + (port_dword << 8), mach);
+        } else {
+            temp = mach_accel_inw(0x02e8 + (port_dword << 8), mach);
+        }
+    } else
+        temp = svga_readw_linear(addr, &mach->svga);
+
+    return temp;
+}
+
+static uint32_t
+mach32_ap_readl(uint32_t addr, void *p)
+{
+    mach_t *mach = (mach_t *) p;
+    uint32_t temp;
+    uint8_t port_dword = addr & 0xfc;
+
+    if (((addr >= ((mach->ap_size << 20) - 0x200)) && (addr < (mach->ap_size << 20)))) {
+        if (addr & 0x100) {
+            temp = mach_accel_inw(0x02ee + (port_dword << 8), mach);
+            temp |= (mach_accel_inw(0x02ee + (port_dword << 8) + 4, mach) << 8);
+        } else {
+            temp = mach_accel_inw(0x02e8 + (port_dword << 8), mach);
+            temp |= (mach_accel_inw(0x02e8 + (port_dword << 8) + 4, mach) << 8);
+        }
+    } else
+        temp = svga_readl_linear(addr, &mach->svga);
+
+    return temp;
+}
+
+static void
+mach32_updatemapping(mach_t *mach)
+{
+    svga_t *svga = &mach->svga;
+
+    if ((mach->pci_bus && (!(mach->pci_regs[PCI_REG_COMMAND] & PCI_COMMAND_MEM)))) {
+        mem_mapping_disable(&svga->mapping);
+        mem_mapping_disable(&mach->mmio_linear_mapping);
+        return;
+    }
+
+    if (mach->regs[0xbd] & 4) {
+        mem_mapping_set_addr(&svga->mapping, 0xa0000, 0x20000);
+        svga->banked_mask = 0xffff;
+        mach_log("Bit 2 of BD.\n");
+    } else {
+        switch (svga->gdcreg[6] & 0x0c) {
+            case 0x0: /*128k at A0000*/
+                mem_mapping_set_addr(&svga->mapping, 0xa0000, 0x20000);
+                svga->banked_mask = 0xffff;
+                break;
+            case 0x4: /*64k at A0000*/
+                mem_mapping_set_addr(&svga->mapping, 0xa0000, 0x10000);
+                svga->banked_mask = 0xffff;
+                break;
+            case 0x8: /*32k at B0000*/
+                mem_mapping_set_addr(&svga->mapping, 0xb0000, 0x08000);
+                svga->banked_mask = 0x7fff;
+                break;
+            case 0xC: /*32k at B8000*/
+                mem_mapping_set_addr(&svga->mapping, 0xb8000, 0x08000);
+                svga->banked_mask = 0x7fff;
+                break;
+        }
+    }
+
+    mach_log("Linear base = %08x, aperture = %04x, localcntl = %02x svgagdc = %x.\n", mach->linear_base, mach->memory_aperture, mach->local_cntl, svga->gdcreg[6] & 0x0c);
+    if (mach->linear_base) {
+        if (((mach->memory_aperture & 3) == 1) && !mach->pci_bus) {
+            /*1 MB aperture*/
+            mach->ap_size = 1;
+            mem_mapping_set_addr(&mach->mmio_linear_mapping, mach->linear_base, mach->ap_size << 20);
+        } else {
+            /*4 MB aperture*/
+            mach->ap_size = 4;
+            mem_mapping_set_addr(&mach->mmio_linear_mapping, mach->linear_base, mach->ap_size << 20);
+        }
+        /*Force IBM/ATI mode on when the MMIO registers are loaded.*/
+        if (mach->local_cntl & 0x20) {
+            if (!ibm8514_on) {
+                ibm8514_on ^= 1;
+                svga->adv_flags |= FLAG_ATI;
+                svga_recalctimings(svga);
+            }
+        }
+    } else {
+        mach->ap_size = 4;
+        mem_mapping_disable(&mach->mmio_linear_mapping);
+    }
+}
+
+static void
+mach32_hwcursor_draw(svga_t *svga, int displine)
+{
+    mach_t   *mach = (mach_t *) svga->p;
+    uint16_t dat;
+    int      comb;
+    int      offset = svga->hwcursor_latch.x - svga->hwcursor_latch.xoff;
+    uint32_t color0, color1;
+
+    if (svga->bpp == 8) {
+        color0 = svga->pallook[mach->cursor_col_0];
+        color1 = svga->pallook[mach->cursor_col_1];
+        mach_log("8BPP: Offset = %x, XOFF = %02x, YOFF = %02x.\n", offset, svga->hwcursor_latch.xoff, svga->hwcursor_latch.yoff);
+    } else if (svga->bpp == 15) {
+        color0 = video_15to32[((mach->ext_cur_col_0_r << 16) | (mach->ext_cur_col_0_g << 8) | mach->cursor_col_0) & 0xffff];
+        color1 = video_15to32[((mach->ext_cur_col_1_r << 16) | (mach->ext_cur_col_1_g << 8) | mach->cursor_col_1) & 0xffff];
+        mach_log("15BPP: Offset = %x, XOFF = %02x, YOFF = %02x.\n", offset, svga->hwcursor_latch.xoff, svga->hwcursor_latch.yoff);
+    } else if (svga->bpp == 16) {
+        color0 = video_16to32[((mach->ext_cur_col_0_r << 16) | (mach->ext_cur_col_0_g << 8) | mach->cursor_col_0) & 0xffff];
+        color1 = video_16to32[((mach->ext_cur_col_1_r << 16) | (mach->ext_cur_col_1_g << 8) | mach->cursor_col_1) & 0xffff];
+        mach_log("16BPP: Offset = %x, XOFF = %02x, YOFF = %02x.\n", offset, svga->hwcursor_latch.xoff, svga->hwcursor_latch.yoff);
+    } else {
+        color0 = ((mach->ext_cur_col_0_r << 16) | (mach->ext_cur_col_0_g << 8) | mach->cursor_col_0);
+        color1 = ((mach->ext_cur_col_1_r << 16) | (mach->ext_cur_col_1_g << 8) | mach->cursor_col_1);
+        mach_log("24BPP: Offset = %x, XOFF = %02x, YOFF = %02x.\n", offset, svga->hwcursor_latch.xoff, svga->hwcursor_latch.yoff);
+    }
+
+    if (svga->interlace && svga->hwcursor_oddeven)
+        svga->hwcursor_latch.addr += 16;
+
+    for (int x = 0; x < 64; x += 8) {
+        dat = svga->vram[svga->hwcursor_latch.addr & svga->vram_mask] | (svga->vram[(svga->hwcursor_latch.addr + 1) & svga->vram_mask] << 8);
+        for (int xx = 0; xx < 8; xx++) {
+            comb = (dat >> (xx << 1)) & 0x03;
+            if (offset >= svga->hwcursor_latch.x) {
+                switch (comb) {
+                    case 0:
+                        ((uint32_t *) svga->monitor->target_buffer->line[displine])[offset + svga->x_add] = color0;
+                        break;
+                    case 1:
+                        ((uint32_t *) svga->monitor->target_buffer->line[displine])[offset + svga->x_add] = color1;
+                        break;
+                    case 3:
+                        ((uint32_t *) svga->monitor->target_buffer->line[displine])[offset + svga->x_add] ^= 0xffffff;
+                        break;
+                }
+            }
+            offset++;
+        }
+        svga->hwcursor_latch.addr += 2;
+    }
+    if (svga->interlace && !svga->hwcursor_oddeven)
+        svga->hwcursor_latch.addr += 16;
+}
+
+static void
+mach_io_remove(mach_t *mach)
+{
+    io_removehandler(0x01ce, 2,
+                mach_in, NULL, NULL,
+                mach_out, NULL, NULL, mach);
+    io_removehandler(0x03c0, 32,
+                mach_in, NULL, NULL,
+                mach_out, NULL, NULL, mach);
+
+    io_removehandler(0x2e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x2ea, 0x0004, mach_in, NULL, NULL, mach_out, NULL, NULL, mach);
+    io_removehandler(0x6e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xae8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xee8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x12e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x16e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x1ae8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x1ee8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x22e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x26e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x2ee8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x42e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x4ae8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x52e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x56e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x5ae8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x5ee8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x82e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x86e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x8ae8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x8ee8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x92e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x96e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x9ae8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x9ee8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xa2e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xa6e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xaae8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xaee8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xb2e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xb6e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xbae8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xbee8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xe2e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+
+    io_removehandler(0xc2e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xc6e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xcae8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xcee8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xd2e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xd6e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xdae8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xdee8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xe6e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xeae8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xeee8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xf2e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xf6e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xfae8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xfee8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+
+    io_removehandler(0x06ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x0aee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x0eee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x12ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x16ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x1aee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x1eee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x26ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x2aee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x2eee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x32ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x36ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x3aee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x3eee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x42ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x46ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x4aee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x52ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x56ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x5aee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x5eee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x62ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x66ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x6aee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x6eee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x72ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x76ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x7aee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x7eee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x82ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x8eee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x92ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x96ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0x9aee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xa2ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xa6ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xaaee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xaeee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xb2ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xb6ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xbaee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xbeee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xc2ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xc6ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xcaee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xceee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xd2ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xd6ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xdaee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xdeee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xe2ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xe6ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xeeee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xf2ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xfaee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_removehandler(0xfeee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+}
+
+static void
+mach_io_set(mach_t *mach)
+{
+    io_sethandler(0x01ce, 2,
+                mach_in, NULL, NULL,
+                mach_out, NULL, NULL, mach);
+    io_sethandler(0x03c0, 32,
+                mach_in, NULL, NULL,
+                mach_out, NULL, NULL, mach);
+
+    io_sethandler(0x2e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x2ea, 0x0004, mach_in, NULL, NULL, mach_out, NULL, NULL, mach);
+    io_sethandler(0x6e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xae8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xee8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x12e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x16e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x1ae8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x1ee8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x22e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x26e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x2ee8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x42e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x4ae8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x52e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x56e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x5ae8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x5ee8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x82e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x86e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x8ae8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x8ee8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x92e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x96e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x9ae8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x9ee8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xa2e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xa6e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xaae8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xaee8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xb2e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xb6e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xbae8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xbee8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xe2e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+
+    io_sethandler(0xc2e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xc6e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xcae8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xcee8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xd2e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xd6e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xdae8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xdee8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xe6e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xeae8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xeee8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xf2e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xf6e8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xfae8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xfee8, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+
+    io_sethandler(0x06ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x0aee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x0eee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x12ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x16ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x1aee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x1eee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x26ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x2aee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x2eee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x32ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x36ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x3aee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x3eee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x42ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x46ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x4aee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x52ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x56ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x5aee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x5eee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x62ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x66ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x6aee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x6eee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x72ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x76ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x7aee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x7eee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x82ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x8eee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x92ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x96ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0x9aee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xa2ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xa6ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xaaee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xaeee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xb2ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xb6ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xbaee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xbeee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xc2ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xc6ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xcaee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xceee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xd2ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xd6ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xdaee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xdeee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xe2ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xe6ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xeeee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xf2ee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xfaee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+    io_sethandler(0xfeee, 0x0002, mach_accel_inb, mach_accel_inw, NULL, mach_accel_outb, mach_accel_outw, NULL, mach);
+}
+
+static uint8_t
+mach32_pci_read(int func, int addr, void *p)
+{
+    mach_t *mach = (mach_t *) p;
+    uint8_t   ret    = 0x00;
+
+    switch (addr) {
+        case 0x00:
+            ret = 0x02; /*ATI*/
+            break;
+        case 0x01:
+            ret = 0x10;
+            break;
+
+        case 0x02:
+            ret = 0x58;
+            break;
+        case 0x03:
+            ret = 0x41;
+            break;
+
+        case PCI_REG_COMMAND:
+            ret = mach->pci_regs[PCI_REG_COMMAND] | 0x80; /*Respond to IO and memory accesses*/
+            break;
+
+        case 0x07:
+            ret = 0x01; /*Medium DEVSEL timing*/
+            break;
+
+        case 0x0a:
+            ret = 0x00; /*Supports VGA interface*/
+            break;
+        case 0x0b:
+            ret = 0x03;
+            break;
+
+        case 0x10:
+            ret = 0x00; /*Linear frame buffer address*/
+            break;
+        case 0x11:
+            ret = 0x00;
+            break;
+        case 0x12:
+            ret = mach->linear_base >> 16;
+            break;
+        case 0x13:
+            ret = mach->linear_base >> 24;
+            break;
+
+        case 0x30:
+            ret = (mach->pci_regs[0x30] & 0x01); /*BIOS ROM address*/
+            break;
+        case 0x31:
+            ret = 0x00;
+            break;
+        case 0x32:
+            ret = mach->pci_regs[0x32];
+            break;
+        case 0x33:
+            ret = mach->pci_regs[0x33];
+            break;
+
+        case 0x3c:
+            ret = mach->int_line;
+            break;
+        case 0x3d:
+            ret = PCI_INTA;
+            break;
+    }
+
+    return ret;
+}
+
+static void
+mach32_pci_write(int func, int addr, uint8_t val, void *p)
+{
+    mach_t *mach = (mach_t *) p;
+
+    switch (addr) {
+        case PCI_REG_COMMAND:
+            mach->pci_regs[PCI_REG_COMMAND] = val & 0x27;
+            mach_io_remove(mach);
+            if (val & PCI_COMMAND_IO) {
+                mach_io_set(mach);
+            }
+            mach32_updatemapping(mach);
+            break;
+
+        case 0x12:
+            mach->linear_base = (mach->linear_base & 0xff000000) | ((val & 0xc0) << 16);
+            mach32_updatemapping(mach);
+            break;
+        case 0x13:
+            mach->linear_base = (mach->linear_base & 0xc00000) | (val << 24);
+            mach32_updatemapping(mach);
+            break;
+
+        case 0x30:
+        case 0x32:
+        case 0x33:
+            mach->pci_regs[addr] = val;
+            if (mach->pci_regs[0x30] & 0x01) {
+                uint32_t bios_addr = (mach->pci_regs[0x32] << 16) | (mach->pci_regs[0x33] << 24);
+                mach_log("Mach32 bios_rom enabled at %08x\n", bios_addr);
+                mem_mapping_set_addr(&mach->bios_rom.mapping, bios_addr, 0x8000);
+            } else {
+                mach_log("Mach32 bios_rom disabled\n");
+                mem_mapping_disable(&mach->bios_rom.mapping);
+            }
+            return;
+
+        case 0x3c:
+            mach->int_line = val;
+            break;
+    }
+}
+
+static void *
+mach8_init(const device_t *info)
+{
+    mach_t *mach;
+    svga_t *svga;
+    ibm8514_t *dev;
+    uint32_t memory;
+
+    mach = malloc(sizeof(mach_t));
+    memset(mach, 0x00, sizeof(mach_t));
+
+    svga = &mach->svga;
+    dev = &svga->dev8514;
+
+    mach->pci_bus    = !!(info->flags & DEVICE_PCI);
+    mach->vlb_bus    = !!(info->flags & DEVICE_VLB);
+    dev->local       = info->local;
+    dev->vram_size   = (1024 << 10);
+    dev->vram        = calloc(dev->vram_size, 1);
+    dev->changedvram = calloc(dev->vram_size >> 12, 1);
+    dev->vram_mask   = dev->vram_size - 1;
+    dev->map8        = svga->pallook;
+    memory           = device_get_config_int("memory");
+
+    if (dev->local) {
+        if (mach->vlb_bus)
+            rom_init(&mach->bios_rom,
+                     BIOS_MACH32_VLB_ROM_PATH,
+                     0xc0000, 0x8000, 0x7fff,
+                     0, MEM_MAPPING_EXTERNAL);
+        else if (mach->pci_bus)
+            rom_init(&mach->bios_rom,
+                     BIOS_MACH32_PCI_ROM_PATH,
+                     0xc0000, 0x8000, 0x7fff,
+                     0, MEM_MAPPING_EXTERNAL);
+        else
+            rom_init(&mach->bios_rom,
+                     BIOS_MACH32_ISA_ROM_PATH,
+                     0xc0000, 0x8000, 0x7fff,
+                     0, MEM_MAPPING_EXTERNAL);
+    } else {
+        rom_init(&mach->bios_rom,
+                 BIOS_MACH8_ROM_PATH,
+                 0xc0000, 0x8000, 0x7fff,
+                 0, MEM_MAPPING_EXTERNAL);
+    }
+
+    svga_init(info, svga, mach, dev->local ? (memory << 10) : (512 << 10), /*default: 512kB for Mach8, 2MB for Mach32*/
+                  mach_recalctimings,
+                  mach_in, mach_out,
+                  dev->local ? mach32_hwcursor_draw : NULL,
+                  NULL);
+
+    if (dev->local) {
+        switch (memory) {
+            case 1024:
+                mach->misc |= 0x04;
+                break;
+            case 2048:
+                mach->misc |= 0x08;
+                break;
+            case 4096:
+                mach->misc |= 0x0c;
+                break;
+        }
+        svga->hwcursor.cur_ysize = 64;
+        mach->config1 = 0x20;
+        mach->config2 = 0x08;
+        /*Fake the RAMDAC to give the VLB/MCA variants full 24-bit support until said RAMDAC is implemented.*/
+        if (mach->vlb_bus) {
+            video_inform(VIDEO_FLAG_TYPE_8514, &timing_mach32_vlb);
+            mach->config1 |= 0x0c;
+            mach->config1 |= 0x0400;
+        } else if (mach->pci_bus) {
+            video_inform(VIDEO_FLAG_TYPE_8514, &timing_mach32_pci);
+            mach->config1 |= 0x0e;
+            mach->config1 |= 0x0a00;
+            mach->config2 |= 0x2000;
+            svga->ramdac = device_add(&ati68860_ramdac_device);
+        } else {
+            video_inform(VIDEO_FLAG_TYPE_8514, &timing_gfxultra_isa);
+        }
+        mem_mapping_add(&mach->mmio_linear_mapping, 0, 0, mach32_ap_readb, mach32_ap_readw, mach32_ap_readl, mach32_ap_writeb, mach32_ap_writew, mach32_ap_writel, NULL, MEM_MAPPING_EXTERNAL, mach);
+        mem_mapping_disable(&mach->mmio_linear_mapping);
+    } else {
+        video_inform(VIDEO_FLAG_TYPE_8514, &timing_gfxultra_isa);
+        mach->config1 = 0x02 | 0x20 | 0x80;
+        mach->config2 = 0x02;
+        dev->ext_pitch = 1024;
+    }
+
+    svga->force_old_addr = 1;
+    svga->miscout = 1;
+    svga->bpp = 8;
+    svga->packed_chain4 = 1;
+    ibm8514_enabled = 1;
+    ibm8514_has_vga = 1;
+    dev->ibm_mode = 1;
+    dev->rowoffset = 128;
+    mach_io_set(mach);
+
+    if (dev->local) {
+        svga->decode_mask = (4 << 20) - 1;
+        mach->cursor_col_1 = 0xff;
+        mach->ext_cur_col_1_r = 0xff;
+        mach->ext_cur_col_1_g = 0xff;
+        dev->ext_crt_pitch = 128;
+        if (mach->vlb_bus)
+            ati_eeprom_load(&mach->eeprom, "mach32_vlb.nvr", 1);
+        else if (mach->pci_bus) {
+            ati_eeprom_load(&mach->eeprom, "mach32_pci.nvr", 1);
+            mem_mapping_disable(&mach->bios_rom.mapping);
+            mach->card = pci_add_card(PCI_ADD_VIDEO, mach32_pci_read, mach32_pci_write, mach);
+            mach->pci_regs[PCI_REG_COMMAND] = 0x83;
+            mach->pci_regs[0x30]            = 0x00;
+            mach->pci_regs[0x32]            = 0x0c;
+            mach->pci_regs[0x33]            = 0x00;
+        } else
+            ati_eeprom_load(&mach->eeprom, "mach32.nvr", 1);
+    } else {
+        ati_eeprom_load_mach8(&mach->eeprom, "mach8.nvr");
+    }
+
+    return (mach);
+}
+
+static int
+mach8_available(void)
+{
+    return rom_present(BIOS_MACH8_ROM_PATH);
+}
+
+static int
+mach32_isa_available(void)
+{
+    return rom_present(BIOS_MACH32_ISA_ROM_PATH);
+}
+
+static int
+mach32_vlb_available(void)
+{
+    return rom_present(BIOS_MACH32_VLB_ROM_PATH);
+}
+
+static int
+mach32_pci_available(void)
+{
+    return rom_present(BIOS_MACH32_PCI_ROM_PATH);
+}
+
+static void
+mach_close(void *p)
+{
+    mach_t *mach = (mach_t *) p;
+    svga_t *svga = &mach->svga;
+    ibm8514_t *dev = &svga->dev8514;
+
+    if (dev) {
+        free(dev->vram);
+        free(dev->changedvram);
+    }
+
+    svga_close(svga);
+    free(mach);
+}
+
+static void
+mach_speed_changed(void *p)
+{
+    mach_t *mach = (mach_t *) p;
+    svga_t *svga = &mach->svga;
+
+    svga_recalctimings(svga);
+}
+
+static void
+mach_force_redraw(void *p)
+{
+    mach_t *mach = (mach_t *) p;
+    svga_t *svga = &mach->svga;
+
+    svga->fullchange = changeframecount;
+}
+
+// clang-format off
+static const device_config_t mach32_config[] = {
+    {
+        .name = "memory",
+        .description = "Memory size",
+        .type = CONFIG_SELECTION,
+        .default_int = 2048,
+        .selection = {
+            {
+                .description = "512 KB",
+                .value = 512
+            },
+            {
+                .description = "1 MB",
+                .value = 1024
+            },
+            {
+                .description = "2 MB",
+                .value = 2048
+            },
+            {
+                .description = "4 MB",
+                .value = 4096
+            },
+            {
+                .description = ""
+            }
+        }
+    },
+    {
+        .type = CONFIG_END
+    }
+};
+
+const device_t mach8_isa_device = {
+    .name = "ATI Mach8 (ISA)",
+    .internal_name = "mach8_isa",
+    .flags = DEVICE_ISA,
+    .local = 0,
+    .init = mach8_init,
+    .close = mach_close,
+    .reset = NULL,
+    { .available = mach8_available },
+    .speed_changed = mach_speed_changed,
+    .force_redraw = mach_force_redraw,
+    .config = NULL
+};
+
+const device_t mach32_isa_device = {
+    .name = "ATI Mach32 (ISA)",
+    .internal_name = "mach32_isa",
+    .flags = DEVICE_ISA,
+    .local = 1,
+    .init = mach8_init,
+    .close = mach_close,
+    .reset = NULL,
+    { .available = mach32_isa_available },
+    .speed_changed = mach_speed_changed,
+    .force_redraw = mach_force_redraw,
+    .config = mach32_config
+};
+
+const device_t mach32_vlb_device = {
+    .name = "ATI Mach32 (VLB)",
+    .internal_name = "mach32_vlb",
+    .flags = DEVICE_VLB,
+    .local = 1,
+    .init = mach8_init,
+    .close = mach_close,
+    .reset = NULL,
+    { .available = mach32_vlb_available },
+    .speed_changed = mach_speed_changed,
+    .force_redraw = mach_force_redraw,
+    .config = mach32_config
+};
+
+const device_t mach32_pci_device = {
+    .name = "ATI Mach32 (PCI)",
+    .internal_name = "mach32_pci",
+    .flags = DEVICE_PCI,
+    .local = 1,
+    .init = mach8_init,
+    .close = mach_close,
+    .reset = NULL,
+    { .available = mach32_pci_available },
+    .speed_changed = mach_speed_changed,
+    .force_redraw = mach_force_redraw,
+    .config = mach32_config
+};
diff --git a/src/video/vid_cga.c b/src/video/vid_cga.c
index 7239790dc..1fd8babb5 100644
--- a/src/video/vid_cga.c
+++ b/src/video/vid_cga.c
@@ -77,7 +77,7 @@ cga_out(uint16_t addr, uint8_t val, void *p)
             cga->cgamode = val;
 
             if (old ^ val) {
-                if ((old ^ val) & 0x05)
+                if ((old ^ val) & 0x07)
                     update_cga16_color(val);
 
                 cga_recalctimings(cga);
@@ -352,10 +352,7 @@ cga_poll(void *p)
             x = (cga->crtc[1] << 4) + 16;
 
         if (cga->composite) {
-            if (cga->cgamode & 0x10)
-                border = 0x00;
-            else
-                border = cga->cgacol & 0x0f;
+            border = ((cga->cgamode & 0x12) == 0x12) ? 0 : (cga->cgacol & 15);
 
             Composite_Process(cga->cgamode, border, x >> 2, buffer32->line[cga->displine << 1]);
             Composite_Process(cga->cgamode, border, x >> 2, buffer32->line[(cga->displine << 1) + 1]);
@@ -536,6 +533,7 @@ cga_standalone_init(const device_t *info)
     cga->rgb_type = device_get_config_int("rgb_type");
     cga_palette   = (cga->rgb_type << 1);
     cgapal_rebuild();
+    update_cga16_color(cga->cgamode);
 
     return cga;
 }
diff --git a/src/video/vid_et4000w32.c b/src/video/vid_et4000w32.c
index e9ea4d111..4016067ab 100644
--- a/src/video/vid_et4000w32.c
+++ b/src/video/vid_et4000w32.c
@@ -277,7 +277,7 @@ et4000w32p_out(uint16_t addr, uint8_t val, void *p)
             svga->hwcursor.ena                                  = !!(et4000->regs[0xF7] & 0x80);
             svga->hwcursor.xoff                                 = et4000->regs[0xE2];
             svga->hwcursor.yoff                                 = et4000->regs[0xE6];
-            svga->hwcursor.cur_xsize = svga->hwcursor.cur_ysize = ((et4000->regs[0xEF] & 4) || ((et4000->type == ET4000W32) && et4000->regs[0xe2] && et4000->regs[0xe6])) ? 128 : 64;
+            svga->hwcursor.cur_xsize = svga->hwcursor.cur_ysize = ((et4000->regs[0xEF] & 4) || ((et4000->type == ET4000W32) && (et4000->regs[0xe2] >= 0x1f) && (et4000->regs[0xe6] >= 0x1f))) ? 128 : 64;
 
             if (et4000->type == ET4000W32) {
                 if ((svga->bpp == 15) || (svga->bpp == 16)) {
diff --git a/src/video/vid_s3.c b/src/video/vid_s3.c
index 453972402..7f840d134 100644
--- a/src/video/vid_s3.c
+++ b/src/video/vid_s3.c
@@ -498,7 +498,7 @@ static void s3_visionx68_video_engine_op(uint32_t cpu_dat, s3_t *s3);
     temp = svga->vram[dword_remap(svga, (s3->accel.dest + s3->accel.cx)) & s3->vram_mask];
 
 #define READ_PIXTRANS_WORD                                                                                 \
-    if (s3->bpp == 0 && !s3->color_16bit) {                                                                \
+    if ((s3->bpp == 0) && !s3->color_16bit) {                                                                \
         temp = svga->vram[dword_remap(svga, (s3->accel.dest + s3->accel.cx)) & s3->vram_mask];             \
         temp |= (svga->vram[dword_remap(svga, (s3->accel.dest + s3->accel.cx + 1)) & s3->vram_mask] << 8); \
     } else {                                                                                               \
@@ -506,7 +506,7 @@ static void s3_visionx68_video_engine_op(uint32_t cpu_dat, s3_t *s3);
     }
 
 #define READ_PIXTRANS_LONG                                                                                       \
-    if (s3->bpp == 0 && !s3->color_16bit) {                                                                      \
+    if ((s3->bpp == 0) && !s3->color_16bit) {                                                                      \
         temp = svga->vram[dword_remap(svga, (s3->accel.dest + s3->accel.cx)) & s3->vram_mask];                   \
         temp |= (svga->vram[dword_remap(svga, (s3->accel.dest + s3->accel.cx + 1)) & s3->vram_mask] << 8);       \
         temp |= (svga->vram[dword_remap(svga, (s3->accel.dest + s3->accel.cx + 2)) & s3->vram_mask] << 16);      \
@@ -2671,6 +2671,7 @@ s3_out(uint16_t addr, uint8_t val, void *p)
                 return;
             if ((s3->chip <= S3_86C924) && (svga->crtcreg >= 0x50))
                 return;
+
             old                       = svga->crtc[svga->crtcreg];
             svga->crtc[svga->crtcreg] = val;
 
@@ -3075,7 +3076,7 @@ s3_recalctimings(svga_t *svga)
             svga->rowoffset |= 0x100;
     }
     if (!svga->rowoffset)
-        svga->rowoffset = 256;
+        svga->rowoffset = 0x100;
 
     if ((s3->chip == S3_VISION964) || (s3->chip == S3_86C928)) {
         if (s3->card_type == S3_ELSAWIN2KPROX_964)
@@ -3120,7 +3121,7 @@ s3_recalctimings(svga_t *svga)
             }
         } else {
             if (s3->card_type == S3_NUMBER9_9FX_531) {
-                if (svga->hdisp == 1600 && s3->width == 1600)
+                if ((svga->hdisp == 1600) && (s3->width == 1600))
                     s3->width = 800;
             }
         }
@@ -3131,10 +3132,10 @@ s3_recalctimings(svga_t *svga)
         }
     }
 
-    if ((svga->crtc[0x43] & 0x08) && (s3->color_16bit == 0) && (s3->chip <= S3_86C805)) {
+    if ((svga->crtc[0x43] & 0x08) && !s3->color_16bit && (s3->chip <= S3_86C805)) {
         s3->color_16bit = 1;
-        s3->width       = 1024;
-    } else if (!(svga->crtc[0x43] & 0x08) && (s3->color_16bit == 1) && (s3->chip <= S3_86C805)) {
+        s3->width = 1024;
+    } else if (!(svga->crtc[0x43] & 0x08) && s3->color_16bit && (s3->chip <= S3_86C805)) {
         s3->color_16bit = 0;
         if (s3->chip <= S3_86C924) {
             if (s3->accel.advfunc_cntl & 4)
@@ -4808,9 +4809,9 @@ polygon_setup(s3_t *s3)
 }
 
 #define READ(addr, dat)                                                 \
-    if (s3->bpp == 0 && !s3->color_16bit)                               \
+    if ((s3->bpp == 0) && !s3->color_16bit)                               \
         dat = svga->vram[dword_remap(svga, addr) & s3->vram_mask];      \
-    else if (s3->bpp == 1 || s3->color_16bit)                           \
+    else if ((s3->bpp == 1) || s3->color_16bit)                           \
         dat = vram_w[dword_remap_w(svga, addr) & (s3->vram_mask >> 1)]; \
     else if (s3->bpp == 2)                                              \
         dat = svga->vram[dword_remap(svga, addr) & s3->vram_mask];      \
@@ -5660,10 +5661,10 @@ polygon_setup(s3_t *s3)
     }
 
 #define WRITE(addr, dat)                                                                                                   \
-    if (s3->bpp == 0 && !s3->color_16bit) {                                                                                \
+    if ((s3->bpp == 0) && !s3->color_16bit) {                                                                                \
         svga->vram[dword_remap(svga, addr) & s3->vram_mask]                = dat;                                          \
         svga->changedvram[(dword_remap(svga, addr) & s3->vram_mask) >> 12] = svga->monitor->mon_changeframecount;          \
-    } else if (s3->bpp == 1 || s3->color_16bit) {                                                                          \
+    } else if ((s3->bpp == 1) || s3->color_16bit) {                                                                          \
         vram_w[dword_remap_w(svga, addr) & (s3->vram_mask >> 1)]                    = dat;                                 \
         svga->changedvram[(dword_remap_w(svga, addr) & (s3->vram_mask >> 1)) >> 11] = svga->monitor->mon_changeframecount; \
     } else if (s3->bpp == 2) {                                                                                             \
@@ -6123,7 +6124,7 @@ s3_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat, s3_
         s3->accel.dat_count = 0;
 
     if (cpu_input && (((s3->accel.multifunc[0xa] & 0xc0) != 0x80) || (!(s3->accel.cmd & 2)))) {
-        if ((s3->bpp == 3) && count == 2) {
+        if ((s3->bpp == 3) && (count == 2)) {
             if (s3->accel.dat_count) {
                 cpu_dat             = ((cpu_dat & 0xffff) << 16) | s3->accel.dat_buf;
                 count               = 4;
@@ -6133,20 +6134,20 @@ s3_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat, s3_
                 s3->accel.dat_count = 1;
             }
         }
-        if (s3->bpp == 1 || s3->color_16bit)
+        if ((s3->bpp == 1) || s3->color_16bit)
             count >>= 1;
         if (s3->bpp == 3)
             count >>= 2;
     }
 
-    if (s3->bpp == 0 && !s3->color_16bit)
+    if ((s3->bpp == 0) && !s3->color_16bit)
         rd_mask &= 0xff;
-    else if (s3->bpp == 1 || s3->color_16bit)
+    else if ((s3->bpp == 1) || s3->color_16bit)
         rd_mask &= 0xffff;
 
-    if (s3->bpp == 0 && !s3->color_16bit)
+    if ((s3->bpp == 0) && !s3->color_16bit)
         compare &= 0xff;
-    if (s3->bpp == 1 || s3->color_16bit)
+    if ((s3->bpp == 1) || s3->color_16bit)
         compare &= 0xffff;
 
     switch (s3->accel.cmd & 0x600) {
@@ -6212,6 +6213,7 @@ s3_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat, s3_
                         cpu_dat >>= 8;
                     else
                         cpu_dat >>= 16;
+
                     if (!s3->accel.ssv_len)
                         break;
 
@@ -6259,18 +6261,16 @@ s3_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat, s3_
                 s3->accel.cx = s3->accel.cur_x & 0x7ff;
                 s3->accel.cy = s3->accel.cur_y & 0x7ff;
 
-                if (s3->accel.cur_x & 0x800) {
+                if (s3->accel.cur_x & 0x800)
                     s3->accel.cx |= ~0x7ff;
-                }
-                if (s3->accel.cur_y & 0x800) {
+
+                if (s3->accel.cur_y & 0x800)
                     s3->accel.cy |= ~0x7ff;
-                }
 
                 s3->accel.sy = s3->accel.maj_axis_pcnt;
 
-                if (s3_cpu_src(s3)) {
+                if (s3_cpu_src(s3))
                     return; /*Wait for data from CPU*/
-                }
             }
             frgd_mix = (s3->accel.frgd_mix >> 5) & 3;
             bkgd_mix = (s3->accel.bkgd_mix >> 5) & 3;
@@ -6305,11 +6305,10 @@ s3_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat, s3_
 
                     mix_dat <<= 1;
                     mix_dat |= 1;
-                    if (s3->bpp == 0 && !s3->color_16bit)
+                    if ((s3->bpp == 0) && !s3->color_16bit)
                         cpu_dat >>= 8;
-                    else {
+                    else
                         cpu_dat >>= 16;
-                    }
 
                     if (!s3->accel.sy) {
                         break;
@@ -6351,13 +6350,13 @@ s3_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat, s3_
                 s3->accel.cur_y = s3->accel.cy;
             } else /*Bresenham*/
             {
-                if (s3->accel.b2e8_pix && s3_cpu_src(s3) && count == 16) { /*Stupid undocumented 0xB2E8 on 911/924*/
+                if (s3->accel.b2e8_pix && s3_cpu_src(s3) && (count == 16)) { /*Stupid undocumented 0xB2E8 on 911/924*/
                     count              = s3->accel.maj_axis_pcnt + 1;
                     s3->accel.temp_cnt = 16;
                 }
 
                 while (count-- && s3->accel.sy >= 0) {
-                    if (s3->accel.b2e8_pix && s3_cpu_src(s3) && s3->accel.temp_cnt == 0) {
+                    if (s3->accel.b2e8_pix && s3_cpu_src(s3) && !s3->accel.temp_cnt) {
                         mix_dat >>= 16;
                         s3->accel.temp_cnt = 16;
                     }
@@ -6567,7 +6566,7 @@ s3_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat, s3_
                     mix_dat |= 1;
                 }
 
-                if (s3->bpp == 0 && !s3->color_16bit)
+                if ((s3->bpp == 0) && !s3->color_16bit)
                     cpu_dat >>= 8;
                 else {
                     cpu_dat >>= 16;
@@ -8021,7 +8020,7 @@ s3_init(const device_t *info)
         switch (vram) {
             case 0: /* 512 kB */
                 svga->vram_mask = (1 << 19) - 1;
-                svga->vram_max  = 2 << 20;
+                svga->vram_max  = 1 << 19;
                 break;
             case 1: /* 1 MB */
                 /* VRAM in first MB, mirrored in 2nd MB, 3rd and 4th MBs are open bus.
@@ -8029,7 +8028,7 @@ s3_init(const device_t *info)
                    This works with the #9 9FX BIOS, and matches how my real Trio64 behaves,
                    but does not work with the Phoenix EDO BIOS. Possibly an FPM/EDO difference? */
                 svga->vram_mask = (1 << 20) - 1;
-                svga->vram_max  = 2 << 20;
+                svga->vram_max  = 1 << 20;
                 break;
             case 2:
             default: /*2 MB */
diff --git a/src/video/vid_s3_virge.c b/src/video/vid_s3_virge.c
index feaba7fe8..1eb1fa855 100644
--- a/src/video/vid_s3_virge.c
+++ b/src/video/vid_s3_virge.c
@@ -3372,7 +3372,7 @@ s3_virge_triangle(virge_t *virge, s3d_t *s3d_tri)
     state.base_w = s3d_tri->tws;
 
     tex_base = s3d_tri->tex_base;
-    for (uint8_t c = 9; c >= 0; c--) {
+    for (int c = 9; c >= 0; c--) {
         state.texture[c] = (uint16_t *) &virge->svga.vram[tex_base];
         if (c <= state.max_d)
             tex_base += ((1 << (c * 2)) * tex_size[(s3d_tri->cmd_set >> 5) & 7]) / 2;
diff --git a/src/video/vid_svga.c b/src/video/vid_svga.c
index d8829eb2f..913a052ff 100644
--- a/src/video/vid_svga.c
+++ b/src/video/vid_svga.c
@@ -41,6 +41,7 @@
 #include <86box/video.h>
 #include <86box/vid_svga.h>
 #include <86box/vid_svga_render.h>
+#include <86box/vid_xga_device.h>
 
 void svga_doblit(int wx, int wy, svga_t *svga);
 
@@ -131,7 +132,7 @@ svga_out(uint16_t addr, uint8_t val, void *p)
                 if (svga->attraddr < 16)
                     svga->fullchange = svga->monitor->mon_changeframecount;
                 if (svga->attraddr == 0x10 || svga->attraddr == 0x14 || svga->attraddr < 0x10) {
-                    for (uint8_t c = 0; c < 16; c++) {
+                    for (int c = 0; c < 16; c++) {
                         if (svga->attrregs[0x10] & 0x80) {
                             svga->egapal[c] = (svga->attrregs[c] & 0xf) | ((svga->attrregs[0x14] & 0xf) << 4);
                         } else {
@@ -165,6 +166,12 @@ svga_out(uint16_t addr, uint8_t val, void *p)
                 io_sethandler(0x03a0, 0x0020, svga->video_in, NULL, NULL, svga->video_out, NULL, NULL, svga->p);
             svga_recalctimings(svga);
             break;
+        case 0x3c3:
+            if (xga_enabled) {
+                svga->xga.on = (val & 0x01) ? 0 : 1;
+                vga_on = !svga->xga.on;
+            }
+            break;
         case 0x3c4:
             svga->seqaddr = val;
             break;
@@ -407,7 +414,7 @@ svga_set_ramdac_type(svga_t *svga, int type)
     if (svga->ramdac_type != type) {
         svga->ramdac_type = type;
 
-        for (uint16_t c = 0; c < 256; c++) {
+        for (int c = 0; c < 256; c++) {
             if (svga->ramdac_type == RAMDAC_8BIT)
                 svga->pallook[c] = makecol32(svga->vgapal[c].r, svga->vgapal[c].g, svga->vgapal[c].b);
             else
@@ -583,8 +590,14 @@ svga_recalctimings(svga_t *svga)
             svga->recalctimings_ex(svga);
         }
     } else {
-        if (ibm8514_enabled)
-            ibm8514_recalctimings(svga);
+        if (ibm8514_enabled) {
+            if (svga->dev8514.local) {
+                if (svga->recalctimings_ex) {
+                    svga->recalctimings_ex(svga);
+                }
+            } else
+                ibm8514_recalctimings(svga);
+        }
         if (xga_enabled)
             xga_recalctimings(svga);
     }
@@ -597,8 +610,13 @@ svga_recalctimings(svga_t *svga)
 
     crtcconst = svga->clock * svga->char_width;
 
-    disptime    = svga->htotal;
-    _dispontime = svga->hdisp_time;
+    if (ibm8514_on && !svga->dev8514.local) {
+        disptime    = svga->dev8514.h_total;
+        _dispontime = svga->dev8514.h_disp;
+    } else {
+        disptime    = svga->htotal;
+        _dispontime = svga->hdisp_time;
+    }
 
     if (svga->seqregs[1] & 8) {
         disptime *= 2;
@@ -678,16 +696,20 @@ void
 svga_poll(void *p)
 {
     svga_t  *svga = (svga_t *) p;
+    ibm8514_t *dev = &svga->dev8514;
     uint32_t x;
     uint32_t blink_delay;
     int      wx;
     int      wy;
     int      ret;
     int      old_ma;
+    int      linecountff = 0;
 
     if (!vga_on && ibm8514_enabled && ibm8514_on) {
-        ibm8514_poll(&svga->dev8514, svga);
-        return;
+        if (!dev->local) {
+            ibm8514_poll(dev, svga);
+            return;
+        }
     } else if (!vga_on && xga_enabled && svga->xga.on) {
         xga_poll(&svga->xga, svga);
         return;
@@ -695,22 +717,22 @@ svga_poll(void *p)
 
     if (!svga->linepos) {
         if (svga->displine == svga->hwcursor_latch.y && svga->hwcursor_latch.ena) {
-            svga->hwcursor_on      = svga->hwcursor.cur_ysize - svga->hwcursor_latch.yoff;
+            svga->hwcursor_on      = svga->hwcursor_latch.cur_ysize - svga->hwcursor_latch.yoff;
             svga->hwcursor_oddeven = 0;
         }
 
         if (svga->displine == (svga->hwcursor_latch.y + 1) && svga->hwcursor_latch.ena && svga->interlace) {
-            svga->hwcursor_on      = svga->hwcursor.cur_ysize - (svga->hwcursor_latch.yoff + 1);
+            svga->hwcursor_on      = svga->hwcursor_latch.cur_ysize - (svga->hwcursor_latch.yoff + 1);
             svga->hwcursor_oddeven = 1;
         }
 
         if (svga->displine == svga->dac_hwcursor_latch.y && svga->dac_hwcursor_latch.ena) {
-            svga->dac_hwcursor_on      = svga->dac_hwcursor.cur_ysize - svga->dac_hwcursor_latch.yoff;
+            svga->dac_hwcursor_on      = svga->dac_hwcursor_latch.cur_ysize - svga->dac_hwcursor_latch.yoff;
             svga->dac_hwcursor_oddeven = 0;
         }
 
         if (svga->displine == (svga->dac_hwcursor_latch.y + 1) && svga->dac_hwcursor_latch.ena && svga->interlace) {
-            svga->dac_hwcursor_on      = svga->dac_hwcursor.cur_ysize - (svga->dac_hwcursor_latch.yoff + 1);
+            svga->dac_hwcursor_on      = svga->dac_hwcursor_latch.cur_ysize - (svga->dac_hwcursor_latch.yoff + 1);
             svga->dac_hwcursor_oddeven = 1;
         }
 
@@ -783,8 +805,14 @@ svga_poll(void *p)
         if ((svga->sc == (svga->crtc[11] & 31)) || (svga->sc == svga->rowcount))
             svga->con = 0;
         if (svga->dispon) {
-            if (svga->linedbl && !svga->linecountff) {
-                svga->linecountff = 1;
+        	/*Real IBM 8514/A or compatibility mode doesn't have linedbl, so skip those.*/
+            if (dev->local && ibm8514_on) {
+                svga->linedbl = 0;
+                svga->linecountff = 0;
+                linecountff = 1;
+            }
+            if (svga->linedbl && !svga->linecountff && !linecountff) {
+            	svga->linecountff = 1;
                 svga->ma          = svga->maback;
             } else if (svga->sc == svga->rowcount) {
                 svga->linecountff = 0;
@@ -793,23 +821,24 @@ svga_poll(void *p)
                 svga->maback += (svga->rowoffset << 3);
                 if (svga->interlace)
                     svga->maback += (svga->rowoffset << 3);
+
                 svga->maback &= svga->vram_display_mask;
                 svga->ma = svga->maback;
             } else {
                 svga->linecountff = 0;
                 svga->sc++;
-                svga->sc &= 31;
+                svga->sc &= 0x1f;
                 svga->ma = svga->maback;
             }
         }
 
-        svga->hsync_divisor = !svga->hsync_divisor;
+        svga->hsync_divisor ^= 1;
 
         if (svga->hsync_divisor && (svga->crtc[0x17] & 4))
             return;
 
         svga->vc++;
-        svga->vc &= 2047;
+        svga->vc &= 0x7ff;
 
         if (svga->vc == svga->split) {
             ret = 1;
@@ -835,6 +864,7 @@ svga_poll(void *p)
         if (svga->vc == svga->dispend) {
             if (svga->vblank_start)
                 svga->vblank_start(svga);
+
             svga->dispon = 0;
             blink_delay  = (svga->crtc[11] & 0x60) >> 5;
             if (svga->crtc[10] & 0x20)
@@ -846,6 +876,7 @@ svga_poll(void *p)
 
             if (!(svga->gdcreg[6] & 1) && !(svga->blink & 15))
                 svga->fullchange = 2;
+
             svga->blink = (svga->blink + 1) & 0x7f;
 
             for (x = 0; x < ((svga->vram_mask + 1) >> 12); x++) {
@@ -888,12 +919,18 @@ svga_poll(void *p)
             svga->monitor->mon_changeframecount = svga->interlace ? 3 : 2;
             svga->vslines    = 0;
 
-            if (svga->interlace && svga->oddeven)
-                svga->ma = svga->maback = svga->ma_latch + (svga->rowoffset << 1) + ((svga->crtc[5] & 0x60) >> 5);
-            else
-                svga->ma = svga->maback = svga->ma_latch + ((svga->crtc[5] & 0x60) >> 5);
+            if ((dev->local && vga_on) || !dev->local) {
+                if (svga->interlace && svga->oddeven)
+                    svga->ma = svga->maback = svga->ma_latch + (svga->rowoffset << 1) + ((svga->crtc[5] & 0x60) >> 5);
+                else
+                    svga->ma = svga->maback = svga->ma_latch + ((svga->crtc[5] & 0x60) >> 5);
+            } else if (dev->local && ibm8514_on) {
+                if (svga->interlace && svga->oddeven)
+                    svga->ma = svga->maback = svga->ma_latch + (svga->rowoffset << 1);
+                else
+                    svga->ma = svga->maback = svga->ma_latch;
+            }
             svga->ca = ((svga->crtc[0xe] << 8) | svga->crtc[0xf]) + ((svga->crtc[0xb] & 0x60) >> 5) + svga->ca_adj;
-
             svga->ma     = (svga->ma << 2);
             svga->maback = (svga->maback << 2);
             svga->ca     = (svga->ca << 2);
@@ -958,9 +995,9 @@ svga_init(const device_t *info, svga_t *svga, void *p, int memsize,
     svga->monitor_index = monitor_index_global;
     svga->monitor = &monitors[svga->monitor_index];
 
-    for (uint16_t c = 0; c < 256; c++) {
+    for (int c = 0; c < 256; c++) {
         e = c;
-        for (uint8_t d = 0; d < 8; d++) {
+        for (int d = 0; d < 8; d++) {
             svga_rotate[d][c] = e;
             e                 = (e >> 1) | ((e & 1) ? 0x80 : 0);
         }
@@ -1100,7 +1137,7 @@ svga_write_common(uint32_t addr, uint8_t val, uint8_t linear, void *p)
 
     if (!linear) {
         if (xga_enabled) {
-            if (((svga->xga.op_mode & 7) >= 4) && (svga->xga.aperture_cntl == 1)) {
+            if (((svga->xga.op_mode & 7) >= 4) && (svga->xga.aperture_cntl >= 1)) {
                 if (val == 0xa5) { /*Memory size test of XGA*/
                     svga->xga.test    = val;
                     svga->xga.a5_test = 1;
@@ -1108,7 +1145,7 @@ svga_write_common(uint32_t addr, uint8_t val, uint8_t linear, void *p)
                 } else if (val == 0x5a) {
                     svga->xga.test = val;
                     return;
-                } else if (val == 0x12 || val == 0x34) {
+                } else if ((val == 0x12) || (val == 0x34)) {
                     addr += svga->xga.write_bank;
                     svga->xga.vram[addr & svga->xga.vram_mask] = val;
                     svga->xga.linear_endian_reverse            = 1;
@@ -1145,10 +1182,18 @@ svga_write_common(uint32_t addr, uint8_t val, uint8_t linear, void *p)
         if (addr & 1)
             writemask2 <<= 1;
         addr &= ~1;
-        addr <<= 2;
-    } else
-        addr <<= 2;
-
+        if (linear && ibm8514_on && (svga->adv_flags & FLAG_ATI)) {
+            addr &= svga->vram_mask;
+        } else
+            addr <<= 2;
+    } else {
+        if (linear && ibm8514_on && (svga->adv_flags & FLAG_ATI)) {
+            writemask2 = 1 << (addr & 3);
+            addr &= ~3;
+            addr &= svga->vram_mask;
+        } else
+            addr <<= 2;
+    }
     addr &= svga->decode_mask;
 
     if (svga->translate_address)
@@ -1303,7 +1348,7 @@ svga_read_common(uint32_t addr, uint8_t linear, void *p)
 
     if (!linear) {
         if (xga_enabled) {
-            if (((svga->xga.op_mode & 7) >= 4) && (svga->xga.aperture_cntl == 1)) {
+            if (((svga->xga.op_mode & 7) >= 4) && (svga->xga.aperture_cntl >= 1)) {
                 if (svga->xga.test == 0xa5) { /*Memory size test of XGA*/
                     svga->xga.on = 1;
                     vga_on = !svga->xga.on;
@@ -1312,7 +1357,7 @@ svga_read_common(uint32_t addr, uint8_t linear, void *p)
                     svga->xga.on = 1;
                     vga_on = !svga->xga.on;
                     return svga->xga.test;
-                } else if (addr == 0xa0000 || addr == 0xa0010) {
+                } else if ((addr == 0xa0000) || (addr == 0xa0010)) {
                     addr += svga->xga.read_bank;
                     return svga->xga.vram[addr & svga->xga.vram_mask];
                 }
@@ -1354,11 +1399,24 @@ svga_read_common(uint32_t addr, uint8_t linear, void *p)
     } else if (svga->chain2_read) {
         readplane = (readplane & 2) | (addr & 1);
         addr &= ~1;
-        addr <<= 2;
-    } else
-        addr <<= 2;
-
+        if (linear && ibm8514_on && (svga->adv_flags & FLAG_ATI))
+            addr &= svga->vram_mask;
+       	else
+            addr <<= 2;
+    } else {
+        if (linear && ibm8514_on && (svga->adv_flags & FLAG_ATI)) {
+            addr &= svga->decode_mask;
+            if (addr >= svga->vram_max)
+                return 0xff;
+            latch_addr = (addr & svga->vram_mask) & ~3;
+            for (uint8_t i = 0; i < count; i++)
+                svga->latch.b[i] = svga->vram[latch_addr | i];
+            return svga->vram[addr & svga->vram_mask];
+        } else
+            addr <<= 2;
+    }
     addr &= svga->decode_mask;
+
     if (svga->translate_address) {
         latch_addr = svga->translate_address(latch_addr, p);
         addr       = svga->translate_address(addr, p);
diff --git a/src/video/vid_table.c b/src/video/vid_table.c
index d4915a056..6aee2e76f 100644
--- a/src/video/vid_table.c
+++ b/src/video/vid_table.c
@@ -36,6 +36,7 @@
 #include <86box/vid_ega.h>
 #include <86box/vid_colorplus.h>
 #include <86box/vid_mda.h>
+#include <86box/vid_xga_device.h>
 
 typedef struct {
     const device_t *device;
@@ -79,6 +80,8 @@ video_cards[] = {
     { &vid_none_device                               },
     { &vid_internal_device                           },
     { &atiega_device                                 },
+    { &mach8_isa_device,        VIDEO_FLAG_TYPE_8514 },
+    { &mach32_isa_device, 		VIDEO_FLAG_TYPE_8514 },
     { &mach64gx_isa_device                           },
     { &ati28800k_device                              },
     { &ati18800_vga88_device                         },
@@ -112,6 +115,7 @@ video_cards[] = {
     { &hercules_device,          VIDEO_FLAG_TYPE_MDA },
     { &herculesplus_device,      VIDEO_FLAG_TYPE_MDA },
     { &incolor_device                                },
+    { &inmos_isa_device,         VIDEO_FLAG_TYPE_XGA },
     { &im1024_device                                 },
     { &iskra_ega_device                              },
     { &et4000_kasan_isa_device                       },
@@ -154,6 +158,7 @@ video_cards[] = {
     { &gd5428_mca_device                             },
     { &et4000_mca_device                             },
     { &radius_svga_multiview_mca_device              },
+    { &mach32_pci_device, 		VIDEO_FLAG_TYPE_8514 },
     { &mach64gx_pci_device                           },
     { &mach64vt2_device                              },
     { &et4000w32p_videomagic_revb_pci_device         },
@@ -212,6 +217,7 @@ video_cards[] = {
     { &voodoo_3_1000_device                          },
     { &voodoo_3_2000_device                          },
     { &voodoo_3_3000_device                          },
+    { &mach32_vlb_device, 		VIDEO_FLAG_TYPE_8514 },
     { &mach64gx_vlb_device                           },
     { &et4000w32i_vlb_device                         },
     { &et4000w32p_videomagic_revb_vlb_device         },
@@ -432,3 +438,15 @@ video_is_ega_vga(void)
 {
     return (video_get_type() == VIDEO_FLAG_TYPE_SPECIAL);
 }
+
+int
+video_is_8514(void)
+{
+    return (video_get_type() == VIDEO_FLAG_TYPE_8514);
+}
+
+int
+video_is_xga(void)
+{
+    return (video_get_type() == VIDEO_FLAG_TYPE_XGA);
+}
diff --git a/src/video/vid_voodoo_banshee.c b/src/video/vid_voodoo_banshee.c
index d2e57de81..4fcd6af61 100644
--- a/src/video/vid_voodoo_banshee.c
+++ b/src/video/vid_voodoo_banshee.c
@@ -323,9 +323,9 @@ banshee_vblank_start(svga_t *svga)
 }
 
 static void
-banshee_out(uint16_t addr, uint8_t val, void *p)
+banshee_out(uint16_t addr, uint8_t val, void *priv)
 {
-    banshee_t *banshee = (banshee_t *) p;
+    banshee_t *banshee = (banshee_t *) priv;
     svga_t    *svga    = &banshee->svga;
     uint8_t    old;
 
@@ -375,13 +375,16 @@ banshee_out(uint16_t addr, uint8_t val, void *p)
 }
 
 static uint8_t
-banshee_in(uint16_t addr, void *p)
+banshee_in(uint16_t addr, void *priv)
 {
-    banshee_t *banshee = (banshee_t *) p;
+    banshee_t *banshee = (banshee_t *) priv;
     svga_t    *svga    = &banshee->svga;
     uint8_t    temp;
 
-    //        if (addr != 0x3da) banshee_log("banshee_in : %04X ", addr);
+#if 0
+    if (addr != 0x3da)
+        banshee_log("banshee_in : %04X ", addr);
+#endif
 
     if (((addr & 0xfff0) == 0x3d0 || (addr & 0xfff0) == 0x3b0) && !(svga->miscout & 1))
         addr ^= 0x60;
@@ -405,7 +408,10 @@ banshee_in(uint16_t addr, void *p)
             temp = svga_in(addr, svga);
             break;
     }
-    //        if (addr != 0x3da) banshee_log("%02X  %04X:%04X %i\n", temp, CS,cpu_state.pc, ins);
+#if 0
+    if (addr != 0x3da)
+        banshee_log("%02X  %04X:%04X %i\n", temp, CS,cpu_state.pc, ins);
+#endif
     return temp;
 }
 
@@ -524,7 +530,9 @@ banshee_recalctimings(svga_t *svga)
         svga->vblankstart += 0x400;
     if (svga->crtc[0x1b] & 0x40)
         svga->vsyncstart += 0x400;
-    //        banshee_log("svga->hdisp=%i\n", svga->hdisp);
+#if 0
+    banshee_log("svga->hdisp=%i\n", svga->hdisp);
+#endif
 
     svga->interlace = 0;
 
@@ -559,7 +567,9 @@ banshee_recalctimings(svga_t *svga)
             svga->rowoffset = (banshee->vidDesktopOverlayStride & 0x3fff) >> 3;
         svga->ma_latch                = banshee->vidDesktopStartAddr >> 2;
         banshee->desktop_stride_tiled = (banshee->vidDesktopOverlayStride & 0x3fff) * 128 * 32;
-        //                banshee_log("Extended shift out %i rowoffset=%i %02x\n", VIDPROCCFG_DESKTOP_PIX_FORMAT, svga->rowoffset, svga->crtc[1]);
+#if 0
+        banshee_log("Extended shift out %i rowoffset=%i %02x\n", VIDPROCCFG_DESKTOP_PIX_FORMAT, svga->rowoffset, svga->crtc[1]);
+#endif
 
         svga->char_width = 8;
         svga->split      = 99999;
@@ -583,11 +593,13 @@ banshee_recalctimings(svga_t *svga)
         if (svga->overlay.cur_xsize <= 0 || svga->overlay.cur_ysize <= 0)
             svga->overlay.ena = 0;
         if (svga->overlay.ena) {
-            /*                        banshee_log("Overlay enabled : start=%i,%i end=%i,%i size=%i,%i pitch=%x\n",
-                                            voodoo->overlay.start_x, voodoo->overlay.start_y,
-                                            voodoo->overlay.end_x, voodoo->overlay.end_y,
-                                            voodoo->overlay.size_x, voodoo->overlay.size_y,
-                                            svga->overlay.pitch);*/
+#if 0
+            banshee_log("Overlay enabled : start=%i,%i end=%i,%i size=%i,%i pitch=%x\n",
+                        voodoo->overlay.start_x, voodoo->overlay.start_y,
+                        voodoo->overlay.end_x, voodoo->overlay.end_y,
+                        voodoo->overlay.size_x, voodoo->overlay.size_y,
+                        svga->overlay.pitch);
+#endif
             if (!voodoo->overlay.start_x && !voodoo->overlay.start_y && svga->hdisp == voodoo->overlay.size_x && svga->dispend == voodoo->overlay.size_y) {
                 /*Overlay is full screen, so don't bother rendering the desktop
                   behind it*/
@@ -596,7 +608,9 @@ banshee_recalctimings(svga_t *svga)
             }
         }
     } else {
-        //                banshee_log("Normal shift out\n");
+#if 0
+        banshee_log("Normal shift out\n");
+#endif
         svga->bpp = 8;
     }
 
@@ -609,19 +623,27 @@ banshee_recalctimings(svga_t *svga)
         double freq = (((double) n + 2) / (((double) m + 2) * (double) (1 << k))) * 14318184.0;
 
         svga->clock = (cpuclock * (float) (1ULL << 32)) / freq;
-        //                svga->clock = cpuclock / freq;
+#if 0
+        svga->clock = cpuclock / freq;
+#endif
 
-        //                banshee_log("svga->clock = %g %g  m=%i k=%i n=%i\n", freq, freq / 1000000.0, m, k, n);
+#if 0
+        banshee_log("svga->clock = %g %g  m=%i k=%i n=%i\n", freq, freq / 1000000.0, m, k, n);
+#endif
     }
 }
 
 static void
-banshee_ext_out(uint16_t addr, uint8_t val, void *p)
+banshee_ext_out(uint16_t addr, uint8_t val, void *priv)
 {
-    //        banshee_t *banshee = (banshee_t *)p;
-    //        svga_t *svga = &banshee->svga;
+#if 0
+    banshee_t *banshee = (banshee_t *)priv;
+    svga_t *svga = &banshee->svga;
+#endif
 
-    //        banshee_log("banshee_ext_out: addr=%04x val=%02x\n", addr, val);
+#if 0
+    banshee_log("banshee_ext_out: addr=%04x val=%02x\n", addr, val);
+#endif
 
     switch (addr & 0xff) {
         case 0xb0:
@@ -672,7 +694,7 @@ banshee_ext_out(uint16_t addr, uint8_t val, void *p)
         case 0xdd:
         case 0xde:
         case 0xdf:
-            banshee_out((addr & 0xff) + 0x300, val, p);
+            banshee_out((addr & 0xff) + 0x300, val, priv);
             break;
 
         default:
@@ -680,13 +702,15 @@ banshee_ext_out(uint16_t addr, uint8_t val, void *p)
     }
 }
 static void
-banshee_ext_outl(uint16_t addr, uint32_t val, void *p)
+banshee_ext_outl(uint16_t addr, uint32_t val, void *priv)
 {
-    banshee_t *banshee = (banshee_t *) p;
+    banshee_t *banshee = (banshee_t *) priv;
     voodoo_t  *voodoo  = banshee->voodoo;
     svga_t    *svga    = &banshee->svga;
 
-    //        banshee_log("banshee_ext_outl: addr=%04x val=%08x %04x(%08x):%08x\n", addr, val, CS,cs,cpu_state.pc);
+#if 0
+    banshee_log("banshee_ext_outl: addr=%04x val=%08x %04x(%08x):%08x\n", addr, val, CS,cs,cpu_state.pc);
+#endif
 
     switch (addr & 0xff) {
         case Init_pciInit0:
@@ -698,7 +722,9 @@ banshee_ext_outl(uint16_t addr, uint32_t val, void *p)
 
         case Init_lfbMemoryConfig:
             banshee->lfbMemoryConfig = val;
-            //                banshee_log("lfbMemoryConfig=%08x\n", val);
+#if 0
+            banshee_log("lfbMemoryConfig=%08x\n", val);
+#endif
             voodoo->tile_base         = (val & 0x1fff) << 12;
             voodoo->tile_stride       = 1024 << ((val >> 13) & 7);
             voodoo->tile_stride_shift = 10 + ((val >> 13) & 7);
@@ -764,7 +790,9 @@ banshee_ext_outl(uint16_t addr, uint32_t val, void *p)
 
         case Video_vidProcCfg:
             banshee->vidProcCfg = val;
-            //                banshee_log("vidProcCfg=%08x\n", val);
+#if 0
+            banshee_log("vidProcCfg=%08x\n", val);
+#endif
             banshee->overlay_pix_fmt = (val & VIDPROCCFG_OVERLAY_PIX_FORMAT_MASK) >> VIDPROCCFG_OVERLAY_PIX_FORMAT_SHIFT;
             svga->hwcursor.ena       = val & VIDPROCCFG_HWCURSOR_ENA;
             svga->fullchange         = changeframecount;
@@ -797,7 +825,9 @@ banshee_ext_outl(uint16_t addr, uint32_t val, void *p)
             svga->hwcursor.addr      = (banshee->hwCurPatAddr & 0xfffff0) + (svga->hwcursor.yoff * 16);
             svga->hwcursor.cur_xsize = 64;
             svga->hwcursor.cur_ysize = 64;
-            //                banshee_log("hwCurLoc %08x %i\n", val, svga->hwcursor.y);
+#if 0
+            banshee_log("hwCurLoc %08x %i\n", val, svga->hwcursor.y);
+#endif
             break;
         case Video_hwCurC0:
             banshee->hwCurC0 = val;
@@ -808,7 +838,9 @@ banshee_ext_outl(uint16_t addr, uint32_t val, void *p)
 
         case Video_vidSerialParallelPort:
             banshee->vidSerialParallelPort = val;
-            //                banshee_log("vidSerialParallelPort: write %08x %08x %04x(%08x):%08x\n", val, val & (VIDSERIAL_DDC_DCK_W | VIDSERIAL_DDC_DDA_W), CS,cs,cpu_state.pc);
+#if 0
+            banshee_log("vidSerialParallelPort: write %08x %08x %04x(%08x):%08x\n", val, val & (VIDSERIAL_DDC_DCK_W | VIDSERIAL_DDC_DDA_W), CS,cs,cpu_state.pc);
+#endif
             i2c_gpio_set(banshee->i2c_ddc, !!(val & VIDSERIAL_DDC_DCK_W), !!(val & VIDSERIAL_DDC_DDA_W));
             i2c_gpio_set(banshee->i2c, !!(val & VIDSERIAL_I2C_SCK_W), !!(val & VIDSERIAL_I2C_SDA_W));
             break;
@@ -836,16 +868,22 @@ banshee_ext_outl(uint16_t addr, uint32_t val, void *p)
             break;
         case Video_vidOverlayDudx:
             voodoo->overlay.vidOverlayDudx = val & VID_DUDX_MASK;
-            //                banshee_log("vidOverlayDudx=%08x\n", val);
+#if 0
+            banshee_log("vidOverlayDudx=%08x\n", val);
+#endif
             break;
         case Video_vidOverlayDudxOffsetSrcWidth:
             voodoo->overlay.vidOverlayDudxOffsetSrcWidth = val;
             voodoo->overlay.overlay_bytes                = (val & OVERLAY_SRC_WIDTH_MASK) >> OVERLAY_SRC_WIDTH_SHIFT;
-            //                banshee_log("vidOverlayDudxOffsetSrcWidth=%08x\n", val);
+#if 0
+            banshee_log("vidOverlayDudxOffsetSrcWidth=%08x\n", val);
+#endif
             break;
         case Video_vidOverlayDvdy:
             voodoo->overlay.vidOverlayDvdy = val & VID_DVDY_MASK;
-            //                banshee_log("vidOverlayDvdy=%08x\n", val);
+#if 0
+            banshee_log("vidOverlayDvdy=%08x\n", val);
+#endif
             break;
         case Video_vidOverlayDvdyOffset:
             voodoo->overlay.vidOverlayDvdyOffset = val;
@@ -853,26 +891,34 @@ banshee_ext_outl(uint16_t addr, uint32_t val, void *p)
 
         case Video_vidDesktopStartAddr:
             banshee->vidDesktopStartAddr = val & 0xffffff;
-            //                banshee_log("vidDesktopStartAddr=%08x\n", val);
+#if 0
+            banshee_log("vidDesktopStartAddr=%08x\n", val);
+#endif
             svga->fullchange = changeframecount;
             svga_recalctimings(svga);
             break;
         case Video_vidDesktopOverlayStride:
             banshee->vidDesktopOverlayStride = val;
-            //                banshee_log("vidDesktopOverlayStride=%08x\n", val);
+#if 0
+            banshee_log("vidDesktopOverlayStride=%08x\n", val);
+#endif
             svga->fullchange = changeframecount;
             svga_recalctimings(svga);
             break;
-            //                default:
-            //                fatal("bad banshee_ext_outl: addr=%04x val=%08x\n", addr, val);
+#if 0
+        default:
+            fatal("bad banshee_ext_outl: addr=%04x val=%08x\n", addr, val);
+#endif
     }
 }
 
 static uint8_t
-banshee_ext_in(uint16_t addr, void *p)
+banshee_ext_in(uint16_t addr, void *priv)
 {
-    banshee_t *banshee = (banshee_t *) p;
-    //        svga_t *svga = &banshee->svga;
+    banshee_t *banshee = (banshee_t *) priv;
+#if 0
+    svga_t *svga = &banshee->svga;
+#endif
     uint8_t ret = 0xff;
 
     switch (addr & 0xff) {
@@ -881,7 +927,9 @@ banshee_ext_in(uint16_t addr, void *p)
         case Init_status + 2:
         case Init_status + 3:
             ret = (banshee_status(banshee) >> ((addr & 3) * 8)) & 0xff;
-            //                banshee_log("Read status reg! %04x(%08x):%08x\n", CS, cs, cpu_state.pc);
+#if 0
+            banshee_log("Read status reg! %04x(%08x):%08x\n", CS, cs, cpu_state.pc);
+#endif
             break;
 
         case 0xb0:
@@ -932,7 +980,7 @@ banshee_ext_in(uint16_t addr, void *p)
         case 0xdd:
         case 0xde:
         case 0xdf:
-            ret = banshee_in((addr & 0xff) + 0x300, p);
+            ret = banshee_in((addr & 0xff) + 0x300, priv);
             break;
 
         default:
@@ -979,15 +1027,17 @@ banshee_status(banshee_t *banshee)
     if (!voodoo->voodoo_busy)
         voodoo_wake_fifo_thread(voodoo);
 
-    //        banshee_log("banshee_status: busy %i  %i (%i %i)  %i   %i %i  %04x(%08x):%08x %08x\n", busy, written, voodoo->cmd_written, voodoo->cmd_written_fifo, voodoo->cmd_read, voodoo->cmdfifo_depth_rd, voodoo->cmdfifo_depth_wr, CS,cs,cpu_state.pc, ret);
+#if 0
+    banshee_log("banshee_status: busy %i  %i (%i %i)  %i   %i %i  %04x(%08x):%08x %08x\n", busy, written, voodoo->cmd_written, voodoo->cmd_written_fifo, voodoo->cmd_read, voodoo->cmdfifo_depth_rd, voodoo->cmdfifo_depth_wr, CS,cs,cpu_state.pc, ret);
+#endif
 
     return ret;
 }
 
 static uint32_t
-banshee_ext_inl(uint16_t addr, void *p)
+banshee_ext_inl(uint16_t addr, void *priv)
 {
-    banshee_t *banshee = (banshee_t *) p;
+    banshee_t *banshee = (banshee_t *) priv;
     voodoo_t  *voodoo  = banshee->voodoo;
     svga_t    *svga    = &banshee->svga;
     uint32_t   ret     = 0xffffffff;
@@ -997,7 +1047,9 @@ banshee_ext_inl(uint16_t addr, void *p)
     switch (addr & 0xff) {
         case Init_status:
             ret = banshee_status(banshee);
-            //                banshee_log("Read status reg! %04x(%08x):%08x\n", CS, cs, cpu_state.pc);
+#if 0
+            banshee_log("Read status reg! %04x(%08x):%08x\n", CS, cs, cpu_state.pc);
+#endif
             break;
         case Init_pciInit0:
             ret = banshee->pciInit0;
@@ -1090,7 +1142,9 @@ banshee_ext_inl(uint16_t addr, void *p)
                 if (i2c_gpio_get_sda(banshee->i2c))
                     ret |= VIDSERIAL_I2C_SDA_R;
             }
-            //                banshee_log("vidSerialParallelPort: read %08x %08x  %04x(%08x):%08x\n", ret, ret & (VIDSERIAL_DDC_DCK_R | VIDSERIAL_DDC_DDA_R), CS,cs,cpu_state.pc);
+#if 0
+            banshee_log("vidSerialParallelPort: read %08x %08x  %04x(%08x):%08x\n", ret, ret & (VIDSERIAL_DDC_DCK_R | VIDSERIAL_DDC_DDA_R), CS,cs,cpu_state.pc);
+#endif
             break;
 
         case Video_vidScreenSize:
@@ -1135,20 +1189,24 @@ banshee_ext_inl(uint16_t addr, void *p)
     return ret;
 }
 
-static uint32_t banshee_reg_readl(uint32_t addr, void *p);
+static uint32_t banshee_reg_readl(uint32_t addr, void *priv);
 
 static uint8_t
-banshee_reg_read(uint32_t addr, void *p)
+banshee_reg_read(uint32_t addr, void *priv)
 {
-    //        banshee_log("banshee_reg_read: addr=%08x\n", addr);
-    return banshee_reg_readl(addr & ~3, p) >> (8 * (addr & 3));
+#if 0
+    banshee_log("banshee_reg_read: addr=%08x\n", addr);
+#endif
+    return banshee_reg_readl(addr & ~3, priv) >> (8 * (addr & 3));
 }
 
 static uint16_t
-banshee_reg_readw(uint32_t addr, void *p)
+banshee_reg_readw(uint32_t addr, void *priv)
 {
-    //        banshee_log("banshee_reg_readw: addr=%08x\n", addr);
-    return banshee_reg_readl(addr & ~3, p) >> (8 * (addr & 2));
+#if 0
+    banshee_log("banshee_reg_readw: addr=%08x\n", addr);
+#endif
+    return banshee_reg_readl(addr & ~3, priv) >> (8 * (addr & 2));
 }
 
 static uint32_t
@@ -1212,9 +1270,9 @@ banshee_cmd_read(banshee_t *banshee, uint32_t addr)
 }
 
 static uint32_t
-banshee_reg_readl(uint32_t addr, void *p)
+banshee_reg_readl(uint32_t addr, void *priv)
 {
-    banshee_t *banshee = (banshee_t *) p;
+    banshee_t *banshee = (banshee_t *) priv;
     voodoo_t  *voodoo  = banshee->voodoo;
     uint32_t   ret     = 0xffffffff;
 
@@ -1372,7 +1430,7 @@ banshee_reg_readl(uint32_t addr, void *p)
 }
 
 static void
-banshee_reg_write(uint32_t addr, uint8_t val, void *p)
+banshee_reg_write(UNUSED(uint32_t addr), UNUSED(uint8_t val), UNUSED(void *priv))
 {
 #if 0
     banshee_log("banshee_reg_writeb: addr=%08x val=%02x\n", addr, val);
@@ -1380,9 +1438,9 @@ banshee_reg_write(uint32_t addr, uint8_t val, void *p)
 }
 
 static void
-banshee_reg_writew(uint32_t addr, uint16_t val, void *p)
+banshee_reg_writew(uint32_t addr, uint16_t val, void *priv)
 {
-    banshee_t *banshee = (banshee_t *) p;
+    banshee_t *banshee = (banshee_t *) priv;
     voodoo_t  *voodoo  = banshee->voodoo;
 
     cycles -= voodoo->write_time;
@@ -1498,9 +1556,9 @@ banshee_cmd_write(banshee_t *banshee, uint32_t addr, uint32_t val)
 }
 
 static void
-banshee_reg_writel(uint32_t addr, uint32_t val, void *p)
+banshee_reg_writel(uint32_t addr, uint32_t val, void *priv)
 {
-    banshee_t *banshee = (banshee_t *) p;
+    banshee_t *banshee = (banshee_t *) priv;
     voodoo_t  *voodoo  = banshee->voodoo;
 
     if (addr == voodoo->last_write_addr + 4)
@@ -1625,9 +1683,9 @@ banshee_reg_writel(uint32_t addr, uint32_t val, void *p)
 }
 
 static uint8_t
-banshee_read_linear(uint32_t addr, void *p)
+banshee_read_linear(uint32_t addr, void *priv)
 {
-    banshee_t *banshee = (banshee_t *) p;
+    banshee_t *banshee = (banshee_t *) priv;
     voodoo_t  *voodoo  = banshee->voodoo;
     svga_t    *svga    = &banshee->svga;
 
@@ -1663,14 +1721,14 @@ banshee_read_linear(uint32_t addr, void *p)
 }
 
 static uint16_t
-banshee_read_linear_w(uint32_t addr, void *p)
+banshee_read_linear_w(uint32_t addr, void *priv)
 {
-    banshee_t *banshee = (banshee_t *) p;
+    banshee_t *banshee = (banshee_t *) priv;
     voodoo_t  *voodoo  = banshee->voodoo;
     svga_t    *svga    = &banshee->svga;
 
     if (addr & 1)
-        return banshee_read_linear(addr, p) | (banshee_read_linear(addr + 1, p) << 8);
+        return banshee_read_linear(addr, priv) | (banshee_read_linear(addr + 1, priv) << 8);
 
     cycles -= voodoo->read_time;
     if ((banshee->pci_regs[0x30] & 0x01) && addr >= banshee->bios_rom.mapping.base && addr < (banshee->bios_rom.mapping.base + banshee->bios_rom.sz)) {
@@ -1703,14 +1761,14 @@ banshee_read_linear_w(uint32_t addr, void *p)
 }
 
 static uint32_t
-banshee_read_linear_l(uint32_t addr, void *p)
+banshee_read_linear_l(uint32_t addr, void *priv)
 {
-    banshee_t *banshee = (banshee_t *) p;
+    banshee_t *banshee = (banshee_t *) priv;
     voodoo_t  *voodoo  = banshee->voodoo;
     svga_t    *svga    = &banshee->svga;
 
     if (addr & 3)
-        return banshee_read_linear_w(addr, p) | (banshee_read_linear_w(addr + 2, p) << 16);
+        return banshee_read_linear_w(addr, priv) | (banshee_read_linear_w(addr + 2, priv) << 16);
 
     cycles -= voodoo->read_time;
 
@@ -1744,9 +1802,9 @@ banshee_read_linear_l(uint32_t addr, void *p)
 }
 
 static void
-banshee_write_linear(uint32_t addr, uint8_t val, void *p)
+banshee_write_linear(uint32_t addr, uint8_t val, void *priv)
 {
-    banshee_t *banshee = (banshee_t *) p;
+    banshee_t *banshee = (banshee_t *) priv;
     voodoo_t  *voodoo  = banshee->voodoo;
     svga_t    *svga    = &banshee->svga;
 
@@ -1779,15 +1837,15 @@ banshee_write_linear(uint32_t addr, uint8_t val, void *p)
 }
 
 static void
-banshee_write_linear_w(uint32_t addr, uint16_t val, void *p)
+banshee_write_linear_w(uint32_t addr, uint16_t val, void *priv)
 {
-    banshee_t *banshee = (banshee_t *) p;
+    banshee_t *banshee = (banshee_t *) priv;
     voodoo_t  *voodoo  = banshee->voodoo;
     svga_t    *svga    = &banshee->svga;
 
     if (addr & 1) {
-        banshee_write_linear(addr, val, p);
-        banshee_write_linear(addr + 1, val >> 8, p);
+        banshee_write_linear(addr, val, priv);
+        banshee_write_linear(addr + 1, val >> 8, priv);
         return;
     }
 
@@ -1819,16 +1877,16 @@ banshee_write_linear_w(uint32_t addr, uint16_t val, void *p)
 }
 
 static void
-banshee_write_linear_l(uint32_t addr, uint32_t val, void *p)
+banshee_write_linear_l(uint32_t addr, uint32_t val, void *priv)
 {
-    banshee_t *banshee = (banshee_t *) p;
+    banshee_t *banshee = (banshee_t *) priv;
     voodoo_t  *voodoo  = banshee->voodoo;
     svga_t    *svga    = &banshee->svga;
     int        timing;
 
     if (addr & 3) {
-        banshee_write_linear_w(addr, val, p);
-        banshee_write_linear_w(addr + 2, val >> 16, p);
+        banshee_write_linear_w(addr, val, priv);
+        banshee_write_linear_w(addr + 2, val >> 16, priv);
         return;
     }
 
@@ -2510,9 +2568,9 @@ banshee_overlay_draw(svga_t *svga, int displine)
 }
 
 void
-banshee_set_overlay_addr(void *p, uint32_t addr)
+banshee_set_overlay_addr(void *priv, UNUSED(uint32_t addr))
 {
-    banshee_t *banshee = (banshee_t *) p;
+    banshee_t *banshee = (banshee_t *) priv;
     voodoo_t  *voodoo  = banshee->voodoo;
 
     banshee->svga.overlay.addr       = banshee->voodoo->leftOverlayBuf & 0xfffffff;
@@ -2548,10 +2606,12 @@ banshee_vsync_callback(svga_t *svga)
 }
 
 static uint8_t
-banshee_pci_read(int func, int addr, void *p)
+banshee_pci_read(int func, int addr, void *priv)
 {
-    banshee_t *banshee = (banshee_t *) p;
-    //        svga_t *svga = &banshee->svga;
+    banshee_t *banshee = (banshee_t *) priv;
+#if 0
+    svga_t *svga = &banshee->svga;
+#endif
     uint8_t ret = 0;
 
     if (func)
@@ -2751,9 +2811,9 @@ banshee_pci_read(int func, int addr, void *p)
 }
 
 static void
-banshee_pci_write(int func, int addr, uint8_t val, void *p)
+banshee_pci_write(int func, int addr, uint8_t val, void *priv)
 {
-    banshee_t *banshee = (banshee_t *) p;
+    banshee_t *banshee = (banshee_t *) priv;
 #if 0
     svga_t *svga = &banshee->svga;
 #endif
@@ -3357,9 +3417,9 @@ velocity_200_available(void)
 }
 
 static void
-banshee_close(void *p)
+banshee_close(void *priv)
 {
-    banshee_t *banshee = (banshee_t *) p;
+    banshee_t *banshee = (banshee_t *) priv;
 
     voodoo_card_close(banshee->voodoo);
     svga_close(&banshee->svga);
@@ -3371,17 +3431,17 @@ banshee_close(void *p)
 }
 
 static void
-banshee_speed_changed(void *p)
+banshee_speed_changed(void *priv)
 {
-    banshee_t *banshee = (banshee_t *) p;
+    banshee_t *banshee = (banshee_t *) priv;
 
     svga_recalctimings(&banshee->svga);
 }
 
 static void
-banshee_force_redraw(void *p)
+banshee_force_redraw(void *priv)
 {
-    banshee_t *banshee = (banshee_t *) p;
+    banshee_t *banshee = (banshee_t *) priv;
 
     banshee->svga.fullchange = changeframecount;
 }
diff --git a/src/video/vid_voodoo_banshee_blitter.c b/src/video/vid_voodoo_banshee_blitter.c
index 175288856..f5009e0b8 100644
--- a/src/video/vid_voodoo_banshee_blitter.c
+++ b/src/video/vid_voodoo_banshee_blitter.c
@@ -207,7 +207,7 @@ get_addr(voodoo_t *voodoo, int x, int y, int src_notdst, uint32_t src_stride)
 }
 
 static void
-PLOT(voodoo_t *voodoo, int x, int y, int pat_x, int pat_y, uint8_t pattern_mask, uint8_t rop, uint32_t src, int src_colorkey)
+PLOT(voodoo_t *voodoo, int x, int y, int pat_x, int pat_y, uint8_t pattern_mask, UNUSED(uint8_t rop), uint32_t src, int src_colorkey)
 {
     switch (voodoo->banshee_blt.dstFormat & DST_FORMAT_COL_MASK) {
         case DST_FORMAT_COL_8_BPP:
@@ -254,7 +254,7 @@ PLOT(voodoo_t *voodoo, int x, int y, int pat_x, int pat_y, uint8_t pattern_mask,
 }
 
 static void
-PLOT_LINE(voodoo_t *voodoo, int x, int y, uint8_t rop, uint32_t pattern, int src_colorkey)
+PLOT_LINE(voodoo_t *voodoo, int x, int y, UNUSED(uint8_t rop), uint32_t pattern, int src_colorkey)
 {
     switch (voodoo->banshee_blt.dstFormat & DST_FORMAT_COL_MASK) {
         case DST_FORMAT_COL_8_BPP:
@@ -984,7 +984,7 @@ banshee_do_screen_to_screen_stretch_blt(voodoo_t *voodoo)
 }
 
 static void
-banshee_do_host_to_screen_stretch_blt(voodoo_t *voodoo, int count, uint32_t data)
+banshee_do_host_to_screen_stretch_blt(voodoo_t *voodoo, UNUSED(int count), uint32_t data)
 {
 #if 0
     if (voodoo->banshee_blt.dstBaseAddr == 0xee5194)
diff --git a/src/video/vid_voodoo_blitter.c b/src/video/vid_voodoo_blitter.c
index a1105795c..ff45fc837 100644
--- a/src/video/vid_voodoo_blitter.c
+++ b/src/video/vid_voodoo_blitter.c
@@ -166,8 +166,10 @@ voodoo_v2_blit_start(voodoo_t *voodoo)
     int      x;
     int      y;
 
-    /*        voodooblt_log("blit_start: command=%08x srcX=%i srcY=%i dstX=%i dstY=%i sizeX=%i sizeY=%i color=%04x,%04x\n",
-                    voodoo->bltCommand, voodoo->bltSrcX, voodoo->bltSrcY, voodoo->bltDstX, voodoo->bltDstY, voodoo->bltSizeX, voodoo->bltSizeY, voodoo->bltColorFg, voodoo->bltColorBg);*/
+#if 0
+    voodooblt_log("blit_start: command=%08x srcX=%i srcY=%i dstX=%i dstY=%i sizeX=%i sizeY=%i color=%04x,%04x\n",
+                  voodoo->bltCommand, voodoo->bltSrcX, voodoo->bltSrcY, voodoo->bltDstX, voodoo->bltDstY, voodoo->bltSizeX, voodoo->bltSizeY, voodoo->bltColorFg, voodoo->bltColorBg);
+#endif
 
     voodoo_wait_for_render_thread_idle(voodoo);
 
diff --git a/src/video/vid_voodoo_display.c b/src/video/vid_voodoo_display.c
index 77d54398f..84f376a8f 100644
--- a/src/video/vid_voodoo_display.c
+++ b/src/video/vid_voodoo_display.c
@@ -141,10 +141,11 @@ voodoo_calc_clutData(voodoo_t *voodoo)
         int r = (c >> 8) & 0xf8;
         int g = (c >> 3) & 0xfc;
         int b = (c << 3) & 0xf8;
-        //                r |= (r >> 5);
-        //                g |= (g >> 6);
-        //                b |= (b >> 5);
-
+#if 0
+        r |= (r >> 5);
+        g |= (g >> 6);
+        b |= (b >> 5);
+#endif
         voodoo->video_16to32[c] = (voodoo->clutData256[r].r << 16) | (voodoo->clutData256[g].g << 8) | voodoo->clutData256[b].b;
     }
 }
@@ -442,7 +443,7 @@ voodoo_filterline_v1(voodoo_t *voodoo, uint8_t *fil, int column, uint16_t *src,
 }
 
 static void
-voodoo_filterline_v2(voodoo_t *voodoo, uint8_t *fil, int column, uint16_t *src, int line)
+voodoo_filterline_v2(voodoo_t *voodoo, uint8_t *fil, int column, uint16_t *src, UNUSED(int line))
 {
     int x;
 
@@ -506,9 +507,9 @@ voodoo_filterline_v2(voodoo_t *voodoo, uint8_t *fil, int column, uint16_t *src,
 }
 
 void
-voodoo_callback(void *p)
+voodoo_callback(void *priv)
 {
-    voodoo_t  *voodoo  = (voodoo_t *) p;
+    voodoo_t  *voodoo  = (voodoo_t *) priv;
     monitor_t *monitor = &monitors[voodoo->monitor_index];
 
     if (voodoo->fbiInit0 & FBIINIT0_VGA_PASS) {
@@ -576,7 +577,9 @@ voodoo_callback(void *p)
     }
 skip_draw:
     if (voodoo->line == voodoo->v_disp) {
-        //                voodoodisp_log("retrace %i %i %08x %i\n", voodoo->retrace_count, voodoo->swap_interval, voodoo->swap_offset, voodoo->swap_pending);
+#if 0
+        voodoodisp_log("retrace %i %i %08x %i\n", voodoo->retrace_count, voodoo->swap_interval, voodoo->swap_offset, voodoo->swap_pending);
+#endif
         voodoo->retrace_count++;
         if (SLI_ENABLED && (voodoo->fbiInit2 & FBIINIT2_SWAP_ALGORITHM_MASK) == FBIINIT2_SWAP_ALGORITHM_SLI_SYNC) {
             if (voodoo == voodoo->set->voodoos[0]) {
diff --git a/src/video/vid_voodoo_fb.c b/src/video/vid_voodoo_fb.c
index 06941eb05..5e06836b1 100644
--- a/src/video/vid_voodoo_fb.c
+++ b/src/video/vid_voodoo_fb.c
@@ -59,9 +59,9 @@ voodoo_fb_log(const char *fmt, ...)
 #endif
 
 uint16_t
-voodoo_fb_readw(uint32_t addr, void *p)
+voodoo_fb_readw(uint32_t addr, void *priv)
 {
-    voodoo_t *voodoo = (voodoo_t *) p;
+    voodoo_t *voodoo = (voodoo_t *) priv;
     int       x;
     int       y;
     uint32_t  read_addr;
@@ -100,9 +100,9 @@ voodoo_fb_readw(uint32_t addr, void *p)
     return temp;
 }
 uint32_t
-voodoo_fb_readl(uint32_t addr, void *p)
+voodoo_fb_readl(uint32_t addr, void *priv)
 {
-    voodoo_t *voodoo = (voodoo_t *) p;
+    voodoo_t *voodoo = (voodoo_t *) priv;
     int       x;
     int       y;
     uint32_t  read_addr;
@@ -168,9 +168,9 @@ do_dither(voodoo_params_t *params, rgba8_t col, int x, int y)
 }
 
 void
-voodoo_fb_writew(uint32_t addr, uint16_t val, void *p)
+voodoo_fb_writew(uint32_t addr, uint16_t val, void *priv)
 {
-    voodoo_t        *voodoo = (voodoo_t *) p;
+    voodoo_t        *voodoo = (voodoo_t *) priv;
     voodoo_params_t *params = &voodoo->params;
     int              x;
     int              y;
@@ -186,10 +186,14 @@ voodoo_fb_writew(uint32_t addr, uint16_t val, void *p)
     depth_data = voodoo->params.zaColor & 0xffff;
     alpha_data = voodoo->params.zaColor >> 24;
 
-    //        while (!RB_EMPTY)
-    //                thread_reset_event(voodoo->not_full_event);
+#if 0
+    while (!RB_EMPTY)
+        thread_reset_event(voodoo->not_full_event);
+#endif
 
-    //        voodoo_fb_log("voodoo_fb_writew : %08X %04X\n", addr, val);
+#if 0
+    voodoo_fb_log("voodoo_fb_writew : %08X %04X\n", addr, val);
+#endif
 
     switch (voodoo->lfbMode & LFB_FORMAT_MASK) {
         case LFB_FORMAT_RGB565:
@@ -304,9 +308,9 @@ skip_pixel:
 }
 
 void
-voodoo_fb_writel(uint32_t addr, uint32_t val, void *p)
+voodoo_fb_writel(uint32_t addr, uint32_t val, void *priv)
 {
-    voodoo_t        *voodoo = (voodoo_t *) p;
+    voodoo_t        *voodoo = (voodoo_t *) priv;
     voodoo_params_t *params = &voodoo->params;
     int              x;
     int              y;
@@ -320,10 +324,14 @@ voodoo_fb_writel(uint32_t addr, uint32_t val, void *p)
 
     depth_data[0] = depth_data[1] = voodoo->params.zaColor & 0xffff;
     alpha_data[0] = alpha_data[1] = voodoo->params.zaColor >> 24;
-    //        while (!RB_EMPTY)
-    //                thread_reset_event(voodoo->not_full_event);
+#if 0
+    while (!RB_EMPTY)
+        thread_reset_event(voodoo->not_full_event);
+#endif
 
-    //        voodoo_fb_log("voodoo_fb_writel : %08X %08X\n", addr, val);
+#if 0
+    voodoo_fb_log("voodoo_fb_writel : %08X %08X\n", addr, val);
+#endif
 
     switch (voodoo->lfbMode & LFB_FORMAT_MASK) {
         case LFB_FORMAT_RGB565:
@@ -393,12 +401,12 @@ voodoo_fb_writel(uint32_t addr, uint32_t val, void *p)
     else
         write_addr_aux = voodoo->params.aux_offset + x + (y * voodoo->row_width);
 
-    //        voodoo_fb_log("fb_writel %08x x=%i y=%i rw=%i %08x wo=%08x\n", addr, x, y, voodoo->row_width, write_addr, voodoo->fb_write_offset);
+#if 0
+    voodoo_fb_log("fb_writel %08x x=%i y=%i rw=%i %08x wo=%08x\n", addr, x, y, voodoo->row_width, write_addr, voodoo->fb_write_offset);
+#endif
 
     if (voodoo->lfbMode & 0x100) {
-        int c;
-
-        for (c = 0; c < count; c++) {
+        for (int c = 0; c < count; c++) {
             rgba8_t  write_data = colour_data[c];
             uint16_t new_depth  = depth_data[c];
 
diff --git a/src/video/vid_voodoo_fifo.c b/src/video/vid_voodoo_fifo.c
index ab82d104e..8a84dabe4 100644
--- a/src/video/vid_voodoo_fifo.c
+++ b/src/video/vid_voodoo_fifo.c
@@ -81,9 +81,9 @@ voodoo_wake_fifo_thread_now(voodoo_t *voodoo)
 }
 
 void
-voodoo_wake_timer(void *p)
+voodoo_wake_timer(void *priv)
 {
-    voodoo_t *voodoo = (voodoo_t *) p;
+    voodoo_t *voodoo = (voodoo_t *) priv;
 
     thread_set_event(voodoo->wake_fifo_thread); /*Wake up FIFO thread if moving from idle*/
 }
@@ -295,21 +295,27 @@ voodoo_fifo_thread(void *param)
             int      num_verticies;
             int      v_num;
 
-            //                        voodoo_fifo_log(" CMDFIFO header %08x at %08x\n", header, voodoo->cmdfifo_rp);
+#if 0
+            voodoo_fifo_log(" CMDFIFO header %08x at %08x\n", header, voodoo->cmdfifo_rp);
+#endif
 
             voodoo->cmd_status &= ~7;
             voodoo->cmd_status |= (header & 7);
             voodoo->cmd_status |= (1 << 11);
             switch (header & 7) {
                 case 0:
-                    //                                voodoo_fifo_log("CMDFIFO0\n");
+#if 0
+                    voodoo_fifo_log("CMDFIFO0\n");
+#endif
                     voodoo->cmd_status = (voodoo->cmd_status & 0xffff8fff) | (((header >> 3) & 7) << 12);
                     switch ((header >> 3) & 7) {
                         case 0: /*NOP*/
                             break;
 
                         case 1: /*JSR*/
-                            //                                        voodoo_fifo_log("JSR %08x\n", (header >> 4) & 0xfffffc);
+#if 0
+                            voodoo_fifo_log("JSR %08x\n", (header >> 4) & 0xfffffc);
+#endif
                             voodoo->cmdfifo_ret_addr = voodoo->cmdfifo_rp;
                             voodoo->cmdfifo_rp       = (header >> 4) & 0xfffffc;
                             voodoo->cmdfifo_in_sub   = 1;
@@ -322,7 +328,9 @@ voodoo_fifo_thread(void *param)
 
                         case 3: /*JMP local frame buffer*/
                             voodoo->cmdfifo_rp = (header >> 4) & 0xfffffc;
-                            //                                        voodoo_fifo_log("JMP to %08x %04x\n", voodoo->cmdfifo_rp, header);
+#if 0
+                            voodoo_fifo_log("JMP to %08x %04x\n", voodoo->cmdfifo_rp, header);
+#endif
                             break;
 
                         default:
@@ -334,13 +342,20 @@ voodoo_fifo_thread(void *param)
                 case 1:
                     num  = header >> 16;
                     addr = (header & 0x7ff8) >> 1;
-                    //                                voodoo_fifo_log("CMDFIFO1 addr=%08x\n",addr);
+#if 0
+                    voodoo_fifo_log("CMDFIFO1 addr=%08x\n",addr);
+#endif
                     while (num--) {
                         uint32_t val = cmdfifo_get(voodoo);
                         if ((addr & (1 << 13)) && voodoo->type >= VOODOO_BANSHEE) {
-                            //                                                if (voodoo->type != VOODOO_BANSHEE)
-                            //                                                        fatal("CMDFIFO1: Not Banshee\n");
-                            //                                                voodoo_fifo_log("CMDFIFO1: write %08x %08x\n", addr, val);
+#if 0
+                            if (voodoo->type != VOODOO_BANSHEE)
+                                fatal("CMDFIFO1: Not Banshee\n");
+#endif
+
+#if 0
+                            voodoo_fifo_log("CMDFIFO1: write %08x %08x\n", addr, val);
+#endif
                             voodoo_2d_reg_writel(voodoo, addr, val);
                         } else {
                             if ((addr & 0x3ff) == SST_triangleCMD || (addr & 0x3ff) == SST_ftriangleCMD || (addr & 0x3ff) == SST_fastfillCMD || (addr & 0x3ff) == SST_nopCMD)
@@ -382,8 +397,10 @@ voodoo_fifo_thread(void *param)
                     v_num         = 0;
                     if (((header >> 3) & 7) == 2)
                         v_num = 1;
-                    //                                voodoo_fifo_log("CMDFIFO3: num=%i verts=%i mask=%02x\n", num, num_verticies, (header >> 10) & 0xff);
-                    //                                voodoo_fifo_log("CMDFIFO3 %02x %i\n", (header >> 10), (header >> 3) & 7);
+#if 0
+                    voodoo_fifo_log("CMDFIFO3: num=%i verts=%i mask=%02x\n", num, num_verticies, (header >> 10) & 0xff);
+                    voodoo_fifo_log("CMDFIFO3 %02x %i\n", (header >> 10), (header >> 3) & 7);
+#endif
 
                     while (num_verticies--) {
                         voodoo->verts[3].sVx = cmdfifo_get_f(voodoo);
@@ -435,7 +452,9 @@ voodoo_fifo_thread(void *param)
                     num  = (header >> 29) & 7;
                     mask = (header >> 15) & 0x3fff;
                     addr = (header & 0x7ff8) >> 1;
-                    //                                voodoo_fifo_log("CMDFIFO4 addr=%08x\n",addr);
+#if 0
+                    voodoo_fifo_log("CMDFIFO4 addr=%08x\n",addr);
+#endif
                     while (mask) {
                         if (mask & 1) {
                             uint32_t val = cmdfifo_get(voodoo);
@@ -443,7 +462,10 @@ voodoo_fifo_thread(void *param)
                             if ((addr & (1 << 13)) && voodoo->type >= VOODOO_BANSHEE) {
                                 if (voodoo->type < VOODOO_BANSHEE)
                                     fatal("CMDFIFO1: Not Banshee\n");
-                                //                                                voodoo_fifo_log("CMDFIFO1: write %08x %08x\n", addr, val);
+#if 0
+                                voodoo_fifo_log("CMDFIFO1: write %08x %08x\n", addr, val);
+#endif
+
                                 voodoo_2d_reg_writel(voodoo, addr, val);
                             } else {
                                 if ((addr & 0x3ff) == SST_triangleCMD || (addr & 0x3ff) == SST_ftriangleCMD || (addr & 0x3ff) == SST_fastfillCMD || (addr & 0x3ff) == SST_nopCMD)
@@ -463,22 +485,30 @@ voodoo_fifo_thread(void *param)
                     break;
 
                 case 5:
-                    //                                if (header & 0x3fc00000)
-                    //                                        fatal("CMDFIFO packet 5 has byte disables set %08x\n", header);
+#if 0
+                    if (header & 0x3fc00000)
+                        fatal("CMDFIFO packet 5 has byte disables set %08x\n", header);
+#endif
                     num  = (header >> 3) & 0x7ffff;
                     addr = cmdfifo_get(voodoo) & 0xffffff;
                     if (!num)
                         num = 1;
-                    //                                voodoo_fifo_log("CMDFIFO5 addr=%08x num=%i\n", addr, num);
+#if 0
+                    voodoo_fifo_log("CMDFIFO5 addr=%08x num=%i\n", addr, num);
+#endif
                     switch (header >> 30) {
                         case 0: /*Linear framebuffer (Banshee)*/
                         case 1: /*Planar YUV*/
                             if (voodoo->texture_present[0][(addr & voodoo->texture_mask) >> TEX_DIRTY_SHIFT]) {
-                                //                                                voodoo_fifo_log("texture_present at %08x %i\n", addr, (addr & voodoo->texture_mask) >> TEX_DIRTY_SHIFT);
+#if 0
+                                voodoo_fifo_log("texture_present at %08x %i\n", addr, (addr & voodoo->texture_mask) >> TEX_DIRTY_SHIFT);
+#endif
                                 flush_texture_cache(voodoo, addr & voodoo->texture_mask, 0);
                             }
                             if (voodoo->texture_present[1][(addr & voodoo->texture_mask) >> TEX_DIRTY_SHIFT]) {
-                                //                                                voodoo_fifo_log("texture_present at %08x %i\n", addr, (addr & voodoo->texture_mask) >> TEX_DIRTY_SHIFT);
+#if 0
+                                voodoo_fifo_log("texture_present at %08x %i\n", addr, (addr & voodoo->texture_mask) >> TEX_DIRTY_SHIFT);
+#endif
                                 flush_texture_cache(voodoo, addr & voodoo->texture_mask, 1);
                             }
                             while (num--) {
diff --git a/src/video/vid_voodoo_reg.c b/src/video/vid_voodoo_reg.c
index 51b9568b8..4a66161f5 100644
--- a/src/video/vid_voodoo_reg.c
+++ b/src/video/vid_voodoo_reg.c
@@ -70,9 +70,9 @@ voodoo_reg_log(const char *fmt, ...)
 #endif
 
 void
-voodoo_reg_writel(uint32_t addr, uint32_t val, void *p)
+voodoo_reg_writel(uint32_t addr, uint32_t val, void *priv)
 {
-    voodoo_t *voodoo                                      = (voodoo_t *) p;
+    voodoo_t *voodoo                                      = (voodoo_t *) priv;
     void (*voodoo_recalc_tex)(voodoo_t * voodoo, int tmu) = NULL;
     union {
         uint32_t i;
@@ -89,7 +89,9 @@ voodoo_reg_writel(uint32_t addr, uint32_t val, void *p)
         voodoo_recalc_tex = voodoo_recalc_tex12;
 
     tempif.i = val;
-    // voodoo_reg_log("voodoo_reg_write_l: addr=%08x val=%08x(%f) chip=%x\n", addr, val, tempif.f, chip);
+#if 0
+    voodoo_reg_log("voodoo_reg_write_l: addr=%08x val=%08x(%f) chip=%x\n", addr, val, tempif.f, chip);
+#endif
     addr &= 0x3fc;
 
     if ((voodoo->fbiInit3 & FBIINIT3_REMAP) && addr < 0x100 && ad21)
@@ -97,7 +99,9 @@ voodoo_reg_writel(uint32_t addr, uint32_t val, void *p)
     switch (addr) {
         case SST_swapbufferCMD:
             if (voodoo->type >= VOODOO_BANSHEE) {
-                //                        voodoo_reg_log("swapbufferCMD %08x %08x\n", val, voodoo->leftOverlayBuf);
+#if 0
+                voodoo_reg_log("swapbufferCMD %08x %08x\n", val, voodoo->leftOverlayBuf);
+#endif
 
                 voodoo_wait_for_render_thread_idle(voodoo);
                 if (!(val & 1)) {
@@ -136,8 +140,10 @@ voodoo_reg_writel(uint32_t addr, uint32_t val, void *p)
 
             voodoo->params.swapbufferCMD = val;
 
-            //                voodoo_reg_log("Swap buffer %08x %d %p %i\n", val, voodoo->swap_count, &voodoo->swap_count, (voodoo == voodoo->set->voodoos[1]) ? 1 : 0);
-            //                voodoo->front_offset = params->front_offset;
+#if 0
+            voodoo_reg_log("Swap buffer %08x %d %p %i\n", val, voodoo->swap_count, &voodoo->swap_count, (voodoo == voodoo->set->voodoos[1]) ? 1 : 0);
+            voodoo->front_offset = params->front_offset;
+#endif
             voodoo_wait_for_render_thread_idle(voodoo);
             if (!(val & 1)) {
                 memset(voodoo->dirty_line, 1, sizeof(voodoo->dirty_line));
@@ -670,7 +676,9 @@ voodoo_reg_writel(uint32_t addr, uint32_t val, void *p)
             if (voodoo->type >= VOODOO_BANSHEE) {
                 voodoo->params.draw_offset = val & 0xfffff0;
                 voodoo->fb_write_offset    = voodoo->params.draw_offset;
-                //                        voodoo_reg_log("colorBufferAddr=%06x\n", voodoo->params.draw_offset);
+#if 0
+                voodoo_reg_log("colorBufferAddr=%06x\n", voodoo->params.draw_offset);
+#endif
             }
             break;
         case SST_colBufferStride:
@@ -679,10 +687,14 @@ voodoo_reg_writel(uint32_t addr, uint32_t val, void *p)
                 voodoo->params.col_tiled = voodoo->col_tiled;
                 if (voodoo->col_tiled) {
                     voodoo->row_width = (val & 0x7f) * 128 * 32;
-                    //                                voodoo_reg_log("colBufferStride tiled = %i bytes, tiled  %08x\n", voodoo->row_width, val);
+#if 0
+                    voodoo_reg_log("colBufferStride tiled = %i bytes, tiled  %08x\n", voodoo->row_width, val);
+#endif
                 } else {
                     voodoo->row_width = val & 0x3fff;
-                    //                                voodoo_reg_log("colBufferStride linear = %i bytes, linear\n", voodoo->row_width);
+#if 0
+                    voodoo_reg_log("colBufferStride linear = %i bytes, linear\n", voodoo->row_width);
+#endif
                 }
                 voodoo->params.row_width = voodoo->row_width;
             }
@@ -690,7 +702,9 @@ voodoo_reg_writel(uint32_t addr, uint32_t val, void *p)
         case SST_auxBufferAddr:
             if (voodoo->type >= VOODOO_BANSHEE) {
                 voodoo->params.aux_offset = val & 0xfffff0;
-                //                        pclog("auxBufferAddr=%06x\n", voodoo->params.aux_offset);
+#if 0
+                pclog("auxBufferAddr=%06x\n", voodoo->params.aux_offset);
+#endif
             }
             break;
         case SST_auxBufferStride:
@@ -699,10 +713,14 @@ voodoo_reg_writel(uint32_t addr, uint32_t val, void *p)
                 voodoo->params.aux_tiled = voodoo->aux_tiled;
                 if (voodoo->aux_tiled) {
                     voodoo->aux_row_width = (val & 0x7f) * 128 * 32;
-                    //                                voodoo_reg_log("auxBufferStride tiled = %i bytes, tiled\n", voodoo->aux_row_width);
+#if 0
+                    voodoo_reg_log("auxBufferStride tiled = %i bytes, tiled\n", voodoo->aux_row_width);
+#endif
                 } else {
                     voodoo->aux_row_width = val & 0x3fff;
-                    //                                voodoo_reg_log("auxBufferStride linear = %i bytes, linear\n", voodoo->aux_row_width);
+#if 0
+                    voodoo_reg_log("auxBufferStride linear = %i bytes, linear\n", voodoo->aux_row_width);
+#endif
                 }
                 voodoo->params.aux_row_width = voodoo->aux_row_width;
             }
@@ -726,12 +744,16 @@ voodoo_reg_writel(uint32_t addr, uint32_t val, void *p)
         case SST_sVx:
             tempif.i             = val;
             voodoo->verts[3].sVx = tempif.f;
-            //                voodoo_reg_log("sVx[%i]=%f\n", voodoo->vertex_num, tempif.f);
+#if 0
+            voodoo_reg_log("sVx[%i]=%f\n", voodoo->vertex_num, tempif.f);
+#endif
             break;
         case SST_sVy:
             tempif.i             = val;
             voodoo->verts[3].sVy = tempif.f;
-            //                voodoo_reg_log("sVy[%i]=%f\n", voodoo->vertex_num, tempif.f);
+#if 0
+            voodoo_reg_log("sVy[%i]=%f\n", voodoo->vertex_num, tempif.f);
+#endif
             break;
         case SST_sARGB:
             voodoo->verts[3].sBlue  = (float) (val & 0xff);
@@ -789,7 +811,9 @@ voodoo_reg_writel(uint32_t addr, uint32_t val, void *p)
             break;
 
         case SST_sBeginTriCMD:
-            //                voodoo_reg_log("sBeginTriCMD %i %f\n", voodoo->vertex_num, voodoo->verts[4].sVx);
+#if 0
+            voodoo_reg_log("sBeginTriCMD %i %f\n", voodoo->vertex_num, voodoo->verts[4].sVx);
+#endif
             voodoo->verts[0]        = voodoo->verts[3];
             voodoo->verts[1]        = voodoo->verts[3];
             voodoo->verts[2]        = voodoo->verts[3];
@@ -800,7 +824,9 @@ voodoo_reg_writel(uint32_t addr, uint32_t val, void *p)
             voodoo->cull_pingpong = 0;
             break;
         case SST_sDrawTriCMD:
-            //                voodoo_reg_log("sDrawTriCMD %i %i\n", voodoo->num_verticies, voodoo->sSetupMode & SETUPMODE_STRIP_MODE);
+#if 0
+            voodoo_reg_log("sDrawTriCMD %i %i\n", voodoo->num_verticies, voodoo->sSetupMode & SETUPMODE_STRIP_MODE);
+#endif
             /*I'm not sure this is the vertex selection algorithm actually used in the 3dfx
               chips, but this works with a number of games that switch between strip and fan
               mode in the middle of a run (eg Black & White, Viper Racing)*/
@@ -840,7 +866,9 @@ voodoo_reg_writel(uint32_t addr, uint32_t val, void *p)
 
             voodoo->num_verticies++;
             if (voodoo->num_verticies == 3) {
-                //                        voodoo_reg_log("triangle_setup\n");
+#if 0
+                voodoo_reg_log("triangle_setup\n");
+#endif
                 voodoo_triangle_setup(voodoo);
                 voodoo->cull_pingpong = !voodoo->cull_pingpong;
 
@@ -852,7 +880,9 @@ voodoo_reg_writel(uint32_t addr, uint32_t val, void *p)
             voodoo->bltSrcBaseAddr = val & 0x3fffff;
             break;
         case SST_bltDstBaseAddr:
-            //                voodoo_reg_log("Write bltDstBaseAddr %08x\n", val);
+#if 0
+            voodoo_reg_log("Write bltDstBaseAddr %08x\n", val);
+#endif
             voodoo->bltDstBaseAddr = val & 0x3fffff;
             break;
         case SST_bltXYStrides:
@@ -969,7 +999,9 @@ voodoo_reg_writel(uint32_t addr, uint32_t val, void *p)
                     voodoo->params.texBaseAddr[0] = val & 0xfffff0;
                 else
                     voodoo->params.texBaseAddr[0] = (val & 0x7ffff) << 3;
-                //                        voodoo_reg_log("texBaseAddr = %08x %08x\n", voodoo->params.texBaseAddr[0], val);
+#if 0
+                voodoo_reg_log("texBaseAddr = %08x %08x\n", voodoo->params.texBaseAddr[0], val);
+#endif
                 voodoo_recalc_tex(voodoo, 0);
             }
             if (chip & CHIP_TREX1) {
diff --git a/src/video/vid_voodoo_render.c b/src/video/vid_voodoo_render.c
index c0e6e0196..052a0b51e 100644
--- a/src/video/vid_voodoo_render.c
+++ b/src/video/vid_voodoo_render.c
@@ -178,7 +178,9 @@ voodoo_fls(uint16_t val)
 {
     int num = 0;
 
-    // voodoo_render_log("fls(%04x) = ", val);
+#if 0
+    voodoo_render_log("fls(%04x) = ", val);
+#endif
     if (!(val & 0xff00)) {
         num += 8;
         val <<= 8;
@@ -195,7 +197,9 @@ voodoo_fls(uint16_t val)
         num += 1;
         val <<= 1;
     }
-    // voodoo_render_log("%i %04x\n", num, val);
+#if 0
+    voodoo_render_log("%i %04x\n", num, val);
+#endif
     return num;
 }
 
@@ -241,7 +245,7 @@ tex_read(voodoo_state_t *state, voodoo_texture_state_t *texture_state, int tmu)
 #define HIGH4(x) ((x & 0xf0) | ((x & 0xf0) >> 4))
 
 static inline void
-tex_read_4(voodoo_state_t *state, voodoo_texture_state_t *texture_state, int s, int t, int *d, int tmu, int x)
+tex_read_4(voodoo_state_t *state, voodoo_texture_state_t *texture_state, int s, int t, int *d, int tmu, UNUSED(int x))
 {
     rgba_u dat[4];
 
@@ -320,36 +324,44 @@ voodoo_get_texture(voodoo_t *voodoo, voodoo_params_t *params, voodoo_state_t *st
 
         s >>= 4;
         t >>= 4;
-        // if (x == 80)
-        // if (voodoo_output)
-        //         voodoo_render_log("s=%08x t=%08x _ds=%02x _dt=%02x\n", s, t, _ds, dt);
+#if 0
+        if (x == 80)
+        if (voodoo_output)
+            voodoo_render_log("s=%08x t=%08x _ds=%02x _dt=%02x\n", s, t, _ds, dt);
+#endif
         d[0] = (16 - _ds) * (16 - dt);
         d[1] = _ds * (16 - dt);
         d[2] = (16 - _ds) * dt;
         d[3] = _ds * dt;
 
-        //                texture_state.s = s;
-        //                texture_state.t = t;
+#if 0
+        texture_state.s = s;
+        texture_state.t = t;
+#endif
         tex_read_4(state, &texture_state, s, t, d, tmu, x);
 
-        /*                state->tex_r = (tex_samples[0].rgba.r * d[0] + tex_samples[1].rgba.r * d[1] + tex_samples[2].rgba.r * d[2] + tex_samples[3].rgba.r * d[3]) >> 8;
-                        state->tex_g = (tex_samples[0].rgba.g * d[0] + tex_samples[1].rgba.g * d[1] + tex_samples[2].rgba.g * d[2] + tex_samples[3].rgba.g * d[3]) >> 8;
-                        state->tex_b = (tex_samples[0].rgba.b * d[0] + tex_samples[1].rgba.b * d[1] + tex_samples[2].rgba.b * d[2] + tex_samples[3].rgba.b * d[3]) >> 8;
-                        state->tex_a = (tex_samples[0].rgba.a * d[0] + tex_samples[1].rgba.a * d[1] + tex_samples[2].rgba.a * d[2] + tex_samples[3].rgba.a * d[3]) >> 8;*/
-        /*                state->tex_r = tex_samples[0].r;
-                        state->tex_g = tex_samples[0].g;
-                        state->tex_b = tex_samples[0].b;
-                        state->tex_a = tex_samples[0].a;*/
+#if 0
+        state->tex_r = (tex_samples[0].rgba.r * d[0] + tex_samples[1].rgba.r * d[1] + tex_samples[2].rgba.r * d[2] + tex_samples[3].rgba.r * d[3]) >> 8;
+        state->tex_g = (tex_samples[0].rgba.g * d[0] + tex_samples[1].rgba.g * d[1] + tex_samples[2].rgba.g * d[2] + tex_samples[3].rgba.g * d[3]) >> 8;
+        state->tex_b = (tex_samples[0].rgba.b * d[0] + tex_samples[1].rgba.b * d[1] + tex_samples[2].rgba.b * d[2] + tex_samples[3].rgba.b * d[3]) >> 8;
+        state->tex_a = (tex_samples[0].rgba.a * d[0] + tex_samples[1].rgba.a * d[1] + tex_samples[2].rgba.a * d[2] + tex_samples[3].rgba.a * d[3]) >> 8;*/
+#endif
+#if 0
+        state->tex_r = tex_samples[0].r;
+        state->tex_g = tex_samples[0].g;
+        state->tex_b = tex_samples[0].b;
+        state->tex_a = tex_samples[0].a;
+#endif
     } else {
-        //        rgba_t tex_samples;
-        //        voodoo_texture_state_t texture_state;
-        //                int s = state->tex_s >> (18+state->lod);
-        //                int t = state->tex_t >> (18+state->lod);
-        //        int s, t;
-
-        //                state->tex_s -= 1 << (17+state->lod);
-        //                state->tex_t -= 1 << (17+state->lod);
+#if 0
+        rgba_t tex_samples;
+        voodoo_texture_state_t texture_state;
+        int s = state->tex_s >> (18+state->lod);
+        int t = state->tex_t >> (18+state->lod);
 
+        state->tex_s -= 1 << (17+state->lod);
+        state->tex_t -= 1 << (17+state->lod);
+#endif
         s = state->tex_s >> (4 + tex_lod);
         t = state->tex_t >> (4 + tex_lod);
 
@@ -357,10 +369,12 @@ voodoo_get_texture(voodoo_t *voodoo, voodoo_params_t *params, voodoo_state_t *st
         texture_state.t = t;
         tex_read(state, &texture_state, tmu);
 
-        /*                state->tex_r = tex_samples[0].rgba.r;
-                        state->tex_g = tex_samples[0].rgba.g;
-                        state->tex_b = tex_samples[0].rgba.b;
-                        state->tex_a = tex_samples[0].rgba.a;*/
+#if 0
+        state->tex_r = tex_samples[0].rgba.r;
+        state->tex_g = tex_samples[0].rgba.g;
+        state->tex_b = tex_samples[0].rgba.b;
+        state->tex_a = tex_samples[0].rgba.a;
+#endif
     }
 }
 
@@ -432,8 +446,10 @@ voodoo_tmu_fetch_and_blend(voodoo_t *voodoo, voodoo_params_t *params, voodoo_sta
         c_reverse = !tc_reverse_blend;
         a_reverse = !tca_reverse_blend;
     }
-    /*        c_reverse1 = c_reverse;
-            a_reverse1 = a_reverse;*/
+#if 0
+    c_reverse1 = c_reverse;
+    a_reverse1 = a_reverse;
+#endif
     if (tc_sub_clocal_1) {
         switch (tc_mselect_1) {
             case TC_MSELECT_ZERO:
@@ -639,30 +655,32 @@ int voodoo_recomp = 0;
 static void
 voodoo_half_triangle(voodoo_t *voodoo, voodoo_params_t *params, voodoo_state_t *state, int ystart, int yend, int odd_even)
 {
-    /*        int rgb_sel                 = params->fbzColorPath & 3;
-            int a_sel                   = (params->fbzColorPath >> 2) & 3;
-            int cc_localselect          = params->fbzColorPath & (1 << 4);
-            int cca_localselect         = (params->fbzColorPath >> 5) & 3;
-            int cc_localselect_override = params->fbzColorPath & (1 << 7);
-            int cc_zero_other           = params->fbzColorPath & (1 << 8);
-            int cc_sub_clocal           = params->fbzColorPath & (1 << 9);
-            int cc_mselect              = (params->fbzColorPath >> 10) & 7;
-            int cc_reverse_blend        = params->fbzColorPath & (1 << 13);
-            int cc_add                  = (params->fbzColorPath >> 14) & 3;
-            int cc_add_alocal           = params->fbzColorPath & (1 << 15);
-            int cc_invert_output        = params->fbzColorPath & (1 << 16);
-            int cca_zero_other          = params->fbzColorPath & (1 << 17);
-            int cca_sub_clocal          = params->fbzColorPath & (1 << 18);
-            int cca_mselect             = (params->fbzColorPath >> 19) & 7;
-            int cca_reverse_blend       = params->fbzColorPath & (1 << 22);
-            int cca_add                 = (params->fbzColorPath >> 23) & 3;
-            int cca_invert_output       = params->fbzColorPath & (1 << 25);
-            int src_afunc = (params->alphaMode >> 8) & 0xf;
-            int dest_afunc = (params->alphaMode >> 12) & 0xf;
-            int alpha_func = (params->alphaMode >> 1) & 7;
-            int a_ref = params->alphaMode >> 24;
-            int depth_op = (params->fbzMode >> 5) & 7;
-            int dither = params->fbzMode & FBZ_DITHER;*/
+#if 0
+    int rgb_sel                 = params->fbzColorPath & 3;
+    int a_sel                   = (params->fbzColorPath >> 2) & 3;
+    int cc_localselect          = params->fbzColorPath & (1 << 4);
+    int cca_localselect         = (params->fbzColorPath >> 5) & 3;
+    int cc_localselect_override = params->fbzColorPath & (1 << 7);
+    int cc_zero_other           = params->fbzColorPath & (1 << 8);
+    int cc_sub_clocal           = params->fbzColorPath & (1 << 9);
+    int cc_mselect              = (params->fbzColorPath >> 10) & 7;
+    int cc_reverse_blend        = params->fbzColorPath & (1 << 13);
+    int cc_add                  = (params->fbzColorPath >> 14) & 3;
+    int cc_add_alocal           = params->fbzColorPath & (1 << 15);
+    int cc_invert_output        = params->fbzColorPath & (1 << 16);
+    int cca_zero_other          = params->fbzColorPath & (1 << 17);
+    int cca_sub_clocal          = params->fbzColorPath & (1 << 18);
+    int cca_mselect             = (params->fbzColorPath >> 19) & 7;
+    int cca_reverse_blend       = params->fbzColorPath & (1 << 22);
+    int cca_add                 = (params->fbzColorPath >> 23) & 3;
+    int cca_invert_output       = params->fbzColorPath & (1 << 25);
+    int src_afunc               = (params->alphaMode >> 8) & 0xf;
+    int dest_afunc              = (params->alphaMode >> 12) & 0xf;
+    int alpha_func              = (params->alphaMode >> 1) & 7;
+    int a_ref                   = params->alphaMode >> 24;
+    int depth_op                = (params->fbzMode >> 5) & 7;
+    int dither                  = params->fbzMode & FBZ_DITHER;*/
+#endif
     int texels;
 #ifndef NO_CODEGEN
     uint8_t (*voodoo_draw)(voodoo_state_t * state, voodoo_params_t * params, int x, int real_y);
@@ -723,7 +741,9 @@ voodoo_half_triangle(voodoo_t *voodoo, voodoo_params_t *params, voodoo_state_t *
         yend = params->clipHighY;
 
     state->y = ystart;
-    //        yend--;
+#if 0
+    yend--;
+#endif
 
     if (SLI_ENABLED) {
         int test_y;
@@ -760,7 +780,9 @@ voodoo_half_triangle(voodoo_t *voodoo, voodoo_params_t *params, voodoo_state_t *
 #endif
 
     voodoo_render_log("dxAB=%08x dxBC=%08x dxAC=%08x\n", state->dxAB, state->dxBC, state->dxAC);
-    //        voodoo_render_log("Start %i %i\n", ystart, voodoo->fbzMode & (1 << 17));
+#if 0
+    voodoo_render_log("Start %i %i\n", ystart, voodoo->fbzMode & (1 << 17));
+#endif
 
     for (; state->y < yend; state->y += y_diff) {
         int       x;
@@ -916,7 +938,9 @@ voodoo_half_triangle(voodoo_t *voodoo, voodoo_params_t *params, voodoo_state_t *
                 voodoo->fbiPixelsIn++;
 
                 voodoo_render_log("  X=%03i T=%08x\n", x, state->tmu0_t);
-                //                        if (voodoo->fbzMode & FBZ_RGB_WMASK)
+#if 0
+                if (voodoo->fbzMode & FBZ_RGB_WMASK)
+#endif
                 {
                     int      update   = 1;
                     uint8_t  cother_r = 0;
@@ -956,7 +980,9 @@ voodoo_half_triangle(voodoo_t *voodoo, voodoo_params_t *params, voodoo_state_t *
                             w_depth = 0xffff;
                     }
 
-                    //                                w_depth = CLAMP16(w_depth);
+#if 0
+                    w_depth = CLAMP16(w_depth);
+#endif
 
                     if (params->fbzMode & FBZ_W_BUFFER)
                         new_depth = w_depth;
diff --git a/src/video/vid_voodoo_texture.c b/src/video/vid_voodoo_texture.c
index 1e9413460..d02041d16 100644
--- a/src/video/vid_voodoo_texture.c
+++ b/src/video/vid_voodoo_texture.c
@@ -550,7 +550,9 @@ flush_texture_cache(voodoo_t *voodoo, uint32_t dirty_addr, int tmu)
     int wait_for_idle = 0;
 
     memset(voodoo->texture_present[tmu], 0, sizeof(voodoo->texture_present[0]));
-    //        voodoo_texture_log("Evict %08x %i\n", dirty_addr, sizeof(voodoo->texture_present));
+#if 0
+    voodoo_texture_log("Evict %08x %i\n", dirty_addr, sizeof(voodoo->texture_present));
+#endif
     for (uint8_t c = 0; c < TEX_CACHE_MAX; c++) {
         if (voodoo->texture_cache[tmu][c].base != -1) {
             for (uint8_t d = 0; d < 4; d++) {
@@ -618,8 +620,10 @@ voodoo_tex_writel(uint32_t addr, uint32_t val, void *p)
         if (lod > LOD_MAX)
             return;
 
-        //                if (addr >= 0x200000)
-        //                        return;
+#if 0
+        if (addr >= 0x200000)
+            return;
+#endif
 
         if (voodoo->params.tformat[tmu] & 8)
             addr = voodoo->params.tex_base[tmu][lod] + s * 2 + (t << voodoo->params.tex_shift[tmu][lod]) * 2;
@@ -629,11 +633,15 @@ voodoo_tex_writel(uint32_t addr, uint32_t val, void *p)
         addr = (addr & 0x1ffffc) + voodoo->params.tex_base[tmu][0];
 
     if (voodoo->texture_present[tmu][(addr & voodoo->texture_mask) >> TEX_DIRTY_SHIFT]) {
-        //                voodoo_texture_log("texture_present at %08x %i\n", addr, (addr & voodoo->texture_mask) >> TEX_DIRTY_SHIFT);
+#if 0
+        voodoo_texture_log("texture_present at %08x %i\n", addr, (addr & voodoo->texture_mask) >> TEX_DIRTY_SHIFT);
+#endif
         flush_texture_cache(voodoo, addr & voodoo->texture_mask, tmu);
     }
     if (voodoo->type == VOODOO_3 && voodoo->texture_present[tmu ^ 1][(addr & voodoo->texture_mask) >> TEX_DIRTY_SHIFT]) {
-        //                voodoo_texture_log("texture_present at %08x %i\n", addr, (addr & voodoo->texture_mask) >> TEX_DIRTY_SHIFT);
+#if 0
+        voodoo_texture_log("texture_present at %08x %i\n", addr, (addr & voodoo->texture_mask) >> TEX_DIRTY_SHIFT);
+#endif
         flush_texture_cache(voodoo, addr & voodoo->texture_mask, tmu ^ 1);
     }
     *(uint32_t *) (&voodoo->tex_mem[tmu][addr & voodoo->texture_mask]) = val;
diff --git a/src/video/vid_xga.c b/src/video/vid_xga.c
index 0c795d132..e4f91bb00 100644
--- a/src/video/vid_xga.c
+++ b/src/video/vid_xga.c
@@ -37,6 +37,7 @@
 
 #define XGA_BIOS_PATH  "roms/video/xga/XGA_37F9576_Ver200.BIN"
 #define XGA2_BIOS_PATH "roms/video/xga/xga2_v300.bin"
+#define INMOS_XGA_BIOS_PATH "roms/video/xga/InMOS XGA - Fairchild NM27C256Q-150.BIN"
 
 static video_timings_t timing_xga_isa = { .type = VIDEO_ISA, .write_b = 3, .write_w = 3, .write_l = 6, .read_b = 5, .read_w = 5, .read_l = 10 };
 static video_timings_t timing_xga_mca = { .type = VIDEO_MCA, .write_b = 4, .write_w = 5, .write_l = 10, .read_b = 5, .read_w = 5, .read_l = 10 };
@@ -44,16 +45,86 @@ static video_timings_t timing_xga_mca = { .type = VIDEO_MCA, .write_b = 4, .writ
 static void    xga_ext_outb(uint16_t addr, uint8_t val, void *p);
 static uint8_t xga_ext_inb(uint16_t addr, void *p);
 
+int xga_has_vga = 0;
+
+void
+svga_xga_out(uint16_t addr, uint8_t val, void *p)
+{
+    svga_t *svga = (svga_t *)p;
+    uint8_t old;
+
+    if (((addr & 0xfff0) == 0x3d0 || (addr & 0xfff0) == 0x3b0) && !(svga->miscout & 1))
+        addr ^= 0x60;
+
+    switch (addr) {
+        case 0x3D4:
+            svga->crtcreg = val & 0x3f;
+            return;
+        case 0x3D5:
+            if (svga->crtcreg & 0x20)
+                return;
+            if ((svga->crtcreg < 7) && (svga->crtc[0x11] & 0x80))
+                return;
+            if ((svga->crtcreg == 7) && (svga->crtc[0x11] & 0x80))
+                val = (svga->crtc[7] & ~0x10) | (val & 0x10);
+            old                       = svga->crtc[svga->crtcreg];
+            svga->crtc[svga->crtcreg] = val;
+            if (old != val) {
+                if (svga->crtcreg < 0xe || svga->crtcreg > 0x10) {
+                    if ((svga->crtcreg == 0xc) || (svga->crtcreg == 0xd)) {
+                        svga->fullchange = 3;
+                        svga->ma_latch   = ((svga->crtc[0xc] << 8) | svga->crtc[0xd]) + ((svga->crtc[8] & 0x60) >> 5);
+                    } else {
+                        svga->fullchange = changeframecount;
+                        svga_recalctimings(svga);
+                    }
+                }
+            }
+            break;
+    }
+    svga_out(addr, val, svga);
+}
+
+uint8_t
+svga_xga_in(uint16_t addr, void *p)
+{
+    svga_t *svga = (svga_t *)p;
+    uint8_t temp;
+
+    if (((addr & 0xfff0) == 0x3d0 || (addr & 0xfff0) == 0x3b0) && !(svga->miscout & 1))
+        addr ^= 0x60;
+
+    switch (addr) {
+        case 0x3D4:
+            temp = svga->crtcreg;
+            break;
+        case 0x3D5:
+            if (svga->crtcreg & 0x20)
+                temp = 0xff;
+            else
+                temp = svga->crtc[svga->crtcreg];
+            break;
+        default:
+            temp = svga_in(addr, svga);
+            break;
+    }
+    return temp;
+}
+
 void
 xga_updatemapping(svga_t *svga)
 {
     xga_t *xga = &svga->xga;
 
-    //pclog("OpMode = %x, linear base = %08x, aperture cntl = %d, opmodereset1 = %d, access mode = %x, map = %x.\n", xga->op_mode, xga->linear_base, xga->aperture_cntl, xga->op_mode_reset, xga->access_mode, svga->gdcreg[6] & 0x0c);
+    //pclog("OpMode = %x, linear base = %08x, aperture cntl = %d, access mode = %x, map = %x, endian reverse = %d, a5test = %d, XGA on = %d.\n", xga->op_mode, xga->linear_base, xga->aperture_cntl, xga->access_mode, svga->gdcreg[6] & 0x0c, xga->linear_endian_reverse, xga->a5_test, xga->on);
     if (((xga->op_mode & 7) >= 4) || ((xga->op_mode & 7) == 0)) {
-        if (xga->aperture_cntl == 1) {
+        if ((xga->aperture_cntl == 1) || (xga->aperture_cntl == 2)) {
             mem_mapping_disable(&svga->mapping);
-            mem_mapping_set_addr(&xga->video_mapping, 0xa0000, 0x10000);
+            if (xga->aperture_cntl == 1)
+                mem_mapping_set_addr(&xga->video_mapping, 0xa0000, 0x10000);
+            else
+                mem_mapping_set_addr(&xga->video_mapping, 0xb0000, 0x10000);
+
             mem_mapping_enable(&xga->video_mapping);
             xga->banked_mask = 0xffff;
             if (!xga->linear_endian_reverse)
@@ -63,38 +134,23 @@ xga_updatemapping(svga_t *svga)
             mem_mapping_set_addr(&xga->video_mapping, 0xa0000, 0x10000);
             mem_mapping_enable(&xga->video_mapping);
             xga->banked_mask = 0xffff;
-            if (xga->pos_regs[4] & 1)
-                mem_mapping_set_addr(&xga->linear_mapping, xga->linear_base, 0x400000);
-            else if (xga->base_addr_1mb)
+            if (xga->base_addr_1mb)
                 mem_mapping_set_addr(&xga->linear_mapping, xga->base_addr_1mb, 0x100000);
             else
                 mem_mapping_set_addr(&xga->linear_mapping, xga->linear_base, 0x400000);
             if (((xga->op_mode & 7) == 4) && ((svga->gdcreg[6] & 0x0c) == 0x0c) && !xga->a5_test && xga->on)
                 xga->linear_endian_reverse = 1;
-            else if (((xga->op_mode & 7) == 0) && ((svga->gdcreg[6] & 0x0c) == 0x0c) && !xga->a5_test && !xga->on)
+            else if (((xga->op_mode & 7) == 0) && ((svga->gdcreg[6] & 0x0c) == 0x0c) && !xga->a5_test && !xga->on) {
                 xga->linear_endian_reverse = 1;
-            xga->on = 0;
-            vga_on  = !xga->on;
-        } else {
-            mem_mapping_disable(&svga->mapping);
-            mem_mapping_set_addr(&xga->video_mapping, 0xb0000, 0x10000);
-            mem_mapping_enable(&xga->video_mapping);
-            xga->banked_mask = 0xffff;
-            mem_mapping_disable(&xga->linear_mapping);
+            }
+            if (xga->a5_test && (xga->access_mode & 8) && !xga->linear_endian_reverse) {
+                xga->on = 0;
+                vga_on  = !xga->on;
+            }
         }
-    } else {
-        xga->on = 0;
-        vga_on  = !xga->on;
-        mem_mapping_disable(&svga->mapping);
-        if (xga->aperture_cntl == 2)
-            mem_mapping_set_addr(&xga->video_mapping, 0xb0000, 0x10000);
-        else
-            mem_mapping_set_addr(&xga->video_mapping, 0xa0000, 0x10000);
-        mem_mapping_enable(&xga->video_mapping);
-        xga->banked_mask = 0xffff;
-        mem_mapping_disable(&xga->linear_mapping);
-        //pclog("XGA opmode (not extended) = %d, disp mode = %d, aperture = %d.\n", xga->op_mode & 7, xga->disp_cntl_2 & 7, xga->aperture_cntl);
+        //pclog("XGA opmode (extended) = %d, disp mode = %d, aperture = %d.\n", xga->op_mode & 7, xga->disp_cntl_2 & 7, xga->aperture_cntl);
     }
+    //pclog("VGA on = %d.\n", vga_on);
 }
 
 void
@@ -126,19 +182,19 @@ xga_recalctimings(svga_t *svga)
 
         xga->ma_latch = xga->disp_start_addr;
 
-        switch (xga->clk_sel_1 & 0x0c) {
+        switch ((xga->clk_sel_1 >> 2) & 3) {
             case 0:
                 if (xga->clk_sel_2 & 0x80) {
-                    svga->clock = (cpuclock * (double) (1ULL << 32)) / 41539000.0;
+                    svga->clock = (cpuclock * (double) (1ull << 32)) / 41539000.0;
                 } else {
-                    svga->clock = (cpuclock * (double) (1ULL << 32)) / 25175000.0;
+                    svga->clock = (cpuclock * (double) (1ull << 32)) / 25175000.0;
                 }
                 break;
-            case 4:
-                svga->clock = (cpuclock * (double) (1ULL << 32)) / 28322000.0;
+            case 1:
+                svga->clock = (cpuclock * (double) (1ull << 32)) / 28322000.0;
                 break;
-            case 0x0c:
-                svga->clock = (cpuclock * (double) (1ULL << 32)) / 44900000.0;
+            case 3:
+                svga->clock = (cpuclock * (double) (1ull << 32)) / 44900000.0;
                 break;
         }
     }
@@ -309,7 +365,7 @@ xga_ext_out_reg(xga_t *xga, svga_t *svga, uint8_t idx, uint8_t val)
             if ((xga->sprite_pos >= 0) && (xga->sprite_pos <= 16)) {
                 if ((xga->op_mode & 7) >= 5)
                     xga->cursor_data_on = 1;
-                else if (xga->sprite_pos >= 1)
+                else if ((xga->sprite_pos >= 1) || ((xga->disp_cntl_2 & 7) > 3))
                     xga->cursor_data_on = 1;
                 else if (xga->aperture_cntl == 0) {
                     if (xga->linear_endian_reverse && !(xga->access_mode & 8))
@@ -474,7 +530,10 @@ xga_ext_inb(uint16_t addr, void *p)
         case 0x0f:
             switch (xga->regs_idx) {
                 case 4:
-                    ret = (xga->bus & DEVICE_MCA) ? 1 : 0;
+                    if (xga->bus & DEVICE_MCA)
+                        ret = 0x01; /*32-bit MCA*/
+                    else
+                        ret = 0x10; /*16-bit ISA*/
                     break;
                 case 0x10:
                     ret = xga->htotal & 0xff;
@@ -653,6 +712,16 @@ xga_ext_inb(uint16_t addr, void *p)
                     ret = xga->clk_sel_2;
                     break;
 
+                case 0x74:
+                    if (xga->bus & DEVICE_MCA)
+                        ret = xga->regs[xga->regs_idx];
+                    else {
+                        ret = (xga->dma_channel << 1);
+                        if (xga->dma_channel)
+                            ret |= 1;
+                    }
+                    break;
+
                 default:
                     ret = xga->regs[xga->regs_idx];
                     break;
@@ -678,15 +747,35 @@ xga_ext_inb(uint16_t addr, void *p)
     dat = xga->vram[(addr + 1) & (xga->vram_mask - 1)] & 0xff; \
     dat |= (xga->vram[(addr) & (xga->vram_mask - 1)] << 8);
 
+#define READL(addr, dat) \
+    dat = *(uint32_t *) &xga->vram[(addr) & (xga->vram_mask)];
+
+#define READL_REVERSE(addr, dat)                               \
+    dat = xga->vram[(addr + 3) & (xga->vram_mask - 3)] & 0xff; \
+    dat |= (xga->vram[(addr + 2) & (xga->vram_mask - 3)] << 8);  \
+    dat |= (xga->vram[(addr + 1) & (xga->vram_mask - 3)] << 16);  \
+    dat |= (xga->vram[(addr) & (xga->vram_mask - 3)] << 24);
+
 #define WRITEW(addr, dat)                                        \
     *(uint16_t *) &xga->vram[((addr)) & (xga->vram_mask)] = dat; \
     xga->changedvram[(((addr)) & (xga->vram_mask)) >> 12] = svga->monitor->mon_changeframecount;
 
+#define WRITEL(addr, dat)                                        \
+    *(uint32_t *) &xga->vram[((addr)) & (xga->vram_mask)] = dat; \
+    xga->changedvram[(((addr)) & (xga->vram_mask)) >> 12] = svga->monitor->mon_changeframecount;
+
 #define WRITEW_REVERSE(addr, dat)                                       \
     xga->vram[((addr + 1)) & (xga->vram_mask - 1)]        = dat & 0xff; \
     xga->vram[((addr)) & (xga->vram_mask - 1)]            = dat >> 8;   \
     xga->changedvram[(((addr)) & (xga->vram_mask)) >> 12] = svga->monitor->mon_changeframecount;
 
+#define WRITEL_REVERSE(addr, dat)                                       \
+    xga->vram[((addr + 3)) & (xga->vram_mask - 3)]        = dat & 0xff; \
+    xga->vram[((addr + 2)) & (xga->vram_mask - 3)]        = dat >> 8;   \
+    xga->vram[((addr + 1)) & (xga->vram_mask - 3)]        = dat >> 16;   \
+    xga->vram[((addr)) & (xga->vram_mask - 3)]        = dat >> 24;   \
+    xga->changedvram[(((addr)) & (xga->vram_mask)) >> 12] = svga->monitor->mon_changeframecount;
+
 #define ROP(mix, d, s)                                                                 \
     {                                                                                  \
         switch ((mix) ? (xga->accel.frgd_mix & 0x1f) : (xga->accel.bkgd_mix & 0x1f)) { \
@@ -859,6 +948,23 @@ xga_accel_read_map_pixel(svga_t *svga, int x, int y, int map, uint32_t base, int
                 byte = mem_readw_phys(addr);
             }
             return byte;
+        case 5: /*24-bit*/
+            addr += (y * (width << 2));
+            addr += (x << 2);
+            if (!skip) {
+                if ((xga->accel.px_map_format[map] & 8)) {
+                    if (xga->linear_endian_reverse) {
+                        READL(addr, byte);
+                    } else {
+                        READL_REVERSE(addr, byte);
+                    }
+                } else {
+                    READL(addr, byte);
+                }
+            } else {
+                byte = mem_readl_phys(addr);
+            }
+            return byte;
     }
 
     return 0;
@@ -936,6 +1042,22 @@ xga_accel_write_map_pixel(svga_t *svga, int x, int y, int map, uint32_t base, ui
             }
             mem_writew_phys(addr, pixel);
             break;
+        case 5: /*24-bit*/
+            addr += (y * (width) << 2);
+            addr += (x << 2);
+            if (!skip) {
+                if ((xga->accel.px_map_format[map] & 8)) {
+                    if (xga->linear_endian_reverse) {
+                        WRITEL(addr, pixel);
+                    } else {
+                        WRITEL_REVERSE(addr, pixel);
+                    }
+                } else {
+                    WRITEL(addr, pixel);
+                }
+            }
+            mem_writel_phys(addr, pixel);
+            break;
     }
 }
 
@@ -1447,6 +1569,10 @@ xga_mem_write(uint32_t addr, uint32_t val, xga_t *xga, svga_t *svga, int len)
 
     if (addr >= 0x1800) {
         switch (addr & 0x7f) {
+            case 0x11:
+                xga->accel.control = val;
+                break;
+
             case 0x12:
                 xga->accel.px_map_idx = val & 3;
                 break;
@@ -1923,11 +2049,21 @@ xga_mem_read(uint32_t addr, xga_t *xga, svga_t *svga)
     uint8_t temp = 0;
 
     addr &= 0x1fff;
-
     if (addr < 0x1800) {
-        temp = xga->bios_rom.rom[addr];
+        if (!xga_has_vga)
+            temp = xga->bios_rom.rom[addr];
+        else
+            temp = xga->vga_bios_rom.rom[addr];
     } else {
         switch (addr & 0x7f) {
+            case 0x11:
+                temp = xga->accel.control;
+                if (xga->accel.control & 0x08)
+                    temp |= 0x10;
+                else
+                    temp &= ~0x10;
+                break;
+
             case 0x20:
                 temp = xga->accel.bres_err_term & 0xff;
                 break;
@@ -2092,8 +2228,9 @@ xga_render_overscan_left(xga_t *xga, svga_t *svga)
     if (svga->scrblank || (xga->h_disp == 0))
         return;
 
+    uint32_t *line_ptr = svga->monitor->target_buffer->line[xga->displine + svga->y_add];
     for (int i = 0; i < svga->x_add; i++)
-        buffer32->line[xga->displine + svga->y_add][i] = svga->overscan_color;
+        *line_ptr++ = svga->overscan_color;
 }
 
 static void
@@ -2107,9 +2244,10 @@ xga_render_overscan_right(xga_t *xga, svga_t *svga)
     if (svga->scrblank || (xga->h_disp == 0))
         return;
 
-    right = (overscan_x >> 1);
+    uint32_t *line_ptr = &svga->monitor->target_buffer->line[xga->displine + svga->y_add][svga->x_add + xga->h_disp];
+    right              = (overscan_x >> 1);
     for (int i = 0; i < right; i++)
-        buffer32->line[xga->displine + svga->y_add][svga->x_add + xga->h_disp + i] = svga->overscan_color;
+        *line_ptr++ = svga->overscan_color;
 }
 
 static void
@@ -2122,7 +2260,7 @@ xga_render_8bpp(xga_t *xga, svga_t *svga)
         return;
 
     if (xga->changedvram[xga->ma >> 12] || xga->changedvram[(xga->ma >> 12) + 1] || svga->fullchange) {
-        p = &buffer32->line[xga->displine + svga->y_add][svga->x_add];
+        p = &svga->monitor->target_buffer->line[xga->displine + svga->y_add][svga->x_add];
 
         if (xga->firstline_draw == 2000)
             xga->firstline_draw = xga->displine;
@@ -2159,7 +2297,7 @@ xga_render_16bpp(xga_t *xga, svga_t *svga)
         return;
 
     if (xga->changedvram[xga->ma >> 12] || xga->changedvram[(xga->ma >> 12) + 1] || svga->fullchange) {
-        p = &buffer32->line[xga->displine + svga->y_add][svga->x_add];
+        p = &svga->monitor->target_buffer->line[xga->displine + svga->y_add][svga->x_add];
 
         if (xga->firstline_draw == 2000)
             xga->firstline_draw = xga->displine;
@@ -2524,7 +2662,7 @@ xga_poll(xga_t *xga, svga_t *svga)
         }
 
         xga->vc++;
-        xga->vc &= 2047;
+        xga->vc &= 0x7ff;
 
         if (xga->vc == xga->split) {
             if (xga->interlace && xga->oddeven)
@@ -2683,8 +2821,149 @@ static uint8_t
 xga_pos_in(uint16_t addr, void *priv)
 {
     svga_t *svga = (svga_t *) priv;
+    xga_t   *xga  = &svga->xga;
+    uint8_t ret = 0xff;
 
-    return (xga_mca_read(addr, svga));
+    if (xga_has_vga) {
+        switch (addr) {
+            case 0x0100:
+            case 0x0101:
+                if (xga->instance_isa == xga->instance_num)
+                    ret = xga->pos_regs[addr & 7];
+                else
+                    ret = 0xff;
+                break;
+            case 0x0102:
+            case 0x0105:
+                ret = xga->pos_regs[addr & 7];
+                break;
+            case 0x0106:
+                ret = xga->pos_idx >> 8;
+                break;
+            case 0x0107:
+                ret = xga->pos_idx & 0xff;
+                break;
+            case 0x0103:
+                if (!(xga->pos_idx & 3)) {
+                    ret = xga->pos_regs[3];
+                } else
+                    ret = 0;
+                //pclog("POS IDX for 0103 = %d, ret = %02x.\n", xga->pos_idx & 3, ret);
+                break;
+            case 0x0104:
+                switch (xga->pos_idx & 3) {
+                    case 0:
+                        ret = xga->pos_regs[4];
+                        break;
+                    case 1:
+                        ret = xga->pos_regs[0];
+                        break;
+                    case 2:
+                        ret = xga->pos_regs[1];
+                        break;
+                    case 3:
+                        ret = 0;
+                        break;
+                }
+                //pclog("POS IDX for 0104 = %d, ret = %02x.\n", xga->pos_idx & 3, ret);
+                break;
+            case 0x0108:
+            case 0x0109:
+            case 0x010a:
+            case 0x010b:
+            case 0x010c:
+            case 0x010d:
+            case 0x010e:
+            case 0x010f:
+                xga->instance_num = addr & 7;
+                if (xga->instance_isa == xga->instance_num)
+                    ret = xga->instance_isa;
+                else
+                    ret = 0;
+
+                ret |= xga->isa_pos_enable;
+                break;
+        }
+    } else {
+        switch (addr) {
+            case 0x0100:
+            case 0x0101:
+                ret = xga->pos_regs[addr & 7];
+                break;
+            case 0x0103:
+                ret = xga->pos_regs[3] | 7;
+                ret |= (xga->dma_channel << 3);
+                break;
+            case 0x0102:
+            case 0x0104:
+            case 0x0105:
+            case 0x0106:
+            case 0x0107:
+                ret = (xga_mca_read(addr, svga));
+                break;
+            case 0x0108:
+            case 0x0109:
+            case 0x010a:
+            case 0x010b:
+            case 0x010c:
+            case 0x010d:
+            case 0x010e:
+            case 0x010f:
+                xga->instance_num = addr & 7;
+                if (xga->instance_isa == xga->instance_num)
+                    ret = xga->instance_isa;
+                else
+                    ret = 0;
+
+                ret |= xga->isa_pos_enable;
+                break;
+        }
+    }
+    return ret;
+}
+
+static void
+xga_pos_out(uint16_t addr, uint8_t val, void *priv)
+{
+    svga_t *svga = (svga_t *) priv;
+    xga_t   *xga  = &svga->xga;
+
+    if (xga_has_vga) {
+        switch (addr) {
+            case 0x0106:
+                xga->pos_idx = (xga->pos_idx & 0x00ff) | (val << 8);
+                break;
+            case 0x0107:
+                xga->pos_idx = (xga->pos_idx & 0xff00) | (val);
+                //pclog("POS IDX Write = %04x.\n", xga->pos_idx);
+                break;
+            case 0x0108:
+            case 0x0109:
+            case 0x010a:
+            case 0x010b:
+            case 0x010c:
+            case 0x010d:
+            case 0x010e:
+            case 0x010f:
+                xga->instance_num = addr & 7;
+                xga->isa_pos_enable = val & 0x08;
+                break;
+        }
+    } else {
+        switch (addr) {
+            case 0x0108:
+            case 0x0109:
+            case 0x010a:
+            case 0x010b:
+            case 0x010c:
+            case 0x010d:
+            case 0x010e:
+            case 0x010f:
+                xga->instance_num = addr & 7;
+                xga->isa_pos_enable = val & 0x08;
+                break;
+        }
+    }
 }
 
 static void
@@ -2700,7 +2979,10 @@ static void
     uint32_t temp;
     uint8_t *rom               = NULL;
 
+    xga->ext_mem_addr = device_get_config_hex16("ext_mem_addr");
+    xga->instance_isa = device_get_config_int("instance");
     xga->type = device_get_config_int("type");
+    xga->dma_channel = device_get_config_int("dma");
     xga->bus  = info->flags;
 
     xga->vram_size             = (1024 << 10);
@@ -2739,12 +3021,16 @@ static void
         xga->rom_addr    = 0;
         rom_init(&xga->bios_rom, xga->type ? XGA2_BIOS_PATH : XGA_BIOS_PATH, 0xc0000, 0x2000, 0x1fff, 0, MEM_MAPPING_EXTERNAL);
     } else {
-        video_inform(VIDEO_FLAG_TYPE_SPECIAL, &timing_xga_isa);
-        xga->pos_regs[2] = 1 | 0x0c | 0xf0;
+        if (xga_has_vga) {
+            rom_init(&xga->vga_bios_rom, INMOS_XGA_BIOS_PATH, 0xc0000, 0x8000, 0x7fff, 0, MEM_MAPPING_EXTERNAL);
+        } else
+            video_inform(VIDEO_FLAG_TYPE_SPECIAL, &timing_xga_isa);
+
+        xga->pos_regs[2] = 1 | (xga->instance_isa << 1) | xga->ext_mem_addr;
         xga->instance    = (xga->pos_regs[2] & 0x0e) >> 1;
         xga->pos_regs[4] = 1 | 2;
         xga->linear_base = ((xga->pos_regs[4] & 0xfe) * 0x1000000) + (xga->instance << 22);
-        xga->rom_addr    = 0xc0000 + (((xga->pos_regs[2] & 0xf0) >> 4) * 0x2000);
+        xga->rom_addr = 0xc0000 + (((xga->pos_regs[2] & 0xf0) >> 4) * 0x2000);
     }
 
     mem_mapping_add(&xga->video_mapping, 0, 0, xga_readb, xga_readw, xga_readl,
@@ -2755,7 +3041,7 @@ static void
                     NULL, MEM_MAPPING_EXTERNAL, svga);
     mem_mapping_add(&xga->memio_mapping, 0, 0, xga_memio_readb, xga_memio_readw, xga_memio_readl,
                     xga_memio_writeb, xga_memio_writew, xga_memio_writel,
-                    xga->bios_rom.rom, MEM_MAPPING_EXTERNAL, svga);
+                    xga_has_vga ? xga->vga_bios_rom.rom : xga->bios_rom.rom, MEM_MAPPING_EXTERNAL, svga);
 
     mem_mapping_disable(&xga->video_mapping);
     mem_mapping_disable(&xga->linear_mapping);
@@ -2768,13 +3054,41 @@ static void
         mca_add(xga_mca_read, xga_mca_write, xga_mca_feedb, xga_mca_reset, svga);
     } else {
         io_sethandler(0x0100, 0x0008, xga_pos_in, NULL, NULL, NULL, NULL, NULL, svga);
+        if (xga_has_vga)
+            io_sethandler(0x0106, 0x0002, NULL, NULL, NULL, xga_pos_out, NULL, NULL, svga);
+
         io_sethandler(0x2100 + (xga->instance << 4), 0x0010, xga_ext_inb, NULL, NULL, xga_ext_outb, NULL, NULL, svga);
+		io_sethandler(0x0108, 0x0008, xga_pos_in, NULL, NULL, xga_pos_out, NULL, NULL, svga);
         mem_mapping_set_addr(&xga->memio_mapping, xga->rom_addr + 0x1c00 + (xga->instance * 0x80), 0x80);
     }
-
     return svga;
 }
 
+static void
+    *
+    svga_xga_init(const device_t *info)
+{
+    svga_t *svga = malloc(sizeof(svga_t));
+    memset(svga, 0, sizeof(svga_t));
+
+    video_inform(VIDEO_FLAG_TYPE_XGA, &timing_xga_isa);
+
+    svga_init(info, svga, svga, 1 << 18, /*256kB*/
+              NULL,
+              svga_xga_in, svga_xga_out,
+              NULL,
+              NULL);
+
+    io_sethandler(0x03c0, 0x0020, svga_xga_in, NULL, NULL, svga_xga_out, NULL, NULL, svga);
+
+    svga->bpp     = 8;
+    svga->miscout = 1;
+    xga_has_vga = 1;
+    xga_enabled = 1;
+
+    return xga_init(info);
+}
+
 static void
 xga_close(void *p)
 {
@@ -2793,6 +3107,12 @@ xga_available(void)
     return rom_present(XGA_BIOS_PATH) && rom_present(XGA2_BIOS_PATH);
 }
 
+static int
+inmos_xga_available(void)
+{
+    return rom_present(INMOS_XGA_BIOS_PATH);
+}
+
 static void
 xga_speed_changed(void *p)
 {
@@ -2809,7 +3129,7 @@ xga_force_redraw(void *p)
     svga->fullchange = svga->monitor->mon_changeframecount;
 }
 
-static const device_config_t xga_configuration[] = {
+static const device_config_t xga_mca_configuration[] = {
   // clang-format off
     {
         .name = "type",
@@ -2835,6 +3155,91 @@ static const device_config_t xga_configuration[] = {
 // clang-format on
 };
 
+static const device_config_t xga_isa_configuration[] = {
+  // clang-format off
+    {
+        .name = "type",
+        .description = "XGA type",
+        .type = CONFIG_SELECTION,
+        .default_string = "",
+        .default_int = 0,
+        .file_filter = "",
+        .spinner = { 0 },
+        .selection = {
+            {
+                .description = "XGA-1",
+                .value = 0
+            },
+            {
+                .description = "XGA-2",
+                .value = 1
+            },
+            { .description = "" }
+        }
+    },
+    {
+        .name = "instance",
+        .description = "Instance",
+        .type = CONFIG_SELECTION,
+        .default_string = "",
+        .default_int = 6,
+        .file_filter = "",
+        .spinner = { 0 },
+        .selection = {
+            { .description = "0 (2100h-210Fh)", .value = 0 },
+            { .description = "1 (2110h-211Fh)", .value = 1 },
+            { .description = "2 (2120h-212Fh)", .value = 2 },
+            { .description = "3 (2130h-213Fh)", .value = 3 },
+            { .description = "4 (2140h-214Fh)", .value = 4 },
+            { .description = "5 (2150h-215Fh)", .value = 5 },
+            { .description = "6 (2160h-216Fh)", .value = 6 },
+            { .description = "7 (2170h-217Fh)", .value = 7 },
+            { .description = ""                      }
+        },
+    },
+    {
+        .name = "ext_mem_addr",
+        .description = "MMIO address",
+        .type = CONFIG_HEX16,
+        .default_string = "",
+        .default_int = 0x00f0,
+        .file_filter = "",
+        .spinner = { 0 },
+        .selection = {
+            { .description = "C800h", .value = 0x0040 },
+            { .description = "CA00h", .value = 0x0050 },
+            { .description = "CC00h", .value = 0x0060 },
+            { .description = "CE00h", .value = 0x0070 },
+            { .description = "D000h", .value = 0x0080 },
+            { .description = "D200h", .value = 0x0090 },
+            { .description = "D400h", .value = 0x00a0 },
+            { .description = "D600h", .value = 0x00b0 },
+            { .description = "D800h", .value = 0x00c0 },
+            { .description = "DA00h", .value = 0x00d0 },
+            { .description = "DC00h", .value = 0x00e0 },
+            { .description = "DE00h", .value = 0x00f0 },
+            { .description = ""                      }
+        },
+    },
+    {
+        .name = "dma",
+        .description = "DMA channel",
+        .type = CONFIG_SELECTION,
+        .default_string = "",
+        .default_int = 7,
+        .file_filter = "",
+        .spinner = { 0 },
+        .selection = {
+            { .description = "Disabled", .value = 0 },
+            { .description = "DMA 6", .value = 6 },
+            { .description = "DMA 7", .value = 7 },
+            { .description = ""                  }
+        },
+    },
+    { .name = "", .description = "", .type = CONFIG_END }
+// clang-format on
+};
+
 const device_t xga_device = {
     .name          = "XGA (MCA)",
     .internal_name = "xga_mca",
@@ -2846,7 +3251,7 @@ const device_t xga_device = {
     { .available = xga_available },
     .speed_changed = xga_speed_changed,
     .force_redraw  = xga_force_redraw,
-    .config        = xga_configuration
+    .config        = xga_mca_configuration
 };
 
 const device_t xga_isa_device = {
@@ -2860,13 +3265,27 @@ const device_t xga_isa_device = {
     { .available = xga_available },
     .speed_changed = xga_speed_changed,
     .force_redraw  = xga_force_redraw,
-    .config        = xga_configuration
+    .config        = xga_isa_configuration
+};
+
+const device_t inmos_isa_device = {
+    .name          = "INMOS XGA (ISA)",
+    .internal_name = "inmos_xga_isa",
+    .flags         = DEVICE_ISA | DEVICE_AT,
+    .local         = 0,
+    .init          = svga_xga_init,
+    .close         = xga_close,
+    .reset         = xga_reset,
+    { .available = inmos_xga_available },
+    .speed_changed = xga_speed_changed,
+    .force_redraw  = xga_force_redraw,
+    .config        = xga_isa_configuration
 };
 
 void
 xga_device_add(void)
 {
-    if (!xga_enabled)
+    if (!xga_enabled || (xga_has_vga && xga_enabled))
         return;
 
     if (machine_has_bus(machine, MACHINE_BUS_MCA))
diff --git a/src/vnc.c b/src/vnc.c
index 8743c1e17..30caff9e2 100644
--- a/src/vnc.c
+++ b/src/vnc.c
@@ -93,7 +93,9 @@ vnc_mouse_poll(void)
         ms.dx = 0;
         ms.dy = 0;
 
-        // pclog("dx=%d, dy=%d, dwheel=%d\n", mouse_x, mouse_y, mouse_z);
+#if 0
+        pclog("dx=%d, dy=%d, dwheel=%d\n", mouse_x, mouse_y, mouse_z);
+#endif
     }
 
     if (b != ms.buttons) {
@@ -132,7 +134,7 @@ vnc_ptrevent(int but, int x, int y, rfbClientPtr cl)
 }
 
 static void
-vnc_clientgone(rfbClientPtr cl)
+vnc_clientgone(UNUSED(rfbClientPtr cl))
 {
     vnc_log("VNC: client disconnected: %s\n", cl->host);
 
@@ -143,7 +145,9 @@ vnc_clientgone(rfbClientPtr cl)
         vnc_log("VNC: no clients, pausing..\n");
 
         /* Disable the mouse. */
-        // plat_mouse_capture(0);
+#if 0
+        plat_mouse_capture(0);
+#endif
         mouse_set_poll_ex(NULL);
 
         plat_pause(1);
@@ -169,7 +173,9 @@ vnc_newclient(rfbClientPtr cl)
         vnc_log("VNC: unpausing..\n");
 
         /* Enable the mouse. */
-        // plat_mouse_capture(1);
+#if 0
+        plat_mouse_capture(1);
+#endif
         mouse_set_poll_ex(vnc_mouse_poll);
 
         plat_pause(0);
@@ -325,7 +331,7 @@ vnc_pause(void)
 }
 
 void
-vnc_take_screenshot(wchar_t *fn)
+vnc_take_screenshot(UNUSED(wchar_t *fn))
 {
     vnc_log("VNC: take_screenshot\n");
 }
diff --git a/src/vnc_keymap.c b/src/vnc_keymap.c
index 923f6ecf0..2e5f3c0d8 100644
--- a/src/vnc_keymap.c
+++ b/src/vnc_keymap.c
@@ -665,8 +665,8 @@ vnc_kbinput(int down, int k)
 
     /* Send this scancode sequence to the PC keyboard. */
     switch (scan >> 8) {
-        case 0x00:
         default:
+        case 0x00:
             if (scan & 0xff)
                 keyboard_input(down, scan & 0xff);
             break;
diff --git a/src/win/Makefile.mingw b/src/win/Makefile.mingw
index 4b60cb930..806f2b45b 100644
--- a/src/win/Makefile.mingw
+++ b/src/win/Makefile.mingw
@@ -242,13 +242,13 @@ PROG := 86Box
 #              Nothing should need changing from here on..              #
 #########################################################################
 VPATH := $(EXPATH) . $(CODEGEN) minitrace cpu cpu/softfloat \
-         cdrom chipset device disk disk/minivhd floppy \
+         cpu/808x cdrom chipset device disk disk/minivhd floppy \
          game machine mem printer \
          sio sound \
          sound/munt sound/munt/c_interface sound/munt/sha1 \
          sound/munt/srchelper sound/munt/srchelper/srctools/src \
          sound/resid-fp sound/ymfm \
-         scsi video network network/slirp win
+         scsi video network win
 
 WINDRES := windres
 STRIP   := strip
@@ -545,13 +545,15 @@ MAINOBJ := 86box.o config.o log.o random.o timer.o io.o acpi.o apm.o dma.o ddma.
            usb.o device.o nvr.o nvr_at.o nvr_ps2.o machine_status.o ini.o \
            $(VNCOBJ)
 
-MEMOBJ := catalyst_flash.o i2c_eeprom.o intel_flash.o mem.o rom.o smram.o spd.o sst_flash.o
+MEMOBJ := catalyst_flash.o i2c_eeprom.o intel_flash.o mem.o rom.o row.o smram.o spd.o sst_flash.o
+
+CPU808XOBJ := queue.o
 
 CPUOBJ := $(DYNARECOBJ) \
           $(CGTOBJ) \
           cpu.o cpu_table.o fpu.o x86.o \
           8080.o 808x.o 386.o 386_common.o 386_dynarec.o 386_dynarec_ops.o \
-          x86seg.o x87.o x87_timings.o \
+          x86_ops_mmx.o x86seg.o x87.o x87_timings.o \
 	  f2xm1.o fpatan.o fprem.o fsincos.o fyl2x.o softfloat_poly.o softfloat.o softfloat16.o \
 	  softfloat-muladd.o softfloat-round-pack.o softfloat-specialize.o softfloatx80.o
 
@@ -654,18 +656,9 @@ SCSIOBJ := scsi.o scsi_device.o \
            scsi_ncr5380.o scsi_ncr53c8xx.o \
            scsi_pcscsi.o scsi_spock.o
 
-SLIRPOBJ := net_slirp.o tinyglib.o \
-          arp_table.o bootp.o cksum.o dhcpv6.o dnssearch.o if.o \
-          ip_icmp.o ip_input.o ip_output.o \
-          ip6_icmp.o ip6_input.o ip6_output.o \
-          mbuf.o misc.o sbuf.o slirp.o socket.o \
-          tcp_input.o tcp_output.o tcp_subr.o tcp_timer.o \
-          udp.o udp6.o \
-          util.o version.o \
-
 NETOBJ := network.o \
           net_pcap.o \
-          ${SLIRPOBJ} \
+          net_slirp.o \
           net_dp8390.o net_3c501.o \
           net_3c503.o net_ne2000.o \
           net_pcnet.o net_wd8003.o \
@@ -720,6 +713,7 @@ VIDOBJ := agpgart.o video.o \
           vid_vga.o \
           vid_ati_eeprom.o \
           vid_ati18800.o vid_ati28800.o \
+	  vid_ati_mach8.o \
           vid_ati_mach64.o vid_ati68860_ramdac.o \
           vid_bt48x_ramdac.o \
           vid_av9194.o vid_icd2061.o vid_ics2494.o vid_ics2595.o \
@@ -781,7 +775,7 @@ ifeq ($(RTMIDI), y)
  SNDOBJ += midi_rtmidi.o
 endif
 
-OBJ := $(MAINOBJ) $(CPUOBJ) $(CHIPSETOBJ) $(MCHOBJ) $(DEVOBJ) $(MEMOBJ) \
+OBJ := $(MAINOBJ) $(CPU808XOBJ) $(CPUOBJ) $(CHIPSETOBJ) $(MCHOBJ) $(DEVOBJ) $(MEMOBJ) \
        $(FDDOBJ) $(GAMEOBJ) $(CDROMOBJ) $(ZIPOBJ) $(MOOBJ) $(HDDOBJ) $(MINIVHDOBJ) \
        $(NETOBJ) $(PRINTOBJ) $(SCSIOBJ) $(SIOOBJ) $(SNDOBJ) $(VIDOBJ) $(VOODOOOBJ) \
        $(PLATOBJ) $(UIOBJ) $(FSYNTHOBJ) $(MUNTOBJ) $(DEVBROBJ) $(MINITRACEOBJ) $(THREADOBJ)
@@ -795,13 +789,24 @@ else
  MWIN := -mwindows
 endif
 
+LIBS := -lfreetype -lfluidsynth -lslirp -lbz2 -lharfbuzz -lgraphite2 -lbrotlidec \
+        -lbrotlicommon -lusp10 -lrpcrt4 -lgomp -lsndfile -lflac -lmp3lame -lmpg123 \
+        -lopus -lvorbis -lvorbisenc -logg -ldsound -lshlwapi -lksuser -lreadline \
+        -ltermcap -lportaudio -lgmodule-2.0 -lglib-2.0 -lintl -liconv
+
 ifeq ($(OPENAL), y)
- LIBS := $(MWIN) -lopenal -lcomctl32 -lSDL2 -limagehlp -ldinput8 -ldxguid -ldxerr8 -luser32 -lgdi32 -lwinmm -limm32 -lole32 -loleaut32 -lshell32 -lversion -luuid -lws2_32
+ LIBS += $(MWIN) -lopenal -lcomctl32 -lSDL2 -limagehlp -ldinput8 -ldxguid -ldxerr8 \
+         -luser32 -lgdi32 -lwinmm -limm32 -lole32 -loleaut32 -lshell32 -lversion \
+         -luuid -lws2_32
 else
  ifeq ($(FAUDIO), y)
-  LIBS := $(MWIN) -lfaudio -lcomctl32 -lSDL2 -limagehlp -ldinput8 -ldxguid -ldxerr8 -luser32 -lgdi32 -lwinmm -limm32 -lole32 -loleaut32 -lshell32 -lversion -luuid -lws2_32
+  LIBS += $(MWIN) -lfaudio -lcomctl32 -lSDL2 -limagehlp -ldinput8 -ldxguid -ldxerr8 \
+          -luser32 -lgdi32 -lwinmm -limm32 -lole32 -loleaut32 -lshell32 -lversion \
+          -luuid -lws2_32
  else
-  LIBS := $(MWIN) -lcomctl32 -lSDL2 -limagehlp -ldinput8 -ldxguid -ldxerr8 -luser32 -lgdi32 -lwinmm -limm32 -lole32 -loleaut32 -lshell32 -lversion -luuid -lws2_32
+  LIBS += $(MWIN) -lcomctl32 -lSDL2 -limagehlp -ldinput8 -ldxguid -ldxerr8 -luser32 \
+          -lgdi32 -lwinmm -limm32 -lole32 -loleaut32 -lshell32 -lversion -luuid \
+          -lws2_32
  endif
 endif
 
diff --git a/src/win/win_settings.c b/src/win/win_settings.c
index 38911b0f8..9ded351b9 100644
--- a/src/win/win_settings.c
+++ b/src/win/win_settings.c
@@ -714,7 +714,8 @@ win_settings_machine_recalc_fpu(HWND hdlg)
         c++;
     }
 
-    settings_set_check(hdlg, IDC_CHECK_SOFTFLOAT, temp_fpu_softfloat);
+    settings_set_check(hdlg, IDC_CHECK_SOFTFLOAT, (machine_has_flags(temp_machine, MACHINE_SOFTFLOAT_ONLY) ? TRUE : temp_fpu_softfloat));
+    settings_enable_window(hdlg, IDC_CHECK_SOFTFLOAT, (machine_has_flags(temp_machine, MACHINE_SOFTFLOAT_ONLY) ? FALSE : TRUE));
 
     settings_enable_window(hdlg, IDC_COMBO_FPU, c > 1);