From d70f43828c86d72258a2d5737531e3a94b45ab73 Mon Sep 17 00:00:00 2001
From: Akamaki <97360908+akmed772@users.noreply.github.com>
Date: Mon, 24 Feb 2025 05:12:54 +0900
Subject: [PATCH 01/58] replace magic numbers with predifined macros

---
 src/video/vid_ps55da2.c | 46 +++++++++++++++++++++--------------------
 1 file changed, 24 insertions(+), 22 deletions(-)

diff --git a/src/video/vid_ps55da2.c b/src/video/vid_ps55da2.c
index 31559cc52..a517b7e6c 100644
--- a/src/video/vid_ps55da2.c
+++ b/src/video/vid_ps55da2.c
@@ -54,9 +54,14 @@
 #define DA2_FONTROM_BASESBCS     0x98000
 #define DA2_GAIJIRAM_SBCS        0x34000
 #define DA2_GAIJIRAM_SBEX        0x3c000
+#define DA2_VM03_BASECHR         0x18000
+#define DA2_VM03_BASEEXATTR      0x10000
 #define DA2_INVALIDACCESS8       0xff
 #define DA2_MASK_MMIO            0x1ffff
-#define DA2_MASK_GRAM            0x1ffff
+#define DA2_SIZE_VRAM            1024 * 1024 /* 0x100000 */
+#define DA2_SIZE_CRAM            4 * 1024 /* 0x1000 */
+#define DA2_SIZE_GAIJIRAM        256 * 1024 /* 0x40000 */
+#define DA2_MASK_A000            0x1ffff
 #define DA2_MASK_CRAM            0xfff
 #define DA2_MASK_GAIJIRAM        0x3ffff
 #define DA2_PIXELCLOCK           58000000.0
@@ -243,7 +248,7 @@
 #define LG_SET_RESET_2          0x10
 
 #ifndef RELEASE_BUILD
-// #define ENABLE_DA2_LOG 1
+#define ENABLE_DA2_LOG 1
 #endif
 
 #ifdef ENABLE_DA2_LOG
@@ -370,7 +375,7 @@ typedef struct da2_t {
     {
         int           enable;
         mem_mapping_t mapping;
-        uint8_t       ram[256 * 1024];
+        uint8_t       ram[DA2_SIZE_GAIJIRAM];
         uint8_t      *font;
         int           charset;
     } mmio;
@@ -1855,10 +1860,10 @@ da2_render_textm3(da2_t *da2)
         int       chr_wide = 0;
         // da2_log("\nda2ma: %x, da2sc: %x\n", da2->ma, da2->sc);
         for (x = 0; x < da2->hdisp; x += 13) {
-            chr     = da2->vram[(0x18000 + (da2->ma)) & da2->vram_mask];
-            attr    = da2->vram[((0x18000 + (da2->ma)) + 1) & da2->vram_mask];
-            extattr = da2->vram[((0x10000 + (da2->ma)) + 1) & da2->vram_mask];
-            // if(chr!=0x20) da2_log("addr: %x, chr: %x, attr: %x    ", (0x18000 + da2->ma << 1) & da2->vram_mask, chr, attr);
+            chr     = da2->vram[(DA2_VM03_BASECHR + (da2->ma)) & da2->vram_mask];
+            attr    = da2->vram[((DA2_VM03_BASECHR + (da2->ma)) + 1) & da2->vram_mask];
+            extattr = da2->vram[((DA2_VM03_BASEEXATTR + (da2->ma)) + 1) & da2->vram_mask];
+            // if(chr!=0x20) da2_log("addr: %x, chr: %x, attr: %x    ", (DA2_VM03_BASECHR + da2->ma << 1) & da2->vram_mask, chr, attr);
             bg = attr >> 4;
             // if (da2->blink) bg &= ~0x8;
             // fg = (da2->blink || (!(attr & 0x80))) ? (attr & 0xf) : bg;
@@ -1875,7 +1880,7 @@ da2_render_textm3(da2_t *da2)
                 /* Stay drawing if the char code is DBCS and not at last column. */
                 if (chr_wide) {
                     /* Get high DBCS code from the next video address */
-                    chr_dbcs = da2->vram[(0x18000 + (da2->ma) + 2) & da2->vram_mask];
+                    chr_dbcs = da2->vram[(DA2_VM03_BASECHR + (da2->ma) + 2) & da2->vram_mask];
                     chr_dbcs <<= 8;
                     chr_dbcs |= chr;
                     /* Get the font pattern */
@@ -1887,11 +1892,9 @@ da2_render_textm3(da2_t *da2)
                     }
                 } else {
                     /* the char code is SBCS (ANK) */
-                    uint32_t fontbase;
-                    fontbase      = DA2_GAIJIRAM_SBCS;
-                    uint16_t font = da2->mmio.ram[fontbase + chr * 0x40 + da2->sc * 2]; /* w13xh29 font */
+                    uint16_t font = da2->mmio.ram[DA2_GAIJIRAM_SBCS + chr * 0x40 + da2->sc * 2]; /* w13xh29 font */
                     font <<= 8;
-                    font |= da2->mmio.ram[fontbase + chr * 0x40 + da2->sc * 2 + 1]; /* w13xh29 font */
+                    font |= da2->mmio.ram[DA2_GAIJIRAM_SBCS+ chr * 0x40 + da2->sc * 2 + 1]; /* w13xh29 font */
                     // if(chr!=0x20) da2_log("ma: %x, sc: %x, chr: %x, font: %x    ", da2->ma, da2->sc, chr, font);
                     for (uint32_t n = 0; n < 13; n++) {
                         p[n] = da2->pallook[da2->egapal[(font & 0x8000) ? fg : bg]];
@@ -2101,10 +2104,10 @@ da2_recalctimings(da2_t *da2)
         if (da2->crtc[LC_VIEWPORT_PRIORITY] & 0x80) {
             da2_log("Set videomode to PS/55 8 bpp graphics.\n");
             da2->render            = da2_render_color_8bpp;
-            da2->vram_display_mask = DA2_MASK_GRAM;
+            da2->vram_display_mask = DA2_MASK_A000;
         } else { /* PS/55 8-color */
             da2_log("Set videomode to PS/55 4 bpp graphics.\n");
-            da2->vram_display_mask = DA2_MASK_GRAM;
+            da2->vram_display_mask = DA2_MASK_A000;
             da2->render            = da2_render_color_4bpp;
         }
     } else {
@@ -3029,12 +3032,11 @@ da2_init(UNUSED(const device_t *info))
 
     da2->dispontime        = 1000ull << 32;
     da2->dispofftime       = 1000ull << 32;
-    int memsize            = 1024 * 1024;
-    da2->vram              = calloc(1, memsize);
-    da2->vram_mask         = memsize - 1;
-    da2->cram              = calloc(1, 0x1000);
+    da2->vram              = calloc(1, DA2_SIZE_VRAM);
+    da2->vram_mask         = DA2_SIZE_VRAM - 1;
+    da2->cram              = calloc(1, DA2_SIZE_CRAM);
     da2->vram_display_mask = DA2_MASK_CRAM;
-    da2->changedvram       = calloc(1, /*(memsize >> 12) << 1*/ 0x1000000 >> 12); /* XX000h */
+    da2->changedvram       = calloc(1, /*(memsize >> 12) << 1*/ DA2_SIZE_VRAM >> 12); /* XX000h */
     da2->monitorid         = device_get_config_int("montype");                 /* Configuration for Monitor ID (aaaa) -> (xxax xxxx, xxxx xaaa) */
 
     da2->mmio.charset = device_get_config_int("charset");
@@ -3095,17 +3097,17 @@ da2_close(void *p)
     FILE *f;
     f = fopen("da2_cram.dmp", "wb");
     if (f != NULL) {
-        fwrite(da2->cram, 0x1000, 1, f);
+        fwrite(da2->cram, DA2_SIZE_CRAM, 1, f);
         fclose(f);
     }
     f = fopen("da2_vram.dmp", "wb");
     if (f != NULL) {
-        fwrite(da2->vram, 1024 * 1024, 1, f);
+        fwrite(da2->vram, DA2_SIZE_VRAM, 1, f);
         fclose(f);
     }
     f = fopen("da2_gram.dmp", "wb");
     if (f != NULL) {
-        fwrite(da2->mmio.ram, 256 * 1024, 1, f);
+        fwrite(da2->mmio.ram, DA2_SIZE_GAIJIRAM, 1, f);
         fclose(f);
     }
     f = fopen("da2_attrpal.dmp", "wb");

From 4bc47894de81c995c36592b2a1eeafb59e9846c1 Mon Sep 17 00:00:00 2001
From: Akamaki <97360908+akmed772@users.noreply.github.com>
Date: Mon, 24 Feb 2025 05:53:34 +0900
Subject: [PATCH 02/58] safety read/write vram memory

---
 src/video/vid_ps55da2.c | 112 ++++++++++++++++++++--------------------
 1 file changed, 57 insertions(+), 55 deletions(-)

diff --git a/src/video/vid_ps55da2.c b/src/video/vid_ps55da2.c
index a517b7e6c..7c7891926 100644
--- a/src/video/vid_ps55da2.c
+++ b/src/video/vid_ps55da2.c
@@ -437,12 +437,19 @@ typedef struct {
 
 /* safety read for internal functions */
 uint32_t
-DA2_readvram_s(uint32_t addr, da2_t *da2)
+DA2_vram_r(uint32_t addr, da2_t *da2)
 {
     if (addr & ~da2->vram_mask)
         return -1;
     return da2->vram[addr];
 }
+/* safety write for internal functions */
+void
+DA2_vram_w(uint32_t addr, uint8_t val, da2_t *da2)
+{
+    da2->vram[addr & da2->vram_mask] = val;
+    return;
+}
 
 void
 DA2_WritePlaneDataWithBitmask(uint32_t destaddr, const uint16_t mask, pixel32 *srcpx, da2_t *da2)
@@ -455,10 +462,10 @@ DA2_WritePlaneDataWithBitmask(uint32_t destaddr, const uint16_t mask, pixel32 *s
     destaddr <<= 3;
     /* read destination data with big endian order */
     for (int i = 0; i < 8; i++)
-        writepx[i] = DA2_readvram_s((destaddr + 24) | i, da2)
-            | (DA2_readvram_s((destaddr + 16) | i, da2) << 8)
-            | (DA2_readvram_s((destaddr + 8) | i, da2) << 16)
-            | (DA2_readvram_s((destaddr + 0) | i, da2) << 24);
+        writepx[i] = DA2_vram_r((destaddr + 24) | i, da2)
+            | (DA2_vram_r((destaddr + 16) | i, da2) << 8)
+            | (DA2_vram_r((destaddr + 8) | i, da2) << 16)
+            | (DA2_vram_r((destaddr + 0) | i, da2) << 24);
 
     DA2_VidSeq32 mask32in;
     mask32in.d = (uint32_t) mask;
@@ -487,8 +494,8 @@ DA2_WritePlaneDataWithBitmask(uint32_t destaddr, const uint16_t mask, pixel32 *s
         }
     }
     for (int i = 0; i < 8; i++) {
-        da2->vram[(da2->vram_mask & destaddr) | i]       = (writepx[i] >> 24) & 0xff;
-        da2->vram[(da2->vram_mask & (destaddr + 8)) | i] = (writepx[i] >> 16) & 0xff;
+        DA2_vram_w(destaddr | i, (writepx[i] >> 24) & 0xff, da2);
+        DA2_vram_w((destaddr + 8) | i, (writepx[i] >> 16) & 0xff, da2);
     }
 }
 
@@ -531,10 +538,10 @@ DA2_CopyPlaneDataWithBitmask(uint32_t srcaddr, uint32_t destaddr, uint16_t mask,
     srcaddr &= 0xfffffffe;
     srcaddr <<= 3;
     for (int i = 0; i < 8; i++)
-        srcpx.p8[i] = DA2_readvram_s((srcaddr + 24) | i, da2)
-            | (DA2_readvram_s((srcaddr + 16) | i, da2) << 8)
-            | (DA2_readvram_s((srcaddr + 8) | i, da2) << 16)
-            | (DA2_readvram_s((srcaddr + 0) | i, da2) << 24);
+        srcpx.p8[i] = DA2_vram_r((srcaddr + 24) | i, da2)
+            | (DA2_vram_r((srcaddr + 16) | i, da2) << 8)
+            | (DA2_vram_r((srcaddr + 8) | i, da2) << 16)
+            | (DA2_vram_r((srcaddr + 0) | i, da2) << 24);
 
     DA2_WritePlaneDataWithBitmask(destaddr, mask, &srcpx, da2);
 }
@@ -560,7 +567,7 @@ pixel1tohex(uint32_t addr, int index, da2_t *da2)
 {
     uint8_t pixeldata = 0;
     for (int j = 0; j < 8; j++) {
-        if (da2->vram[(da2->vram_mask & (addr << 3)) | j] & (1 << (7 - index)))
+        if (DA2_vram_r(((addr << 3) | j) & (1 << (7 - index)), da2))
             pixeldata++;
     }
     return pixeldata;
@@ -1724,8 +1731,8 @@ da2_render_text(da2_t *da2)
         int       chr_wide = 0;
         // da2_log("\nda2ma: %x, da2sc: %x\n", da2->ma, da2->sc);
         for (x = 0; x < da2->hdisp; x += 13) {
-            chr  = da2->cram[(da2->ma) & da2->vram_display_mask];
-            attr = da2->cram[((da2->ma) + 1) & da2->vram_display_mask];
+            chr  = da2->cram[(da2->ma) & DA2_MASK_CRAM];
+            attr = da2->cram[((da2->ma) + 1) & DA2_MASK_CRAM];
             // if(chr!=0x20) da2_log("chr: %x, attr: %x    ", chr, attr);
             if (da2->attrc[LV_PAS_STATUS_CNTRL] & 0x80) /* IO 3E8h, Index 1Dh */
             {                                           /* --Parse attribute byte in color mode-- */
@@ -1860,9 +1867,9 @@ da2_render_textm3(da2_t *da2)
         int       chr_wide = 0;
         // da2_log("\nda2ma: %x, da2sc: %x\n", da2->ma, da2->sc);
         for (x = 0; x < da2->hdisp; x += 13) {
-            chr     = da2->vram[(DA2_VM03_BASECHR + (da2->ma)) & da2->vram_mask];
-            attr    = da2->vram[((DA2_VM03_BASECHR + (da2->ma)) + 1) & da2->vram_mask];
-            extattr = da2->vram[((DA2_VM03_BASEEXATTR + (da2->ma)) + 1) & da2->vram_mask];
+            chr     = DA2_vram_r(DA2_VM03_BASECHR + (da2->ma), da2);
+            attr    = DA2_vram_r(DA2_VM03_BASECHR + (da2->ma) + 1, da2);
+            extattr = DA2_vram_r(DA2_VM03_BASEEXATTR + (da2->ma) + 1, da2);
             // if(chr!=0x20) da2_log("addr: %x, chr: %x, attr: %x    ", (DA2_VM03_BASECHR + da2->ma << 1) & da2->vram_mask, chr, attr);
             bg = attr >> 4;
             // if (da2->blink) bg &= ~0x8;
@@ -1880,7 +1887,7 @@ da2_render_textm3(da2_t *da2)
                 /* Stay drawing if the char code is DBCS and not at last column. */
                 if (chr_wide) {
                     /* Get high DBCS code from the next video address */
-                    chr_dbcs = da2->vram[(DA2_VM03_BASECHR + (da2->ma) + 2) & da2->vram_mask];
+                    chr_dbcs = DA2_vram_r(DA2_VM03_BASECHR + (da2->ma) + 2, da2);
                     chr_dbcs <<= 8;
                     chr_dbcs |= chr;
                     /* Get the font pattern */
@@ -2116,9 +2123,7 @@ da2_recalctimings(da2_t *da2)
             da2_log("Set videomode to PS/55 Mode 03 text.\n");
             da2->render            = da2_render_textm3;
             da2->vram_display_mask = DA2_MASK_CRAM;
-        }
-        /* PS/55 text(color/mono) */
-        else {
+        } else { /* PS/55 text(color/mono) */
             da2_log("Set videomode to PS/55 Mode 8/E text.\n");
             da2->render            = da2_render_text;
             da2->vram_display_mask = DA2_MASK_CRAM;
@@ -2243,19 +2248,19 @@ da2_gdcropB(uint32_t addr,uint8_t bitmask, da2_t *da2)
                 case 0: /*Set*/
                     // da2->vram[addr | i] = (da2->gdcinput[i] & bitmask) | (da2->gdcsrc[i] & ~bitmask);
                     // da2->vram[addr | i] = (da2->gdcinput[i] & bitmask) | (da2->vram[addr | i] & ~bitmask);
-                    da2->vram[addr | i] = (da2->gdcinput[i] & bitmask) | (da2->vram[addr | i] & ~bitmask);
+                    DA2_vram_w(addr | i, (da2->gdcinput[i] & bitmask) | (da2->vram[addr | i] & ~bitmask), da2);
                     break;
                 case 1: /*AND*/
                     // da2->vram[addr | i] = (da2->gdcinput[i] | ~bitmask) & da2->gdcsrc[i];
-                    da2->vram[addr | i] = ((da2->gdcinput[i] & da2->gdcsrc[i]) & bitmask) | (da2->vram[addr | i] & ~bitmask);
+                    DA2_vram_w(addr | i, ((da2->gdcinput[i] & da2->gdcsrc[i]) & bitmask) | (da2->vram[addr | i] & ~bitmask), da2);
                     break;
                 case 2: /*OR*/
                     // da2->vram[addr | i] = (da2->gdcinput[i] & bitmask) | da2->gdcsrc[i];
-                    da2->vram[addr | i] = ((da2->gdcinput[i] | da2->gdcsrc[i]) & bitmask) | (da2->vram[addr | i] & ~bitmask);
+                    DA2_vram_w(addr | i,  ((da2->gdcinput[i] | da2->gdcsrc[i]) & bitmask) | (da2->vram[addr | i] & ~bitmask), da2);
                     break;
                 case 3: /*XOR*/
                     // da2->vram[addr | i] = (da2->gdcinput[i] & bitmask) ^ da2->gdcsrc[i];
-                    da2->vram[addr | i] = ((da2->gdcinput[i] ^ da2->gdcsrc[i]) & bitmask) | (da2->vram[addr | i] & ~bitmask);
+                    DA2_vram_w(addr | i,  ((da2->gdcinput[i] ^ da2->gdcsrc[i]) & bitmask) | (da2->vram[addr | i] & ~bitmask), da2);
                     break;
             }
         }
@@ -2273,30 +2278,30 @@ da2_gdcropW(uint32_t addr, uint16_t bitmask, da2_t *da2)
                 case 0: /*Set*/
                     // da2->vram[addr | i]       = (da2->gdcinput[i] & bitmask_l) | (da2->gdcsrc[i] & ~bitmask_l);
                     // da2->vram[(addr + 8) | i] = ((da2->gdcinput[i] >> 8) & bitmask_h) | ((da2->gdcsrc[i] >> 8) & ~bitmask_h);
-                    da2->vram[addr | i]       = (da2->gdcinput[i] & bitmask_l) | (da2->vram[addr | i] & ~bitmask_l);
-                    da2->vram[(addr + 8) | i] = ((da2->gdcinput[i] >> 8) & bitmask_h)
-                        | (da2->vram[(addr + 8) | i] & ~bitmask_h);
+                    DA2_vram_w(addr | i, (da2->gdcinput[i] & bitmask_l) | (da2->vram[addr | i] & ~bitmask_l), da2);
+                    DA2_vram_w((addr + 8) | i, ((da2->gdcinput[i] >> 8) & bitmask_h)
+                        | (da2->vram[(addr + 8) | i] & ~bitmask_h), da2);
                     break;
                 case 1: /*AND*/
                     // da2->vram[addr | i]       = (da2->gdcinput[i] | ~bitmask_l) & da2->gdcsrc[i];
                     // da2->vram[(addr + 8) | i] = ((da2->gdcinput[i] >> 8) | ~bitmask_h) & (da2->gdcsrc[i] >> 8);
-                    da2->vram[addr | i]       = ((da2->gdcinput[i] & da2->gdcsrc[i]) & bitmask_l) | (da2->vram[addr | i] & ~bitmask_l);
-                    da2->vram[(addr + 8) | i] = (((da2->gdcinput[i] >> 8) & (da2->gdcsrc[i] >> 8)) & bitmask_h)
-                        | (da2->vram[(addr + 8) | i] & ~bitmask_h);
+                    DA2_vram_w(addr | i, ((da2->gdcinput[i] & da2->gdcsrc[i]) & bitmask_l) | (da2->vram[addr | i] & ~bitmask_l), da2);
+                    DA2_vram_w((addr + 8) | i, (((da2->gdcinput[i] >> 8) & (da2->gdcsrc[i] >> 8)) & bitmask_h)
+                        | (da2->vram[(addr + 8) | i] & ~bitmask_h), da2);
                     break;
                 case 2: /*OR*/
                     // da2->vram[addr | i]       = (da2->gdcinput[i] & bitmask_l) | da2->gdcsrc[i];
                     // da2->vram[(addr + 8) | i] = ((da2->gdcinput[i] >> 8) & bitmask_h) | (da2->gdcsrc[i] >> 8);
-                    da2->vram[addr | i]       = ((da2->gdcinput[i] | da2->gdcsrc[i]) & bitmask_l) | (da2->vram[addr | i] & ~bitmask_l);
-                    da2->vram[(addr + 8) | i] = (((da2->gdcinput[i] >> 8) | (da2->gdcsrc[i] >> 8)) & bitmask_h)
-                        | (da2->vram[(addr + 8) | i] & ~bitmask_h);
+                    DA2_vram_w(addr | i, ((da2->gdcinput[i] | da2->gdcsrc[i]) & bitmask_l) | (da2->vram[addr | i] & ~bitmask_l), da2);
+                    DA2_vram_w((addr + 8) | i, (((da2->gdcinput[i] >> 8) | (da2->gdcsrc[i] >> 8)) & bitmask_h)
+                        | (da2->vram[(addr + 8) | i] & ~bitmask_h), da2);
                     break;
                 case 3: /*XOR*/
                     // da2->vram[addr | i]       = (da2->gdcinput[i] & bitmask_l) ^ da2->gdcsrc[i];
                     // da2->vram[(addr + 8) | i] = ((da2->gdcinput[i] >> 8) & bitmask_h) ^ (da2->gdcsrc[i] >> 8);
-                    da2->vram[addr | i]       = ((da2->gdcinput[i] ^ da2->gdcsrc[i]) & bitmask_l) | (da2->vram[addr | i] & ~bitmask_l);
-                    da2->vram[(addr + 8) | i] = (((da2->gdcinput[i] >> 8) ^ (da2->gdcsrc[i] >> 8)) & bitmask_h)
-                        | (da2->vram[(addr + 8) | i] & ~bitmask_h);
+                    DA2_vram_w(addr | i, ((da2->gdcinput[i] ^ da2->gdcsrc[i]) & bitmask_l) | (da2->vram[addr | i] & ~bitmask_l), da2);
+                    DA2_vram_w((addr + 8) | i, (((da2->gdcinput[i] >> 8) ^ (da2->gdcsrc[i] >> 8)) & bitmask_h)
+                        | (da2->vram[(addr + 8) | i] & ~bitmask_h), da2);
                     break;
             }
         }
@@ -2359,7 +2364,6 @@ da2_mmio_read(uint32_t addr, void *p)
         } else
             return da2->gdcla[da2->readplane];
     } else { /* text mode 3 */
-
         cycles -= video_timing_read_b;
         return da2->vram[addr];
     }
@@ -2371,8 +2375,7 @@ da2_mmio_readw(uint32_t addr, void *p)
     //da2_log("da2_readW: %x %x %x %x %x\n", da2->ioctl[LS_MMIO], da2->fctl[LF_MMIO_SEL], da2->fctl[LF_MMIO_MODE], da2->fctl[LF_MMIO_ADDR], addr);
     if (da2->ioctl[LS_MMIO] & 0x10) {
         return (uint16_t) da2_mmio_read(addr, da2) | (uint16_t) (da2_mmio_read(addr + 1, da2) << 8);
-    } else if (!(da2->ioctl[LS_MODE] & 1)) /* 16 color or 256 color mode */
-    {
+    } else if (!(da2->ioctl[LS_MODE] & 1)) {/* 16 color or 256 color mode */
         cycles -= video_timing_read_w;
         addr &= DA2_MASK_MMIO;
         for (int i = 0; i < 8; i++)
@@ -2417,8 +2420,8 @@ da2_mmio_write(uint32_t addr, uint8_t val, void *p)
 {
     da2_t *da2 = (da2_t *) p;
     // da2_log("da2_mmio_write %x %x\n", addr, val);
-    if ((addr & ~DA2_MASK_MMIO) != 0xA0000)
-        return;
+    // if ((addr & ~DA2_MASK_MMIO) != 0xA0000)
+    //     return;
     addr &= DA2_MASK_MMIO;
 
     if (da2->ioctl[LS_MMIO] & 0x10) {
@@ -2434,8 +2437,7 @@ da2_mmio_write(uint32_t addr, uint8_t val, void *p)
         }
         switch (da2->fctl[LF_MMIO_MODE]) {
             case 0xb0:                     /* Gaiji RAM 1011 0000 */
-                addr &= DA2_MASK_GAIJIRAM; /* safety access */
-                da2->mmio.ram[addr] = val;
+                da2->mmio.ram[addr & DA2_MASK_GAIJIRAM] = val;
                 break;
             case 0x10: /* Font ROM  0001 0000 */
                 /* Read-Only */
@@ -2509,7 +2511,7 @@ da2_mmio_write(uint32_t addr, uint8_t val, void *p)
             case 2: /* equiv to vga write mode 1 */
                 for (int i = 0; i < 8; i++)
                     if (da2->writemask & (1 << i))
-                        da2->vram[addr | i] = da2->gdcsrc[i];
+                        DA2_vram_w(addr | i, da2->gdcsrc[i], da2);
                 break;
             case 0:/* equiv to vga write mode 0 */
                 if (da2->gdcreg[LG_DATA_ROTATION] & 7)
@@ -2517,7 +2519,7 @@ da2_mmio_write(uint32_t addr, uint8_t val, void *p)
                 if (bitmask == 0xff && !(da2->gdcreg[LG_COMMAND] & 0x03) && (!da2->gdcreg[LG_ENABLE_SRJ])) {
                     for (int i = 0; i < 8; i++)
                         if (da2->writemask & (1 << i))
-                            da2->vram[addr | i] = val;
+                            DA2_vram_w(addr | i, val, da2);
                 } else {
                     for (int i = 0; i < 8; i++)
                         if (da2->gdcreg[LG_ENABLE_SRJ] & (1 << i))
@@ -2537,7 +2539,7 @@ da2_mmio_write(uint32_t addr, uint8_t val, void *p)
                 if (!(da2->gdcreg[LG_COMMAND] & 0x03) && (!da2->gdcreg[LG_ENABLE_SRJ])) {
                     for (int i = 0; i < 8; i++)
                         if (da2->writemask & (1 << i))
-                            da2->vram[addr | i] = (((val & (1 << i)) ? 0xff : 0) & bitmask) | (da2->gdcsrc[i] & ~bitmask);
+                            DA2_vram_w(addr | i, (((val & (1 << i)) ? 0xff : 0) & bitmask) | (da2->gdcsrc[i] & ~bitmask), da2);
                     //fprintf(da2->mmdbg_fp, "m1-1");
                 } else {
                     for (int i = 0; i < 8; i++)
@@ -2559,7 +2561,7 @@ da2_mmio_write(uint32_t addr, uint8_t val, void *p)
         // da2_log("%02x%02x%02x%02x\n", da2->vram[addr + 0], da2->vram[addr + 1], da2->vram[addr + 2], da2->vram[addr + 3]);
     } else { /*  mode 3h text */
         cycles -= video_timing_write_b;
-        da2->vram[addr] = val;
+        DA2_vram_w(addr, val, da2);
         da2->fullchange = 2;
     }
 }
@@ -2622,8 +2624,8 @@ da2_mmio_gc_writeW(uint32_t addr, uint16_t val, void *p)
         case 2:
             for (int i = 0; i < 8; i++)
                 if (da2->writemask & (1 << i)) {
-                    da2->vram[addr | i]       = da2->gdcsrc[i] & 0xff;
-                    da2->vram[(addr + 8) | i] = da2->gdcsrc[i] >> 8;
+                    DA2_vram_w(addr | i, da2->gdcsrc[i] & 0xff, da2);
+                    DA2_vram_w((addr + 8) | i, da2->gdcsrc[i] >> 8, da2);
                 }
             break;
         case 0:
@@ -2632,8 +2634,8 @@ da2_mmio_gc_writeW(uint32_t addr, uint16_t val, void *p)
             if (bitmask == 0xffff && !(da2->gdcreg[LG_COMMAND] & 0x03) && (!da2->gdcreg[LG_ENABLE_SRJ])) {
                 for (int i = 0; i < 8; i++)
                     if (da2->writemask & (1 << i)) {
-                        da2->vram[addr | i]       = val & 0xff;
-                        da2->vram[(addr + 8) | i] = val >> 8;
+                        DA2_vram_w(addr | i, val & 0xff, da2);
+                        DA2_vram_w((addr + 8) | i, val >> 8, da2);
                     }
             } else {
                 for (int i = 0; i < 8; i++)
@@ -2649,9 +2651,9 @@ da2_mmio_gc_writeW(uint32_t addr, uint16_t val, void *p)
             if (!(da2->gdcreg[LG_COMMAND] & 0x03) && (!da2->gdcreg[LG_ENABLE_SRJ])) {
                 for (int i = 0; i < 8; i++)
                     if (da2->writemask & (1 << i)) {
-                        uint16_t wdata            = (((val & (1 << i)) ? 0xffff : 0) & bitmask) | (da2->gdcsrc[i] & ~bitmask);
-                        da2->vram[addr | i]       = wdata & 0xff;
-                        da2->vram[(addr + 8) | i] = wdata >> 8;
+                        uint16_t wdata = (((val & (1 << i)) ? 0xffff : 0) & bitmask) | (da2->gdcsrc[i] & ~bitmask);
+                        DA2_vram_w(addr | i, wdata & 0xff, da2);
+                        DA2_vram_w((addr + 8) | i, wdata >> 8, da2);
                     }
             } else {
                 for (int i = 0; i < 8; i++)

From 50300a1d9a949ccfb59b2b00a8bd9b794c802527 Mon Sep 17 00:00:00 2001
From: Akamaki <97360908+akmed772@users.noreply.github.com>
Date: Tue, 25 Feb 2025 10:08:12 +0900
Subject: [PATCH 03/58] Add line drawing function (incomplete)

remains a bug that the second hand doesn't appear in the Clock app of Windows 3.1.
---
 src/video/vid_ps55da2.c | 163 ++++++++++++++++++++++++++++++++++------
 1 file changed, 138 insertions(+), 25 deletions(-)

diff --git a/src/video/vid_ps55da2.c b/src/video/vid_ps55da2.c
index 7c7891926..f157d8fca 100644
--- a/src/video/vid_ps55da2.c
+++ b/src/video/vid_ps55da2.c
@@ -87,8 +87,9 @@
 #define DA2_BLT_CCOPYF            3
 #define DA2_BLT_CCOPYR            4
 #define DA2_BLT_CPUTCHAR          5
-#define DA2_BLT_CDONE             6
-#define DA2_BLT_CLOAD             7
+#define DA2_BLT_CLINE             6
+#define DA2_BLT_CDONE             7
+#define DA2_BLT_CLOAD             8
 /* POS ID = 0xeffe : Display Adapter II, III, V  */
 #define DA2_POSID_H 0xef
 #define DA2_POSID_L 0xfe
@@ -253,7 +254,9 @@
 
 #ifdef ENABLE_DA2_LOG
 #    define ENABLE_DA2_DEBUGBLT 1
+// #    define ENABLE_DA2_DEBUGVRAM 1
 // #    define ENABLE_DA2_DEBUGFULLSCREEN 1
+// #    define ENABLE_DA2_DEBUGMONWAIT 1
 int da2_do_log = ENABLE_DA2_LOG;
 
 static void
@@ -388,6 +391,7 @@ typedef struct da2_t {
 #ifdef ENABLE_DA2_DEBUGBLT
         int32_t   *debug_reg;            // for debug
         int        debug_reg_ip;         // for debug
+        int        debug_exesteps;
 #endif
         int        payload_addr;
         pc_timer_t timer;
@@ -398,14 +402,20 @@ typedef struct da2_t {
         int32_t    srcaddr;
         int32_t    size_x, tile_w;
         int32_t    size_y;
+        int32_t    dest_x;
+        int32_t    dest_y;
         int16_t    destpitch;
         int16_t    srcpitch;
         int32_t    fcolor;
         int32_t    maskl, maskr;
+        int32_t    count;
+        // int32_t    countj;
+        int32_t    d;
+        int        octdir;
         int        x, y;
     } bitblt;
 
-#ifdef ENABLE_DA2_DEBUGBLT
+#ifdef ENABLE_DA2_DEBUGVRAM
     FILE    *mmdbg_fp;
     FILE    *mmrdbg_fp;
     uint32_t mmdbg_vidaddr;
@@ -520,7 +530,7 @@ Param   Desc
 0A      Plane Mask?
 0B      ROP?(8h or 200h + 0-3h)
 0D
-20      Exec (1)
+20      Exec (1) or Exec without reset regs (21h)
 21      ?
 22      ?
 23      Tile W
@@ -718,6 +728,7 @@ da2_bitblt_load(da2_t *da2)
         da2->bitblt.debug_reg_ip++;
         if (da2->bitblt.debug_reg_ip >= DA2_DEBUG_BLTLOG_MAX)
             da2->bitblt.debug_reg_ip = 0;
+        da2->bitblt.debug_exesteps = 0;
 #endif
         da2->bitblt.bitshift_destr = ((da2->bitblt.reg[0x3] >> 4) & 0x0f); /* set bit shift */
         da2->bitblt.raster_op      = da2->bitblt.reg[0x0b] & 0x03;         /* 01 AND, 03 XOR */
@@ -749,11 +760,14 @@ da2_bitblt_load(da2_t *da2)
         da2->bitblt.exec   = DA2_BLT_CDONE;
         timer_set_delay_u64(&da2->bitblt.timer, da2->bitblt.timerspeed);
 
-        if (da2->bitblt.reg[0x2f] < 0x80) /* MS Paint 3.1 will cause hang up in 256 color mode */
-        {
-            da2_log("bitblt not executed 2f:%x\n", da2->bitblt.reg[0x2f]);
-            da2->bitblt.exec = DA2_BLT_CDONE;
-        } else if (da2->bitblt.reg[0x10] == 0xbc04) { /* Put char used by OS/2 (i'm not sure what the condition is) */
+        // if (da2->bitblt.reg[0x2f] < 0x80) /* MS Paint 3.1 will cause hang up in 256 color mode */
+        // {
+        //     da2_log("bitblt not executed 2f:%x\n", da2->bitblt.reg[0x2f]);
+        //     da2->bitblt.exec = DA2_BLT_CDONE;
+        // } else
+
+        /* Put DBCS char used by OS/2 (i'm not sure what the condition is) */
+        if (da2->bitblt.reg[0x10] == 0xbc04) {
             da2->bitblt.exec = DA2_BLT_CPUTCHAR;
             /* Todo: addressing */
             // if (da2->bitblt.reg[0x2F] == 0x90) /* destaddr -= 2, length += 1; */
@@ -776,7 +790,9 @@ da2_bitblt_load(da2_t *da2)
                     da2->bitblt.reg[0x29] % (da2->rowoffset * 2), da2->bitblt.reg[0x29] / (da2->rowoffset * 2),
                     da2->bitblt.size_x, da2->bitblt.size_y, sjis_h, sjis_l);
 #endif
-        } else if (da2->bitblt.reg[0x10] == 0x0004 || da2->bitblt.reg[0x10] == 0x0E04) {
+        }
+        /* Put SBCS char used by OS/2 */
+        else if (da2->bitblt.reg[0x10] == 0x0004 || da2->bitblt.reg[0x10] == 0x0E04) {
             da2->bitblt.exec    = DA2_BLT_CPUTCHAR;
             da2->bitblt.srcaddr = da2->bitblt.reg[0x12] * 64 + 2 + DA2_FONTROM_BASESBCS;
             da2->bitblt.destaddr += 2;
@@ -792,7 +808,36 @@ da2_bitblt_load(da2_t *da2)
                     da2->bitblt.size_x, da2->bitblt.size_y, sjis_h, sjis_l);
 #endif
         }
-        /* Fill a rectangle(or draw a line) */
+        /* Draw a line */
+        else if (da2->bitblt.reg[0x5] == 0x43) {
+            // da2_log("drawline x=%d, y=%d, 24=%d, 2A=%d, 2B=%d, 2D=%d\n",
+            //         da2->bitblt.reg[0x29] % (da2->rowoffset * 2), da2->bitblt.reg[0x29] / (da2->rowoffset * 2),
+            //         n24, n2a, n2b, da2->bitblt.reg[0x2D]);
+            da2->bitblt.exec = DA2_BLT_CLINE;
+            da2->bitblt.dest_x    = (da2->bitblt.reg[0x32] & 0x7ff);
+            da2->bitblt.dest_y    = (da2->bitblt.reg[0x34] & 0x7ff);
+            da2->bitblt.size_x    = abs((da2->bitblt.reg[0x33] & 0x7ff) - (da2->bitblt.reg[0x32] & 0x7ff));
+            da2->bitblt.size_y    = abs((da2->bitblt.reg[0x35] & 0x7ff) - (da2->bitblt.reg[0x34] & 0x7ff));
+            // da2->bitblt.raster_op = 0;
+            da2->bitblt.count    = 0;
+            da2->bitblt.octdir    = da2->bitblt.reg[0x2D];
+            da2->bitblt.bitshift_destr = 0;
+            if (da2->bitblt.octdir & 0x04) /* dX > dY */
+                da2->bitblt.d = 2 * da2->bitblt.size_y - da2->bitblt.size_x;
+            else
+                da2->bitblt.d = 2 * da2->bitblt.size_x - da2->bitblt.size_y;
+            da2_log("drawline x=%d, y=%d, dx=%d, dy=%d, oct=%d\n",
+                    da2->bitblt.dest_x, da2->bitblt.dest_y,
+                    da2->bitblt.size_x, da2->bitblt.size_y, da2->bitblt.octdir);
+            // FILE *f = fopen("da2_drawline.csv", "a");
+            // if (f != NULL) {
+            //     fprintf(f,"drawline,%d,%d,%d,%d,%d,%d\n",
+            //         da2->bitblt.reg[0x29] % (da2->rowoffset * 2), da2->bitblt.reg[0x29] / (da2->rowoffset * 2),
+            //         n24, n2a, n2b, da2->bitblt.reg[0x2D]);
+            //     fclose(f);
+            // }
+        }
+        /* Fill a rectangle (or draw a horizontal / vertical line) */
         else if ((da2->bitblt.reg[0x5] & 0xfff0) == 0x40 && da2->bitblt.reg[0x3D] == 0) {
             da2_log("fillrect x=%d, y=%d, w=%d, h=%d, c=%d, 2f=%x, rowcount=%x\n",
                     da2->bitblt.reg[0x29] % (da2->rowoffset * 2), da2->bitblt.reg[0x29] / (da2->rowoffset * 2),
@@ -824,7 +869,7 @@ da2_bitblt_load(da2_t *da2)
                     da2->bitblt.reg[0x29] % (da2->rowoffset * 2), da2->bitblt.reg[0x29] / (da2->rowoffset * 2),
                     da2->bitblt.size_x, da2->bitblt.size_y);
         }
-        /* Block transfer (range copy) */
+        /* Block copy */
         else if ((da2->bitblt.reg[0x5] & 0xfff0) == 0x1040 && da2->bitblt.reg[0x3D] == 0x00) {
             da2->bitblt.exec    = DA2_BLT_CCOPYF;
             da2->bitblt.srcaddr = da2->bitblt.reg[0x2A];
@@ -855,7 +900,13 @@ da2_bitblt_exec(void *p)
 {
     da2_t *da2 = (da2_t *) p;
     // timer_set_delay_u64(&da2->bitblt.timer, da2->bitblt.timerspeed);
+#ifdef ENABLE_DA2_DEBUGBLT
+    if(!(da2->bitblt.debug_exesteps & 0xf))
+        da2_log("bitblt_exec: %d %d\n", da2->bitblt.exec, da2->bitblt.debug_exesteps);
+    da2->bitblt.debug_exesteps++;
+#else
     da2_log("bitblt_exec: %d\n", da2->bitblt.exec);
+#endif
     // da2_log("blt %x %x %x %x\n", da2->bitblt.destaddr, da2->bitblt.x, da2->bitblt.y);
     switch (da2->bitblt.exec) {
         case DA2_BLT_CIDLE:
@@ -865,6 +916,60 @@ da2_bitblt_exec(void *p)
             da2->bitblt.indata = 0;
             da2_bitblt_load(da2);
             break;
+        case DA2_BLT_CLINE:
+            int pos_x = da2->bitblt.dest_x + da2->bitblt.x;
+            int pos_y = da2->bitblt.dest_y + da2->bitblt.y;
+
+            /* Draw a dot */
+            // outb(0x680, da2->bitblt.octdir);
+            // da2_log("point: %d %d %d %d %d\n", pos_x, pos_y, da2->bitblt.d, da2->bitblt.x, da2->bitblt.y);
+            int destaddr = pos_y * (da2->rowoffset * 2) + pos_x / 8;
+            int pixelmask = pos_x % 16;
+            if (pixelmask >= 8)
+                pixelmask = (0x8000 >> (pixelmask - 8));
+            else
+                pixelmask = (0x80 >> pixelmask);
+            DA2_DrawColorWithBitmask(destaddr, da2->bitblt.fcolor, pixelmask, da2);
+            // da2_log("draw: %x %x %x\n", destaddr, da2->bitblt.fcolor, pixelmask);
+            da2->bitblt.count++;
+
+            /* Bresenham's line */
+            if (da2->bitblt.octdir & 0x04) { /* dX > dY */
+                if (da2->bitblt.octdir & 0x02) {
+                    da2->bitblt.x++;
+                } else {
+                    da2->bitblt.x--;
+                }
+                if (da2->bitblt.d >= 0) {
+                    da2->bitblt.d -= (2 * da2->bitblt.size_x);
+                    if (da2->bitblt.octdir & 0x01) {
+                        da2->bitblt.y++;
+                    } else {
+                        da2->bitblt.y--;
+                    }
+                }
+                da2->bitblt.d += (2 * da2->bitblt.size_y);
+                if (da2->bitblt.count >= da2->bitblt.size_x)
+                    da2->bitblt.exec = DA2_BLT_CDONE;
+            } else {
+                if (da2->bitblt.octdir & 0x01) {
+                    da2->bitblt.y++;
+                } else {
+                    da2->bitblt.y--;
+                }
+                if (da2->bitblt.d >= 0) {
+                    da2->bitblt.d -= (2 * da2->bitblt.size_y);
+                    if (da2->bitblt.octdir & 0x02) {
+                        da2->bitblt.x++;
+                    } else {
+                        da2->bitblt.x--;
+                    }
+                }
+                da2->bitblt.d += (2 * da2->bitblt.size_x);
+                if (da2->bitblt.count >= da2->bitblt.size_y)
+                    da2->bitblt.exec = DA2_BLT_CDONE;
+            }
+            break;
         case DA2_BLT_CFILLRECT:
             // da2_log("%x %x %x\n", da2->bitblt.destaddr, da2->bitblt.x, da2->bitblt.y);
             if (da2->bitblt.x >= da2->bitblt.size_x - 1) {
@@ -982,11 +1087,14 @@ da2_bitblt_exec(void *p)
             // da2->bitblt.srcaddr = da2->bitblt.reg[0x12] * 72 + (da2->bitblt.x * 2 ) + (da2->bitblt.y * 3) + 2;
             break;
         case DA2_BLT_CDONE:
-            /* initialize regs for debug dump */
-            for (int i = 0; i < DA2_BLT_REGSIZE; i++) {
-                if (da2->bitblt.reg[i] != DA2_DEBUG_BLT_NEVERUSED)
-                    da2->bitblt.reg[i] = DA2_DEBUG_BLT_USEDRESET;
-            }
+            if (!(da2->bitblt.reg[0x20] & 0x20)) {
+                /* initialize regs and set magic value for debug dump */
+                for (int i = 0; i < DA2_BLT_REGSIZE; i++) {
+                    if (da2->bitblt.reg[i] != DA2_DEBUG_BLT_NEVERUSED)
+                        da2->bitblt.reg[i] = DA2_DEBUG_BLT_USEDRESET;
+                }
+            } else /* without init regs */
+                da2->bitblt.reg[0x20] = 0; /* need to stop execution */
             if (da2->bitblt.indata)
                 da2->bitblt.exec = DA2_BLT_CLOAD;
             else
@@ -1013,7 +1121,7 @@ da2_bitblt_dopayload(void *priv)
     // }
     if (da2->bitblt.exec != DA2_BLT_CIDLE)
         da2_bitblt_exec(da2);
-    else if (da2->bitblt.indata) /* for OS/2 J1.3 */
+    else if (da2->bitblt.indata)
     {
         if (da2->bitblt.exec == DA2_BLT_CIDLE) {
             da2->bitblt.exec = DA2_BLT_CLOAD;
@@ -1339,9 +1447,10 @@ da2_in(uint16_t addr, void *p)
                 }
                 if (da2->bitblt.indata)
                     temp |= 0x08;
-                da2_log("DA2 In %04X(%02X) %04X %04X:%04X\n", addr, da2->ioctladdr,temp, cs >> 4, cpu_state.pc);
+#ifdef ENABLE_DA2_DEBUGMONWAIT
+                da2_log("DA2 In %04X(%02X) %04X %04X:%04X\n", addr, da2->ioctladdr, temp, cs >> 4, cpu_state.pc);
+#endif
             }
-            // da2_log("DA2 In %04X(%02X) %04X %04X:%04X\n", addr, da2->ioctladdr,temp, cs >> 4, cpu_state.pc);
             break;
         case LF_INDEX:
             temp = da2->fctladdr;
@@ -2381,7 +2490,7 @@ da2_mmio_readw(uint32_t addr, void *p)
         for (int i = 0; i < 8; i++)
             da2->gdcla[i] = (uint16_t) (da2->vram[(addr << 3) | i]) | ((uint16_t) (da2->vram[((addr << 3) + 8) | i]) << 8); /* read vram into latch */
 
-#ifdef ENABLE_DA2_DEBUGBLT
+#ifdef ENABLE_DA2_DEBUGVRAM
         ////debug
         // if (((int)addr - (int)da2->mmrdbg_vidaddr) > 2 || (((int)da2->mmrdbg_vidaddr - (int)addr) > 2) || da2->mmrdbg_vidaddr == addr)
         //{
@@ -2465,7 +2574,7 @@ da2_mmio_write(uint32_t addr, uint8_t val, void *p)
         /* align bitmask to even address */
         if (addr & 1) bitmask = da2->gdcreg[LG_BIT_MASK_HIGH];
         else  bitmask = da2->gdcreg[LG_BIT_MASK_LOW];
-#ifdef ENABLE_DA2_DEBUGBLT
+#ifdef ENABLE_DA2_DEBUGVRAM
         da2_log("da2_wB %x %02x\n", addr, val);
         // if (!(da2->gdcreg[LG_COMMAND] & 0x08))
         //{
@@ -2577,7 +2686,7 @@ da2_mmio_gc_writeW(uint32_t addr, uint16_t val, void *p)
     uint16_t bitmask = da2->gdcreg[LG_BIT_MASK_HIGH];
     bitmask <<= 8;
     bitmask |= (uint16_t) da2->gdcreg[LG_BIT_MASK_LOW];
-#ifdef ENABLE_DA2_DEBUGBLT
+#ifdef ENABLE_DA2_DEBUGVRAM
     // if (!(da2->gdcreg[LG_COMMAND] & 0x08))
     //{
     if (((int) addr - (int) da2->mmdbg_vidaddr) > 2 || (((int) da2->mmdbg_vidaddr - (int) addr) > 2) || da2->mmdbg_vidaddr == addr) {
@@ -3061,9 +3170,11 @@ da2_init(UNUSED(const device_t *info))
     memset(da2->bitblt.reg, 0xfe, DA2_BLT_REGSIZE * sizeof(uint32_t)); /* clear memory */
 #ifdef ENABLE_DA2_DEBUGBLT
     da2->bitblt.debug_reg    = malloc(DA2_DEBUG_BLTLOG_MAX * DA2_DEBUG_BLTLOG_SIZE);
+    da2->bitblt.debug_reg_ip = 0;
+#endif
+#ifdef ENABLE_DA2_DEBUGVRAM
     da2->mmdbg_fp            = fopen("da2_mmiowdat.txt", "w");
     da2->mmrdbg_fp           = fopen("da2_mmiordat.txt", "w");
-    da2->bitblt.debug_reg_ip = 0;
 #endif
     da2->bitblt.payload_addr = 0;
     da2_reset(da2);
@@ -3162,11 +3273,13 @@ da2_close(void *p)
         }
         fclose(f);
     }
+    free(da2->bitblt.debug_reg);
+#endif
+#ifdef ENABLE_DA2_DEBUGVRAM
     if (da2->mmdbg_fp != NULL)
         fclose(da2->mmdbg_fp);
     if (da2->mmrdbg_fp != NULL)
         fclose(da2->mmrdbg_fp);
-    free(da2->bitblt.debug_reg);
 #endif
     free(da2->cram);
     free(da2->vram);

From e9b8bbd0be4c575e5d420fbb24956c660776bcdf Mon Sep 17 00:00:00 2001
From: Akamaki <97360908+akmed772@users.noreply.github.com>
Date: Tue, 25 Feb 2025 11:59:46 +0900
Subject: [PATCH 04/58] add bit invert operation in bitblt

---
 src/video/vid_ps55da2.c | 16 ++++++++--------
 1 file changed, 8 insertions(+), 8 deletions(-)

diff --git a/src/video/vid_ps55da2.c b/src/video/vid_ps55da2.c
index f157d8fca..4922c6d30 100644
--- a/src/video/vid_ps55da2.c
+++ b/src/video/vid_ps55da2.c
@@ -502,6 +502,8 @@ DA2_WritePlaneDataWithBitmask(uint32_t destaddr, const uint16_t mask, pixel32 *s
                 writepx[i] ^= srcpx->p8[i] & mask32.d;
                 break;
         }
+        if (da2->bitblt.raster_op & 0x20) /* NOT */
+            writepx[i] ^= 0xFFFFFFFF & mask32.d;
     }
     for (int i = 0; i < 8; i++) {
         DA2_vram_w(destaddr | i, (writepx[i] >> 24) & 0xff, da2);
@@ -524,11 +526,11 @@ Param   Desc
 01      Color
 03      Bit Shift
 04      Select plane?
-05      Dir(10 or 11) + Command?(40 or 48)
+05      Dir(1000h or 1100h) + Command?(40 or 48)
 08      Mask Left
 09      Mask Right
 0A      Plane Mask?
-0B      ROP?(8h or 200h + 0-3h)
+0B      ROP?(8h or 200h) + Bitop (0 None, 1 AND, 2 OR, 3 XOR)
 0D
 20      Exec (1) or Exec without reset regs (21h)
 21      ?
@@ -709,9 +711,9 @@ da2_bitblt_load(da2_t *da2)
         i++;
     }
     da2->bitblt.exec = DA2_BLT_CIDLE;
-    // /* clear payload memory */
-    // memset(da2->bitblt.payload, 0x00, DA2_BLT_MEMSIZE);
-    // da2->bitblt.payload_addr = 0;
+    /* clear payload memory */
+    memset(da2->bitblt.payload, 0x00, DA2_BLT_MEMSIZE);
+    da2->bitblt.payload_addr = 0;
     /* [89] 20: 0001 (1) then execute payload */
     if (da2->bitblt.reg[0x20] & 0x1) {
         /* clear payload memory */
@@ -731,7 +733,7 @@ da2_bitblt_load(da2_t *da2)
         da2->bitblt.debug_exesteps = 0;
 #endif
         da2->bitblt.bitshift_destr = ((da2->bitblt.reg[0x3] >> 4) & 0x0f); /* set bit shift */
-        da2->bitblt.raster_op      = da2->bitblt.reg[0x0b] & 0x03;         /* 01 AND, 03 XOR */
+        da2->bitblt.raster_op      = da2->bitblt.reg[0x0b] & 0x23;         /* 01 AND, 03 XOR */
         da2_log("bltload_exec: %x, rop: %x CS:PC=%4x:%4x\n", da2->bitblt.reg[0x5], da2->bitblt.reg[0x0b], CS, cpu_state.pc);
         // for (int i = 0; i <= 0xb; i++)
         // {
@@ -780,7 +782,6 @@ da2_bitblt_load(da2_t *da2)
             da2->bitblt.srcaddr = da2->bitblt.reg[0x12] * 72 + 2;
             da2->bitblt.destaddr += 2;
             da2->bitblt.srcpitch  = 0;
-            da2->bitblt.raster_op = da2->bitblt.reg[0x05] & 0x03; /* XOR */
             da2->bitblt.bitshift_destr += 1;
 #ifdef ENABLE_DA2_DEBUGBLT
             uint32_t sjis_h = IBMJtoSJIS(da2->bitblt.reg[0x12]) >> 8;
@@ -797,7 +798,6 @@ da2_bitblt_load(da2_t *da2)
             da2->bitblt.srcaddr = da2->bitblt.reg[0x12] * 64 + 2 + DA2_FONTROM_BASESBCS;
             da2->bitblt.destaddr += 2;
             da2->bitblt.srcpitch  = 0;
-            da2->bitblt.raster_op = da2->bitblt.reg[0x05] & 0x03; /* XOR */
             da2->bitblt.bitshift_destr += 1;
 #ifdef ENABLE_DA2_DEBUGBLT
             uint32_t sjis_h = IBMJtoSJIS(da2->bitblt.reg[0x12]) >> 8;

From 813afb8562d8cab16c7fdc88ad066015504584d7 Mon Sep 17 00:00:00 2001
From: Akamaki <97360908+akmed772@users.noreply.github.com>
Date: Wed, 26 Feb 2025 04:16:36 +0900
Subject: [PATCH 05/58] change bitblt exec speed slower

probably the real machine is slower than this
---
 src/video/vid_ps55da2.c | 40 ++++++----------------------------------
 1 file changed, 6 insertions(+), 34 deletions(-)

diff --git a/src/video/vid_ps55da2.c b/src/video/vid_ps55da2.c
index 4922c6d30..fa43ebeb6 100644
--- a/src/video/vid_ps55da2.c
+++ b/src/video/vid_ps55da2.c
@@ -532,7 +532,7 @@ Param   Desc
 0A      Plane Mask?
 0B      ROP?(8h or 200h) + Bitop (0 None, 1 AND, 2 OR, 3 XOR)
 0D
-20      Exec (1) or Exec without reset regs (21h)
+20      Exec (1) or Exec without reset regs? (21h)
 21      ?
 22      ?
 23      Tile W
@@ -762,12 +762,6 @@ da2_bitblt_load(da2_t *da2)
         da2->bitblt.exec   = DA2_BLT_CDONE;
         timer_set_delay_u64(&da2->bitblt.timer, da2->bitblt.timerspeed);
 
-        // if (da2->bitblt.reg[0x2f] < 0x80) /* MS Paint 3.1 will cause hang up in 256 color mode */
-        // {
-        //     da2_log("bitblt not executed 2f:%x\n", da2->bitblt.reg[0x2f]);
-        //     da2->bitblt.exec = DA2_BLT_CDONE;
-        // } else
-
         /* Put DBCS char used by OS/2 (i'm not sure what the condition is) */
         if (da2->bitblt.reg[0x10] == 0xbc04) {
             da2->bitblt.exec = DA2_BLT_CPUTCHAR;
@@ -810,15 +804,11 @@ da2_bitblt_load(da2_t *da2)
         }
         /* Draw a line */
         else if (da2->bitblt.reg[0x5] == 0x43) {
-            // da2_log("drawline x=%d, y=%d, 24=%d, 2A=%d, 2B=%d, 2D=%d\n",
-            //         da2->bitblt.reg[0x29] % (da2->rowoffset * 2), da2->bitblt.reg[0x29] / (da2->rowoffset * 2),
-            //         n24, n2a, n2b, da2->bitblt.reg[0x2D]);
             da2->bitblt.exec = DA2_BLT_CLINE;
             da2->bitblt.dest_x    = (da2->bitblt.reg[0x32] & 0x7ff);
             da2->bitblt.dest_y    = (da2->bitblt.reg[0x34] & 0x7ff);
             da2->bitblt.size_x    = abs((da2->bitblt.reg[0x33] & 0x7ff) - (da2->bitblt.reg[0x32] & 0x7ff));
             da2->bitblt.size_y    = abs((da2->bitblt.reg[0x35] & 0x7ff) - (da2->bitblt.reg[0x34] & 0x7ff));
-            // da2->bitblt.raster_op = 0;
             da2->bitblt.count    = 0;
             da2->bitblt.octdir    = da2->bitblt.reg[0x2D];
             da2->bitblt.bitshift_destr = 0;
@@ -829,13 +819,6 @@ da2_bitblt_load(da2_t *da2)
             da2_log("drawline x=%d, y=%d, dx=%d, dy=%d, oct=%d\n",
                     da2->bitblt.dest_x, da2->bitblt.dest_y,
                     da2->bitblt.size_x, da2->bitblt.size_y, da2->bitblt.octdir);
-            // FILE *f = fopen("da2_drawline.csv", "a");
-            // if (f != NULL) {
-            //     fprintf(f,"drawline,%d,%d,%d,%d,%d,%d\n",
-            //         da2->bitblt.reg[0x29] % (da2->rowoffset * 2), da2->bitblt.reg[0x29] / (da2->rowoffset * 2),
-            //         n24, n2a, n2b, da2->bitblt.reg[0x2D]);
-            //     fclose(f);
-            // }
         }
         /* Fill a rectangle (or draw a horizontal / vertical line) */
         else if ((da2->bitblt.reg[0x5] & 0xfff0) == 0x40 && da2->bitblt.reg[0x3D] == 0) {
@@ -922,7 +905,7 @@ da2_bitblt_exec(void *p)
 
             /* Draw a dot */
             // outb(0x680, da2->bitblt.octdir);
-            // da2_log("point: %d %d %d %d %d\n", pos_x, pos_y, da2->bitblt.d, da2->bitblt.x, da2->bitblt.y);
+            da2_log("point: %d %d %d %d %d\n", pos_x, pos_y, da2->bitblt.d, da2->bitblt.x, da2->bitblt.y);
             int destaddr = pos_y * (da2->rowoffset * 2) + pos_x / 8;
             int pixelmask = pos_x % 16;
             if (pixelmask >= 8)
@@ -1107,26 +1090,15 @@ da2_bitblt_dopayload(void *priv)
 {
     da2_t *da2 = (da2_t *) priv;
     timer_set_delay_u64(&da2->bitblt.timer, da2->bitblt.timerspeed);
-    // if (da2->bitblt.indata) /* for OS/2 J1.3 */
-    // {
-    //     if (da2->bitblt.exec == DA2_BLT_CIDLE) {
-    //         da2->bitblt.exec = DA2_BLT_CLOAD;
-    //         /* do all queues (ignore async executing) for OS/2 J1.3 commannd prompt that doesn't wait for idle */
-    //         da2_log("da2 Do bitblt\n");
-    //         while (da2->bitblt.exec != DA2_BLT_CIDLE) {
-    //             da2_bitblt_exec(da2);
-    //         }
-    //         da2_log("da2 End bitblt %x\n", da2->bitblt.exec);
-    //     }
-    // }
     if (da2->bitblt.exec != DA2_BLT_CIDLE)
         da2_bitblt_exec(da2);
-    else if (da2->bitblt.indata)
-    {
+    else if (da2->bitblt.indata) {
         if (da2->bitblt.exec == DA2_BLT_CIDLE) {
             da2->bitblt.exec = DA2_BLT_CLOAD;
             da2_bitblt_exec(da2);
         }
+    } else {
+        // timer_disable(&da2->bitblt.timer);
     }
 }
 
@@ -3189,7 +3161,7 @@ da2_init(UNUSED(const device_t *info))
     mem_mapping_disable(&da2->cmapping);
 
     timer_add(&da2->timer, da2_poll, da2, 0);
-    da2->bitblt.timerspeed = 1 * TIMER_USEC;
+    da2->bitblt.timerspeed = 10 * TIMER_USEC;
     timer_add(&da2->bitblt.timer, da2_bitblt_dopayload, da2, 0);
 
     return da2;

From 168910b72ff1820a2524b642496483a910ecc81b Mon Sep 17 00:00:00 2001
From: Cacodemon345 <wahil1976@outlook.com>
Date: Wed, 26 Feb 2025 11:39:04 +0600
Subject: [PATCH 06/58] All held-down keys are now released when Alt-Tab'ing
 out

Only applies when Raw Input is in use on Windows
---
 src/device/keyboard.c        | 14 ++++++++++++++
 src/include/86box/keyboard.h |  1 +
 src/qt/qt_mainwindow.cpp     |  2 ++
 3 files changed, 17 insertions(+)

diff --git a/src/device/keyboard.c b/src/device/keyboard.c
index 582dcf0b1..58d5a4724 100644
--- a/src/device/keyboard.c
+++ b/src/device/keyboard.c
@@ -332,6 +332,20 @@ keyboard_input(int down, uint16_t scan)
     }
 }
 
+void
+keyboard_all_up(void)
+{
+    for (unsigned short i = 0; i < 0x200; i++) {
+        if (recv_key_ui[i]) {
+            recv_key_ui[i] = 0;
+        }
+        if (recv_key[i]) {
+            recv_key[i] = 0;
+            key_process(i, 0);
+        }
+    }
+}
+
 static uint8_t
 keyboard_do_break(uint16_t scan)
 {
diff --git a/src/include/86box/keyboard.h b/src/include/86box/keyboard.h
index ec5c05775..110e4f760 100644
--- a/src/include/86box/keyboard.h
+++ b/src/include/86box/keyboard.h
@@ -269,6 +269,7 @@ extern void     keyboard_poll_host(void);
 extern void     keyboard_process(void);
 extern uint16_t keyboard_convert(int ch);
 extern void     keyboard_input(int down, uint16_t scan);
+extern void     keyboard_all_up(void);
 extern void     keyboard_update_states(uint8_t cl, uint8_t nl, uint8_t sl);
 extern uint8_t  keyboard_get_shift(void);
 extern void     keyboard_get_states(uint8_t *cl, uint8_t *nl, uint8_t *sl);
diff --git a/src/qt/qt_mainwindow.cpp b/src/qt/qt_mainwindow.cpp
index 3ee58f23b..62ac94f30 100644
--- a/src/qt/qt_mainwindow.cpp
+++ b/src/qt/qt_mainwindow.cpp
@@ -277,6 +277,8 @@ MainWindow::MainWindow(QWidget *parent)
             if (mouse_capture)
                 emit setMouseCapture(false);
 
+            keyboard_all_up();
+
             if (do_auto_pause && !dopause) {
                 auto_paused = 1;
                 plat_pause(1);

From 6d3816df64aac6cb0e604bab30239c3b4b888265 Mon Sep 17 00:00:00 2001
From: Cacodemon345 <wahil1976@outlook.com>
Date: Thu, 27 Feb 2025 01:03:00 +0600
Subject: [PATCH 07/58] x87: Fix Final Reality discolored screen for
 interpreter

---
 src/cpu/x87_ops.h | 41 +++++++++++++++++++++++++++++++++++++++++
 1 file changed, 41 insertions(+)

diff --git a/src/cpu/x87_ops.h b/src/cpu/x87_ops.h
index f1362bf76..276c32876 100644
--- a/src/cpu/x87_ops.h
+++ b/src/cpu/x87_ops.h
@@ -113,12 +113,33 @@ typedef union {
 static __inline void
 x87_push(double i)
 {
+#ifdef X87_INLINE_ASM
+    unsigned char buffer[10];
+#else
+    x87_conv_t test;
+#endif
 #ifdef USE_NEW_DYNAREC
     cpu_state.TOP--;
 #else
     cpu_state.TOP                    = (cpu_state.TOP - 1) & 7;
 #endif
     cpu_state.ST[cpu_state.TOP & 7] = i;
+
+#ifdef X87_INLINE_ASM
+    __asm volatile(""
+        :
+        :
+        : "memory");
+
+    __asm volatile("fldl %1\n"
+                    "fstpt %0\n" : "=m"(buffer) : "m"(i));
+
+    cpu_state.MM[cpu_state.TOP & 7].q = (*(uint64_t*)buffer);
+#else
+    x87_to80(i, &test);
+    cpu_state.MM[cpu_state.TOP & 7].q = test.eind.ll;
+#endif
+
 #ifdef USE_NEW_DYNAREC
     cpu_state.tag[cpu_state.TOP & 7] = TAG_VALID;
 #else
@@ -129,6 +150,11 @@ x87_push(double i)
 static __inline void
 x87_push_u64(uint64_t i)
 {
+#ifdef X87_INLINE_ASM
+    unsigned char buffer[10];
+#else
+    x87_conv_t test;
+#endif
     union {
         double   d;
         uint64_t ll;
@@ -142,6 +168,21 @@ x87_push_u64(uint64_t i)
     cpu_state.TOP                    = (cpu_state.TOP - 1) & 7;
 #endif
     cpu_state.ST[cpu_state.TOP & 7] = td.d;
+
+#ifdef X87_INLINE_ASM
+    __asm volatile(""
+        :
+        :
+        : "memory");
+
+    __asm volatile("fldl %1\n"
+                    "fstpt %0\n" : "=m"(buffer) : "m"(td.d));
+
+    cpu_state.MM[cpu_state.TOP & 7].q = (*(uint64_t*)buffer);
+#else
+    x87_to80(td.d, &test);
+    cpu_state.MM[cpu_state.TOP & 7].q = test.eind.ll;
+#endif
 #ifdef USE_NEW_DYNAREC
     cpu_state.tag[cpu_state.TOP & 7] = TAG_VALID;
 #else

From 3cde104ff683bb91c71768d32fbd8c8a26bf643b Mon Sep 17 00:00:00 2001
From: Akamaki <97360908+akmed772@users.noreply.github.com>
Date: Thu, 27 Feb 2025 08:29:06 +0900
Subject: [PATCH 08/58] add inversion ROPs (still in debug)

---
 src/video/vid_ps55da2.c | 75 ++++++++++++++++++++---------------------
 1 file changed, 37 insertions(+), 38 deletions(-)

diff --git a/src/video/vid_ps55da2.c b/src/video/vid_ps55da2.c
index fa43ebeb6..91752a7a8 100644
--- a/src/video/vid_ps55da2.c
+++ b/src/video/vid_ps55da2.c
@@ -387,13 +387,13 @@ typedef struct da2_t {
         int        bitshift_destr;
         int        raster_op;
         uint8_t    payload[DA2_BLT_MEMSIZE];
+        int        payload_addr;
         int32_t    reg[DA2_BLT_REGSIZE]; // must be signed int
 #ifdef ENABLE_DA2_DEBUGBLT
         int32_t   *debug_reg;            // for debug
         int        debug_reg_ip;         // for debug
         int        debug_exesteps;
 #endif
-        int        payload_addr;
         pc_timer_t timer;
         int64_t    timerspeed;
         int        exec;
@@ -409,7 +409,6 @@ typedef struct da2_t {
         int32_t    fcolor;
         int32_t    maskl, maskr;
         int32_t    count;
-        // int32_t    countj;
         int32_t    d;
         int        octdir;
         int        x, y;
@@ -460,7 +459,7 @@ DA2_vram_w(uint32_t addr, uint8_t val, da2_t *da2)
     da2->vram[addr & da2->vram_mask] = val;
     return;
 }
-
+/* write pixel data with rop (Note: bitmask must be in big endian) */
 void
 DA2_WritePlaneDataWithBitmask(uint32_t destaddr, const uint16_t mask, pixel32 *srcpx, da2_t *da2)
 {
@@ -487,7 +486,11 @@ DA2_WritePlaneDataWithBitmask(uint32_t destaddr, const uint16_t mask, pixel32 *s
     for (int i = 0; i < 8; i++) {
         if (da2->bitblt.bitshift_destr > 0)
             srcpx->p8[i] <<= 16 - da2->bitblt.bitshift_destr;
-        switch (da2->bitblt.raster_op) {
+        if (da2->bitblt.raster_op & 0x2010) /* NOT Src or NOT Pattern */
+            srcpx->p8[i] = ~srcpx->p8[i] & mask32.d;
+        if (da2->bitblt.raster_op & 0x20) /* Dest NOT */
+            writepx[i] = (~writepx[i] & mask32.d) | (writepx[i] & ~mask32.d);
+        switch (da2->bitblt.raster_op & 0x03) {
             case 0x00: /* None */
                 writepx[i] &= ~mask32.d;
                 writepx[i] |= srcpx->p8[i] & mask32.d;
@@ -502,8 +505,6 @@ DA2_WritePlaneDataWithBitmask(uint32_t destaddr, const uint16_t mask, pixel32 *s
                 writepx[i] ^= srcpx->p8[i] & mask32.d;
                 break;
         }
-        if (da2->bitblt.raster_op & 0x20) /* NOT */
-            writepx[i] ^= 0xFFFFFFFF & mask32.d;
     }
     for (int i = 0; i < 8; i++) {
         DA2_vram_w(destaddr | i, (writepx[i] >> 24) & 0xff, da2);
@@ -530,7 +531,13 @@ Param   Desc
 08      Mask Left
 09      Mask Right
 0A      Plane Mask?
-0B      ROP?(8h or 200h) + Bitop (0 None, 1 AND, 2 OR, 3 XOR)
+0B      ROP
+            2000 NOT Source
+            200  Source / Pattern
+            20   NOT Dest
+            10   NOT Pattern
+            8    ?
+            0-3  Bit op (0 None, 1 AND, 2 OR, 3 XOR)
 0D
 20      Exec (1) or Exec without reset regs? (21h)
 21      ?
@@ -602,7 +609,7 @@ print_bytetobin(uint8_t b)
         b <<= 1;
     }
 }
-// Convert internal char code to Shift JIS code
+/* Convert internal char code to Shift JIS code */
 inline int
 isKanji1(uint8_t chr)
 {
@@ -649,17 +656,19 @@ da2_bitblt_load(da2_t *da2)
 {
     uint32_t value32;
     uint64_t value64;
+#ifdef ENABLE_DA2_DEBUGBLT
     da2_log("bltload: loading params\n");
-    // da2_log("BitBlt memory:\n");
-    // if (da2->bitblt.payload[0] != 0)
-    //     for (int j = 0; j < DA2_BLT_MEMSIZE / 8; j++)
-    //     {
-    //         int i = j * 8;
-    //         da2_log("%02x %02x %02x %02x %02x %02x %02x %02x \n", da2->bitblt.payload[i], da2->bitblt.payload[i + 1], da2->bitblt.payload[i + 2], da2->bitblt.payload[i + 3],
-    //             da2->bitblt.payload[i + 4], da2->bitblt.payload[i + 5], da2->bitblt.payload[i + 6], da2->bitblt.payload[i + 7]);
-    //     }
+    da2_log("BitBlt memory:\n");
+    if (da2->bitblt.payload[0] != 0)
+        for (int j = 0; j < DA2_BLT_MEMSIZE / 8; j++)
+        {
+            int i = j * 8;
+            da2_log("%02x %02x %02x %02x %02x %02x %02x %02x \n", da2->bitblt.payload[i], da2->bitblt.payload[i + 1], da2->bitblt.payload[i + 2], da2->bitblt.payload[i + 3],
+                da2->bitblt.payload[i + 4], da2->bitblt.payload[i + 5], da2->bitblt.payload[i + 6], da2->bitblt.payload[i + 7]);
+        }
+#endif
     int i = 0;
-    while (i < DA2_BLT_MEMSIZE) {
+    while (i < da2->bitblt.payload_addr) {
         if (da2->bitblt.reg[0x20] & 0x1)
             break;
         switch (da2->bitblt.payload[i]) {
@@ -712,36 +721,26 @@ da2_bitblt_load(da2_t *da2)
     }
     da2->bitblt.exec = DA2_BLT_CIDLE;
     /* clear payload memory */
-    memset(da2->bitblt.payload, 0x00, DA2_BLT_MEMSIZE);
-    da2->bitblt.payload_addr = 0;
+    // memset(da2->bitblt.payload, 0x00, DA2_BLT_MEMSIZE);
+    // da2->bitblt.payload_addr = 0;
     /* [89] 20: 0001 (1) then execute payload */
     if (da2->bitblt.reg[0x20] & 0x1) {
         /* clear payload memory */
         memset(da2->bitblt.payload, 0x00, DA2_BLT_MEMSIZE);
         da2->bitblt.payload_addr = 0;
 #ifdef ENABLE_DA2_DEBUGBLT
-        for (i = 0; i < DA2_DEBUG_BLTLOG_SIZE; i++) {
-            // if(da2->bitblt.reg[i] != 0xfefe && da2->bitblt.reg[i] != DA2_DEBUG_BLT_NEVERUSED) da2_log("%02x: %04x (%d)\n",i, da2->bitblt.reg[i], da2->bitblt.reg[i]);
-        }
         for (i = 0; i < DA2_DEBUG_BLTLOG_SIZE; i++) {
             da2->bitblt.debug_reg[DA2_DEBUG_BLTLOG_SIZE * da2->bitblt.debug_reg_ip + i] = da2->bitblt.reg[i];
         }
         da2->bitblt.debug_reg[DA2_DEBUG_BLTLOG_SIZE * (da2->bitblt.debug_reg_ip + 1) - 1] = 0;
         da2->bitblt.debug_reg_ip++;
-        if (da2->bitblt.debug_reg_ip >= DA2_DEBUG_BLTLOG_MAX)
+        if ((DA2_DEBUG_BLTLOG_SIZE * da2->bitblt.debug_reg_ip) >= DA2_DEBUG_BLTLOG_MAX)
             da2->bitblt.debug_reg_ip = 0;
         da2->bitblt.debug_exesteps = 0;
 #endif
-        da2->bitblt.bitshift_destr = ((da2->bitblt.reg[0x3] >> 4) & 0x0f); /* set bit shift */
-        da2->bitblt.raster_op      = da2->bitblt.reg[0x0b] & 0x23;         /* 01 AND, 03 XOR */
         da2_log("bltload_exec: %x, rop: %x CS:PC=%4x:%4x\n", da2->bitblt.reg[0x5], da2->bitblt.reg[0x0b], CS, cpu_state.pc);
-        // for (int i = 0; i <= 0xb; i++)
-        // {
-        //     da2_log("%02x ", da2->gdcreg[i]);
-        //     da2->gdcreg[i] = da2->bitblt.reg[i] & 0xff;
-        // }
-        // da2_log("\n");
-
+        da2->bitblt.bitshift_destr = ((da2->bitblt.reg[0x3] >> 4) & 0x0f); /* set bit shift */
+        da2->bitblt.raster_op      = da2->bitblt.reg[0x0b];
         da2->bitblt.destaddr  = da2->bitblt.reg[0x29];
         da2->bitblt.size_x    = da2->bitblt.reg[0x33];
         da2->bitblt.size_y    = da2->bitblt.reg[0x35];
@@ -890,14 +889,13 @@ da2_bitblt_exec(void *p)
 #else
     da2_log("bitblt_exec: %d\n", da2->bitblt.exec);
 #endif
-    // da2_log("blt %x %x %x %x\n", da2->bitblt.destaddr, da2->bitblt.x, da2->bitblt.y);
     switch (da2->bitblt.exec) {
         case DA2_BLT_CIDLE:
             // timer_disable(&da2->bitblt.timer);
             break;
         case DA2_BLT_CLOAD:
-            da2->bitblt.indata = 0;
             da2_bitblt_load(da2);
+            da2->bitblt.indata = 0;
             break;
         case DA2_BLT_CLINE:
             int pos_x = da2->bitblt.dest_x + da2->bitblt.x;
@@ -1512,7 +1510,7 @@ da2_outb(uint16_t addr, uint8_t val, void *p)
 void
 da2_outw(uint16_t addr, uint16_t val, void *p)
 {
-    da2_log("DA2 Outw addr %03X val %04X %04X:%04X\n", addr, val, cs >> 4, cpu_state.pc);
+    // da2_log("DA2 Outw addr %03X val %04X %04X:%04X\n", addr, val, cs >> 4, cpu_state.pc);
     da2_t *da2      = (da2_t *) p;
     da2->inflipflop = 0;
     switch (addr) {
@@ -2534,8 +2532,10 @@ da2_mmio_write(uint32_t addr, uint8_t val, void *p)
                 da2->bitblt.indata = 1;
                 if (da2->bitblt.payload_addr >= DA2_BLT_MEMSIZE)
                     da2_log("da2_mmio_write payload overflow! mem %x, addr %x, val %x\n", da2->fctl[LF_MMIO_MODE], addr, val);
-                da2->bitblt.payload[da2->bitblt.payload_addr] = val;
-                da2->bitblt.payload_addr++;
+                else {
+                    da2->bitblt.payload[da2->bitblt.payload_addr] = val;
+                    da2->bitblt.payload_addr++;
+                }
                 break;
             default:
                 da2_log("da2_mmio_write failed mem %x, addr %x, val %x\n", da2->fctl[LF_MMIO_MODE], addr, val);
@@ -3008,7 +3008,6 @@ da2_loadfont(char *fname, void *p)
 {
     da2_t  *da2 = (da2_t *) p;
     uint8_t buf;
-    // uint32_t code = 0;
     uint64_t fsize;
     if (!fname)
         return;

From e49be34d3070730b63053a0df9bdffa833c01e8e Mon Sep 17 00:00:00 2001
From: Akamaki <97360908+akmed772@users.noreply.github.com>
Date: Thu, 27 Feb 2025 11:15:57 +0900
Subject: [PATCH 09/58] fix an issue when loading bitblt data in async

---
 src/video/vid_ps55da2.c | 34 +++++++++++++++-------------------
 1 file changed, 15 insertions(+), 19 deletions(-)

diff --git a/src/video/vid_ps55da2.c b/src/video/vid_ps55da2.c
index 91752a7a8..f8c72833e 100644
--- a/src/video/vid_ps55da2.c
+++ b/src/video/vid_ps55da2.c
@@ -273,6 +273,11 @@ da2_log(const char *fmt, ...)
 #else
 #    define da2_log(fmt, ...)
 #endif
+#ifdef ENABLE_DA2_DEBUGBLT
+#    define da2_bltlog da2_log
+#else
+#    define da2_bltlog(fmt, ...)
+#endif
 
 typedef struct da2_t {
     // mem_mapping_t vmapping;
@@ -678,7 +683,7 @@ da2_bitblt_load(da2_t *da2)
                 value32 = da2->bitblt.payload[i + 3];
                 value32 <<= 8;
                 value32 |= da2->bitblt.payload[i + 2];
-                da2_log("[%02x] %02x: %04x (%d)\n", da2->bitblt.payload[i], da2->bitblt.payload[i + 1], value32, value32);
+                da2_bltlog("[%02x] %02x: %04x (%d)\n", da2->bitblt.payload[i], da2->bitblt.payload[i + 1], value32, value32);
                 da2->bitblt.reg[da2->bitblt.payload[i + 1]] = value32;
                 i += 3;
                 break;
@@ -690,7 +695,7 @@ da2_bitblt_load(da2_t *da2)
                 value32 |= da2->bitblt.payload[i + 3];
                 value32 <<= 8;
                 value32 |= da2->bitblt.payload[i + 2];
-                da2_log("[%02x] %02x: %08x (%d)\n", da2->bitblt.payload[i], da2->bitblt.payload[i + 1], value32, value32);
+                da2_bltlog("[%02x] %02x: %08x (%d)\n", da2->bitblt.payload[i], da2->bitblt.payload[i + 1], value32, value32);
                 da2->bitblt.reg[da2->bitblt.payload[i + 1]] = value32;
                 i += 5;
                 break;
@@ -706,7 +711,7 @@ da2_bitblt_load(da2_t *da2)
                 value64 |= da2->bitblt.payload[i + 3];
                 value64 <<= 8;
                 value64 |= da2->bitblt.payload[i + 2];
-                da2_log("[%02x] %02x: %02x %02x %02x %02x %02x %02x\n", da2->bitblt.payload[i], da2->bitblt.payload[i + 1], da2->bitblt.payload[i + 2], da2->bitblt.payload[i + 3],
+                da2_bltlog("[%02x] %02x: %02x %02x %02x %02x %02x %02x\n", da2->bitblt.payload[i], da2->bitblt.payload[i + 1], da2->bitblt.payload[i + 2], da2->bitblt.payload[i + 3],
                         da2->bitblt.payload[i + 4], da2->bitblt.payload[i + 5], da2->bitblt.payload[i + 6], da2->bitblt.payload[i + 7]);
                 da2->bitblt.reg[da2->bitblt.payload[i + 1]] = value64;
                 i += 7;
@@ -721,13 +726,10 @@ da2_bitblt_load(da2_t *da2)
     }
     da2->bitblt.exec = DA2_BLT_CIDLE;
     /* clear payload memory */
-    // memset(da2->bitblt.payload, 0x00, DA2_BLT_MEMSIZE);
-    // da2->bitblt.payload_addr = 0;
+    memset(da2->bitblt.payload, 0x00, DA2_BLT_MEMSIZE);
+    da2->bitblt.payload_addr = 0;
     /* [89] 20: 0001 (1) then execute payload */
     if (da2->bitblt.reg[0x20] & 0x1) {
-        /* clear payload memory */
-        memset(da2->bitblt.payload, 0x00, DA2_BLT_MEMSIZE);
-        da2->bitblt.payload_addr = 0;
 #ifdef ENABLE_DA2_DEBUGBLT
         for (i = 0; i < DA2_DEBUG_BLTLOG_SIZE; i++) {
             da2->bitblt.debug_reg[DA2_DEBUG_BLTLOG_SIZE * da2->bitblt.debug_reg_ip + i] = da2->bitblt.reg[i];
@@ -759,7 +761,6 @@ da2_bitblt_load(da2_t *da2)
         da2->bitblt.x      = 0;
         da2->bitblt.y      = 0;
         da2->bitblt.exec   = DA2_BLT_CDONE;
-        timer_set_delay_u64(&da2->bitblt.timer, da2->bitblt.timerspeed);
 
         /* Put DBCS char used by OS/2 (i'm not sure what the condition is) */
         if (da2->bitblt.reg[0x10] == 0xbc04) {
@@ -886,8 +887,6 @@ da2_bitblt_exec(void *p)
     if(!(da2->bitblt.debug_exesteps & 0xf))
         da2_log("bitblt_exec: %d %d\n", da2->bitblt.exec, da2->bitblt.debug_exesteps);
     da2->bitblt.debug_exesteps++;
-#else
-    da2_log("bitblt_exec: %d\n", da2->bitblt.exec);
 #endif
     switch (da2->bitblt.exec) {
         case DA2_BLT_CIDLE:
@@ -903,7 +902,7 @@ da2_bitblt_exec(void *p)
 
             /* Draw a dot */
             // outb(0x680, da2->bitblt.octdir);
-            da2_log("point: %d %d %d %d %d\n", pos_x, pos_y, da2->bitblt.d, da2->bitblt.x, da2->bitblt.y);
+            da2_bltlog("point: %d %d %d %d %d\n", pos_x, pos_y, da2->bitblt.d, da2->bitblt.x, da2->bitblt.y);
             int destaddr = pos_y * (da2->rowoffset * 2) + pos_x / 8;
             int pixelmask = pos_x % 16;
             if (pixelmask >= 8)
@@ -1090,11 +1089,9 @@ da2_bitblt_dopayload(void *priv)
     timer_set_delay_u64(&da2->bitblt.timer, da2->bitblt.timerspeed);
     if (da2->bitblt.exec != DA2_BLT_CIDLE)
         da2_bitblt_exec(da2);
-    else if (da2->bitblt.indata) {
-        if (da2->bitblt.exec == DA2_BLT_CIDLE) {
+    else if (da2->bitblt.indata && !(da2->ioctl[LS_MMIO] & 0x10) && (da2->bitblt.exec == DA2_BLT_CIDLE)) {
             da2->bitblt.exec = DA2_BLT_CLOAD;
             da2_bitblt_exec(da2);
-        }
     } else {
         // timer_disable(&da2->bitblt.timer);
     }
@@ -2656,6 +2653,7 @@ da2_mmio_gc_writeW(uint32_t addr, uint16_t val, void *p)
 {
     da2_t   *da2     = (da2_t *) p;
     uint16_t bitmask = da2->gdcreg[LG_BIT_MASK_HIGH];
+    addr &= DA2_MASK_MMIO;
     bitmask <<= 8;
     bitmask |= (uint16_t) da2->gdcreg[LG_BIT_MASK_LOW];
 #ifdef ENABLE_DA2_DEBUGVRAM
@@ -2676,7 +2674,6 @@ da2_mmio_gc_writeW(uint32_t addr, uint16_t val, void *p)
     //}
 #endif
     cycles -= video_timing_write_w;
-    // cycles_lost += video_timing_write_w;
 
     // da2_log("da2_gcW m%d a%x d%x\n", da2->writemode, addr, val);
     // da2_log("da2_gcW %05X %02X %04X:%04X esdi %04X:%04X dssi %04X:%04X\n", addr, val, cs >> 4, cpu_state.pc, ES, DI, DS, SI);
@@ -2765,7 +2762,6 @@ da2_mmio_writew(uint32_t addr, uint16_t val, void *p)
         da2_mmio_write(addr, val & 0xff, da2);
         da2_mmio_write(addr + 1, val >> 8, da2);
     } else if (!(da2->ioctl[LS_MODE] & 1)) { /* 16 color or 256 color mode */
-        addr &= DA2_MASK_MMIO;
         // return;
         // da2_log("da2_mmio_writeGW %x %x %x %x %x %x\n", da2->ioctl[LS_MMIO], da2->fctl[LF_MMIO_SEL], da2->fctl[LF_MMIO_MODE], da2->fctl[LF_MMIO_ADDR], addr, val);
         da2_mmio_gc_writeW(addr, val, da2);
@@ -2808,8 +2804,8 @@ static uint8_t
 da2_code_read(uint32_t addr, void *p)
 {
     da2_t *da2 = (da2_t *) p;
-    if ((addr & ~DA2_MASK_CRAM) != 0xE0000)
-        return DA2_INVALIDACCESS8;
+    // if ((addr & ~DA2_MASK_CRAM) != 0xE0000)
+    //     return DA2_INVALIDACCESS8;
     addr &= DA2_MASK_CRAM;
     return da2->cram[addr];
 }

From 03dd94f361259b102fad8cfce2592fa81b801ee8 Mon Sep 17 00:00:00 2001
From: Cacodemon345 <wahil1976@outlook.com>
Date: Thu, 27 Feb 2025 13:50:45 +0600
Subject: [PATCH 10/58] x87: Fix Final Reality discolored screen for all
 dynarecs

---
 src/codegen/codegen_ops_x86-64.h      | 14 +++++++++
 src/codegen/codegen_ops_x86.h         |  7 +++++
 src/codegen_new/codegen_ops_helpers.h |  3 ++
 src/cpu/386_dynarec.c                 | 45 +++++++++++++++++++++++++++
 src/cpu/cpu.h                         |  6 ++++
 5 files changed, 75 insertions(+)

diff --git a/src/codegen/codegen_ops_x86-64.h b/src/codegen/codegen_ops_x86-64.h
index e46f55a05..01fb292cf 100644
--- a/src/codegen/codegen_ops_x86-64.h
+++ b/src/codegen/codegen_ops_x86-64.h
@@ -3571,6 +3571,8 @@ FP_FLD(int reg)
     addbyte(0x89); /*MOV [TOP], EBX*/
     addbyte(0x5d);
     addbyte((uint8_t) cpu_state_offset(TOP));
+
+    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 
 static __inline void
@@ -3688,6 +3690,8 @@ FP_LOAD_S(void)
     addbyte(0x44);
     addbyte(0x1d);
     addbyte((uint8_t) cpu_state_offset(tag));
+    
+    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 static __inline void
 FP_LOAD_D(void)
@@ -3717,6 +3721,8 @@ FP_LOAD_D(void)
     addbyte(0x44);
     addbyte(0x1d);
     addbyte((uint8_t) cpu_state_offset(tag));
+
+    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 
 static __inline void
@@ -3754,6 +3760,8 @@ FP_LOAD_IW(void)
     addbyte(0x44);
     addbyte(0x1d);
     addbyte((uint8_t) cpu_state_offset(tag));
+
+    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 static __inline void
 FP_LOAD_IL(void)
@@ -3787,6 +3795,8 @@ FP_LOAD_IL(void)
     addbyte(0x44);
     addbyte(0x1d);
     addbyte((uint8_t) cpu_state_offset(tag));
+
+    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 static __inline void
 FP_LOAD_IQ(void)
@@ -3831,6 +3841,8 @@ FP_LOAD_IQ(void)
     addbyte(0x44);
     addbyte(0x1d);
     addbyte((uint8_t) cpu_state_offset(tag));
+
+    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 
 static __inline void
@@ -3863,6 +3875,8 @@ FP_LOAD_IMM_Q(uint64_t v)
     addbyte(0x1d);
     addbyte((uint8_t) cpu_state_offset(tag));
     addbyte(v ? 0 : 1);
+
+    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 
 static __inline void
diff --git a/src/codegen/codegen_ops_x86.h b/src/codegen/codegen_ops_x86.h
index c48324c2a..3b47d81a9 100644
--- a/src/codegen/codegen_ops_x86.h
+++ b/src/codegen/codegen_ops_x86.h
@@ -1794,6 +1794,7 @@ FP_FLD(int reg)
         addbyte(0x5d);
         addbyte((uint8_t) cpu_state_offset(TOP));
     }
+    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 
 static __inline void
@@ -2037,6 +2038,7 @@ FP_LOAD_S(void)
         addbyte(0x1d);
         addbyte((uint8_t) cpu_state_offset(tag[0]));
     }
+    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 static __inline void
 FP_LOAD_D(void)
@@ -2096,6 +2098,7 @@ FP_LOAD_D(void)
         addbyte(0x1d);
         addbyte((uint8_t) cpu_state_offset(tag[0]));
     }
+    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 static __inline void
 FP_LOAD_IW(void)
@@ -2154,6 +2157,7 @@ FP_LOAD_IW(void)
         addbyte(0x1d);
         addbyte((uint8_t) cpu_state_offset(tag[0]));
     }
+    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 static __inline void
 FP_LOAD_IL(void)
@@ -2210,6 +2214,7 @@ FP_LOAD_IL(void)
         addbyte(0x1d);
         addbyte((uint8_t) cpu_state_offset(tag[0]));
     }
+    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 static __inline void
 FP_LOAD_IQ(void)
@@ -2285,6 +2290,7 @@ FP_LOAD_IQ(void)
         addbyte(0x1d);
         addbyte((uint8_t) cpu_state_offset(tag[0]));
     }
+    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 
 static __inline void
@@ -2336,6 +2342,7 @@ FP_LOAD_IMM_Q(uint64_t v)
         addbyte(0x5d);
         addbyte((uint8_t) cpu_state_offset(TOP));
     }
+    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 
 static __inline int
diff --git a/src/codegen_new/codegen_ops_helpers.h b/src/codegen_new/codegen_ops_helpers.h
index 92b721099..2304aa542 100644
--- a/src/codegen_new/codegen_ops_helpers.h
+++ b/src/codegen_new/codegen_ops_helpers.h
@@ -64,6 +64,9 @@ fpu_POP2(codeblock_t *block, ir_data_t *ir)
 static inline void
 fpu_PUSH(codeblock_t *block, ir_data_t *ir)
 {
+    uop_LOAD_FUNC_ARG_IMM(ir, 0, ((uint16_t)cpu_state.TOP - 1));
+    uop_CALL_FUNC(ir, x87_to_mmxreg);
+
     if (block->flags & CODEBLOCK_STATIC_TOP)
         uop_MOV_IMM(ir, IREG_FPU_TOP, cpu_state.TOP - 1);
     else
diff --git a/src/cpu/386_dynarec.c b/src/cpu/386_dynarec.c
index 5f41c416a..5cb385122 100644
--- a/src/cpu/386_dynarec.c
+++ b/src/cpu/386_dynarec.c
@@ -74,6 +74,51 @@ x386_dynarec_log(const char *fmt, ...)
 #    define x386_dynarec_log(fmt, ...)
 #endif
 
+/* Deliberately stashed here; this function is only relevant for dynamic recompilers. */
+#if defined(_MSC_VER) && !defined(__clang__)
+#    if defined i386 || defined __i386 || defined __i386__ || defined _X86_ || defined _M_IX86
+#        define X87_INLINE_ASM
+#    endif
+#else
+#    if defined i386 || defined __i386 || defined __i386__ || defined _X86_ || defined _M_IX86 || defined _M_X64 || defined __amd64__
+#        define X87_INLINE_ASM
+#    endif
+#endif
+
+#ifdef USE_NEW_DYNAREC
+void
+x87_to_mmxreg(uint16_t reg)
+#else
+void
+x87_to_mmxreg(void)
+#endif
+{
+#ifndef USE_NEW_DYNAREC
+    uint32_t reg = cpu_state.TOP & 7;
+#endif
+    double val = cpu_state.ST[reg & 7];
+#ifdef X87_INLINE_ASM
+    unsigned char buffer[10];
+#else
+    x87_conv_t test;
+#endif
+
+#ifdef X87_INLINE_ASM
+    __asm volatile(""
+        :
+        :
+        : "memory");
+
+    __asm volatile("fldl %1\n"
+                    "fstpt %0\n" : "=m"(buffer) : "m"(val));
+
+    cpu_state.MM[reg & 7].q = (*(uint64_t*)buffer);
+#else
+    x87_to80(val, &test);
+    cpu_state.MM[reg & 7].q = test.eind.ll;
+#endif
+}
+
 static __inline void
 fetch_ea_32_long(uint32_t rmdat)
 {
diff --git a/src/cpu/cpu.h b/src/cpu/cpu.h
index 1d003ddc9..98c6f4938 100644
--- a/src/cpu/cpu.h
+++ b/src/cpu/cpu.h
@@ -829,6 +829,12 @@ extern uint16_t prefetch_queue_get_ip(void);
 extern int      prefetch_queue_get_prefetching(void);
 extern int      prefetch_queue_get_size(void);
 
+#ifdef USE_NEW_DYNAREC
+extern void     x87_to_mmxreg(uint16_t reg);
+#else
+extern void     x87_to_mmxreg(void);
+#endif
+
 #define prefetch_queue_set_suspended(s) prefetch_queue_set_prefetching(!s)
 #define prefetch_queue_get_suspended !prefetch_queue_get_prefetching
 

From c7153916eb21c67647e4df2512cda9b8776e5653 Mon Sep 17 00:00:00 2001
From: Cacodemon345 <wahil1976@outlook.com>
Date: Thu, 27 Feb 2025 14:33:44 +0600
Subject: [PATCH 11/58] Fix compile on ARM64

---
 src/cpu/386_dynarec.c | 70 +++++++++++++++++++++----------------------
 1 file changed, 35 insertions(+), 35 deletions(-)

diff --git a/src/cpu/386_dynarec.c b/src/cpu/386_dynarec.c
index 5cb385122..04681c9dc 100644
--- a/src/cpu/386_dynarec.c
+++ b/src/cpu/386_dynarec.c
@@ -74,7 +74,6 @@ x386_dynarec_log(const char *fmt, ...)
 #    define x386_dynarec_log(fmt, ...)
 #endif
 
-/* Deliberately stashed here; this function is only relevant for dynamic recompilers. */
 #if defined(_MSC_VER) && !defined(__clang__)
 #    if defined i386 || defined __i386 || defined __i386__ || defined _X86_ || defined _M_IX86
 #        define X87_INLINE_ASM
@@ -85,40 +84,6 @@ x386_dynarec_log(const char *fmt, ...)
 #    endif
 #endif
 
-#ifdef USE_NEW_DYNAREC
-void
-x87_to_mmxreg(uint16_t reg)
-#else
-void
-x87_to_mmxreg(void)
-#endif
-{
-#ifndef USE_NEW_DYNAREC
-    uint32_t reg = cpu_state.TOP & 7;
-#endif
-    double val = cpu_state.ST[reg & 7];
-#ifdef X87_INLINE_ASM
-    unsigned char buffer[10];
-#else
-    x87_conv_t test;
-#endif
-
-#ifdef X87_INLINE_ASM
-    __asm volatile(""
-        :
-        :
-        : "memory");
-
-    __asm volatile("fldl %1\n"
-                    "fstpt %0\n" : "=m"(buffer) : "m"(val));
-
-    cpu_state.MM[reg & 7].q = (*(uint64_t*)buffer);
-#else
-    x87_to80(val, &test);
-    cpu_state.MM[reg & 7].q = test.eind.ll;
-#endif
-}
-
 static __inline void
 fetch_ea_32_long(uint32_t rmdat)
 {
@@ -278,6 +243,41 @@ fetch_ea_16_long(uint32_t rmdat)
 #    define CACHE_ON() (!(cr0 & (1 << 30)) && !(cpu_state.flags & T_FLAG))
 #endif
 
+/* Deliberately stashed here; this function is only relevant for dynamic recompilers. */
+#ifdef USE_NEW_DYNAREC
+void
+x87_to_mmxreg(uint16_t reg)
+#else
+void
+x87_to_mmxreg(void)
+#endif
+{
+#ifndef USE_NEW_DYNAREC
+    uint32_t reg = cpu_state.TOP & 7;
+#endif
+    double val = cpu_state.ST[reg & 7];
+#ifdef X87_INLINE_ASM
+    unsigned char buffer[10];
+#else
+    x87_conv_t test;
+#endif
+
+#ifdef X87_INLINE_ASM
+    __asm volatile(""
+        :
+        :
+        : "memory");
+
+    __asm volatile("fldl %1\n"
+                    "fstpt %0\n" : "=m"(buffer) : "m"(val));
+
+    cpu_state.MM[reg & 7].q = (*(uint64_t*)buffer);
+#else
+    x87_to80(val, &test);
+    cpu_state.MM[reg & 7].q = test.eind.ll;
+#endif
+}
+
 #ifdef USE_DYNAREC
 int32_t         cycles_main = 0;
 static int32_t  cycles_old  = 0;

From fc656cbe0519b1a108aef25a333bf61cd5131cf8 Mon Sep 17 00:00:00 2001
From: Cacodemon345 <wahil1976@outlook.com>
Date: Thu, 27 Feb 2025 16:31:27 +0600
Subject: [PATCH 12/58] Mouse cursor now properly appears after uncapture

---
 src/qt/qt_mainwindow.cpp    | 4 +++-
 src/qt/qt_rendererstack.cpp | 2 +-
 2 files changed, 4 insertions(+), 2 deletions(-)

diff --git a/src/qt/qt_mainwindow.cpp b/src/qt/qt_mainwindow.cpp
index 62ac94f30..89d17a07c 100644
--- a/src/qt/qt_mainwindow.cpp
+++ b/src/qt/qt_mainwindow.cpp
@@ -262,8 +262,10 @@ MainWindow::MainWindow(QWidget *parent)
                 ui->stackedWidget->mouse_capture_func(this->windowHandle());
         } else {
             this->releaseKeyboard();
-            if (ui->stackedWidget->mouse_uncapture_func)
+            if (ui->stackedWidget->mouse_uncapture_func) {
                 ui->stackedWidget->mouse_uncapture_func();
+            }
+            ui->stackedWidget->unsetCursor();
         }
     });
 
diff --git a/src/qt/qt_rendererstack.cpp b/src/qt/qt_rendererstack.cpp
index a8bd47a79..8c31da2b2 100644
--- a/src/qt/qt_rendererstack.cpp
+++ b/src/qt/qt_rendererstack.cpp
@@ -155,7 +155,7 @@ RendererStack::mouseReleaseEvent(QMouseEvent *event)
     }
     if (mouse_capture && (event->button() == Qt::MiddleButton) && (mouse_get_buttons() < 3)) {
         plat_mouse_capture(0);
-        this->setCursor(Qt::ArrowCursor);
+        this->unsetCursor();
         isMouseDown &= ~1;
         return;
     }

From 95f12bc206e955324731bfcac18025f034d6db09 Mon Sep 17 00:00:00 2001
From: Akamaki <97360908+akmed772@users.noreply.github.com>
Date: Fri, 28 Feb 2025 00:50:01 +0900
Subject: [PATCH 13/58] fix glitch in Bitblt's line drawing function

fix a line glitch issue occurs in MS Paint and QEMM Manifest apps.
---
 src/video/vid_ps55da2.c | 58 ++++++++++++++++++++++++++---------------
 1 file changed, 37 insertions(+), 21 deletions(-)

diff --git a/src/video/vid_ps55da2.c b/src/video/vid_ps55da2.c
index f8c72833e..c476d12b9 100644
--- a/src/video/vid_ps55da2.c
+++ b/src/video/vid_ps55da2.c
@@ -405,10 +405,11 @@ typedef struct da2_t {
         int        indata;
         int32_t    destaddr;
         int32_t    srcaddr;
-        int32_t    size_x, tile_w;
-        int32_t    size_y;
-        int32_t    dest_x;
-        int32_t    dest_y;
+        int16_t    size_x;
+        int16_t    tile_w;
+        int16_t    size_y;
+        int16_t    dest_x;
+        int16_t    dest_y;
         int16_t    destpitch;
         int16_t    srcpitch;
         int32_t    fcolor;
@@ -416,7 +417,7 @@ typedef struct da2_t {
         int32_t    count;
         int32_t    d;
         int        octdir;
-        int        x, y;
+        int        x, y, wx1, wx2, wy1, wy2;
     } bitblt;
 
 #ifdef ENABLE_DA2_DEBUGVRAM
@@ -805,20 +806,32 @@ da2_bitblt_load(da2_t *da2)
         /* Draw a line */
         else if (da2->bitblt.reg[0x5] == 0x43) {
             da2->bitblt.exec = DA2_BLT_CLINE;
-            da2->bitblt.dest_x    = (da2->bitblt.reg[0x32] & 0x7ff);
-            da2->bitblt.dest_y    = (da2->bitblt.reg[0x34] & 0x7ff);
-            da2->bitblt.size_x    = abs((da2->bitblt.reg[0x33] & 0x7ff) - (da2->bitblt.reg[0x32] & 0x7ff));
-            da2->bitblt.size_y    = abs((da2->bitblt.reg[0x35] & 0x7ff) - (da2->bitblt.reg[0x34] & 0x7ff));
-            da2->bitblt.count    = 0;
-            da2->bitblt.octdir    = da2->bitblt.reg[0x2D];
+            da2->bitblt.dest_x    = (da2->bitblt.reg[0x32] & 0xffff);
+            da2->bitblt.dest_y    = (da2->bitblt.reg[0x34] & 0xffff);
+            da2->bitblt.wx1 = (da2->bitblt.reg[0x32]) >> 16;
+            da2->bitblt.wx2 = (da2->bitblt.reg[0x33]) >> 16;
+            da2->bitblt.wy1       = (da2->bitblt.reg[0x34]) >> 16;
+            da2->bitblt.wy2       = (da2->bitblt.reg[0x35]) >> 16;
+            da2->bitblt.size_x    = abs((int16_t)(da2->bitblt.reg[0x33] & 0xffff) - da2->bitblt.dest_x);
+            da2->bitblt.size_y    = abs((int16_t)(da2->bitblt.reg[0x35] & 0xffff) - da2->bitblt.dest_y);
+            da2->bitblt.count     = 0;
+            da2->bitblt.octdir         = da2->bitblt.reg[0x2D];
             da2->bitblt.bitshift_destr = 0;
             if (da2->bitblt.octdir & 0x04) /* dX > dY */
                 da2->bitblt.d = 2 * da2->bitblt.size_y - da2->bitblt.size_x;
             else
                 da2->bitblt.d = 2 * da2->bitblt.size_x - da2->bitblt.size_y;
-            da2_log("drawline x=%d, y=%d, dx=%d, dy=%d, oct=%d\n",
+            da2->bitblt.x = da2->bitblt.dest_x;
+            da2->bitblt.y = da2->bitblt.dest_y;
+            da2_log("drawline x=%d, y=%d, dx=%d, dy=%d, oct=%dn",
                     da2->bitblt.dest_x, da2->bitblt.dest_y,
                     da2->bitblt.size_x, da2->bitblt.size_y, da2->bitblt.octdir);
+            da2_log("         x1=%d, x2=%d, y1=%d, y2=%d\n",
+                    da2->bitblt.reg[0x32] & 0x7ff, da2->bitblt.reg[0x33] & 0x7ff,
+                    da2->bitblt.reg[0x34] & 0x7ff, da2->bitblt.reg[0x35] & 0x7ff);
+            da2_log("        ux1=%d,ux2=%d,uy1=%d,uy2=%d\n",
+                    (da2->bitblt.reg[0x32] >> 16) & 0x7ff, (da2->bitblt.reg[0x33] >> 16) & 0x7ff,
+                    (da2->bitblt.reg[0x34] >> 16) & 0x7ff, (da2->bitblt.reg[0x35] >> 16) & 0x7ff);
         }
         /* Fill a rectangle (or draw a horizontal / vertical line) */
         else if ((da2->bitblt.reg[0x5] & 0xfff0) == 0x40 && da2->bitblt.reg[0x3D] == 0) {
@@ -884,7 +897,7 @@ da2_bitblt_exec(void *p)
     da2_t *da2 = (da2_t *) p;
     // timer_set_delay_u64(&da2->bitblt.timer, da2->bitblt.timerspeed);
 #ifdef ENABLE_DA2_DEBUGBLT
-    if(!(da2->bitblt.debug_exesteps & 0xf))
+    if(!(da2->bitblt.debug_exesteps & 0xff))
         da2_log("bitblt_exec: %d %d\n", da2->bitblt.exec, da2->bitblt.debug_exesteps);
     da2->bitblt.debug_exesteps++;
 #endif
@@ -897,23 +910,26 @@ da2_bitblt_exec(void *p)
             da2->bitblt.indata = 0;
             break;
         case DA2_BLT_CLINE:
-            int pos_x = da2->bitblt.dest_x + da2->bitblt.x;
-            int pos_y = da2->bitblt.dest_y + da2->bitblt.y;
-
             /* Draw a dot */
             // outb(0x680, da2->bitblt.octdir);
-            da2_bltlog("point: %d %d %d %d %d\n", pos_x, pos_y, da2->bitblt.d, da2->bitblt.x, da2->bitblt.y);
-            int destaddr = pos_y * (da2->rowoffset * 2) + pos_x / 8;
-            int pixelmask = pos_x % 16;
+            da2_bltlog("point: %d %d %d %d %d\n", da2->bitblt.x, da2->bitblt.y, da2->bitblt.d, da2->bitblt.x, da2->bitblt.y);
+            int destaddr = da2->bitblt.y * (da2->rowoffset * 2) + da2->bitblt.x / 8;
+            int pixelmask = da2->bitblt.x % 16;
             if (pixelmask >= 8)
                 pixelmask = (0x8000 >> (pixelmask - 8));
             else
                 pixelmask = (0x80 >> pixelmask);
-            DA2_DrawColorWithBitmask(destaddr, da2->bitblt.fcolor, pixelmask, da2);
+                
+            /* check the current position is inside the window */
+            if (da2->bitblt.x < da2->bitblt.wx1 || da2->bitblt.x > da2->bitblt.wx2
+                || da2->bitblt.y < da2->bitblt.wy1 || da2->bitblt.y > da2->bitblt.wy2)
+                ;
+            else
+                DA2_DrawColorWithBitmask(destaddr, da2->bitblt.fcolor, pixelmask, da2);
             // da2_log("draw: %x %x %x\n", destaddr, da2->bitblt.fcolor, pixelmask);
             da2->bitblt.count++;
 
-            /* Bresenham's line */
+            /* calculate the next position with Bresenham's line algorithm */
             if (da2->bitblt.octdir & 0x04) { /* dX > dY */
                 if (da2->bitblt.octdir & 0x02) {
                     da2->bitblt.x++;

From 95f30192e5cb7b20755c1235625c107a3391fe93 Mon Sep 17 00:00:00 2001
From: Cacodemon345 <wahil1976@outlook.com>
Date: Thu, 27 Feb 2025 22:09:55 +0600
Subject: [PATCH 14/58] USB multimedia keys now work on Windows

---
 src/qt/qt_main.cpp | 115 ++++++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 114 insertions(+), 1 deletion(-)

diff --git a/src/qt/qt_main.cpp b/src/qt/qt_main.cpp
index 72b0eef53..7723261b4 100644
--- a/src/qt/qt_main.cpp
+++ b/src/qt/qt_main.cpp
@@ -211,8 +211,121 @@ emu_LowLevelKeyboardProc(int nCode, WPARAM wParam, LPARAM lParam)
                          (GetForegroundWindow() == ((HWND) secondaryRenderer->winId())));
     }
 
-    if ((nCode < 0) || (nCode != HC_ACTION) || !is_over_window)
+    bool skip = ((nCode < 0) || (nCode != HC_ACTION) || !is_over_window);
+
+    if (skip)
         return CallNextHookEx(NULL, nCode, wParam, lParam);
+
+    /* USB keyboards send a scancode of 0x00 for multimedia keys. */
+    if (lpKdhs->scanCode == 0x00) {
+        /* Handle USB keyboard multimedia keys where possible.
+           Only a handful of keys can be handled via Virtual Key
+           detection; rest can't be reliably detected. */
+        DWORD vkCode = lpKdhs->vkCode;
+        bool up = !!(lpKdhs->flags & LLKHF_UP);
+        ret = CallNextHookEx(NULL, nCode, wParam, lParam);;
+
+        switch (vkCode)
+        {
+            case VK_MEDIA_PLAY_PAUSE:
+            {
+                win_keyboard_handle(0x22, up, 1, 0);
+                break;
+            }
+            case VK_MEDIA_STOP:
+            {
+                win_keyboard_handle(0x24, up, 1, 0);
+                break;
+            }
+            case VK_VOLUME_UP:
+            {
+                win_keyboard_handle(0x30, up, 1, 0);
+                break;
+            }
+            case VK_VOLUME_DOWN:
+            {
+                win_keyboard_handle(0x2E, up, 1, 0);
+                break;
+            }
+            case VK_VOLUME_MUTE:
+            {
+                win_keyboard_handle(0x20, up, 1, 0);
+                break;
+            }
+            case VK_MEDIA_NEXT_TRACK:
+            {
+                win_keyboard_handle(0x19, up, 1, 0);
+                break;
+            }
+            case VK_MEDIA_PREV_TRACK:
+            {
+                win_keyboard_handle(0x10, up, 1, 0);
+                break;
+            }
+            case VK_LAUNCH_MEDIA_SELECT:
+            {
+                win_keyboard_handle(0x6D, up, 1, 0);
+                break;
+            }
+            case VK_LAUNCH_MAIL:
+            {
+                win_keyboard_handle(0x6C, up, 1, 0);
+                break;
+            }
+            case VK_LAUNCH_APP1:
+            {
+                win_keyboard_handle(0x6B, up, 1, 0);
+                break;
+            }
+            case VK_LAUNCH_APP2:
+            {
+                win_keyboard_handle(0x21, up, 1, 0);
+                break;
+            }
+            case VK_BROWSER_BACK:
+            {
+                win_keyboard_handle(0x6A, up, 1, 0);
+                break;
+            }
+            case VK_BROWSER_FORWARD:
+            {
+                win_keyboard_handle(0x69, up, 1, 0);
+                break;
+            }
+            case VK_BROWSER_STOP:
+            {
+                win_keyboard_handle(0x68, up, 1, 0);
+                break;
+            }
+            case VK_BROWSER_HOME:
+            {
+                win_keyboard_handle(0x32, up, 1, 0);
+                break;
+            }
+            case VK_BROWSER_SEARCH:
+            {
+                win_keyboard_handle(0x65, up, 1, 0);
+                break;
+            }
+            case VK_BROWSER_REFRESH:
+            {
+                win_keyboard_handle(0x67, up, 1, 0);
+                break;
+            }
+            case VK_BROWSER_FAVORITES:
+            {
+                win_keyboard_handle(0x66, up, 1, 0);
+                break;
+            }
+            case VK_HELP:
+            {
+                win_keyboard_handle(0x3b, up, 1, 0);
+                break;
+            }
+        }
+
+        return ret;
+    }
     else if ((lpKdhs->scanCode == 0x01) && (lpKdhs->flags & LLKHF_ALTDOWN) &&
         !(lpKdhs->flags & (LLKHF_UP | LLKHF_EXTENDED)))
         ret = TRUE;

From 86342bfffe057d4ff01b18b05cd89816228c636c Mon Sep 17 00:00:00 2001
From: Cacodemon345 <wahil1976@outlook.com>
Date: Thu, 27 Feb 2025 23:26:42 +0600
Subject: [PATCH 15/58] OS/2 3.0 icon backgrounds are now drawn properly

---
 src/video/vid_mga.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/src/video/vid_mga.c b/src/video/vid_mga.c
index 3937c2a7d..f11983a20 100644
--- a/src/video/vid_mga.c
+++ b/src/video/vid_mga.c
@@ -4677,6 +4677,7 @@ blit_trap(mystique_t *mystique)
     int       err_l = (int32_t)mystique->dwgreg.ar[1];
     int       err_r = (int32_t)mystique->dwgreg.ar[4];
     const int trans_sel = (mystique->dwgreg.dwgctrl_running & DWGCTRL_TRANS_MASK) >> DWGCTRL_TRANS_SHIFT;
+    bool transc = !!(mystique->dwgreg.dwgctrl_running & DWGCTRL_TRANSC);
 
     switch (mystique->dwgreg.dwgctrl_running & DWGCTRL_ATYPE_MASK) {
         case DWGCTRL_ATYPE_BLK:
@@ -4699,6 +4700,7 @@ blit_trap(mystique_t *mystique)
                         int      pattern = mystique->dwgreg.pattern[yoff][xoff];
                         uint32_t dst;
 
+                        if (!transc || (transc && pattern))
                         switch (mystique->maccess_running & MACCESS_PWIDTH_MASK) {
                             case MACCESS_PWIDTH_8:
                                 svga->vram[(mystique->dwgreg.ydst_lin + x_l) & mystique->vram_mask]                = (pattern ? mystique->dwgreg.fcol : mystique->dwgreg.bcol) & 0xff;
@@ -4770,6 +4772,7 @@ blit_trap(mystique_t *mystique)
                         uint32_t dst;
                         uint32_t old_dst;
 
+                        if (!transc || (transc && pattern))
                         switch (mystique->maccess_running & MACCESS_PWIDTH_MASK) {
                             case MACCESS_PWIDTH_8:
                                 dst = svga->vram[(mystique->dwgreg.ydst_lin + x_l) & mystique->vram_mask];

From 726ce6c574076807772b4cbcfaedcfb309b2555d Mon Sep 17 00:00:00 2001
From: Akamaki <97360908+akmed772@users.noreply.github.com>
Date: Fri, 28 Feb 2025 10:29:17 +0900
Subject: [PATCH 16/58] modify video memory mask and screen updating

---
 src/video/vid_ps55da2.c | 109 ++++++++++++++++++++--------------------
 1 file changed, 54 insertions(+), 55 deletions(-)

diff --git a/src/video/vid_ps55da2.c b/src/video/vid_ps55da2.c
index c476d12b9..441df620b 100644
--- a/src/video/vid_ps55da2.c
+++ b/src/video/vid_ps55da2.c
@@ -50,25 +50,29 @@
 
 #define DA2_FONTROM_PATH_JPAN    "roms/video/da2/94X1320.BIN"
 #define DA2_FONTROM_PATH_HANT    "roms/video/da2/23F2698.BIN"
-#define DA2_FONTROM_SIZE         1536 * 1024
+#define DA2_FONTROM_SIZE         (1536 * 1024)
 #define DA2_FONTROM_BASESBCS     0x98000
 #define DA2_GAIJIRAM_SBCS        0x34000
 #define DA2_GAIJIRAM_SBEX        0x3c000
 #define DA2_VM03_BASECHR         0x18000
 #define DA2_VM03_BASEEXATTR      0x10000
-#define DA2_INVALIDACCESS8       0xff
+#define DA2_INVALIDACCESS8       0xffu
+#define DA2_INVALIDACCESS16      0xffffu
+#define DA2_INVALIDACCESS32      0xffffffffu
 #define DA2_MASK_MMIO            0x1ffff
-#define DA2_SIZE_VRAM            1024 * 1024 /* 0x100000 */
-#define DA2_SIZE_CRAM            4 * 1024 /* 0x1000 */
-#define DA2_SIZE_GAIJIRAM        256 * 1024 /* 0x40000 */
-#define DA2_MASK_A000            0x1ffff
-#define DA2_MASK_CRAM            0xfff
-#define DA2_MASK_GAIJIRAM        0x3ffff
-#define DA2_PIXELCLOCK           58000000.0
+#define DA2_SIZE_VRAM            (1024 * 1024) /* 0x100000 */
+#define DA2_SIZE_CRAM            (4 * 1024)    /* 0x1000 */
+#define DA2_SIZE_GAIJIRAM        (256 * 1024)  /* 0x40000 */
+#define DA2_MASK_A000            0x1ffff       /* 0x1FFFF */
+#define DA2_MASK_CRAM            0xfff         /* 0xFFF */
+#define DA2_MASK_GAIJIRAM        0x3ffff       /* 0x3FFFF */
+#define DA2_MASK_VRAM            0xfffff       /* 0xFFFFF */
+#define DA2_MASK_VRAMPLANE       0x1ffff       /* 0x1FFFF */
+#define DA2_PIXELCLOCK           29000000.0    /* 58 MHz interlaced */
 #define DA2_BLT_MEMSIZE          0x100
 #define DA2_BLT_REGSIZE          0x40
 #define DA2_DEBUG_BLTLOG_SIZE    (DA2_BLT_REGSIZE + 1)
-#define DA2_DEBUG_BLTLOG_MAX     256 * 1024
+#define DA2_DEBUG_BLTLOG_MAX     (256 * 1024)
 #define DA2_DEBUG_BLT_NEVERUSED  0xfefefefe
 #define DA2_DEBUG_BLT_USEDRESET  0xfefefe
 #define DA2_DCONFIG_CHARSET_JPAN 0 /* for Code page 932 Japanese */
@@ -253,7 +257,7 @@
 #endif
 
 #ifdef ENABLE_DA2_LOG
-#    define ENABLE_DA2_DEBUGBLT 1
+// #    define ENABLE_DA2_DEBUGBLT 1
 // #    define ENABLE_DA2_DEBUGVRAM 1
 // #    define ENABLE_DA2_DEBUGFULLSCREEN 1
 // #    define ENABLE_DA2_DEBUGMONWAIT 1
@@ -304,7 +308,6 @@ typedef struct da2_t {
 
     uint32_t decode_mask;
     uint32_t vram_max;
-    uint32_t vram_mask;
 
     uint32_t gdcla[8];
     uint32_t gdcinput[8];
@@ -363,13 +366,11 @@ typedef struct da2_t {
     // uint32_t cram_display_mask;
     /* APA Buffer A0000-BFFFFh (128 KB) */
     uint8_t *vram;
-    /* addr >> 12 = xx000h */
+    /* xxh */
     uint8_t *changedvram;
     /* (vram size - 1) >> 3 = 0x1FFFF */
     uint32_t vram_display_mask;
 
-    // uint32_t write_bank, read_bank;
-
     int fullchange;
 
     void (*render)(struct da2_t *da2);
@@ -454,15 +455,17 @@ typedef struct {
 uint32_t
 DA2_vram_r(uint32_t addr, da2_t *da2)
 {
-    if (addr & ~da2->vram_mask)
-        return -1;
+    if (addr & ~DA2_MASK_VRAM)
+        return DA2_INVALIDACCESS32;
     return da2->vram[addr];
 }
 /* safety write for internal functions */
 void
 DA2_vram_w(uint32_t addr, uint8_t val, da2_t *da2)
 {
-    da2->vram[addr & da2->vram_mask] = val;
+    if (addr & ~DA2_MASK_VRAM)
+        return;
+    da2->vram[addr] = val;
     return;
 }
 /* write pixel data with rop (Note: bitmask must be in big endian) */
@@ -472,8 +475,7 @@ DA2_WritePlaneDataWithBitmask(uint32_t destaddr, const uint16_t mask, pixel32 *s
     uint32_t writepx[8];
     destaddr &= 0xfffffffe; /* align to word address to work bit shift correctly */
     // da2_log("DA2_WPDWB addr %x mask %x rop %x shift %d\n", destaddr, mask, da2->bitblt.raster_op, da2->bitblt.bitshift_destr);
-    da2->changedvram[(da2->vram_display_mask & destaddr) >> 12]       = changeframecount;
-    da2->changedvram[(da2->vram_display_mask & (destaddr + 1)) >> 12] = changeframecount;
+    da2->changedvram[(DA2_MASK_VRAMPLANE & destaddr) >> 9]       = changeframecount;
     destaddr <<= 3;
     /* read destination data with big endian order */
     for (int i = 0; i < 8; i++)
@@ -485,7 +487,6 @@ DA2_WritePlaneDataWithBitmask(uint32_t destaddr, const uint16_t mask, pixel32 *s
     DA2_VidSeq32 mask32in;
     mask32in.d = (uint32_t) mask;
     DA2_VidSeq32 mask32;
-    mask32.d    = 0;
     mask32.b[3] = mask32in.b[0];
     mask32.b[2] = mask32in.b[1];
     mask32.d &= 0xffff0000;
@@ -1395,7 +1396,7 @@ da2_in(uint16_t addr, void *p)
         case LS_DATA:
             // da2->ioctl[3] = 0x80; /* 3E1h:3 bit 7 color monitor, bit 3 busy(GC) bit 0 busy (IO?) */
             if (da2->ioctladdr > 0xf)
-                return DA2_INVALIDACCESS8;
+                return DA2_INVALIDACCESS16;
             temp = da2->ioctl[da2->ioctladdr];
             if (da2->ioctladdr == LS_STATUS) {             /* Status register */
                 if (da2->attrc[LV_COMPATIBILITY] & 0x08) { /* for detecting monitor type and cable wiring */
@@ -1440,7 +1441,7 @@ da2_in(uint16_t addr, void *p)
             break;
         case LF_DATA:
             if (da2->fctladdr > 0x1f)
-                return DA2_INVALIDACCESS8;
+                return DA2_INVALIDACCESS16;
             temp = da2->fctl[da2->fctladdr];
             break;
         case LC_INDEX:
@@ -1448,7 +1449,7 @@ da2_in(uint16_t addr, void *p)
             break;
         case LC_DATA:
             if (da2->crtcaddr > 0x1f)
-                return DA2_INVALIDACCESS8;
+                return DA2_INVALIDACCESS16;
             temp = da2->crtc[da2->crtcaddr];
             break;
         case LV_PORT:
@@ -1741,14 +1742,13 @@ getfont_ps55dbcs(int32_t code, int32_t line, void *p)
     if (code >= 0x8000 && code <= 0x8183)
         code -= 0x6000; /* shift for IBM extended characters (I don't know how the real card works.) */
     if (code < DA2_FONTROM_SIZE / 72 && fline >= 0 && fline < 24) {
-        font = da2->mmio.font[code * 72 + fline * 3];             /* 1111 1111 */
-        font <<= 8;                                               /* 1111 1111 0000 0000 */
-        font |= da2->mmio.font[code * 72 + fline * 3 + 1] & 0xf0; /* 1111 1111 2222 0000 */
-        font >>= 1;                                               /* 0111 1111 1222 2000 */
-        font <<= 4;                                               /* 0111 1111 1222 2000 0000 */
-        font |= da2->mmio.font[code * 72 + fline * 3 + 1] & 0x0f; /* 0111 1111 1222 2000 2222 */
-        font <<= 8;                                               /* 0111 1111 1222 2000 2222 0000 0000 */
-        font |= da2->mmio.font[code * 72 + fline * 3 + 2];        /* 0111 1111 1222 2000 2222 3333 3333 */
+        font = da2->mmio.font[code * 72 + fline * 3];             /* 0000 0000 0000 0000 0000 0000 1111 1111 */
+        font <<= 8;                                               /* 0000 0000 0000 0000 1111 1111 0000 0000 */
+        font |= da2->mmio.font[code * 72 + fline * 3 + 1] & 0xf0; /* 0000 0000 0000 0000 1111 1111 2222 0000 */
+        font <<= 3;                                               /* 0000 0000 0000 0111 1111 1222 2000 0000 */
+        font |= da2->mmio.font[code * 72 + fline * 3 + 1] & 0x0f; /* 0000 0000 0000 0111 1111 1222 2000 2222 */
+        font <<= 8;                                               /* 0000 0111 1111 1222 2000 2222 0000 0000 */
+        font |= da2->mmio.font[code * 72 + fline * 3 + 2];        /* 0000 0111 1111 1222 2000 2222 3333 3333 */
         font <<= 4;                                               /* 0111 1111 1222 2000 2222 3333 3333 0000 */
         /* font >>= 1;//put blank at column 1 (and 26) */
     } else if (code >= 0xb000 && code <= 0xb75f) {
@@ -1824,7 +1824,7 @@ da2_render_text(da2_t *da2)
         // da2_log("\nda2ma: %x, da2sc: %x\n", da2->ma, da2->sc);
         for (x = 0; x < da2->hdisp; x += 13) {
             chr  = da2->cram[(da2->ma) & DA2_MASK_CRAM];
-            attr = da2->cram[((da2->ma) + 1) & DA2_MASK_CRAM];
+            attr = da2->cram[(da2->ma + 1) & DA2_MASK_CRAM];
             // if(chr!=0x20) da2_log("chr: %x, attr: %x    ", chr, attr);
             if (da2->attrc[LV_PAS_STATUS_CNTRL] & 0x80) /* IO 3E8h, Index 1Dh */
             {                                           /* --Parse attribute byte in color mode-- */
@@ -1864,7 +1864,7 @@ da2_render_text(da2_t *da2)
                 /* Stay drawing If the char code is DBCS and not at last column. */
                 if (chr_wide) {
                     /* Get high DBCS code from the next video address */
-                    chr_dbcs = da2->cram[((da2->ma) + 2) & da2->vram_display_mask];
+                    chr_dbcs = da2->cram[(da2->ma + 2) & DA2_MASK_CRAM];
                     chr_dbcs <<= 8;
                     chr_dbcs |= chr;
                     /* Get the font pattern */
@@ -1935,7 +1935,7 @@ da2_render_text(da2_t *da2)
             da2->ma += 2;
             p += 13;
         }
-        da2->ma &= da2->vram_display_mask;
+        // da2->ma &= DA2_MASK_CRAM;
         // da2->writelines++;
     }
 }
@@ -1959,9 +1959,9 @@ da2_render_textm3(da2_t *da2)
         int       chr_wide = 0;
         // da2_log("\nda2ma: %x, da2sc: %x\n", da2->ma, da2->sc);
         for (x = 0; x < da2->hdisp; x += 13) {
-            chr     = DA2_vram_r(DA2_VM03_BASECHR + (da2->ma), da2);
-            attr    = DA2_vram_r(DA2_VM03_BASECHR + (da2->ma) + 1, da2);
-            extattr = DA2_vram_r(DA2_VM03_BASEEXATTR + (da2->ma) + 1, da2);
+            chr     = DA2_vram_r(DA2_VM03_BASECHR + da2->ma, da2);
+            attr    = DA2_vram_r(DA2_VM03_BASECHR + da2->ma + 1, da2);
+            extattr = DA2_vram_r(DA2_VM03_BASEEXATTR + da2->ma + 1, da2);
             // if(chr!=0x20) da2_log("addr: %x, chr: %x, attr: %x    ", (DA2_VM03_BASECHR + da2->ma << 1) & da2->vram_mask, chr, attr);
             bg = attr >> 4;
             // if (da2->blink) bg &= ~0x8;
@@ -1979,7 +1979,7 @@ da2_render_textm3(da2_t *da2)
                 /* Stay drawing if the char code is DBCS and not at last column. */
                 if (chr_wide) {
                     /* Get high DBCS code from the next video address */
-                    chr_dbcs = DA2_vram_r(DA2_VM03_BASECHR + (da2->ma) + 2, da2);
+                    chr_dbcs = DA2_vram_r(DA2_VM03_BASECHR + da2->ma + 2, da2);
                     chr_dbcs <<= 8;
                     chr_dbcs |= chr;
                     /* Get the font pattern */
@@ -2027,7 +2027,7 @@ da2_render_textm3(da2_t *da2)
             da2->ma += 2;
             p += 13;
         }
-        da2->ma &= da2->vram_display_mask;
+        // da2->ma &= DA2_MASK_CRAM;
         // da2->writelines++;
     }
 }
@@ -2035,7 +2035,7 @@ da2_render_textm3(da2_t *da2)
 void
 da2_render_color_4bpp(da2_t *da2)
 {
-    int changed_offset = da2->ma >> 12;
+    int changed_offset = da2->ma >> 9;
     // da2_log("ma %x cf %x\n", da2->ma, changed_offset);
     da2->plane_mask &= 0x0f; /*safety */
 
@@ -2084,7 +2084,7 @@ da2_render_color_4bpp(da2_t *da2)
 void
 da2_render_color_8bpp(da2_t *da2)
 {
-    int changed_offset = da2->ma >> 12;
+    int changed_offset = da2->ma >> 9;
     // da2_log("ma %x cf %x\n", da2->ma, changed_offset);
 
     if (da2->changedvram[changed_offset] || da2->changedvram[changed_offset + 1] || da2->fullchange) {
@@ -2255,8 +2255,8 @@ da2_recalctimings(da2_t *da2)
     if (da2->dispofftime < TIMER_USEC)
         da2->dispofftime = TIMER_USEC;
     da2_log("da2 horiz total %i display end %i vidclock %f\n", da2->crtc[0], da2->crtc[1], da2->clock);
-    // da2_log("da2 vert total %i display end %i max row %i vsync %i\n",da2->vtotal,da2->dispend,(da2->crtc[9]&31)+1,da2->vsyncstart);
-    // da2_log("total %f on %i cycles off %i cycles frame %i sec %i\n",disptime*crtcconst,da2->dispontime,da2->dispofftime,(da2->dispontime+da2->dispofftime)*da2->vtotal,(da2->dispontime+da2->dispofftime)*da2->vtotal*70);
+    da2_log("da2 vert total %i display end %i max row %i vsync %i\n",da2->vtotal,da2->dispend,(da2->crtc[9]&31)+1,da2->vsyncstart);
+    da2_log("total %f on %i cycles off %i cycles frame %i sec %i\n",disptime*crtcconst,da2->dispontime,da2->dispofftime,(da2->dispontime+da2->dispofftime)*da2->vtotal,(da2->dispontime+da2->dispofftime)*da2->vtotal*70);
 
     // da2_log("da2->render %08X\n", da2->render);
 }
@@ -2580,7 +2580,7 @@ da2_mmio_write(uint32_t addr, uint8_t val, void *p)
 
         cycles -= video_timing_write_b;
 
-        da2->changedvram[addr >> 12] = changeframecount;
+        da2->changedvram[addr >> 9] = changeframecount;/* 0x1FFFF -> 0x1F */
         addr <<= 3;
 
         for (int i = 0; i < 8; i++)
@@ -2694,8 +2694,8 @@ da2_mmio_gc_writeW(uint32_t addr, uint16_t val, void *p)
     // da2_log("da2_gcW m%d a%x d%x\n", da2->writemode, addr, val);
     // da2_log("da2_gcW %05X %02X %04X:%04X esdi %04X:%04X dssi %04X:%04X\n", addr, val, cs >> 4, cpu_state.pc, ES, DI, DS, SI);
 
-    da2->changedvram[addr >> 12]       = changeframecount;
-    da2->changedvram[(addr + 1) >> 12] = changeframecount;
+    da2->changedvram[addr >> 9]       = changeframecount;
+    // da2->changedvram[(addr + 1) >> 12] = changeframecount;
     addr <<= 3;
 
     for (int i = 0; i < 8; i++)
@@ -2948,7 +2948,7 @@ da2_poll(void *priv)
             }
             da2->blink++;
 
-            for (x = 0; x < ((da2->vram_mask + 1) >> 12); x++) {
+            for (x = 0; x <= (DA2_MASK_VRAMPLANE >> 9); x++) {
                 if (da2->changedvram[x])
                     da2->changedvram[x]--;
             }
@@ -3089,7 +3089,7 @@ da2_reset(void *priv)
 
     da2->pos_regs[0]       = DA2_POSID_L;                    /* Adapter Identification Byte (Low byte) */
     da2->pos_regs[1]       = DA2_POSID_H;                    /* Adapter Identification Byte (High byte) */
-    da2->pos_regs[2]       = 0x40;                           /* Bit 7-5: 010=Mono, 100=Color, Bit 0 : Card Enable (they are changed by system software) */
+    da2->pos_regs[2]       = 0x40;                           /* Bit 7-5: 010=Mono, 100=Color, Bit 0 : Card Enable (set by reference diskette) */
     da2->ioctl[LS_CONFIG1] = OldLSI | Page_Two;              /* Configuration 1 : DA-II, 1024 KB */
     da2->ioctl[LS_CONFIG1] |= ((da2->monitorid & 0x8) << 1); /* Configuration 1 : Monitor ID 3 */
     da2->ioctl[LS_CONFIG2]         = (da2->monitorid & 0x7); /* Configuration 2: Monitor ID 0-2 */
@@ -3127,11 +3127,10 @@ da2_init(UNUSED(const device_t *info))
     da2->dispontime        = 1000ull << 32;
     da2->dispofftime       = 1000ull << 32;
     da2->vram              = calloc(1, DA2_SIZE_VRAM);
-    da2->vram_mask         = DA2_SIZE_VRAM - 1;
     da2->cram              = calloc(1, DA2_SIZE_CRAM);
     da2->vram_display_mask = DA2_MASK_CRAM;
-    da2->changedvram       = calloc(1, /*(memsize >> 12) << 1*/ DA2_SIZE_VRAM >> 12); /* XX000h */
-    da2->monitorid         = device_get_config_int("montype");                 /* Configuration for Monitor ID (aaaa) -> (xxax xxxx, xxxx xaaa) */
+    da2->monitorid         = device_get_config_int("montype");       
+    da2->changedvram       = calloc(1,  (DA2_MASK_VRAMPLANE + 1) >> 9); /* XX000h */
 
     da2->mmio.charset = device_get_config_int("charset");
     da2->mmio.font    = malloc(DA2_FONTROM_SIZE);
@@ -3148,7 +3147,7 @@ da2_init(UNUSED(const device_t *info))
     }
 
     mca_add(da2_mca_read, da2_mca_write, da2_mca_feedb, da2_mca_reset, da2);
-    da2->da2const = (uint64_t) ((cpuclock / DA2_PIXELCLOCK) * (float) (1ull << 32));
+    da2->da2const = (uint64_t) ((cpuclock / DA2_PIXELCLOCK) * (double) (1ull << 32));
     memset(da2->bitblt.payload, 0x00, DA2_BLT_MEMSIZE);
     memset(da2->bitblt.reg, 0xfe, DA2_BLT_REGSIZE * sizeof(uint32_t)); /* clear memory */
 #ifdef ENABLE_DA2_DEBUGBLT
@@ -3172,7 +3171,7 @@ da2_init(UNUSED(const device_t *info))
     mem_mapping_disable(&da2->cmapping);
 
     timer_add(&da2->timer, da2_poll, da2, 0);
-    da2->bitblt.timerspeed = 10 * TIMER_USEC;
+    da2->bitblt.timerspeed = 10ull * TIMER_USEC;
     timer_add(&da2->bitblt.timer, da2_bitblt_dopayload, da2, 0);
 
     return da2;
@@ -3275,7 +3274,7 @@ void
 da2_speed_changed(void *p)
 {
     da2_t *da2    = (da2_t *) p;
-    da2->da2const = (uint64_t) ((cpuclock / DA2_PIXELCLOCK) * (float) (1ull << 32));
+    da2->da2const = (uint64_t) ((cpuclock / DA2_PIXELCLOCK) * (double) (1ull << 32));
     da2_recalctimings(da2);
 }
 

From 0bb89be0ad0e718841ca99ef2405cb42f9666f6f Mon Sep 17 00:00:00 2001
From: Cacodemon345 <wahil1976@outlook.com>
Date: Fri, 28 Feb 2025 16:51:13 +0600
Subject: [PATCH 17/58] Revert "Fix compile on ARM64"

This reverts commit c7153916eb21c67647e4df2512cda9b8776e5653.
---
 src/cpu/386_dynarec.c | 70 +++++++++++++++++++++----------------------
 1 file changed, 35 insertions(+), 35 deletions(-)

diff --git a/src/cpu/386_dynarec.c b/src/cpu/386_dynarec.c
index 04681c9dc..5cb385122 100644
--- a/src/cpu/386_dynarec.c
+++ b/src/cpu/386_dynarec.c
@@ -74,6 +74,7 @@ x386_dynarec_log(const char *fmt, ...)
 #    define x386_dynarec_log(fmt, ...)
 #endif
 
+/* Deliberately stashed here; this function is only relevant for dynamic recompilers. */
 #if defined(_MSC_VER) && !defined(__clang__)
 #    if defined i386 || defined __i386 || defined __i386__ || defined _X86_ || defined _M_IX86
 #        define X87_INLINE_ASM
@@ -84,6 +85,40 @@ x386_dynarec_log(const char *fmt, ...)
 #    endif
 #endif
 
+#ifdef USE_NEW_DYNAREC
+void
+x87_to_mmxreg(uint16_t reg)
+#else
+void
+x87_to_mmxreg(void)
+#endif
+{
+#ifndef USE_NEW_DYNAREC
+    uint32_t reg = cpu_state.TOP & 7;
+#endif
+    double val = cpu_state.ST[reg & 7];
+#ifdef X87_INLINE_ASM
+    unsigned char buffer[10];
+#else
+    x87_conv_t test;
+#endif
+
+#ifdef X87_INLINE_ASM
+    __asm volatile(""
+        :
+        :
+        : "memory");
+
+    __asm volatile("fldl %1\n"
+                    "fstpt %0\n" : "=m"(buffer) : "m"(val));
+
+    cpu_state.MM[reg & 7].q = (*(uint64_t*)buffer);
+#else
+    x87_to80(val, &test);
+    cpu_state.MM[reg & 7].q = test.eind.ll;
+#endif
+}
+
 static __inline void
 fetch_ea_32_long(uint32_t rmdat)
 {
@@ -243,41 +278,6 @@ fetch_ea_16_long(uint32_t rmdat)
 #    define CACHE_ON() (!(cr0 & (1 << 30)) && !(cpu_state.flags & T_FLAG))
 #endif
 
-/* Deliberately stashed here; this function is only relevant for dynamic recompilers. */
-#ifdef USE_NEW_DYNAREC
-void
-x87_to_mmxreg(uint16_t reg)
-#else
-void
-x87_to_mmxreg(void)
-#endif
-{
-#ifndef USE_NEW_DYNAREC
-    uint32_t reg = cpu_state.TOP & 7;
-#endif
-    double val = cpu_state.ST[reg & 7];
-#ifdef X87_INLINE_ASM
-    unsigned char buffer[10];
-#else
-    x87_conv_t test;
-#endif
-
-#ifdef X87_INLINE_ASM
-    __asm volatile(""
-        :
-        :
-        : "memory");
-
-    __asm volatile("fldl %1\n"
-                    "fstpt %0\n" : "=m"(buffer) : "m"(val));
-
-    cpu_state.MM[reg & 7].q = (*(uint64_t*)buffer);
-#else
-    x87_to80(val, &test);
-    cpu_state.MM[reg & 7].q = test.eind.ll;
-#endif
-}
-
 #ifdef USE_DYNAREC
 int32_t         cycles_main = 0;
 static int32_t  cycles_old  = 0;

From 6bb2b447fda8011e992472fa0959d3ee4cd8deed Mon Sep 17 00:00:00 2001
From: Cacodemon345 <wahil1976@outlook.com>
Date: Fri, 28 Feb 2025 16:51:33 +0600
Subject: [PATCH 18/58] Revert "x87: Fix Final Reality discolored screen for
 all dynarecs"

This reverts commit 03dd94f361259b102fad8cfce2592fa81b801ee8.
---
 src/codegen/codegen_ops_x86-64.h      | 14 ---------
 src/codegen/codegen_ops_x86.h         |  7 -----
 src/codegen_new/codegen_ops_helpers.h |  3 --
 src/cpu/386_dynarec.c                 | 45 ---------------------------
 src/cpu/cpu.h                         |  6 ----
 5 files changed, 75 deletions(-)

diff --git a/src/codegen/codegen_ops_x86-64.h b/src/codegen/codegen_ops_x86-64.h
index 01fb292cf..e46f55a05 100644
--- a/src/codegen/codegen_ops_x86-64.h
+++ b/src/codegen/codegen_ops_x86-64.h
@@ -3571,8 +3571,6 @@ FP_FLD(int reg)
     addbyte(0x89); /*MOV [TOP], EBX*/
     addbyte(0x5d);
     addbyte((uint8_t) cpu_state_offset(TOP));
-
-    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 
 static __inline void
@@ -3690,8 +3688,6 @@ FP_LOAD_S(void)
     addbyte(0x44);
     addbyte(0x1d);
     addbyte((uint8_t) cpu_state_offset(tag));
-    
-    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 static __inline void
 FP_LOAD_D(void)
@@ -3721,8 +3717,6 @@ FP_LOAD_D(void)
     addbyte(0x44);
     addbyte(0x1d);
     addbyte((uint8_t) cpu_state_offset(tag));
-
-    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 
 static __inline void
@@ -3760,8 +3754,6 @@ FP_LOAD_IW(void)
     addbyte(0x44);
     addbyte(0x1d);
     addbyte((uint8_t) cpu_state_offset(tag));
-
-    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 static __inline void
 FP_LOAD_IL(void)
@@ -3795,8 +3787,6 @@ FP_LOAD_IL(void)
     addbyte(0x44);
     addbyte(0x1d);
     addbyte((uint8_t) cpu_state_offset(tag));
-
-    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 static __inline void
 FP_LOAD_IQ(void)
@@ -3841,8 +3831,6 @@ FP_LOAD_IQ(void)
     addbyte(0x44);
     addbyte(0x1d);
     addbyte((uint8_t) cpu_state_offset(tag));
-
-    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 
 static __inline void
@@ -3875,8 +3863,6 @@ FP_LOAD_IMM_Q(uint64_t v)
     addbyte(0x1d);
     addbyte((uint8_t) cpu_state_offset(tag));
     addbyte(v ? 0 : 1);
-
-    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 
 static __inline void
diff --git a/src/codegen/codegen_ops_x86.h b/src/codegen/codegen_ops_x86.h
index 3b47d81a9..c48324c2a 100644
--- a/src/codegen/codegen_ops_x86.h
+++ b/src/codegen/codegen_ops_x86.h
@@ -1794,7 +1794,6 @@ FP_FLD(int reg)
         addbyte(0x5d);
         addbyte((uint8_t) cpu_state_offset(TOP));
     }
-    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 
 static __inline void
@@ -2038,7 +2037,6 @@ FP_LOAD_S(void)
         addbyte(0x1d);
         addbyte((uint8_t) cpu_state_offset(tag[0]));
     }
-    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 static __inline void
 FP_LOAD_D(void)
@@ -2098,7 +2096,6 @@ FP_LOAD_D(void)
         addbyte(0x1d);
         addbyte((uint8_t) cpu_state_offset(tag[0]));
     }
-    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 static __inline void
 FP_LOAD_IW(void)
@@ -2157,7 +2154,6 @@ FP_LOAD_IW(void)
         addbyte(0x1d);
         addbyte((uint8_t) cpu_state_offset(tag[0]));
     }
-    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 static __inline void
 FP_LOAD_IL(void)
@@ -2214,7 +2210,6 @@ FP_LOAD_IL(void)
         addbyte(0x1d);
         addbyte((uint8_t) cpu_state_offset(tag[0]));
     }
-    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 static __inline void
 FP_LOAD_IQ(void)
@@ -2290,7 +2285,6 @@ FP_LOAD_IQ(void)
         addbyte(0x1d);
         addbyte((uint8_t) cpu_state_offset(tag[0]));
     }
-    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 
 static __inline void
@@ -2342,7 +2336,6 @@ FP_LOAD_IMM_Q(uint64_t v)
         addbyte(0x5d);
         addbyte((uint8_t) cpu_state_offset(TOP));
     }
-    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 
 static __inline int
diff --git a/src/codegen_new/codegen_ops_helpers.h b/src/codegen_new/codegen_ops_helpers.h
index 2304aa542..92b721099 100644
--- a/src/codegen_new/codegen_ops_helpers.h
+++ b/src/codegen_new/codegen_ops_helpers.h
@@ -64,9 +64,6 @@ fpu_POP2(codeblock_t *block, ir_data_t *ir)
 static inline void
 fpu_PUSH(codeblock_t *block, ir_data_t *ir)
 {
-    uop_LOAD_FUNC_ARG_IMM(ir, 0, ((uint16_t)cpu_state.TOP - 1));
-    uop_CALL_FUNC(ir, x87_to_mmxreg);
-
     if (block->flags & CODEBLOCK_STATIC_TOP)
         uop_MOV_IMM(ir, IREG_FPU_TOP, cpu_state.TOP - 1);
     else
diff --git a/src/cpu/386_dynarec.c b/src/cpu/386_dynarec.c
index 5cb385122..5f41c416a 100644
--- a/src/cpu/386_dynarec.c
+++ b/src/cpu/386_dynarec.c
@@ -74,51 +74,6 @@ x386_dynarec_log(const char *fmt, ...)
 #    define x386_dynarec_log(fmt, ...)
 #endif
 
-/* Deliberately stashed here; this function is only relevant for dynamic recompilers. */
-#if defined(_MSC_VER) && !defined(__clang__)
-#    if defined i386 || defined __i386 || defined __i386__ || defined _X86_ || defined _M_IX86
-#        define X87_INLINE_ASM
-#    endif
-#else
-#    if defined i386 || defined __i386 || defined __i386__ || defined _X86_ || defined _M_IX86 || defined _M_X64 || defined __amd64__
-#        define X87_INLINE_ASM
-#    endif
-#endif
-
-#ifdef USE_NEW_DYNAREC
-void
-x87_to_mmxreg(uint16_t reg)
-#else
-void
-x87_to_mmxreg(void)
-#endif
-{
-#ifndef USE_NEW_DYNAREC
-    uint32_t reg = cpu_state.TOP & 7;
-#endif
-    double val = cpu_state.ST[reg & 7];
-#ifdef X87_INLINE_ASM
-    unsigned char buffer[10];
-#else
-    x87_conv_t test;
-#endif
-
-#ifdef X87_INLINE_ASM
-    __asm volatile(""
-        :
-        :
-        : "memory");
-
-    __asm volatile("fldl %1\n"
-                    "fstpt %0\n" : "=m"(buffer) : "m"(val));
-
-    cpu_state.MM[reg & 7].q = (*(uint64_t*)buffer);
-#else
-    x87_to80(val, &test);
-    cpu_state.MM[reg & 7].q = test.eind.ll;
-#endif
-}
-
 static __inline void
 fetch_ea_32_long(uint32_t rmdat)
 {
diff --git a/src/cpu/cpu.h b/src/cpu/cpu.h
index 98c6f4938..1d003ddc9 100644
--- a/src/cpu/cpu.h
+++ b/src/cpu/cpu.h
@@ -829,12 +829,6 @@ extern uint16_t prefetch_queue_get_ip(void);
 extern int      prefetch_queue_get_prefetching(void);
 extern int      prefetch_queue_get_size(void);
 
-#ifdef USE_NEW_DYNAREC
-extern void     x87_to_mmxreg(uint16_t reg);
-#else
-extern void     x87_to_mmxreg(void);
-#endif
-
 #define prefetch_queue_set_suspended(s) prefetch_queue_set_prefetching(!s)
 #define prefetch_queue_get_suspended !prefetch_queue_get_prefetching
 

From 23b89d88c4e4fee4914d4c20b92a741e25e7c0ca Mon Sep 17 00:00:00 2001
From: Cacodemon345 <wahil1976@outlook.com>
Date: Fri, 28 Feb 2025 16:51:48 +0600
Subject: [PATCH 19/58] Revert "x87: Fix Final Reality discolored screen for
 interpreter"

This reverts commit 6d3816df64aac6cb0e604bab30239c3b4b888265.
---
 src/cpu/x87_ops.h | 41 -----------------------------------------
 1 file changed, 41 deletions(-)

diff --git a/src/cpu/x87_ops.h b/src/cpu/x87_ops.h
index 276c32876..f1362bf76 100644
--- a/src/cpu/x87_ops.h
+++ b/src/cpu/x87_ops.h
@@ -113,33 +113,12 @@ typedef union {
 static __inline void
 x87_push(double i)
 {
-#ifdef X87_INLINE_ASM
-    unsigned char buffer[10];
-#else
-    x87_conv_t test;
-#endif
 #ifdef USE_NEW_DYNAREC
     cpu_state.TOP--;
 #else
     cpu_state.TOP                    = (cpu_state.TOP - 1) & 7;
 #endif
     cpu_state.ST[cpu_state.TOP & 7] = i;
-
-#ifdef X87_INLINE_ASM
-    __asm volatile(""
-        :
-        :
-        : "memory");
-
-    __asm volatile("fldl %1\n"
-                    "fstpt %0\n" : "=m"(buffer) : "m"(i));
-
-    cpu_state.MM[cpu_state.TOP & 7].q = (*(uint64_t*)buffer);
-#else
-    x87_to80(i, &test);
-    cpu_state.MM[cpu_state.TOP & 7].q = test.eind.ll;
-#endif
-
 #ifdef USE_NEW_DYNAREC
     cpu_state.tag[cpu_state.TOP & 7] = TAG_VALID;
 #else
@@ -150,11 +129,6 @@ x87_push(double i)
 static __inline void
 x87_push_u64(uint64_t i)
 {
-#ifdef X87_INLINE_ASM
-    unsigned char buffer[10];
-#else
-    x87_conv_t test;
-#endif
     union {
         double   d;
         uint64_t ll;
@@ -168,21 +142,6 @@ x87_push_u64(uint64_t i)
     cpu_state.TOP                    = (cpu_state.TOP - 1) & 7;
 #endif
     cpu_state.ST[cpu_state.TOP & 7] = td.d;
-
-#ifdef X87_INLINE_ASM
-    __asm volatile(""
-        :
-        :
-        : "memory");
-
-    __asm volatile("fldl %1\n"
-                    "fstpt %0\n" : "=m"(buffer) : "m"(td.d));
-
-    cpu_state.MM[cpu_state.TOP & 7].q = (*(uint64_t*)buffer);
-#else
-    x87_to80(td.d, &test);
-    cpu_state.MM[cpu_state.TOP & 7].q = test.eind.ll;
-#endif
 #ifdef USE_NEW_DYNAREC
     cpu_state.tag[cpu_state.TOP & 7] = TAG_VALID;
 #else

From 843dee57076a28d4028b9105c87ead95e64d7933 Mon Sep 17 00:00:00 2001
From: Cacodemon345 <wahil1976@outlook.com>
Date: Sat, 1 Mar 2025 15:05:39 +0600
Subject: [PATCH 20/58] x64 NDR: Properly address the entire cpu_state struct

All missing edge cases are now handled where possible
---
 src/codegen_new/codegen_backend_x86-64_ops.c | 23 ++++++++++++++++++++
 1 file changed, 23 insertions(+)

diff --git a/src/codegen_new/codegen_backend_x86-64_ops.c b/src/codegen_new/codegen_backend_x86-64_ops.c
index 236a86ce7..2f65b46e3 100644
--- a/src/codegen_new/codegen_backend_x86-64_ops.c
+++ b/src/codegen_new/codegen_backend_x86-64_ops.c
@@ -509,6 +509,11 @@ host_x86_MOV8_ABS_IMM(codeblock_t *block, void *p, uint32_t imm_data)
         codegen_alloc_bytes(block, 4);
         codegen_addbyte3(block, 0xc6, 0x45, offset); /*MOVB offset[RBP], imm_data*/
         codegen_addbyte(block, imm_data);
+    } else if (offset < (1ULL << 32)) {
+        codegen_alloc_bytes(block, 7);
+        codegen_addbyte2(block, 0xc6, 0x85); /*MOVB offset[RBP], imm_data*/
+        codegen_addlong(block, offset);
+        codegen_addbyte(block, imm_data);
     } else {
         if ((uintptr_t) p >> 32)
             fatal("host_x86_MOV8_ABS_IMM - out of range %p\n", p);
@@ -527,6 +532,11 @@ host_x86_MOV16_ABS_IMM(codeblock_t *block, void *p, uint16_t imm_data)
         codegen_alloc_bytes(block, 6);
         codegen_addbyte4(block, 0x66, 0xc7, 0x45, offset); /*MOV offset[RBP], imm_data*/
         codegen_addword(block, imm_data);
+    } else if (offset < (1ULL << 32)) {
+        codegen_alloc_bytes(block, 8);
+        codegen_addbyte3(block, 0x66, 0xc7, 0x85); /*MOV offset[RBP], imm_data*/
+        codegen_addlong(block, offset);
+        codegen_addword(block, imm_data);
     } else {
         if ((uintptr_t) p >> 32)
             fatal("host_x86_MOV32_ABS_IMM - out of range %p\n", p);
@@ -545,6 +555,11 @@ host_x86_MOV32_ABS_IMM(codeblock_t *block, void *p, uint32_t imm_data)
         codegen_alloc_bytes(block, 7);
         codegen_addbyte3(block, 0xc7, 0x45, offset); /*MOV offset[RBP], imm_data*/
         codegen_addlong(block, imm_data);
+    } else if (offset < (1ULL << 32)) {
+        codegen_alloc_bytes(block, 10);
+        codegen_addbyte2(block, 0xc7, 0x85); /*MOV offset[RBP], imm_data*/
+        codegen_addlong(block, offset);
+        codegen_addlong(block, imm_data);
     } else {
         if ((uintptr_t) p >> 32)
             fatal("host_x86_MOV32_ABS_IMM - out of range %p\n", p);
@@ -566,6 +581,10 @@ host_x86_MOV8_ABS_REG(codeblock_t *block, void *p, int src_reg)
     if (offset >= -128 && offset < 127) {
         codegen_alloc_bytes(block, 3);
         codegen_addbyte3(block, 0x88, 0x45 | ((src_reg & 7) << 3), offset); /*MOVB offset[RBP], src_reg*/
+    } else if (offset < (1ULL << 32)) {
+        codegen_alloc_bytes(block, 6);
+        codegen_addbyte2(block, 0x88, 0x85 | ((src_reg & 7) << 3)); /*MOVB offset[RBP], src_reg*/
+        codegen_addlong(block, offset);
     } else {
         if ((uintptr_t) p >> 32)
             fatal("host_x86_MOV8_ABS_REG - out of range %p\n", p);
@@ -630,6 +649,10 @@ host_x86_MOV64_ABS_REG(codeblock_t *block, void *p, int src_reg)
     if (offset >= -128 && offset < 127) {
         codegen_alloc_bytes(block, 4);
         codegen_addbyte4(block, 0x48, 0x89, 0x45 | ((src_reg & 7) << 3), offset); /*MOV offset[RBP], src_reg*/
+    } else if (offset < (1ULL << 32)) {
+        codegen_alloc_bytes(block, 7);
+        codegen_addbyte3(block, 0x48, 0x89, 0x85 | ((src_reg & 7) << 3)); /*MOV offset[RBP], src_reg*/
+        codegen_addlong(block, offset);
     } else {
         if ((uintptr_t) p >> 32)
             fatal("host_x86_MOV64_ABS_REG - out of range %p\n", p);

From d15def050aba3e2652355c718918fb78e9c79b53 Mon Sep 17 00:00:00 2001
From: OBattler <oubattler@gmail.com>
Date: Sat, 1 Mar 2025 10:11:45 +0100
Subject: [PATCH 21/58] Fix an unclosed comment in device/keyboard_at.c.

---
 src/device/keyboard_at.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/device/keyboard_at.c b/src/device/keyboard_at.c
index d04ab07e9..e61b8547a 100644
--- a/src/device/keyboard_at.c
+++ b/src/device/keyboard_at.c
@@ -2526,7 +2526,7 @@ static const scancode scancode_set82[512] = {
     { .mk = {0xe0, 0x17, 0 }, .brk = { 0xe0, 0xF0, 0x17, 0 } }, /* 15a */
     { .mk = {      0x67, 0 }, .brk = {       0xf0, 0x67, 0 } }, /* 15b 0x33 LGUI->Muhenkan (in emulator only) */
     { .mk = {      0x64, 0 }, .brk = {       0xf0, 0x64, 0 } }, /* 15c 0x35 RGUI->Henkan (in emulator only) */
-    { .mk = {0xe0, 0x11, 0 }, .brk = { 0xe0, 0xf0, 0x11, 0 } }, /* 15d 0x36 APPLICATION->Kana (in emulator 
+    { .mk = {0xe0, 0x11, 0 }, .brk = { 0xe0, 0xf0, 0x11, 0 } }, /* 15d 0x36 APPLICATION->Kana (in emulator only) */
     { .mk = {0xe0, 0x37, 0 }, .brk = { 0xe0, 0xF0, 0x37, 0 } }, /* 15e */
     { .mk = {0xe0, 0x3F, 0 }, .brk = { 0xe0, 0xF0, 0x3F, 0 } }, /* 15f */
     { .mk = {            0 }, .brk = {                   0 } }, /* 160 */

From 31d05b3e45a60ab1bf0eb057bac8d012d788e031 Mon Sep 17 00:00:00 2001
From: Akamaki <97360908+akmed772@users.noreply.github.com>
Date: Sat, 1 Mar 2025 23:58:31 +0900
Subject: [PATCH 22/58] modify blank render

---
 src/video/vid_ps55da2.c | 34 ++++++++++++++++++++++------------
 1 file changed, 22 insertions(+), 12 deletions(-)

diff --git a/src/video/vid_ps55da2.c b/src/video/vid_ps55da2.c
index 441df620b..46ef4caf6 100644
--- a/src/video/vid_ps55da2.c
+++ b/src/video/vid_ps55da2.c
@@ -257,9 +257,9 @@
 #endif
 
 #ifdef ENABLE_DA2_LOG
-// #    define ENABLE_DA2_DEBUGBLT 1
-// #    define ENABLE_DA2_DEBUGVRAM 1
-// #    define ENABLE_DA2_DEBUGFULLSCREEN 1
+#    define ENABLE_DA2_DEBUGBLT 1
+#    define ENABLE_DA2_DEBUGVRAM 1
+#    define ENABLE_DA2_DEBUGFULLSCREEN 1
 // #    define ENABLE_DA2_DEBUGMONWAIT 1
 int da2_do_log = ENABLE_DA2_LOG;
 
@@ -1208,8 +1208,8 @@ da2_out(uint16_t addr, uint16_t val, void *p)
         case LC_DATA:
             if (da2->crtcaddr > 0x1f)
                 return;
-            // if (!(da2->crtcaddr == LC_CURSOR_LOC_HIGH || da2->crtcaddr == LC_CURSOR_LOC_LOWJ))
-            //     da2_log("DA2 Out addr %03X idx %02X val %02X %04X:%04X\n", addr, da2->crtcaddr, val, cs >> 4, cpu_state.pc);
+            if (!(da2->crtcaddr == LC_CURSOR_LOC_HIGH || da2->crtcaddr == LC_CURSOR_LOC_LOWJ))
+                da2_log("DA2 Out addr %03X idx %02X val %02X %04X:%04X\n", addr, da2->crtcaddr, val, cs >> 4, cpu_state.pc);
             if (!(da2->crtc[da2->crtcaddr] ^ val))
                 return;
             switch (da2->crtcaddr) {
@@ -1325,6 +1325,7 @@ da2_out(uint16_t addr, uint16_t val, void *p)
             break;
         case LG_DATA:
             // if(da2->gdcaddr != 8 && da2->gdcaddr != 9) da2_log("DA2 Out addr %03X idx %02X val %02X\n", addr, da2->gdcaddr, val);
+            da2_log("DA2 Out addr %03X idx %02X val %02X\n", addr, da2->gdcaddr, val);
             // if(da2->gdcaddr != 8 && da2->gdcaddr != 9) da2_log("DA2 GCOut idx %X val %02X %04X:%04X esdi %04X:%04X\n", da2->gdcaddr, val, cs >> 4, cpu_state.pc, ES, DI);
             switch (da2->gdcaddr & 0x1f) {
                 case LG_READ_MAP_SELECT:
@@ -1332,6 +1333,11 @@ da2_out(uint16_t addr, uint16_t val, void *p)
                     break;
                 case LG_MODE:
                     da2->writemode = val & 3;
+                     /* reset bit mask and plane mask (? idk when) */
+                    da2->gdcreg[LG_BIT_MASK_LOW] = 0xff;
+                    da2->gdcreg[LG_BIT_MASK_HIGH] = 0xff;
+                    da2->writemask = 0xff;
+                    da2->gdcreg[LG_MAP_MASKJ] = 0xff;
                     break;
                 case LG_MAP_MASKJ:
                     da2->writemask = val & 0xff;
@@ -1429,8 +1435,7 @@ da2_in(uint16_t addr, void *p)
                     //     timer_advance_u64(&da2->bitblt.timer, da2->bitblt.timerspeed);
                     // }
                 }
-                if (da2->bitblt.indata)
-                    temp |= 0x08;
+                if (da2->bitblt.indata) temp |= 0x09;
 #ifdef ENABLE_DA2_DEBUGMONWAIT
                 da2_log("DA2 In %04X(%02X) %04X %04X:%04X\n", addr, da2->ioctladdr, temp, cs >> 4, cpu_state.pc);
 #endif
@@ -1524,7 +1529,7 @@ da2_outb(uint16_t addr, uint8_t val, void *p)
 void
 da2_outw(uint16_t addr, uint16_t val, void *p)
 {
-    // da2_log("DA2 Outw addr %03X val %04X %04X:%04X\n", addr, val, cs >> 4, cpu_state.pc);
+    da2_log("DA2 Outw addr %03X val %04X %04X:%04X\n", addr, val, cs >> 4, cpu_state.pc);
     da2_t *da2      = (da2_t *) p;
     da2->inflipflop = 0;
     switch (addr) {
@@ -1794,14 +1799,17 @@ void
 da2_render_blank(da2_t *da2)
 {
     int x, xx;
+    int cwidth;
+    if (da2->ioctl[LS_MODE] & 0x01) cwidth = 13; /* in character mode */
+    else cwidth = 16;
 
     if (da2->firstline_draw == 2000)
         da2->firstline_draw = da2->displine;
     da2->lastline_draw = da2->displine;
 
     for (x = 0; x < da2->hdisp; x++) {
-        for (xx = 0; xx < 13; xx++)
-            ((uint32_t *) buffer32->line[da2->displine])[(x * 13) + xx + 32] = 0;
+        for (xx = 0; xx < cwidth; xx++)
+            ((uint32_t *) buffer32->line[da2->displine])[(x * cwidth) + xx + 32] = 0;
     }
 }
 /* Display Adapter Mode 8, E Drawing */
@@ -2152,8 +2160,10 @@ da2_recalctimings(da2_t *da2)
     double _dispontime, _dispofftime, disptime;
     
     /* if output disabled or VGA passthrough */
-    if (!(da2->attrc[LV_COMPATIBILITY] & 0x08) || da2->ioctl[LS_MODE] & 0x02)
+    if (da2->ioctl[LS_MODE] & 0x02 || !(da2->attrc[LV_COMPATIBILITY] & 0x08)) {
+        da2->render = da2_render_blank;
         return;
+    }
 
     da2->vtotal      = da2->crtc[LC_VERTICAL_TOTALJ] & 0xfff;
     da2->dispend     = da2->crtc[LC_V_DISPLAY_ENABLE_END] & 0xfff;
@@ -3171,7 +3181,7 @@ da2_init(UNUSED(const device_t *info))
     mem_mapping_disable(&da2->cmapping);
 
     timer_add(&da2->timer, da2_poll, da2, 0);
-    da2->bitblt.timerspeed = 10ull * TIMER_USEC;
+    da2->bitblt.timerspeed = 8ull * TIMER_USEC; /* Bitblt execution speed */
     timer_add(&da2->bitblt.timer, da2_bitblt_dopayload, da2, 0);
 
     return da2;

From f290cc017309494ec6c7bc952b63b2de1e8b1935 Mon Sep 17 00:00:00 2001
From: Cacodemon345 <wahil1976@outlook.com>
Date: Sat, 1 Mar 2025 22:31:10 +0600
Subject: [PATCH 23/58] x64 NDR: Fix crashes/undefined behaviour with 2+ GB of
 allocated RAM

---
 src/codegen_new/codegen_backend_x86-64.c      |  2 +-
 src/codegen_new/codegen_backend_x86-64_defs.h |  2 +-
 src/codegen_new/codegen_backend_x86-64_ops.c  | 64 +++++++++----------
 3 files changed, 34 insertions(+), 34 deletions(-)

diff --git a/src/codegen_new/codegen_backend_x86-64.c b/src/codegen_new/codegen_backend_x86-64.c
index 3cbca28f8..f4a5d024a 100644
--- a/src/codegen_new/codegen_backend_x86-64.c
+++ b/src/codegen_new/codegen_backend_x86-64.c
@@ -360,7 +360,7 @@ codegen_backend_prologue(codeblock_t *block)
         host_x86_MOV32_BASE_OFFSET_REG(block, REG_RSP, IREG_TOP_diff_stack_offset, REG_EAX);
     }
     if (block->flags & CODEBLOCK_NO_IMMEDIATES)
-        host_x86_MOV64_REG_IMM(block, REG_R12, (uintptr_t) ram);
+        host_x86_MOV64_REG_IMM(block, REG_R12, ((uintptr_t) ram) + 2147483648ULL);
 }
 
 void
diff --git a/src/codegen_new/codegen_backend_x86-64_defs.h b/src/codegen_new/codegen_backend_x86-64_defs.h
index 12f05f01c..7c58fec31 100644
--- a/src/codegen_new/codegen_backend_x86-64_defs.h
+++ b/src/codegen_new/codegen_backend_x86-64_defs.h
@@ -1,5 +1,5 @@
 /*RBP = cpu_state + 128
-  R12 = ram (if block->flags & CODEBLOCK_NO_IMMEDIATES)*/
+  R12 = ram + 2147483648 (if block->flags & CODEBLOCK_NO_IMMEDIATES)*/
 #define REG_AX               0
 #define REG_CX               1
 #define REG_DX               2
diff --git a/src/codegen_new/codegen_backend_x86-64_ops.c b/src/codegen_new/codegen_backend_x86-64_ops.c
index 2f65b46e3..e53c58083 100644
--- a/src/codegen_new/codegen_backend_x86-64_ops.c
+++ b/src/codegen_new/codegen_backend_x86-64_ops.c
@@ -505,7 +505,7 @@ host_x86_MOV8_ABS_IMM(codeblock_t *block, void *p, uint32_t imm_data)
 {
     int64_t offset = (uintptr_t) p - (((uintptr_t) &cpu_state) + 128);
 
-    if (offset >= -128 && offset < 127) {
+    if (offset >= -128 && offset <= 127) {
         codegen_alloc_bytes(block, 4);
         codegen_addbyte3(block, 0xc6, 0x45, offset); /*MOVB offset[RBP], imm_data*/
         codegen_addbyte(block, imm_data);
@@ -528,7 +528,7 @@ host_x86_MOV16_ABS_IMM(codeblock_t *block, void *p, uint16_t imm_data)
 {
     int64_t offset = (uintptr_t) p - (((uintptr_t) &cpu_state) + 128);
 
-    if (offset >= -128 && offset < 127) {
+    if (offset >= -128 && offset <= 127) {
         codegen_alloc_bytes(block, 6);
         codegen_addbyte4(block, 0x66, 0xc7, 0x45, offset); /*MOV offset[RBP], imm_data*/
         codegen_addword(block, imm_data);
@@ -551,7 +551,7 @@ host_x86_MOV32_ABS_IMM(codeblock_t *block, void *p, uint32_t imm_data)
 {
     int64_t offset = (uintptr_t) p - (((uintptr_t) &cpu_state) + 128);
 
-    if (offset >= -128 && offset < 127) {
+    if (offset >= -128 && offset <= 127) {
         codegen_alloc_bytes(block, 7);
         codegen_addbyte3(block, 0xc7, 0x45, offset); /*MOV offset[RBP], imm_data*/
         codegen_addlong(block, imm_data);
@@ -578,7 +578,7 @@ host_x86_MOV8_ABS_REG(codeblock_t *block, void *p, int src_reg)
     if (src_reg & 8)
         fatal("host_x86_MOV8_ABS_REG - bad reg\n");
 
-    if (offset >= -128 && offset < 127) {
+    if (offset >= -128 && offset <= 127) {
         codegen_alloc_bytes(block, 3);
         codegen_addbyte3(block, 0x88, 0x45 | ((src_reg & 7) << 3), offset); /*MOVB offset[RBP], src_reg*/
     } else if (offset < (1ULL << 32)) {
@@ -602,7 +602,7 @@ host_x86_MOV16_ABS_REG(codeblock_t *block, void *p, int src_reg)
     if (src_reg & 8)
         fatal("host_x86_MOV16_ABS_REG - bad reg\n");
 
-    if (offset >= -128 && offset < 127) {
+    if (offset >= -128 && offset <= 127) {
         codegen_alloc_bytes(block, 4);
         codegen_addbyte4(block, 0x66, 0x89, 0x45 | ((src_reg & 7) << 3), offset); /*MOV offset[RBP], src_reg*/
     } else if (offset < (1ULL << 32)) {
@@ -622,7 +622,7 @@ host_x86_MOV32_ABS_REG(codeblock_t *block, void *p, int src_reg)
     if (src_reg & 8)
         fatal("host_x86_MOV32_ABS_REG - bad reg\n");
 
-    if (offset >= -128 && offset < 127) {
+    if (offset >= -128 && offset <= 127) {
         codegen_alloc_bytes(block, 3);
         codegen_addbyte3(block, 0x89, 0x45 | ((src_reg & 7) << 3), offset); /*MOV offset[RBP], src_reg*/
     } else if (offset < (1ULL << 32)) {
@@ -646,7 +646,7 @@ host_x86_MOV64_ABS_REG(codeblock_t *block, void *p, int src_reg)
     if (src_reg & 8)
         fatal("host_x86_MOV64_ABS_REG - bad reg\n");
 
-    if (offset >= -128 && offset < 127) {
+    if (offset >= -128 && offset <= 127) {
         codegen_alloc_bytes(block, 4);
         codegen_addbyte4(block, 0x48, 0x89, 0x45 | ((src_reg & 7) << 3), offset); /*MOV offset[RBP], src_reg*/
     } else if (offset < (1ULL << 32)) {
@@ -706,19 +706,19 @@ void
 host_x86_MOV8_REG_ABS(codeblock_t *block, int dst_reg, void *p)
 {
     int64_t offset     = (uintptr_t) p - (((uintptr_t) &cpu_state) + 128);
-    int64_t ram_offset = (uintptr_t) p - (uintptr_t) ram;
+    int64_t ram_offset = (uintptr_t) p - (((uintptr_t) ram) + 2147483648ULL);
 
     if (dst_reg & 8)
         fatal("host_x86_MOV8_REG_ABS reg & 8\n");
 
-    if (offset >= -128 && offset < 127) {
+    if (offset >= -128 && offset <= 127) {
         codegen_alloc_bytes(block, 3);
         codegen_addbyte3(block, 0x8a, 0x45 | ((dst_reg & 7) << 3), offset); /*MOV dst_reg, offset[RBP]*/
     } else if (offset < (1ULL << 32)) {
         codegen_alloc_bytes(block, 6);
         codegen_addbyte2(block, 0x8a, 0x85 | ((dst_reg & 7) << 3)); /*MOV dst_reg, offset[RBP]*/
         codegen_addlong(block, offset);
-    } else if ((ram_offset < (1ULL << 32)) && (block->flags & CODEBLOCK_NO_IMMEDIATES)) {
+    } else if ((ram_offset >= -2147483648LL) && (ram_offset <= 2147483647LL) && (block->flags & CODEBLOCK_NO_IMMEDIATES)) {
         codegen_alloc_bytes(block, 8);
         codegen_addbyte4(block, 0x41, 0x8a, 0x84 | ((dst_reg & 7) << 3), 0x24); /*MOV dst_reg, ram_offset[R12]*/
         codegen_addlong(block, ram_offset);
@@ -730,19 +730,19 @@ void
 host_x86_MOV16_REG_ABS(codeblock_t *block, int dst_reg, void *p)
 {
     int64_t offset     = (uintptr_t) p - (((uintptr_t) &cpu_state) + 128);
-    int64_t ram_offset = (uintptr_t) p - (uintptr_t) ram;
+    int64_t ram_offset = (uintptr_t) p - (((uintptr_t) ram) + 2147483648ULL);
 
     if (dst_reg & 8)
         fatal("host_x86_MOV16_REG_ABS reg & 8\n");
 
-    if (offset >= -128 && offset < 127) {
+    if (offset >= -128 && offset <= 127) {
         codegen_alloc_bytes(block, 4);
         codegen_addbyte4(block, 0x66, 0x8b, 0x45 | ((dst_reg & 7) << 3), offset); /*MOV dst_reg, offset[RBP]*/
     } else if (offset < (1ULL << 32)) {
         codegen_alloc_bytes(block, 7);
         codegen_addbyte3(block, 0x66, 0x8b, 0x85 | ((dst_reg & 7) << 3)); /*MOV dst_reg, offset[RBP]*/
         codegen_addlong(block, offset);
-    } else if ((ram_offset < (1ULL << 32)) && (block->flags & CODEBLOCK_NO_IMMEDIATES)) {
+    } else if ((ram_offset >= -2147483648LL) && (ram_offset <= 2147483647LL) && (block->flags & CODEBLOCK_NO_IMMEDIATES)) {
         codegen_alloc_bytes(block, 9);
         codegen_addbyte4(block, 0x66, 0x41, 0x8b, 0x84 | ((dst_reg & 7) << 3)); /*MOV dst_reg, ram_offset[R12]*/
         codegen_addbyte(block, 0x24);
@@ -760,19 +760,19 @@ void
 host_x86_MOV32_REG_ABS(codeblock_t *block, int dst_reg, void *p)
 {
     int64_t offset     = (uintptr_t) p - (((uintptr_t) &cpu_state) + 128);
-    int64_t ram_offset = (uintptr_t) p - (uintptr_t) ram;
+    int64_t ram_offset = (uintptr_t) p - (((uintptr_t) ram) + 2147483648ULL);
 
     if (dst_reg & 8)
         fatal("host_x86_MOV32_REG_ABS reg & 8\n");
 
-    if (offset >= -128 && offset < 127) {
+    if (offset >= -128 && offset <= 127) {
         codegen_alloc_bytes(block, 3);
         codegen_addbyte3(block, 0x8b, 0x45 | ((dst_reg & 7) << 3), offset); /*MOV dst_reg, offset[RBP]*/
     } else if (offset < (1ULL << 32)) {
         codegen_alloc_bytes(block, 6);
         codegen_addbyte2(block, 0x8b, 0x85 | ((dst_reg & 7) << 3)); /*MOV dst_reg, offset[RBP]*/
         codegen_addlong(block, offset);
-    } else if ((ram_offset < (1ULL << 32)) && (block->flags & CODEBLOCK_NO_IMMEDIATES)) {
+    } else if ((ram_offset >= -2147483648LL) && (ram_offset <= 2147483647LL) && (block->flags & CODEBLOCK_NO_IMMEDIATES)) {
         codegen_alloc_bytes(block, 8);
         codegen_addbyte4(block, 0x41, 0x8b, 0x84 | ((dst_reg & 7) << 3), 0x24); /*MOV dst_reg, ram_offset[R12]*/
         codegen_addlong(block, ram_offset);
@@ -792,7 +792,7 @@ host_x86_MOV64_REG_ABS(codeblock_t *block, int dst_reg, void *p)
     if (dst_reg & 8)
         fatal("host_x86_MOV64_REG_ABS reg & 8\n");
 
-    if (offset >= -128 && offset < 127) {
+    if (offset >= -128 && offset <= 127) {
         codegen_alloc_bytes(block, 4);
         codegen_addbyte4(block, 0x48, 0x8b, 0x45 | ((dst_reg & 7) << 3), offset); /*MOV dst_reg, offset[RBP]*/
     } else if (offset < (1ULL << 32)) {
@@ -845,7 +845,7 @@ host_x86_MOV16_REG_BASE_OFFSET(codeblock_t *block, int dst_reg, int base_reg, in
     if ((dst_reg & 8) || (base_reg & 8))
         fatal("host_x86_MOV16_REG_BASE_OFFSET reg & 8\n");
 
-    if (offset >= -128 && offset < 127) {
+    if (offset >= -128 && offset <= 127) {
         if (base_reg == REG_RSP) {
             codegen_alloc_bytes(block, 5);
             codegen_addbyte(block, 0x66);
@@ -863,7 +863,7 @@ host_x86_MOV32_REG_BASE_OFFSET(codeblock_t *block, int dst_reg, int base_reg, in
     if ((dst_reg & 8) || (base_reg & 8))
         fatal("host_x86_MOV32_REG_BASE_OFFSET reg & 8\n");
 
-    if (offset >= -128 && offset < 127) {
+    if (offset >= -128 && offset <= 127) {
         if (base_reg == REG_RSP) {
             codegen_alloc_bytes(block, 4);
             codegen_addbyte4(block, 0x8b, 0x40 | base_reg | (dst_reg << 3), 0x24, offset);
@@ -880,7 +880,7 @@ host_x86_MOV64_REG_BASE_OFFSET(codeblock_t *block, int dst_reg, int base_reg, in
     if ((dst_reg & 8) || (base_reg & 8))
         fatal("host_x86_MOV64_REG_BASE_OFFSET reg & 8\n");
 
-    if (offset >= -128 && offset < 127) {
+    if (offset >= -128 && offset <= 127) {
         if (base_reg == REG_RSP) {
             codegen_alloc_bytes(block, 5);
             codegen_addbyte(block, 0x48);
@@ -899,7 +899,7 @@ host_x86_MOV32_BASE_OFFSET_REG(codeblock_t *block, int base_reg, int offset, int
     if ((src_reg & 8) || (base_reg & 8))
         fatal("host_x86_MOV32_BASE_OFFSET_REG reg & 8\n");
 
-    if (offset >= -128 && offset < 127) {
+    if (offset >= -128 && offset <= 127) {
         if (base_reg == REG_RSP) {
             codegen_alloc_bytes(block, 4);
             codegen_addbyte4(block, 0x89, 0x40 | base_reg | (src_reg << 3), 0x24, offset);
@@ -916,7 +916,7 @@ host_x86_MOV64_BASE_OFFSET_REG(codeblock_t *block, int base_reg, int offset, int
     if ((src_reg & 8) || (base_reg & 8))
         fatal("host_x86_MOV64_BASE_OFFSET_REG reg & 8\n");
 
-    if (offset >= -128 && offset < 127) {
+    if (offset >= -128 && offset <= 127) {
         if (base_reg == REG_RSP) {
             codegen_alloc_bytes(block, 5);
             codegen_addbyte(block, 0x48);
@@ -935,7 +935,7 @@ host_x86_MOV32_BASE_OFFSET_IMM(codeblock_t *block, int base_reg, int offset, uin
     if (base_reg & 8)
         fatal("host_x86_MOV32_BASE_OFFSET_IMM reg & 8\n");
 
-    if (offset >= -128 && offset < 127) {
+    if (offset >= -128 && offset <= 127) {
         if (base_reg == REG_RSP) {
             codegen_alloc_bytes(block, 8);
             codegen_addbyte4(block, 0xc7, 0x40 | base_reg, 0x24, offset);
@@ -1107,16 +1107,16 @@ void
 host_x86_MOVZX_REG_ABS_16_8(codeblock_t *block, int dst_reg, void *p)
 {
     int64_t offset     = (uintptr_t) p - (((uintptr_t) &cpu_state) + 128);
-    int64_t ram_offset = (uintptr_t) p - (uintptr_t) ram;
+    int64_t ram_offset = (uintptr_t) p - (((uintptr_t) ram) + 2147483648ULL);
 
     if (dst_reg & 8)
         fatal("host_x86_MOVZX_REG_ABS_16_8 - bad reg\n");
 
-    if (offset >= -128 && offset < 127) {
+    if (offset >= -128 && offset <= 127) {
         codegen_alloc_bytes(block, 5);
         codegen_addbyte(block, 0x66);
         codegen_addbyte4(block, 0x0f, 0xb6, 0x45 | ((dst_reg & 7) << 3), offset); /*MOVZX dst_reg, offset[RBP]*/
-    } else if ((ram_offset < (1ULL << 32)) && (block->flags & CODEBLOCK_NO_IMMEDIATES)) {
+    } else if ((ram_offset >= -2147483648LL) && (ram_offset <= 2147483647LL) && (block->flags & CODEBLOCK_NO_IMMEDIATES)) {
         codegen_alloc_bytes(block, 10);
         codegen_addbyte2(block, 0x66, 0x41);
         codegen_addbyte4(block, 0x0f, 0xb6, 0x84 | ((dst_reg & 7) << 3), 0x24); /*MOVZX dst_reg, ram_offset[R12]*/
@@ -1134,14 +1134,14 @@ void
 host_x86_MOVZX_REG_ABS_32_8(codeblock_t *block, int dst_reg, void *p)
 {
     int64_t offset     = (uintptr_t) p - (((uintptr_t) &cpu_state) + 128);
-    int64_t ram_offset = (uintptr_t) p - (uintptr_t) ram;
+    int64_t ram_offset = (uintptr_t) p - (((uintptr_t) ram) + 2147483648ULL);
 
 #if 0
     if (dst_reg & 8)
         fatal("host_x86_MOVZX_REG_ABS_32_8 - bad reg\n");
 #endif
 
-    if (offset >= -128 && offset < 127) {
+    if (offset >= -128 && offset <= 127) {
         if (dst_reg & 8) {
             codegen_alloc_bytes(block, 5);
             codegen_addbyte(block, 0x44);
@@ -1150,7 +1150,7 @@ host_x86_MOVZX_REG_ABS_32_8(codeblock_t *block, int dst_reg, void *p)
             codegen_alloc_bytes(block, 4);
             codegen_addbyte4(block, 0x0f, 0xb6, 0x45 | ((dst_reg & 7) << 3), offset); /*MOVZX dst_reg, offset[RBP]*/
         }
-    } else if ((ram_offset < (1ULL << 32)) && (block->flags & CODEBLOCK_NO_IMMEDIATES)) {
+    } else if ((ram_offset >= -2147483648LL) && (ram_offset <= 2147483647LL) && (block->flags & CODEBLOCK_NO_IMMEDIATES)) {
         if (dst_reg & 8)
             fatal("host_x86_MOVZX_REG_ABS_32_8 - bad reg\n");
 
@@ -1173,15 +1173,15 @@ void
 host_x86_MOVZX_REG_ABS_32_16(codeblock_t *block, int dst_reg, void *p)
 {
     int64_t offset     = (uintptr_t) p - (((uintptr_t) &cpu_state) + 128);
-    int64_t ram_offset = (uintptr_t) p - (uintptr_t) ram;
+    int64_t ram_offset = (uintptr_t) p - (((uintptr_t) ram) + 2147483648ULL);
 
     if (dst_reg & 8)
         fatal("host_x86_MOVZX_REG_ABS_32_16 - bad reg\n");
 
-    if (offset >= -128 && offset < 127) {
+    if (offset >= -128 && offset <= 127) {
         codegen_alloc_bytes(block, 4);
         codegen_addbyte4(block, 0x0f, 0xb7, 0x45 | ((dst_reg & 7) << 3), offset); /*MOVZX dst_reg, offset[RBP]*/
-    } else if ((ram_offset < (1ULL << 32)) && (block->flags & CODEBLOCK_NO_IMMEDIATES)) {
+    } else if ((ram_offset >= -2147483648LL) && (ram_offset <= 2147483647LL) && (block->flags & CODEBLOCK_NO_IMMEDIATES)) {
         codegen_alloc_bytes(block, 9);
         codegen_addbyte(block, 0x41);
         codegen_addbyte4(block, 0x0f, 0xb7, 0x84 | ((dst_reg & 7) << 3), 0x24); /*MOVZX dst_reg, ram_offset[R12]*/

From 6b55b2383a3fd6096f0e39a4c749272ecddc5650 Mon Sep 17 00:00:00 2001
From: Cacodemon345 <wahil1976@outlook.com>
Date: Sun, 2 Mar 2025 23:14:10 +0600
Subject: [PATCH 24/58] Misc NDR changes

Force NDR on ARMv7/ARM64

Make GCC/Clang not use anything in the red zone

Compile with `-fno-omit-frame-pointer`
---
 CMakeLists.txt                               |  7 +++++-
 src/CMakeLists.txt                           |  3 +++
 src/codegen_new/codegen_backend_x86-64.c     | 18 +++++++++++++++
 src/codegen_new/codegen_backend_x86-64_ops.c | 24 ++++++++++----------
 4 files changed, 39 insertions(+), 13 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index aa687222a..5c33bb3e9 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -132,13 +132,18 @@ option(RTMIDI       "RtMidi"
 option(FLUIDSYNTH   "FluidSynth"                                                 ON)
 option(MUNT         "MUNT"                                                       ON)
 option(VNC          "VNC renderer"                                               OFF)
-option(NEW_DYNAREC  "Use the PCem v15 (\"new\") dynamic recompiler"              OFF)
 option(MINITRACE    "Enable Chrome tracing using the modified minitrace library" OFF)
 option(GDBSTUB      "Enable GDB stub server for debugging"                       OFF)
 option(DEV_BRANCH   "Development branch"                                         OFF)
 option(DISCORD      "Discord Rich Presence support"                              ON)
 option(DEBUGREGS486 "Enable debug register opeartion on 486+ CPUs"               OFF)
 
+if((ARCH STREQUAL "arm64") OR (ARCH STREQUAL "arm"))
+    set(NEW_DYNAREC ON)
+else()
+    option(NEW_DYNAREC "Use the PCem v15 (\"new\") dynamic recompiler" OFF)
+endif()
+
 if(WIN32)
     set(QT ON)
     option(CPPTHREADS "C++11 threads" OFF)
diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
index 7d6daff9c..ee17a7a21 100644
--- a/src/CMakeLists.txt
+++ b/src/CMakeLists.txt
@@ -223,6 +223,9 @@ elseif(APPLE AND NOT QT)
     COMPONENT Runtime)
 endif()
 
+if(NEW_DYNAREC AND (ARCH STREQUAL "x86_64") AND (NOT WIN32))
+    add_compile_options(-mno-red-zone -fno-omit-frame-pointer)
+endif()
 
 # Install the PDB file on Windows builds
 if(MSVC)
diff --git a/src/codegen_new/codegen_backend_x86-64.c b/src/codegen_new/codegen_backend_x86-64.c
index f4a5d024a..6242ea40b 100644
--- a/src/codegen_new/codegen_backend_x86-64.c
+++ b/src/codegen_new/codegen_backend_x86-64.c
@@ -315,13 +315,19 @@ codegen_backend_init(void)
 #    endif
     host_x86_CALL(block, (void *) x86gpf);
     codegen_exit_rout = &codeblock[block_current].data[block_pos];
+#ifdef _WIN64
     host_x86_ADD64_REG_IMM(block, REG_RSP, 0x38);
+#else
+    host_x86_ADD64_REG_IMM(block, REG_RSP, 0x48);
+#endif
     host_x86_POP(block, REG_R15);
     host_x86_POP(block, REG_R14);
     host_x86_POP(block, REG_R13);
     host_x86_POP(block, REG_R12);
+#ifdef _WIN64
     host_x86_POP(block, REG_RDI);
     host_x86_POP(block, REG_RSI);
+#endif
     host_x86_POP(block, REG_RBP);
     host_x86_POP(block, REG_RDX);
     host_x86_RET(block);
@@ -346,13 +352,19 @@ codegen_backend_prologue(codeblock_t *block)
     block_pos = BLOCK_START; /*Entry code*/
     host_x86_PUSH(block, REG_RBX);
     host_x86_PUSH(block, REG_RBP);
+#ifdef _WIN64
     host_x86_PUSH(block, REG_RSI);
     host_x86_PUSH(block, REG_RDI);
+#endif
     host_x86_PUSH(block, REG_R12);
     host_x86_PUSH(block, REG_R13);
     host_x86_PUSH(block, REG_R14);
     host_x86_PUSH(block, REG_R15);
+#ifdef _WIN64
     host_x86_SUB64_REG_IMM(block, REG_RSP, 0x38);
+#else
+    host_x86_SUB64_REG_IMM(block, REG_RSP, 0x48);
+#endif
     host_x86_MOV64_REG_IMM(block, REG_RBP, ((uintptr_t) &cpu_state) + 128);
     if (block->flags & CODEBLOCK_HAS_FPU) {
         host_x86_MOV32_REG_ABS(block, REG_EAX, &cpu_state.TOP);
@@ -366,13 +378,19 @@ codegen_backend_prologue(codeblock_t *block)
 void
 codegen_backend_epilogue(codeblock_t *block)
 {
+#ifdef _WIN64
     host_x86_ADD64_REG_IMM(block, REG_RSP, 0x38);
+#else
+    host_x86_ADD64_REG_IMM(block, REG_RSP, 0x48);
+#endif
     host_x86_POP(block, REG_R15);
     host_x86_POP(block, REG_R14);
     host_x86_POP(block, REG_R13);
     host_x86_POP(block, REG_R12);
+#ifdef _WIN64
     host_x86_POP(block, REG_RDI);
     host_x86_POP(block, REG_RSI);
+#endif
     host_x86_POP(block, REG_RBP);
     host_x86_POP(block, REG_RDX);
     host_x86_RET(block);
diff --git a/src/codegen_new/codegen_backend_x86-64_ops.c b/src/codegen_new/codegen_backend_x86-64_ops.c
index e53c58083..39173505b 100644
--- a/src/codegen_new/codegen_backend_x86-64_ops.c
+++ b/src/codegen_new/codegen_backend_x86-64_ops.c
@@ -848,11 +848,11 @@ host_x86_MOV16_REG_BASE_OFFSET(codeblock_t *block, int dst_reg, int base_reg, in
     if (offset >= -128 && offset <= 127) {
         if (base_reg == REG_RSP) {
             codegen_alloc_bytes(block, 5);
-            codegen_addbyte(block, 0x66);
+            codegen_addbyte(block, 0x66); /* MOV dst_reg, [RSP + offset] */
             codegen_addbyte4(block, 0x8b, 0x40 | base_reg | (dst_reg << 3), 0x24, offset);
         } else {
             codegen_alloc_bytes(block, 4);
-            codegen_addbyte4(block, 0x66, 0x8b, 0x40 | base_reg | (dst_reg << 3), offset);
+            codegen_addbyte4(block, 0x66, 0x8b, 0x40 | base_reg | (dst_reg << 3), offset);  /* MOV dst_reg, [base_reg + offset] */
         }
     } else
         fatal("MOV16_REG_BASE_OFFSET - offset %i\n", offset);
@@ -866,10 +866,10 @@ host_x86_MOV32_REG_BASE_OFFSET(codeblock_t *block, int dst_reg, int base_reg, in
     if (offset >= -128 && offset <= 127) {
         if (base_reg == REG_RSP) {
             codegen_alloc_bytes(block, 4);
-            codegen_addbyte4(block, 0x8b, 0x40 | base_reg | (dst_reg << 3), 0x24, offset);
+            codegen_addbyte4(block, 0x8b, 0x40 | base_reg | (dst_reg << 3), 0x24, offset); /* MOV dst_reg, [RSP + offset] */
         } else {
             codegen_alloc_bytes(block, 3);
-            codegen_addbyte3(block, 0x8b, 0x40 | base_reg | (dst_reg << 3), offset);
+            codegen_addbyte3(block, 0x8b, 0x40 | base_reg | (dst_reg << 3), offset);  /* MOV dst_reg, [base_reg + offset] */
         }
     } else
         fatal("MOV32_REG_BASE_OFFSET - offset %i\n", offset);
@@ -883,11 +883,11 @@ host_x86_MOV64_REG_BASE_OFFSET(codeblock_t *block, int dst_reg, int base_reg, in
     if (offset >= -128 && offset <= 127) {
         if (base_reg == REG_RSP) {
             codegen_alloc_bytes(block, 5);
-            codegen_addbyte(block, 0x48);
+            codegen_addbyte(block, 0x48); /* MOV dst_reg, [RSP + offset] */
             codegen_addbyte4(block, 0x8b, 0x40 | base_reg | (dst_reg << 3), 0x24, offset);
         } else {
             codegen_alloc_bytes(block, 4);
-            codegen_addbyte4(block, 0x48, 0x8b, 0x40 | base_reg | (dst_reg << 3), offset);
+            codegen_addbyte4(block, 0x48, 0x8b, 0x40 | base_reg | (dst_reg << 3), offset);  /* MOV dst_reg, [base_reg + offset] */
         }
     } else
         fatal("MOV32_REG_BASE_OFFSET - offset %i\n", offset);
@@ -901,11 +901,11 @@ host_x86_MOV32_BASE_OFFSET_REG(codeblock_t *block, int base_reg, int offset, int
 
     if (offset >= -128 && offset <= 127) {
         if (base_reg == REG_RSP) {
-            codegen_alloc_bytes(block, 4);
+            codegen_alloc_bytes(block, 4); /* MOV [RSP + offset], src_reg*/
             codegen_addbyte4(block, 0x89, 0x40 | base_reg | (src_reg << 3), 0x24, offset);
         } else {
             codegen_alloc_bytes(block, 3);
-            codegen_addbyte3(block, 0x89, 0x40 | base_reg | (src_reg << 3), offset);
+            codegen_addbyte3(block, 0x89, 0x40 | base_reg | (src_reg << 3), offset);  /* MOV [base_reg + offset], src_reg*/
         }
     } else
         fatal("MOV32_BASE_OFFSET_REG - offset %i\n", offset);
@@ -919,11 +919,11 @@ host_x86_MOV64_BASE_OFFSET_REG(codeblock_t *block, int base_reg, int offset, int
     if (offset >= -128 && offset <= 127) {
         if (base_reg == REG_RSP) {
             codegen_alloc_bytes(block, 5);
-            codegen_addbyte(block, 0x48);
+            codegen_addbyte(block, 0x48); /* MOV [RSP + offset], src_reg*/
             codegen_addbyte4(block, 0x89, 0x40 | base_reg | (src_reg << 3), 0x24, offset);
         } else {
             codegen_alloc_bytes(block, 4);
-            codegen_addbyte4(block, 0x48, 0x89, 0x40 | base_reg | (src_reg << 3), offset);
+            codegen_addbyte4(block, 0x48, 0x89, 0x40 | base_reg | (src_reg << 3), offset);  /* MOV [base_reg + offset], src_reg*/
         }
     } else
         fatal("MOV64_BASE_OFFSET_REG - offset %i\n", offset);
@@ -938,11 +938,11 @@ host_x86_MOV32_BASE_OFFSET_IMM(codeblock_t *block, int base_reg, int offset, uin
     if (offset >= -128 && offset <= 127) {
         if (base_reg == REG_RSP) {
             codegen_alloc_bytes(block, 8);
-            codegen_addbyte4(block, 0xc7, 0x40 | base_reg, 0x24, offset);
+            codegen_addbyte4(block, 0xc7, 0x40 | base_reg, 0x24, offset); /* MOV [RSP + offset], imm_data */
             codegen_addlong(block, imm_data);
         } else {
             codegen_alloc_bytes(block, 7);
-            codegen_addbyte3(block, 0xc7, 0x40 | base_reg, offset);
+            codegen_addbyte3(block, 0xc7, 0x40 | base_reg, offset);  /* MOV [base_reg + offset], src_reg*/
             codegen_addlong(block, imm_data);
         }
     } else

From 8d6087a016ba1eb034c9c6d6d99533d124cb418e Mon Sep 17 00:00:00 2001
From: Cacodemon345 <wahil1976@outlook.com>
Date: Mon, 3 Mar 2025 01:06:03 +0600
Subject: [PATCH 25/58] Revert some unnecessary changes for NDR

---
 src/CMakeLists.txt | 4 ----
 1 file changed, 4 deletions(-)

diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
index ee17a7a21..fb3bf9f50 100644
--- a/src/CMakeLists.txt
+++ b/src/CMakeLists.txt
@@ -223,10 +223,6 @@ elseif(APPLE AND NOT QT)
     COMPONENT Runtime)
 endif()
 
-if(NEW_DYNAREC AND (ARCH STREQUAL "x86_64") AND (NOT WIN32))
-    add_compile_options(-mno-red-zone -fno-omit-frame-pointer)
-endif()
-
 # Install the PDB file on Windows builds
 if(MSVC)
     # CMake fully supports PDB files on MSVC-compatible compilers

From 54650963f02a995264d2b4cb3fb103c9bc0dac52 Mon Sep 17 00:00:00 2001
From: The Dax <thedax@users.noreply.github.com>
Date: Sun, 2 Mar 2025 14:12:25 -0500
Subject: [PATCH 26/58] Fix Acer 100T machine not showing up on Linux due to
 case-sensitivity.

---
 src/machine/m_at_286_386sx.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/machine/m_at_286_386sx.c b/src/machine/m_at_286_386sx.c
index 24cb88f09..9086f7c8a 100644
--- a/src/machine/m_at_286_386sx.c
+++ b/src/machine/m_at_286_386sx.c
@@ -742,7 +742,7 @@ machine_at_acer100t_init(const machine_t *model)
 {
     int ret;
 
-    ret = bios_load_linear("roms/machines/acer100t/acer386.bin",
+    ret = bios_load_linear("roms/machines/acer100t/acer386.BIN",
                            0x000f0000, 65536, 0);
 
     if (bios_only || !ret)

From fae26729f103010ddc7e621a37083bc4ae6d5736 Mon Sep 17 00:00:00 2001
From: Cacodemon345 <wahil1976@outlook.com>
Date: Mon, 3 Mar 2025 22:03:52 +0600
Subject: [PATCH 27/58] S3 ViRGE: Make IRQs happen in main thread

FIFO is also no longer immediately woken up upon writes
---
 src/video/vid_s3_virge.c | 57 +++++++++++++++++++++++++++++++++++-----
 1 file changed, 51 insertions(+), 6 deletions(-)

diff --git a/src/video/vid_s3_virge.c b/src/video/vid_s3_virge.c
index 91b186240..05cd42459 100644
--- a/src/video/vid_s3_virge.c
+++ b/src/video/vid_s3_virge.c
@@ -338,6 +338,7 @@ typedef struct virge_t {
     event_t *    fifo_not_full_event;
 
     atomic_int   virge_busy;
+    atomic_uint  virge_irq_req;
 
     uint8_t      subsys_stat;
     uint8_t      subsys_cntl;
@@ -373,6 +374,9 @@ typedef struct virge_t {
     uint32_t     dma_mmio_addr;
     uint32_t     dma_data_type;
 
+    pc_timer_t   wake_timer;
+    pc_timer_t   irq_timer;
+
     int          pci;
     int          is_agp;
 } virge_t;
@@ -383,6 +387,28 @@ wake_fifo_thread(virge_t *virge) {
     thread_set_event(virge->wake_fifo_thread);
 }
 
+void
+s3_virge_wake_fifo_timer(void* priv)
+{
+    virge_t *virge = (virge_t *) priv;
+
+    thread_set_event(virge->wake_fifo_thread); /*Wake up FIFO thread if moving from idle*/
+}
+
+void
+s3_virge_wake_fifo_thread(virge_t *virge)
+{
+    if (!timer_is_enabled(&virge->wake_timer)) {
+        /*Don't wake FIFO thread immediately - if we do that it will probably
+          process one word and go back to sleep, requiring it to be woken on
+          almost every write. Instead, wait a short while so that the CPU
+          emulation writes more data so we have more batched-up work.*/
+        timer_on_auto(&virge->wake_timer, 100.0);
+    }
+    if (!timer_is_enabled(&virge->irq_timer))
+        timer_on_auto(&virge->irq_timer, 100.0);
+}
+
 static virge_t         *reset_state = NULL;
 
 static video_timings_t timing_diamond_stealth3d_2000_pci = { .type = VIDEO_PCI, .write_b = 2, .write_w = 2, .write_l = 3, .read_b = 28, .read_w = 28, .read_l = 45 };
@@ -471,6 +497,19 @@ s3_virge_update_irqs(virge_t *virge)
         pci_clear_irq(virge->pci_slot, PCI_INTA, &virge->irq_state);
 }
 
+void
+s3_virge_update_irq_timer(void* priv)
+{
+    virge_t *virge = (virge_t *) priv;
+
+    if (virge->virge_irq_req) {
+        virge->virge_irq_req--;
+        s3_virge_update_irqs(virge);
+    }
+
+    timer_on_auto(&virge->irq_timer, 100.);
+}
+
 static void
 s3_virge_out(uint16_t addr, uint8_t val, void *priv)
 {
@@ -1101,6 +1140,9 @@ static void
 s3_virge_vblank_start(svga_t *svga) {
     virge_t *virge = (virge_t *) svga->priv;
 
+    if (virge->virge_irq_req)
+        virge->virge_irq_req--;
+
     virge->subsys_stat |= INT_VSY;
     s3_virge_update_irqs(virge);
 }
@@ -1795,7 +1837,8 @@ fifo_thread(void *param)
          if (virge->cmd_dma)
             virge->subsys_stat |= (INT_HOST_DONE | INT_CMD_DONE);
 
-         s3_virge_update_irqs(virge);
+         //s3_virge_update_irqs(virge);
+         virge->virge_irq_req++;
     }
 }
 
@@ -1835,10 +1878,8 @@ s3_virge_queue(virge_t *virge, uint32_t addr, uint32_t val, uint32_t type)
 
     virge->fifo_write_idx++;
 
-    if (FIFO_ENTRIES > 0xe000)
-        wake_fifo_thread(virge);
-    if (FIFO_ENTRIES > 0xe000 || FIFO_ENTRIES < 8)
-        wake_fifo_thread(virge);
+    if (FIFO_ENTRIES > 0xe000 || FIFO_ENTRIES < 16)
+        s3_virge_wake_fifo_thread(virge);
 }
 
 static void
@@ -3642,7 +3683,8 @@ render_thread(void *param)
         }
         virge->s3d_busy = 0;
         virge->subsys_stat |= INT_S3D_DONE;
-        s3_virge_update_irqs(virge);
+        //s3_virge_update_irqs(virge);
+        virge->virge_irq_req++;
     }
 }
 
@@ -4571,6 +4613,9 @@ s3_virge_init(const device_t *info)
     virge->fifo_not_full_event = thread_create_event();
     virge->fifo_thread = thread_create(fifo_thread, virge);
 
+    timer_add(&virge->wake_timer, s3_virge_wake_fifo_timer, virge, 0);
+    timer_add(&virge->irq_timer, s3_virge_update_irq_timer, virge, 0);
+
     virge->local = info->local;
 
     *reset_state = *virge;

From e24c9d51633730b165aa7c332600df8aee140c79 Mon Sep 17 00:00:00 2001
From: Cacodemon345 <wahil1976@outlook.com>
Date: Mon, 3 Mar 2025 23:22:24 +0600
Subject: [PATCH 28/58] S3 ViRGE: Use a significantly faster version of the ROP
 code

DirectDraw tests now run significantly faster, and Direct3D tests also now run smoothly.
---
 src/video/vid_s3_virge.c | 788 ++++++++++++++++++++++++++++++++++++++-
 1 file changed, 775 insertions(+), 13 deletions(-)

diff --git a/src/video/vid_s3_virge.c b/src/video/vid_s3_virge.c
index 05cd42459..62ffcff0c 100644
--- a/src/video/vid_s3_virge.c
+++ b/src/video/vid_s3_virge.c
@@ -2216,19 +2216,781 @@ s3_virge_mmio_write_l(uint32_t addr, uint32_t val, void *priv)
             }                                       \
         } while (0)
 
-#define MIX()                                  \
-    do {                                       \
-        int c;                                 \
-        for (c = 0; c < 24; c++) {             \
-            int d = (dest & (1 << c)) ? 1 : 0; \
-            if (source & (1 << c))             \
-                d |= 2;                        \
-            if (pattern & (1 << c))            \
-                d |= 4;                        \
-            if (virge->s3d.rop & (1 << d))     \
-                out |= (1 << c);               \
-        }                                      \
-    } while (0)
+#define ROPMIX(R, D, P, S, out)                    \
+    {                                              \
+        switch (R) {                               \
+            case 0x00:                             \
+                out = 0;                           \
+                break;                             \
+            case 0x01:                             \
+                out = ~(D | (P | S));              \
+                break;                             \
+            case 0x02:                             \
+                out = D & ~(P | S);                \
+                break;                             \
+            case 0x03:                             \
+                out = ~(P | S);                    \
+                break;                             \
+            case 0x04:                             \
+                out = S & ~(D | P);                \
+                break;                             \
+            case 0x05:                             \
+                out = ~(D | P);                    \
+                break;                             \
+            case 0x06:                             \
+                out = ~(P | ~(D ^ S));             \
+                break;                             \
+            case 0x07:                             \
+                out = ~(P | (D & S));              \
+                break;                             \
+            case 0x08:                             \
+                out = S & (D & ~P);                \
+                break;                             \
+            case 0x09:                             \
+                out = ~(P | (D ^ S));              \
+                break;                             \
+            case 0x0a:                             \
+                out = D & ~P;                      \
+                break;                             \
+            case 0x0b:                             \
+                out = ~(P | (S & ~D));             \
+                break;                             \
+            case 0x0c:                             \
+                out = S & ~P;                      \
+                break;                             \
+            case 0x0d:                             \
+                out = ~(P | (D & ~S));             \
+                break;                             \
+            case 0x0e:                             \
+                out = ~(P | ~(D | S));             \
+                break;                             \
+            case 0x0f:                             \
+                out = ~P;                          \
+                break;                             \
+            case 0x10:                             \
+                out = P & ~(D | S);                \
+                break;                             \
+            case 0x11:                             \
+                out = ~(D | S);                    \
+                break;                             \
+            case 0x12:                             \
+                out = ~(S | ~(D ^ P));             \
+                break;                             \
+            case 0x13:                             \
+                out = ~(S | (D & P));              \
+                break;                             \
+            case 0x14:                             \
+                out = ~(D | ~(P ^ S));             \
+                break;                             \
+            case 0x15:                             \
+                out = ~(D | (P & S));              \
+                break;                             \
+            case 0x16:                             \
+                out = P ^ (S ^ (D & ~(P & S)));    \
+                break;                             \
+            case 0x17:                             \
+                out = ~(S ^ ((S ^ P) & (D ^ S)));  \
+                break;                             \
+            case 0x18:                             \
+                out = (S ^ P) & (P ^ D);           \
+                break;                             \
+            case 0x19:                             \
+                out = ~(S ^ (D & ~(P & S)));       \
+                break;                             \
+            case 0x1a:                             \
+                out = P ^ (D | (S & P));           \
+                break;                             \
+            case 0x1b:                             \
+                out = ~(S ^ (D & (P ^ S)));        \
+                break;                             \
+            case 0x1c:                             \
+                out = P ^ (S | (D & P));           \
+                break;                             \
+            case 0x1d:                             \
+                out = ~(D ^ (S & (P ^ D)));        \
+                break;                             \
+            case 0x1e:                             \
+                out = P ^ (D | S);                 \
+                break;                             \
+            case 0x1f:                             \
+                out = ~(P & (D | S));              \
+                break;                             \
+            case 0x20:                             \
+                out = D & (P & ~S);                \
+                break;                             \
+            case 0x21:                             \
+                out = ~(S | (D ^ P));              \
+                break;                             \
+            case 0x22:                             \
+                out = D & ~S;                      \
+                break;                             \
+            case 0x23:                             \
+                out = ~(S | (P & ~D));             \
+                break;                             \
+            case 0x24:                             \
+                out = (S ^ P) & (D ^ S);           \
+                break;                             \
+            case 0x25:                             \
+                out = ~(P ^ (D & ~(S & P)));       \
+                break;                             \
+            case 0x26:                             \
+                out = S ^ (D | (P & S));           \
+                break;                             \
+            case 0x27:                             \
+                out = S ^ (D | ~(P ^ S));          \
+                break;                             \
+            case 0x28:                             \
+                out = D & (P ^ S);                 \
+                break;                             \
+            case 0x29:                             \
+                out = ~(P ^ (S ^ (D | (P & S))));  \
+                break;                             \
+            case 0x2a:                             \
+                out = D & ~(P & S);                \
+                break;                             \
+            case 0x2b:                             \
+                out = ~(S ^ ((S ^ P) & (P ^ D)));  \
+                break;                             \
+            case 0x2c:                             \
+                out = S ^ (P & (D | S));           \
+                break;                             \
+            case 0x2d:                             \
+                out = P ^ (S | ~D);                \
+                break;                             \
+            case 0x2e:                             \
+                out = P ^ (S | (D ^ P));           \
+                break;                             \
+            case 0x2f:                             \
+                out = ~(P & (S | ~D));             \
+                break;                             \
+            case 0x30:                             \
+                out = P & ~S;                      \
+                break;                             \
+            case 0x31:                             \
+                out = ~(S | (D & ~P));             \
+                break;                             \
+            case 0x32:                             \
+                out = S ^ (D | (P | S));           \
+                break;                             \
+            case 0x33:                             \
+                out = ~S;                          \
+                break;                             \
+            case 0x34:                             \
+                out = S ^ (P | (D & S));           \
+                break;                             \
+            case 0x35:                             \
+                out = S ^ (P | ~(D ^ S));          \
+                break;                             \
+            case 0x36:                             \
+                out = S ^ (D | P);                 \
+                break;                             \
+            case 0x37:                             \
+                out = ~(S & (D | P));              \
+                break;                             \
+            case 0x38:                             \
+                out = P ^ (S & (D | P));           \
+                break;                             \
+            case 0x39:                             \
+                out = S ^ (P | ~D);                \
+                break;                             \
+            case 0x3a:                             \
+                out = S ^ (P | (D ^ S));           \
+                break;                             \
+            case 0x3b:                             \
+                out = ~(S & (P | ~D));             \
+                break;                             \
+            case 0x3c:                             \
+                out = P ^ S;                       \
+                break;                             \
+            case 0x3d:                             \
+                out = S ^ (P | ~(D | S));          \
+                break;                             \
+            case 0x3e:                             \
+                out = S ^ (P | (D & ~S));          \
+                break;                             \
+            case 0x3f:                             \
+                out = ~(P & S);                    \
+                break;                             \
+            case 0x40:                             \
+                out = P & (S & ~D);                \
+                break;                             \
+            case 0x41:                             \
+                out = ~(D | (P ^ S));              \
+                break;                             \
+            case 0x42:                             \
+                out = (S ^ D) & (P ^ D);           \
+                break;                             \
+            case 0x43:                             \
+                out = ~(S ^ (P & ~(D & S)));       \
+                break;                             \
+            case 0x44:                             \
+                out = S & ~D;                      \
+                break;                             \
+            case 0x45:                             \
+                out = ~(D | (P & ~S));             \
+                break;                             \
+            case 0x46:                             \
+                out = D ^ (S | (P & D));           \
+                break;                             \
+            case 0x47:                             \
+                out = ~(P ^ (S & (D ^ P)));        \
+                break;                             \
+            case 0x48:                             \
+                out = S & (D ^ P);                 \
+                break;                             \
+            case 0x49:                             \
+                out = ~(P ^ (D ^ (S | (P & D))));  \
+                break;                             \
+            case 0x4a:                             \
+                out = D ^ (P & (S | D));           \
+                break;                             \
+            case 0x4b:                             \
+                out = P ^ (D | ~S);                \
+                break;                             \
+            case 0x4c:                             \
+                out = S & ~(D & P);                \
+                break;                             \
+            case 0x4d:                             \
+                out = ~(S ^ ((S ^ P) | (D ^ S)));  \
+                break;                             \
+            case 0x4e:                             \
+                out = P ^ (D | (S ^ P));           \
+                break;                             \
+            case 0x4f:                             \
+                out = ~(P & (D | ~S));             \
+                break;                             \
+            case 0x50:                             \
+                out = P & ~D;                      \
+                break;                             \
+            case 0x51:                             \
+                out = ~(D | (S & ~P));             \
+                break;                             \
+            case 0x52:                             \
+                out = D ^ (P | (S & D));           \
+                break;                             \
+            case 0x53:                             \
+                out = ~(S ^ (P & (D ^ S)));        \
+                break;                             \
+            case 0x54:                             \
+                out = ~(D | ~(P | S));             \
+                break;                             \
+            case 0x55:                             \
+                out = ~D;                          \
+                break;                             \
+            case 0x56:                             \
+                out = D ^ (P | S);                 \
+                break;                             \
+            case 0x57:                             \
+                out = ~(D & (P | S));              \
+                break;                             \
+            case 0x58:                             \
+                out = P ^ (D & (S | P));           \
+                break;                             \
+            case 0x59:                             \
+                out = D ^ (P | ~S);                \
+                break;                             \
+            case 0x5a:                             \
+                out = D ^ P;                       \
+                break;                             \
+            case 0x5b:                             \
+                out = D ^ (P | ~(S | D));          \
+                break;                             \
+            case 0x5c:                             \
+                out = D ^ (P | (S ^ D));           \
+                break;                             \
+            case 0x5d:                             \
+                out = ~(D & (P | ~S));             \
+                break;                             \
+            case 0x5e:                             \
+                out = D ^ (P | (S & ~D));          \
+                break;                             \
+            case 0x5f:                             \
+                out = ~(D & P);                    \
+                break;                             \
+            case 0x60:                             \
+                out = P & (D ^ S);                 \
+                break;                             \
+            case 0x61:                             \
+                out = ~(D ^ (S ^ (P | (D & S))));  \
+                break;                             \
+            case 0x62:                             \
+                out = D ^ (S & (P | D));           \
+                break;                             \
+            case 0x63:                             \
+                out = S ^ (D | ~P);                \
+                break;                             \
+            case 0x64:                             \
+                out = S ^ (D & (P | S));           \
+                break;                             \
+            case 0x65:                             \
+                out = D ^ (S | ~P);                \
+                break;                             \
+            case 0x66:                             \
+                out = D ^ S;                       \
+                break;                             \
+            case 0x67:                             \
+                out = S ^ (D | ~(P | S));          \
+                break;                             \
+            case 0x68:                             \
+                out = ~(D ^ (S ^ (P | ~(D | S)))); \
+                break;                             \
+            case 0x69:                             \
+                out = ~(P ^ (D ^ S));              \
+                break;                             \
+            case 0x6a:                             \
+                out = D ^ (P & S);                 \
+                break;                             \
+            case 0x6b:                             \
+                out = ~(P ^ (S ^ (D & (P | S))));  \
+                break;                             \
+            case 0x6c:                             \
+                out = S ^ (D & P);                 \
+                break;                             \
+            case 0x6d:                             \
+                out = ~(P ^ (D ^ (S & (P | D))));  \
+                break;                             \
+            case 0x6e:                             \
+                out = S ^ (D & (P | ~S));          \
+                break;                             \
+            case 0x6f:                             \
+                out = ~(P & ~(D ^ S));             \
+                break;                             \
+            case 0x70:                             \
+                out = P & ~(D & S);                \
+                break;                             \
+            case 0x71:                             \
+                out = ~(S ^ ((S ^ D) & (P ^ D)));  \
+                break;                             \
+            case 0x72:                             \
+                out = S ^ (D | (P ^ S));           \
+                break;                             \
+            case 0x73:                             \
+                out = ~(S & (D | ~P));             \
+                break;                             \
+            case 0x74:                             \
+                out = D ^ (S | (P ^ D));           \
+                break;                             \
+            case 0x75:                             \
+                out = ~(D & (S | ~P));             \
+                break;                             \
+            case 0x76:                             \
+                out = S ^ (D | (P & ~S));          \
+                break;                             \
+            case 0x77:                             \
+                out = ~(D & S);                    \
+                break;                             \
+            case 0x78:                             \
+                out = P ^ (D & S);                 \
+                break;                             \
+            case 0x79:                             \
+                out = ~(D ^ (S ^ (P & (D | S))));  \
+                break;                             \
+            case 0x7a:                             \
+                out = D ^ (P & (S | ~D));          \
+                break;                             \
+            case 0x7b:                             \
+                out = ~(S & ~(D ^ P));             \
+                break;                             \
+            case 0x7c:                             \
+                out = S ^ (P & (D | ~S));          \
+                break;                             \
+            case 0x7d:                             \
+                out = ~(D & ~(P ^ S));             \
+                break;                             \
+            case 0x7e:                             \
+                out = (S ^ P) | (D ^ S);           \
+                break;                             \
+            case 0x7f:                             \
+                out = ~(D & (P & S));              \
+                break;                             \
+            case 0x80:                             \
+                out = D & (P & S);                 \
+                break;                             \
+            case 0x81:                             \
+                out = ~((S ^ P) | (D ^ S));        \
+                break;                             \
+            case 0x82:                             \
+                out = D & ~(P ^ S);                \
+                break;                             \
+            case 0x83:                             \
+                out = ~(S ^ (P & (D | ~S)));       \
+                break;                             \
+            case 0x84:                             \
+                out = S & ~(D ^ P);                \
+                break;                             \
+            case 0x85:                             \
+                out = ~(P ^ (D & (S | ~P)));       \
+                break;                             \
+            case 0x86:                             \
+                out = D ^ (S ^ (P & (D | S)));     \
+                break;                             \
+            case 0x87:                             \
+                out = ~(P ^ (D & S));              \
+                break;                             \
+            case 0x88:                             \
+                out = D & S;                       \
+                break;                             \
+            case 0x89:                             \
+                out = ~(S ^ (D | (P & ~S)));       \
+                break;                             \
+            case 0x8a:                             \
+                out = D & (S | ~P);                \
+                break;                             \
+            case 0x8b:                             \
+                out = ~(D ^ (S | (P ^ D)));        \
+                break;                             \
+            case 0x8c:                             \
+                out = S & (D | ~P);                \
+                break;                             \
+            case 0x8d:                             \
+                out = ~(S ^ (D | (P ^ S)));        \
+                break;                             \
+            case 0x8e:                             \
+                out = S ^ ((S ^ D) & (P ^ D));     \
+                break;                             \
+            case 0x8f:                             \
+                out = ~(P & ~(D & S));             \
+                break;                             \
+            case 0x90:                             \
+                out = P & ~(D ^ S);                \
+                break;                             \
+            case 0x91:                             \
+                out = ~(S ^ (D & (P | ~S)));       \
+                break;                             \
+            case 0x92:                             \
+                out = D ^ (P ^ (S & (D | P)));     \
+                break;                             \
+            case 0x93:                             \
+                out = ~(S ^ (P & D));              \
+                break;                             \
+            case 0x94:                             \
+                out = P ^ (S ^ (D & (P | S)));     \
+                break;                             \
+            case 0x95:                             \
+                out = ~(D ^ (P & S));              \
+                break;                             \
+            case 0x96:                             \
+                out = D ^ (P ^ S);                 \
+                break;                             \
+            case 0x97:                             \
+                out = P ^ (S ^ (D | ~(P | S)));    \
+                break;                             \
+            case 0x98:                             \
+                out = ~(S ^ (D | ~(P | S)));       \
+                break;                             \
+            case 0x99:                             \
+                out = ~(D ^ S);                    \
+                break;                             \
+            case 0x9a:                             \
+                out = D ^ (P & ~S);                \
+                break;                             \
+            case 0x9b:                             \
+                out = ~(S ^ (D & (P | S)));        \
+                break;                             \
+            case 0x9c:                             \
+                out = S ^ (P & ~D);                \
+                break;                             \
+            case 0x9d:                             \
+                out = ~(D ^ (S & (P | D)));        \
+                break;                             \
+            case 0x9e:                             \
+                out = D ^ (S ^ (P | (D & S)));     \
+                break;                             \
+            case 0x9f:                             \
+                out = ~(P & (D ^ S));              \
+                break;                             \
+            case 0xa0:                             \
+                out = D & P;                       \
+                break;                             \
+            case 0xa1:                             \
+                out = ~(P ^ (D | (S & ~P)));       \
+                break;                             \
+            case 0xa2:                             \
+                out = D & (P | ~S);                \
+                break;                             \
+            case 0xa3:                             \
+                out = ~(D ^ (P | (S ^ D)));        \
+                break;                             \
+            case 0xa4:                             \
+                out = ~(P ^ (D | ~(S | P)));       \
+                break;                             \
+            case 0xa5:                             \
+                out = ~(P ^ D);                    \
+                break;                             \
+            case 0xa6:                             \
+                out = D ^ (S & ~P);                \
+                break;                             \
+            case 0xa7:                             \
+                out = ~(P ^ (D & (S | P)));        \
+                break;                             \
+            case 0xa8:                             \
+                out = D & (P | S);                 \
+                break;                             \
+            case 0xa9:                             \
+                out = ~(D ^ (P | S));              \
+                break;                             \
+            case 0xaa:                             \
+                out = D;                           \
+                break;                             \
+            case 0xab:                             \
+                out = D | ~(P | S);                \
+                break;                             \
+            case 0xac:                             \
+                out = S ^ (P & (D ^ S));           \
+                break;                             \
+            case 0xad:                             \
+                out = ~(D ^ (P | (S & D)));        \
+                break;                             \
+            case 0xae:                             \
+                out = D | (S & ~P);                \
+                break;                             \
+            case 0xaf:                             \
+                out = D | ~P;                      \
+                break;                             \
+            case 0xb0:                             \
+                out = P & (D | ~S);                \
+                break;                             \
+            case 0xb1:                             \
+                out = ~(P ^ (D | (S ^ P)));        \
+                break;                             \
+            case 0xb2:                             \
+                out = S ^ ((S ^ P) | (D ^ S));     \
+                break;                             \
+            case 0xb3:                             \
+                out = ~(S & ~(D & P));             \
+                break;                             \
+            case 0xb4:                             \
+                out = P ^ (S & ~D);                \
+                break;                             \
+            case 0xb5:                             \
+                out = ~(D ^ (P & (S | D)));        \
+                break;                             \
+            case 0xb6:                             \
+                out = D ^ (P ^ (S | (D & P)));     \
+                break;                             \
+            case 0xb7:                             \
+                out = ~(S & (D ^ P));              \
+                break;                             \
+            case 0xb8:                             \
+                out = P ^ (S & (D ^ P));           \
+                break;                             \
+            case 0xb9:                             \
+                out = ~(D ^ (S | (P & D)));        \
+                break;                             \
+            case 0xba:                             \
+                out = D | (P & ~S);                \
+                break;                             \
+            case 0xbb:                             \
+                out = D | ~S;                      \
+                break;                             \
+            case 0xbc:                             \
+                out = S ^ (P & ~(D & S));          \
+                break;                             \
+            case 0xbd:                             \
+                out = ~((S ^ D) & (P ^ D));        \
+                break;                             \
+            case 0xbe:                             \
+                out = D | (P ^ S);                 \
+                break;                             \
+            case 0xbf:                             \
+                out = D | ~(P & S);                \
+                break;                             \
+            case 0xc0:                             \
+                out = P & S;                       \
+                break;                             \
+            case 0xc1:                             \
+                out = ~(S ^ (P | (D & ~S)));       \
+                break;                             \
+            case 0xc2:                             \
+                out = ~(S ^ (P | ~(D | S)));       \
+                break;                             \
+            case 0xc3:                             \
+                out = ~(P ^ S);                    \
+                break;                             \
+            case 0xc4:                             \
+                out = S & (P | ~D);                \
+                break;                             \
+            case 0xc5:                             \
+                out = ~(S ^ (P | (D ^ S)));        \
+                break;                             \
+            case 0xc6:                             \
+                out = S ^ (D & ~P);                \
+                break;                             \
+            case 0xc7:                             \
+                out = ~(P ^ (S & (D | P)));        \
+                break;                             \
+            case 0xc8:                             \
+                out = S & (D | P);                 \
+                break;                             \
+            case 0xc9:                             \
+                out = ~(S ^ (P | D));              \
+                break;                             \
+            case 0xca:                             \
+                out = D ^ (P & (S ^ D));           \
+                break;                             \
+            case 0xcb:                             \
+                out = ~(S ^ (P | (D & S)));        \
+                break;                             \
+            case 0xcc:                             \
+                out = S;                           \
+                break;                             \
+            case 0xcd:                             \
+                out = S | ~(D | P);                \
+                break;                             \
+            case 0xce:                             \
+                out = S | (D & ~P);                \
+                break;                             \
+            case 0xcf:                             \
+                out = S | ~P;                      \
+                break;                             \
+            case 0xd0:                             \
+                out = P & (S | ~D);                \
+                break;                             \
+            case 0xd1:                             \
+                out = ~(P ^ (S | (D ^ P)));        \
+                break;                             \
+            case 0xd2:                             \
+                out = P ^ (D & ~S);                \
+                break;                             \
+            case 0xd3:                             \
+                out = ~(S ^ (P & (D | S)));        \
+                break;                             \
+            case 0xd4:                             \
+                out = S ^ ((S ^ P) & (P ^ D));     \
+                break;                             \
+            case 0xd5:                             \
+                out = ~(D & ~(P & S));             \
+                break;                             \
+            case 0xd6:                             \
+                out = P ^ (S ^ (D | (P & S)));     \
+                break;                             \
+            case 0xd7:                             \
+                out = ~(D & (P ^ S));              \
+                break;                             \
+            case 0xd8:                             \
+                out = P ^ (D & (S ^ P));           \
+                break;                             \
+            case 0xd9:                             \
+                out = ~(S ^ (D | (P & S)));        \
+                break;                             \
+            case 0xda:                             \
+                out = D ^ (P & ~(S & D));          \
+                break;                             \
+            case 0xdb:                             \
+                out = ~((S ^ P) & (D ^ S));        \
+                break;                             \
+            case 0xdc:                             \
+                out = S | (P & ~D);                \
+                break;                             \
+            case 0xdd:                             \
+                out = S | ~D;                      \
+                break;                             \
+            case 0xde:                             \
+                out = S | (D ^ P);                 \
+                break;                             \
+            case 0xdf:                             \
+                out = S | ~(D & P);                \
+                break;                             \
+            case 0xe0:                             \
+                out = P & (D | S);                 \
+                break;                             \
+            case 0xe1:                             \
+                out = ~(P ^ (D | S));              \
+                break;                             \
+            case 0xe2:                             \
+                out = D ^ (S & (P ^ D));           \
+                break;                             \
+            case 0xe3:                             \
+                out = ~(P ^ (S | (D & P)));        \
+                break;                             \
+            case 0xe4:                             \
+                out = S ^ (D & (P ^ S));           \
+                break;                             \
+            case 0xe5:                             \
+                out = ~(P ^ (D | (S & P)));        \
+                break;                             \
+            case 0xe6:                             \
+                out = S ^ (D & ~(P & S));          \
+                break;                             \
+            case 0xe7:                             \
+                out = ~((S ^ P) & (P ^ D));        \
+                break;                             \
+            case 0xe8:                             \
+                out = S ^ ((S ^ P) & (D ^ S));     \
+                break;                             \
+            case 0xe9:                             \
+                out = ~(D ^ (S ^ (P & ~(D & S)))); \
+                break;                             \
+            case 0xea:                             \
+                out = D | (P & S);                 \
+                break;                             \
+            case 0xeb:                             \
+                out = D | ~(P ^ S);                \
+                break;                             \
+            case 0xec:                             \
+                out = S | (D & P);                 \
+                break;                             \
+            case 0xed:                             \
+                out = S | ~(D ^ P);                \
+                break;                             \
+            case 0xee:                             \
+                out = D | S;                       \
+                break;                             \
+            case 0xef:                             \
+                out = S | (D | ~P);                \
+                break;                             \
+            case 0xf0:                             \
+                out = P;                           \
+                break;                             \
+            case 0xf1:                             \
+                out = P | ~(D | S);                \
+                break;                             \
+            case 0xf2:                             \
+                out = P | (D & ~S);                \
+                break;                             \
+            case 0xf3:                             \
+                out = P | ~S;                      \
+                break;                             \
+            case 0xf4:                             \
+                out = P | (S & ~D);                \
+                break;                             \
+            case 0xf5:                             \
+                out = P | ~D;                      \
+                break;                             \
+            case 0xf6:                             \
+                out = P | (D ^ S);                 \
+                break;                             \
+            case 0xf7:                             \
+                out = P | ~(D & S);                \
+                break;                             \
+            case 0xf8:                             \
+                out = P | (D & S);                 \
+                break;                             \
+            case 0xf9:                             \
+                out = P | ~(D ^ S);                \
+                break;                             \
+            case 0xfa:                             \
+                out = D | P;                       \
+                break;                             \
+            case 0xfb:                             \
+                out = D | (P | ~S);                \
+                break;                             \
+            case 0xfc:                             \
+                out = P | S;                       \
+                break;                             \
+            case 0xfd:                             \
+                out = P | (S | ~D);                \
+                break;                             \
+            case 0xfe:                             \
+                out = D | (P | S);                 \
+                break;                             \
+            case 0xff:                             \
+                out = ~0;                          \
+                break;                             \
+        }                                          \
+    }
+
+#define MIX() do { ROPMIX(virge->s3d.rop & 0xFF, dest, pattern, source, out); out &= 0xFFFFFF; } while (0)
 
 #define WRITE(addr, val)                                                         \
     do {                                                                         \

From 21f5415328cd5c101d020f3d1c601ba16322c53f Mon Sep 17 00:00:00 2001
From: Cacodemon345 <wahil1976@outlook.com>
Date: Tue, 4 Mar 2025 01:31:19 +0600
Subject: [PATCH 29/58] Trident TGUI: Use a significantly faster version of the
 ROP code

At least 30%-40% performance improvement seen
---
 src/video/vid_tgui9440.c | 790 ++++++++++++++++++++++++++++++++++++++-
 1 file changed, 778 insertions(+), 12 deletions(-)

diff --git a/src/video/vid_tgui9440.c b/src/video/vid_tgui9440.c
index d1391f1d1..c6edabac5 100644
--- a/src/video/vid_tgui9440.c
+++ b/src/video/vid_tgui9440.c
@@ -1457,18 +1457,784 @@ enum {
     else                                               \
         dat = vram_l[(addr) & (tgui->vram_mask >> 2)];
 
-#define MIX()                                 \
-    do {                                      \
-        out = 0;                              \
-        for (c = 0; c < 32; c++) {            \
-            d = (dst_dat & (1 << c)) ? 1 : 0; \
-            if (src_dat & (1 << c))           \
-                d |= 2;                       \
-            if (pat_dat & (1 << c))           \
-                d |= 4;                       \
-            if (tgui->accel.rop & (1 << d))   \
-                out |= (1 << c);              \
-        }                                     \
+#define ROPMIX(R, D, P, S, out)                    \
+    {                                              \
+        switch (R) {                               \
+            case 0x00:                             \
+                out = 0;                           \
+                break;                             \
+            case 0x01:                             \
+                out = ~(D | (P | S));              \
+                break;                             \
+            case 0x02:                             \
+                out = D & ~(P | S);                \
+                break;                             \
+            case 0x03:                             \
+                out = ~(P | S);                    \
+                break;                             \
+            case 0x04:                             \
+                out = S & ~(D | P);                \
+                break;                             \
+            case 0x05:                             \
+                out = ~(D | P);                    \
+                break;                             \
+            case 0x06:                             \
+                out = ~(P | ~(D ^ S));             \
+                break;                             \
+            case 0x07:                             \
+                out = ~(P | (D & S));              \
+                break;                             \
+            case 0x08:                             \
+                out = S & (D & ~P);                \
+                break;                             \
+            case 0x09:                             \
+                out = ~(P | (D ^ S));              \
+                break;                             \
+            case 0x0a:                             \
+                out = D & ~P;                      \
+                break;                             \
+            case 0x0b:                             \
+                out = ~(P | (S & ~D));             \
+                break;                             \
+            case 0x0c:                             \
+                out = S & ~P;                      \
+                break;                             \
+            case 0x0d:                             \
+                out = ~(P | (D & ~S));             \
+                break;                             \
+            case 0x0e:                             \
+                out = ~(P | ~(D | S));             \
+                break;                             \
+            case 0x0f:                             \
+                out = ~P;                          \
+                break;                             \
+            case 0x10:                             \
+                out = P & ~(D | S);                \
+                break;                             \
+            case 0x11:                             \
+                out = ~(D | S);                    \
+                break;                             \
+            case 0x12:                             \
+                out = ~(S | ~(D ^ P));             \
+                break;                             \
+            case 0x13:                             \
+                out = ~(S | (D & P));              \
+                break;                             \
+            case 0x14:                             \
+                out = ~(D | ~(P ^ S));             \
+                break;                             \
+            case 0x15:                             \
+                out = ~(D | (P & S));              \
+                break;                             \
+            case 0x16:                             \
+                out = P ^ (S ^ (D & ~(P & S)));    \
+                break;                             \
+            case 0x17:                             \
+                out = ~(S ^ ((S ^ P) & (D ^ S)));  \
+                break;                             \
+            case 0x18:                             \
+                out = (S ^ P) & (P ^ D);           \
+                break;                             \
+            case 0x19:                             \
+                out = ~(S ^ (D & ~(P & S)));       \
+                break;                             \
+            case 0x1a:                             \
+                out = P ^ (D | (S & P));           \
+                break;                             \
+            case 0x1b:                             \
+                out = ~(S ^ (D & (P ^ S)));        \
+                break;                             \
+            case 0x1c:                             \
+                out = P ^ (S | (D & P));           \
+                break;                             \
+            case 0x1d:                             \
+                out = ~(D ^ (S & (P ^ D)));        \
+                break;                             \
+            case 0x1e:                             \
+                out = P ^ (D | S);                 \
+                break;                             \
+            case 0x1f:                             \
+                out = ~(P & (D | S));              \
+                break;                             \
+            case 0x20:                             \
+                out = D & (P & ~S);                \
+                break;                             \
+            case 0x21:                             \
+                out = ~(S | (D ^ P));              \
+                break;                             \
+            case 0x22:                             \
+                out = D & ~S;                      \
+                break;                             \
+            case 0x23:                             \
+                out = ~(S | (P & ~D));             \
+                break;                             \
+            case 0x24:                             \
+                out = (S ^ P) & (D ^ S);           \
+                break;                             \
+            case 0x25:                             \
+                out = ~(P ^ (D & ~(S & P)));       \
+                break;                             \
+            case 0x26:                             \
+                out = S ^ (D | (P & S));           \
+                break;                             \
+            case 0x27:                             \
+                out = S ^ (D | ~(P ^ S));          \
+                break;                             \
+            case 0x28:                             \
+                out = D & (P ^ S);                 \
+                break;                             \
+            case 0x29:                             \
+                out = ~(P ^ (S ^ (D | (P & S))));  \
+                break;                             \
+            case 0x2a:                             \
+                out = D & ~(P & S);                \
+                break;                             \
+            case 0x2b:                             \
+                out = ~(S ^ ((S ^ P) & (P ^ D)));  \
+                break;                             \
+            case 0x2c:                             \
+                out = S ^ (P & (D | S));           \
+                break;                             \
+            case 0x2d:                             \
+                out = P ^ (S | ~D);                \
+                break;                             \
+            case 0x2e:                             \
+                out = P ^ (S | (D ^ P));           \
+                break;                             \
+            case 0x2f:                             \
+                out = ~(P & (S | ~D));             \
+                break;                             \
+            case 0x30:                             \
+                out = P & ~S;                      \
+                break;                             \
+            case 0x31:                             \
+                out = ~(S | (D & ~P));             \
+                break;                             \
+            case 0x32:                             \
+                out = S ^ (D | (P | S));           \
+                break;                             \
+            case 0x33:                             \
+                out = ~S;                          \
+                break;                             \
+            case 0x34:                             \
+                out = S ^ (P | (D & S));           \
+                break;                             \
+            case 0x35:                             \
+                out = S ^ (P | ~(D ^ S));          \
+                break;                             \
+            case 0x36:                             \
+                out = S ^ (D | P);                 \
+                break;                             \
+            case 0x37:                             \
+                out = ~(S & (D | P));              \
+                break;                             \
+            case 0x38:                             \
+                out = P ^ (S & (D | P));           \
+                break;                             \
+            case 0x39:                             \
+                out = S ^ (P | ~D);                \
+                break;                             \
+            case 0x3a:                             \
+                out = S ^ (P | (D ^ S));           \
+                break;                             \
+            case 0x3b:                             \
+                out = ~(S & (P | ~D));             \
+                break;                             \
+            case 0x3c:                             \
+                out = P ^ S;                       \
+                break;                             \
+            case 0x3d:                             \
+                out = S ^ (P | ~(D | S));          \
+                break;                             \
+            case 0x3e:                             \
+                out = S ^ (P | (D & ~S));          \
+                break;                             \
+            case 0x3f:                             \
+                out = ~(P & S);                    \
+                break;                             \
+            case 0x40:                             \
+                out = P & (S & ~D);                \
+                break;                             \
+            case 0x41:                             \
+                out = ~(D | (P ^ S));              \
+                break;                             \
+            case 0x42:                             \
+                out = (S ^ D) & (P ^ D);           \
+                break;                             \
+            case 0x43:                             \
+                out = ~(S ^ (P & ~(D & S)));       \
+                break;                             \
+            case 0x44:                             \
+                out = S & ~D;                      \
+                break;                             \
+            case 0x45:                             \
+                out = ~(D | (P & ~S));             \
+                break;                             \
+            case 0x46:                             \
+                out = D ^ (S | (P & D));           \
+                break;                             \
+            case 0x47:                             \
+                out = ~(P ^ (S & (D ^ P)));        \
+                break;                             \
+            case 0x48:                             \
+                out = S & (D ^ P);                 \
+                break;                             \
+            case 0x49:                             \
+                out = ~(P ^ (D ^ (S | (P & D))));  \
+                break;                             \
+            case 0x4a:                             \
+                out = D ^ (P & (S | D));           \
+                break;                             \
+            case 0x4b:                             \
+                out = P ^ (D | ~S);                \
+                break;                             \
+            case 0x4c:                             \
+                out = S & ~(D & P);                \
+                break;                             \
+            case 0x4d:                             \
+                out = ~(S ^ ((S ^ P) | (D ^ S)));  \
+                break;                             \
+            case 0x4e:                             \
+                out = P ^ (D | (S ^ P));           \
+                break;                             \
+            case 0x4f:                             \
+                out = ~(P & (D | ~S));             \
+                break;                             \
+            case 0x50:                             \
+                out = P & ~D;                      \
+                break;                             \
+            case 0x51:                             \
+                out = ~(D | (S & ~P));             \
+                break;                             \
+            case 0x52:                             \
+                out = D ^ (P | (S & D));           \
+                break;                             \
+            case 0x53:                             \
+                out = ~(S ^ (P & (D ^ S)));        \
+                break;                             \
+            case 0x54:                             \
+                out = ~(D | ~(P | S));             \
+                break;                             \
+            case 0x55:                             \
+                out = ~D;                          \
+                break;                             \
+            case 0x56:                             \
+                out = D ^ (P | S);                 \
+                break;                             \
+            case 0x57:                             \
+                out = ~(D & (P | S));              \
+                break;                             \
+            case 0x58:                             \
+                out = P ^ (D & (S | P));           \
+                break;                             \
+            case 0x59:                             \
+                out = D ^ (P | ~S);                \
+                break;                             \
+            case 0x5a:                             \
+                out = D ^ P;                       \
+                break;                             \
+            case 0x5b:                             \
+                out = D ^ (P | ~(S | D));          \
+                break;                             \
+            case 0x5c:                             \
+                out = D ^ (P | (S ^ D));           \
+                break;                             \
+            case 0x5d:                             \
+                out = ~(D & (P | ~S));             \
+                break;                             \
+            case 0x5e:                             \
+                out = D ^ (P | (S & ~D));          \
+                break;                             \
+            case 0x5f:                             \
+                out = ~(D & P);                    \
+                break;                             \
+            case 0x60:                             \
+                out = P & (D ^ S);                 \
+                break;                             \
+            case 0x61:                             \
+                out = ~(D ^ (S ^ (P | (D & S))));  \
+                break;                             \
+            case 0x62:                             \
+                out = D ^ (S & (P | D));           \
+                break;                             \
+            case 0x63:                             \
+                out = S ^ (D | ~P);                \
+                break;                             \
+            case 0x64:                             \
+                out = S ^ (D & (P | S));           \
+                break;                             \
+            case 0x65:                             \
+                out = D ^ (S | ~P);                \
+                break;                             \
+            case 0x66:                             \
+                out = D ^ S;                       \
+                break;                             \
+            case 0x67:                             \
+                out = S ^ (D | ~(P | S));          \
+                break;                             \
+            case 0x68:                             \
+                out = ~(D ^ (S ^ (P | ~(D | S)))); \
+                break;                             \
+            case 0x69:                             \
+                out = ~(P ^ (D ^ S));              \
+                break;                             \
+            case 0x6a:                             \
+                out = D ^ (P & S);                 \
+                break;                             \
+            case 0x6b:                             \
+                out = ~(P ^ (S ^ (D & (P | S))));  \
+                break;                             \
+            case 0x6c:                             \
+                out = S ^ (D & P);                 \
+                break;                             \
+            case 0x6d:                             \
+                out = ~(P ^ (D ^ (S & (P | D))));  \
+                break;                             \
+            case 0x6e:                             \
+                out = S ^ (D & (P | ~S));          \
+                break;                             \
+            case 0x6f:                             \
+                out = ~(P & ~(D ^ S));             \
+                break;                             \
+            case 0x70:                             \
+                out = P & ~(D & S);                \
+                break;                             \
+            case 0x71:                             \
+                out = ~(S ^ ((S ^ D) & (P ^ D)));  \
+                break;                             \
+            case 0x72:                             \
+                out = S ^ (D | (P ^ S));           \
+                break;                             \
+            case 0x73:                             \
+                out = ~(S & (D | ~P));             \
+                break;                             \
+            case 0x74:                             \
+                out = D ^ (S | (P ^ D));           \
+                break;                             \
+            case 0x75:                             \
+                out = ~(D & (S | ~P));             \
+                break;                             \
+            case 0x76:                             \
+                out = S ^ (D | (P & ~S));          \
+                break;                             \
+            case 0x77:                             \
+                out = ~(D & S);                    \
+                break;                             \
+            case 0x78:                             \
+                out = P ^ (D & S);                 \
+                break;                             \
+            case 0x79:                             \
+                out = ~(D ^ (S ^ (P & (D | S))));  \
+                break;                             \
+            case 0x7a:                             \
+                out = D ^ (P & (S | ~D));          \
+                break;                             \
+            case 0x7b:                             \
+                out = ~(S & ~(D ^ P));             \
+                break;                             \
+            case 0x7c:                             \
+                out = S ^ (P & (D | ~S));          \
+                break;                             \
+            case 0x7d:                             \
+                out = ~(D & ~(P ^ S));             \
+                break;                             \
+            case 0x7e:                             \
+                out = (S ^ P) | (D ^ S);           \
+                break;                             \
+            case 0x7f:                             \
+                out = ~(D & (P & S));              \
+                break;                             \
+            case 0x80:                             \
+                out = D & (P & S);                 \
+                break;                             \
+            case 0x81:                             \
+                out = ~((S ^ P) | (D ^ S));        \
+                break;                             \
+            case 0x82:                             \
+                out = D & ~(P ^ S);                \
+                break;                             \
+            case 0x83:                             \
+                out = ~(S ^ (P & (D | ~S)));       \
+                break;                             \
+            case 0x84:                             \
+                out = S & ~(D ^ P);                \
+                break;                             \
+            case 0x85:                             \
+                out = ~(P ^ (D & (S | ~P)));       \
+                break;                             \
+            case 0x86:                             \
+                out = D ^ (S ^ (P & (D | S)));     \
+                break;                             \
+            case 0x87:                             \
+                out = ~(P ^ (D & S));              \
+                break;                             \
+            case 0x88:                             \
+                out = D & S;                       \
+                break;                             \
+            case 0x89:                             \
+                out = ~(S ^ (D | (P & ~S)));       \
+                break;                             \
+            case 0x8a:                             \
+                out = D & (S | ~P);                \
+                break;                             \
+            case 0x8b:                             \
+                out = ~(D ^ (S | (P ^ D)));        \
+                break;                             \
+            case 0x8c:                             \
+                out = S & (D | ~P);                \
+                break;                             \
+            case 0x8d:                             \
+                out = ~(S ^ (D | (P ^ S)));        \
+                break;                             \
+            case 0x8e:                             \
+                out = S ^ ((S ^ D) & (P ^ D));     \
+                break;                             \
+            case 0x8f:                             \
+                out = ~(P & ~(D & S));             \
+                break;                             \
+            case 0x90:                             \
+                out = P & ~(D ^ S);                \
+                break;                             \
+            case 0x91:                             \
+                out = ~(S ^ (D & (P | ~S)));       \
+                break;                             \
+            case 0x92:                             \
+                out = D ^ (P ^ (S & (D | P)));     \
+                break;                             \
+            case 0x93:                             \
+                out = ~(S ^ (P & D));              \
+                break;                             \
+            case 0x94:                             \
+                out = P ^ (S ^ (D & (P | S)));     \
+                break;                             \
+            case 0x95:                             \
+                out = ~(D ^ (P & S));              \
+                break;                             \
+            case 0x96:                             \
+                out = D ^ (P ^ S);                 \
+                break;                             \
+            case 0x97:                             \
+                out = P ^ (S ^ (D | ~(P | S)));    \
+                break;                             \
+            case 0x98:                             \
+                out = ~(S ^ (D | ~(P | S)));       \
+                break;                             \
+            case 0x99:                             \
+                out = ~(D ^ S);                    \
+                break;                             \
+            case 0x9a:                             \
+                out = D ^ (P & ~S);                \
+                break;                             \
+            case 0x9b:                             \
+                out = ~(S ^ (D & (P | S)));        \
+                break;                             \
+            case 0x9c:                             \
+                out = S ^ (P & ~D);                \
+                break;                             \
+            case 0x9d:                             \
+                out = ~(D ^ (S & (P | D)));        \
+                break;                             \
+            case 0x9e:                             \
+                out = D ^ (S ^ (P | (D & S)));     \
+                break;                             \
+            case 0x9f:                             \
+                out = ~(P & (D ^ S));              \
+                break;                             \
+            case 0xa0:                             \
+                out = D & P;                       \
+                break;                             \
+            case 0xa1:                             \
+                out = ~(P ^ (D | (S & ~P)));       \
+                break;                             \
+            case 0xa2:                             \
+                out = D & (P | ~S);                \
+                break;                             \
+            case 0xa3:                             \
+                out = ~(D ^ (P | (S ^ D)));        \
+                break;                             \
+            case 0xa4:                             \
+                out = ~(P ^ (D | ~(S | P)));       \
+                break;                             \
+            case 0xa5:                             \
+                out = ~(P ^ D);                    \
+                break;                             \
+            case 0xa6:                             \
+                out = D ^ (S & ~P);                \
+                break;                             \
+            case 0xa7:                             \
+                out = ~(P ^ (D & (S | P)));        \
+                break;                             \
+            case 0xa8:                             \
+                out = D & (P | S);                 \
+                break;                             \
+            case 0xa9:                             \
+                out = ~(D ^ (P | S));              \
+                break;                             \
+            case 0xaa:                             \
+                out = D;                           \
+                break;                             \
+            case 0xab:                             \
+                out = D | ~(P | S);                \
+                break;                             \
+            case 0xac:                             \
+                out = S ^ (P & (D ^ S));           \
+                break;                             \
+            case 0xad:                             \
+                out = ~(D ^ (P | (S & D)));        \
+                break;                             \
+            case 0xae:                             \
+                out = D | (S & ~P);                \
+                break;                             \
+            case 0xaf:                             \
+                out = D | ~P;                      \
+                break;                             \
+            case 0xb0:                             \
+                out = P & (D | ~S);                \
+                break;                             \
+            case 0xb1:                             \
+                out = ~(P ^ (D | (S ^ P)));        \
+                break;                             \
+            case 0xb2:                             \
+                out = S ^ ((S ^ P) | (D ^ S));     \
+                break;                             \
+            case 0xb3:                             \
+                out = ~(S & ~(D & P));             \
+                break;                             \
+            case 0xb4:                             \
+                out = P ^ (S & ~D);                \
+                break;                             \
+            case 0xb5:                             \
+                out = ~(D ^ (P & (S | D)));        \
+                break;                             \
+            case 0xb6:                             \
+                out = D ^ (P ^ (S | (D & P)));     \
+                break;                             \
+            case 0xb7:                             \
+                out = ~(S & (D ^ P));              \
+                break;                             \
+            case 0xb8:                             \
+                out = P ^ (S & (D ^ P));           \
+                break;                             \
+            case 0xb9:                             \
+                out = ~(D ^ (S | (P & D)));        \
+                break;                             \
+            case 0xba:                             \
+                out = D | (P & ~S);                \
+                break;                             \
+            case 0xbb:                             \
+                out = D | ~S;                      \
+                break;                             \
+            case 0xbc:                             \
+                out = S ^ (P & ~(D & S));          \
+                break;                             \
+            case 0xbd:                             \
+                out = ~((S ^ D) & (P ^ D));        \
+                break;                             \
+            case 0xbe:                             \
+                out = D | (P ^ S);                 \
+                break;                             \
+            case 0xbf:                             \
+                out = D | ~(P & S);                \
+                break;                             \
+            case 0xc0:                             \
+                out = P & S;                       \
+                break;                             \
+            case 0xc1:                             \
+                out = ~(S ^ (P | (D & ~S)));       \
+                break;                             \
+            case 0xc2:                             \
+                out = ~(S ^ (P | ~(D | S)));       \
+                break;                             \
+            case 0xc3:                             \
+                out = ~(P ^ S);                    \
+                break;                             \
+            case 0xc4:                             \
+                out = S & (P | ~D);                \
+                break;                             \
+            case 0xc5:                             \
+                out = ~(S ^ (P | (D ^ S)));        \
+                break;                             \
+            case 0xc6:                             \
+                out = S ^ (D & ~P);                \
+                break;                             \
+            case 0xc7:                             \
+                out = ~(P ^ (S & (D | P)));        \
+                break;                             \
+            case 0xc8:                             \
+                out = S & (D | P);                 \
+                break;                             \
+            case 0xc9:                             \
+                out = ~(S ^ (P | D));              \
+                break;                             \
+            case 0xca:                             \
+                out = D ^ (P & (S ^ D));           \
+                break;                             \
+            case 0xcb:                             \
+                out = ~(S ^ (P | (D & S)));        \
+                break;                             \
+            case 0xcc:                             \
+                out = S;                           \
+                break;                             \
+            case 0xcd:                             \
+                out = S | ~(D | P);                \
+                break;                             \
+            case 0xce:                             \
+                out = S | (D & ~P);                \
+                break;                             \
+            case 0xcf:                             \
+                out = S | ~P;                      \
+                break;                             \
+            case 0xd0:                             \
+                out = P & (S | ~D);                \
+                break;                             \
+            case 0xd1:                             \
+                out = ~(P ^ (S | (D ^ P)));        \
+                break;                             \
+            case 0xd2:                             \
+                out = P ^ (D & ~S);                \
+                break;                             \
+            case 0xd3:                             \
+                out = ~(S ^ (P & (D | S)));        \
+                break;                             \
+            case 0xd4:                             \
+                out = S ^ ((S ^ P) & (P ^ D));     \
+                break;                             \
+            case 0xd5:                             \
+                out = ~(D & ~(P & S));             \
+                break;                             \
+            case 0xd6:                             \
+                out = P ^ (S ^ (D | (P & S)));     \
+                break;                             \
+            case 0xd7:                             \
+                out = ~(D & (P ^ S));              \
+                break;                             \
+            case 0xd8:                             \
+                out = P ^ (D & (S ^ P));           \
+                break;                             \
+            case 0xd9:                             \
+                out = ~(S ^ (D | (P & S)));        \
+                break;                             \
+            case 0xda:                             \
+                out = D ^ (P & ~(S & D));          \
+                break;                             \
+            case 0xdb:                             \
+                out = ~((S ^ P) & (D ^ S));        \
+                break;                             \
+            case 0xdc:                             \
+                out = S | (P & ~D);                \
+                break;                             \
+            case 0xdd:                             \
+                out = S | ~D;                      \
+                break;                             \
+            case 0xde:                             \
+                out = S | (D ^ P);                 \
+                break;                             \
+            case 0xdf:                             \
+                out = S | ~(D & P);                \
+                break;                             \
+            case 0xe0:                             \
+                out = P & (D | S);                 \
+                break;                             \
+            case 0xe1:                             \
+                out = ~(P ^ (D | S));              \
+                break;                             \
+            case 0xe2:                             \
+                out = D ^ (S & (P ^ D));           \
+                break;                             \
+            case 0xe3:                             \
+                out = ~(P ^ (S | (D & P)));        \
+                break;                             \
+            case 0xe4:                             \
+                out = S ^ (D & (P ^ S));           \
+                break;                             \
+            case 0xe5:                             \
+                out = ~(P ^ (D | (S & P)));        \
+                break;                             \
+            case 0xe6:                             \
+                out = S ^ (D & ~(P & S));          \
+                break;                             \
+            case 0xe7:                             \
+                out = ~((S ^ P) & (P ^ D));        \
+                break;                             \
+            case 0xe8:                             \
+                out = S ^ ((S ^ P) & (D ^ S));     \
+                break;                             \
+            case 0xe9:                             \
+                out = ~(D ^ (S ^ (P & ~(D & S)))); \
+                break;                             \
+            case 0xea:                             \
+                out = D | (P & S);                 \
+                break;                             \
+            case 0xeb:                             \
+                out = D | ~(P ^ S);                \
+                break;                             \
+            case 0xec:                             \
+                out = S | (D & P);                 \
+                break;                             \
+            case 0xed:                             \
+                out = S | ~(D ^ P);                \
+                break;                             \
+            case 0xee:                             \
+                out = D | S;                       \
+                break;                             \
+            case 0xef:                             \
+                out = S | (D | ~P);                \
+                break;                             \
+            case 0xf0:                             \
+                out = P;                           \
+                break;                             \
+            case 0xf1:                             \
+                out = P | ~(D | S);                \
+                break;                             \
+            case 0xf2:                             \
+                out = P | (D & ~S);                \
+                break;                             \
+            case 0xf3:                             \
+                out = P | ~S;                      \
+                break;                             \
+            case 0xf4:                             \
+                out = P | (S & ~D);                \
+                break;                             \
+            case 0xf5:                             \
+                out = P | ~D;                      \
+                break;                             \
+            case 0xf6:                             \
+                out = P | (D ^ S);                 \
+                break;                             \
+            case 0xf7:                             \
+                out = P | ~(D & S);                \
+                break;                             \
+            case 0xf8:                             \
+                out = P | (D & S);                 \
+                break;                             \
+            case 0xf9:                             \
+                out = P | ~(D ^ S);                \
+                break;                             \
+            case 0xfa:                             \
+                out = D | P;                       \
+                break;                             \
+            case 0xfb:                             \
+                out = D | (P | ~S);                \
+                break;                             \
+            case 0xfc:                             \
+                out = P | S;                       \
+                break;                             \
+            case 0xfd:                             \
+                out = P | (S | ~D);                \
+                break;                             \
+            case 0xfe:                             \
+                out = D | (P | S);                 \
+                break;                             \
+            case 0xff:                             \
+                out = ~0;                          \
+                break;                             \
+        }                                          \
+    }
+
+#define MIX()                                                    \
+    do {                                                         \
+        out = 0;                                                 \
+        ROPMIX(tgui->accel.rop, dst_dat, pat_dat, src_dat, out); \
     } while (0)
 
 #define WRITE(addr, dat)                                                                                  \

From a9653e5a99393924f09ab6f6c72fb8afc47310bf Mon Sep 17 00:00:00 2001
From: OBattler <oubattler@gmail.com>
Date: Mon, 3 Mar 2025 21:21:02 +0100
Subject: [PATCH 30/58] TGUI9440: Fix two newly introduced warnings.

---
 src/video/vid_tgui9440.c | 2 --
 1 file changed, 2 deletions(-)

diff --git a/src/video/vid_tgui9440.c b/src/video/vid_tgui9440.c
index c6edabac5..af203f327 100644
--- a/src/video/vid_tgui9440.c
+++ b/src/video/vid_tgui9440.c
@@ -2256,8 +2256,6 @@ tgui_accel_command(int count, uint32_t cpu_dat, tgui_t *tgui)
     const uint32_t *pattern_data;
     int             x;
     int             y;
-    int             c;
-    int             d;
     uint32_t        out;
     uint32_t        src_dat   = 0;
     uint32_t        dst_dat;

From 6f53e69efe666ac6824b03f098edc11cf8db97ff Mon Sep 17 00:00:00 2001
From: Akamaki <97360908+akmed772@users.noreply.github.com>
Date: Tue, 4 Mar 2025 07:47:15 +0900
Subject: [PATCH 31/58] added bit mask reset and alignment

for compatibility with OS/2 J2.1 (beta), Windows 3.1 and Multitool Chart K3.1 in J-DOS.
---
 src/video/vid_ps55da2.c | 125 ++++++++++++++++++++++++----------------
 1 file changed, 74 insertions(+), 51 deletions(-)

diff --git a/src/video/vid_ps55da2.c b/src/video/vid_ps55da2.c
index 46ef4caf6..1b9ec473e 100644
--- a/src/video/vid_ps55da2.c
+++ b/src/video/vid_ps55da2.c
@@ -327,7 +327,7 @@ typedef struct da2_t {
     int      fast;
     uint8_t  colourcompare, colournocare;
     int      readmode, writemode, readplane;
-    uint8_t  writemask;
+    uint8_t  planemask;
     uint32_t charseta, charsetb;
 
     uint8_t  egapal[16];
@@ -441,6 +441,7 @@ void        da2_bitblt_exec(void *p);
 void        da2_updatevidselector(da2_t *da2);
 void        da2_reset_ioctl(da2_t *da2);
 static void da2_reset(void *priv);
+uint16_t    rightRotate(uint16_t data, uint8_t count);
 
 typedef union {
     uint32_t d;
@@ -687,6 +688,7 @@ da2_bitblt_load(da2_t *da2)
                 value32 |= da2->bitblt.payload[i + 2];
                 da2_bltlog("[%02x] %02x: %04x (%d)\n", da2->bitblt.payload[i], da2->bitblt.payload[i + 1], value32, value32);
                 da2->bitblt.reg[da2->bitblt.payload[i + 1]] = value32;
+                // da2->gdcreg[da2->bitblt.payload[i + 1]] = value32;
                 i += 3;
                 break;
             case 0x91:
@@ -699,6 +701,7 @@ da2_bitblt_load(da2_t *da2)
                 value32 |= da2->bitblt.payload[i + 2];
                 da2_bltlog("[%02x] %02x: %08x (%d)\n", da2->bitblt.payload[i], da2->bitblt.payload[i + 1], value32, value32);
                 da2->bitblt.reg[da2->bitblt.payload[i + 1]] = value32;
+                // da2->gdcreg[da2->bitblt.payload[i + 1]] = value32;
                 i += 5;
                 break;
             case 0x99:
@@ -716,6 +719,7 @@ da2_bitblt_load(da2_t *da2)
                 da2_bltlog("[%02x] %02x: %02x %02x %02x %02x %02x %02x\n", da2->bitblt.payload[i], da2->bitblt.payload[i + 1], da2->bitblt.payload[i + 2], da2->bitblt.payload[i + 3],
                         da2->bitblt.payload[i + 4], da2->bitblt.payload[i + 5], da2->bitblt.payload[i + 6], da2->bitblt.payload[i + 7]);
                 da2->bitblt.reg[da2->bitblt.payload[i + 1]] = value64;
+                // da2->gdcreg[da2->bitblt.payload[i + 1]] = value64;
                 i += 7;
                 break;
             case 0x00:
@@ -1321,26 +1325,27 @@ da2_out(uint16_t addr, uint16_t val, void *p)
             da2->attrc[da2->attraddr & 0x3f] = val;
             break;
         case LG_INDEX:
+            da2_log("DA2 Out addr %03X val %02X\n", addr, val);
             da2->gdcaddr = val;
             break;
         case LG_DATA:
-            // if(da2->gdcaddr != 8 && da2->gdcaddr != 9) da2_log("DA2 Out addr %03X idx %02X val %02X\n", addr, da2->gdcaddr, val);
+        // if(da2->gdcaddr != 8 && da2->gdcaddr != 9) da2_log("DA2 GCOut idx %X val %02X %04X:%04X esdi %04X:%04X\n", da2->gdcaddr, val, cs >> 4, cpu_state.pc, ES, DI);
             da2_log("DA2 Out addr %03X idx %02X val %02X\n", addr, da2->gdcaddr, val);
-            // if(da2->gdcaddr != 8 && da2->gdcaddr != 9) da2_log("DA2 GCOut idx %X val %02X %04X:%04X esdi %04X:%04X\n", da2->gdcaddr, val, cs >> 4, cpu_state.pc, ES, DI);
+            da2->gdcreg[da2->gdcaddr & 0x0f] = val & 0xff;
             switch (da2->gdcaddr & 0x1f) {
                 case LG_READ_MAP_SELECT:
                     da2->readplane = val & 0x7;
                     break;
                 case LG_MODE:
                     da2->writemode = val & 3;
-                     /* reset bit mask and plane mask (? idk when) */
-                    da2->gdcreg[LG_BIT_MASK_LOW] = 0xff;
-                    da2->gdcreg[LG_BIT_MASK_HIGH] = 0xff;
-                    da2->writemask = 0xff;
-                    da2->gdcreg[LG_MAP_MASKJ] = 0xff;
+                     /* Resettting masks here gliches the screen in IBM Multitool Chart K3.1 */
+                    // da2->gdcreg[LG_BIT_MASK_LOW] = 0xff;
+                    // da2->gdcreg[LG_BIT_MASK_HIGH] = 0xff;
+                    // da2->planemask = 0xff;
+                    // da2->gdcreg[LG_MAP_MASKJ] = 0xff;
                     break;
                 case LG_MAP_MASKJ:
-                    da2->writemask = val & 0xff;
+                    da2->planemask = val & 0xff;
                     break;
                 case LG_COMMAND:
                     break;
@@ -1348,7 +1353,6 @@ da2_out(uint16_t addr, uint16_t val, void *p)
                     da2_log("!!!DA2 GC Out addr %03X idx 10 val %02X\n", addr, val);
                     return;
             }
-            da2->gdcreg[da2->gdcaddr & 0x0f] = val & 0xff;
             break;
         // case 0x3ed: /* used by Windows 3.1 display driver */
         //     da2->gdcreg[5] = val & 0xff;
@@ -1424,7 +1428,7 @@ da2_in(uint16_t addr, void *p)
                     temp |= 0x80;
                 }
                 temp &= 0xf6;           /* clear busy bit */
-                // if (da2->bitblt.indata) /* for OS/2 J1.3 */
+                // if (da2->bitblt.indata) /* for OS/2 J1.3 command prompt scrolling  */
                 //     da2_bitblt_dopayload(da2);
                 if (da2->bitblt.exec != DA2_BLT_CIDLE) {
                     // da2_log("exec:%x\n", da2->bitblt.exec);
@@ -1435,7 +1439,7 @@ da2_in(uint16_t addr, void *p)
                     //     timer_advance_u64(&da2->bitblt.timer, da2->bitblt.timerspeed);
                     // }
                 }
-                if (da2->bitblt.indata) temp |= 0x09;
+                if (da2->bitblt.indata) temp |= 0x08;
 #ifdef ENABLE_DA2_DEBUGMONWAIT
                 da2_log("DA2 In %04X(%02X) %04X %04X:%04X\n", addr, da2->ioctladdr, temp, cs >> 4, cpu_state.pc);
 #endif
@@ -1529,7 +1533,7 @@ da2_outb(uint16_t addr, uint8_t val, void *p)
 void
 da2_outw(uint16_t addr, uint16_t val, void *p)
 {
-    da2_log("DA2 Outw addr %03X val %04X %04X:%04X\n", addr, val, cs >> 4, cpu_state.pc);
+    da2_log("DA2 Outw addr %03X val %04X\n", addr, val);
     da2_t *da2      = (da2_t *) p;
     da2->inflipflop = 0;
     switch (addr) {
@@ -1550,7 +1554,15 @@ da2_outw(uint16_t addr, uint16_t val, void *p)
             break;
         case 0x3EC:
             // da2_log("DA2 Outw addr %03X val %04X %04X:%04X\n", addr, val, cs >> 4, cpu_state.pc);
+            da2_log("  ");
+            // val = rightRotate(val, 8);
+            // da2_out(LG_DATA, val, da2);
             da2_out(LG_DATA, val >> 8, da2);
+            /* reset masks */
+            da2->gdcreg[LG_BIT_MASK_LOW] = 0xff;
+            da2->gdcreg[LG_BIT_MASK_HIGH] = 0xff;
+            da2->planemask = 0xff;
+            da2->gdcreg[LG_MAP_MASKJ] = 0xff;
             break;
         case 0x3ED:
             da2->gdcaddr = LG_MODE;
@@ -2340,11 +2352,12 @@ da2_mca_reset(void *p)
     da2_mca_write(0x102, 0, da2);
 }
 
+/* ROP gdcinput and gdcsrc, and write the result with bitmask at addr (byte) */
 static void
 da2_gdcropB(uint32_t addr,uint8_t bitmask, da2_t *da2)
 {
     for (int i = 0; i < 8; i++) {
-        if (da2->writemask & (1 << i)) {
+        if (da2->planemask & (1 << i)) {
             // da2_log("da2_gdcropB o%x a%x d%x p%d m%x\n", da2->gdcreg[LG_COMMAND] & 0x03, addr, da2->gdcinput[i], i, bitmask);
             switch (da2->gdcreg[LG_COMMAND] & 0x03) {
                 case 0: /*Set*/
@@ -2368,13 +2381,15 @@ da2_gdcropB(uint32_t addr,uint8_t bitmask, da2_t *da2)
         }
     }
 }
+/* ROP gdcinput and gdcsrc, and write the result with bitmask at addr (word) */
 static void
 da2_gdcropW(uint32_t addr, uint16_t bitmask, da2_t *da2)
 {
+    if((addr & 8) && !(da2->gdcreg[LG_COMMAND] & 0x08)) bitmask = rightRotate(bitmask, 8);
     uint8_t bitmask_l = bitmask & 0xff;
     uint8_t bitmask_h = bitmask >> 8;
     for (int i = 0; i < 8; i++) {
-        if (da2->writemask & (1 << i)) {
+        if (da2->planemask & (1 << i)) {
             // da2_log("da2_gdcropW m%x a%x d%x i%d ml%x mh%x\n", da2->gdcreg[LG_COMMAND] & 0x03, addr, da2->gdcinput[i], i, da2->gdcreg[LG_BIT_MASK_LOW], da2->gdcreg[LG_BIT_MASK_HIGH]);
             switch (da2->gdcreg[LG_COMMAND] & 0x03) {
                 case 0: /*Set*/
@@ -2415,7 +2430,6 @@ da2_mmio_read(uint32_t addr, void *p)
 {
     da2_t *da2 = (da2_t *) p;
     addr &= DA2_MASK_MMIO;
-
     if (da2->ioctl[LS_MMIO] & 0x10) {
         if (da2->fctl[LF_MMIO_SEL] == 0x80)
             /* linear access */
@@ -2455,7 +2469,7 @@ da2_mmio_read(uint32_t addr, void *p)
         cycles -= video_timing_read_b;
         for (int i = 0; i < 8; i++)
             da2->gdcla[i] = da2->vram[(addr << 3) | i]; /* read in byte */
-        // da2_log("da2_Rb: %05x=%02x\n", addr, da2->gdcla[da2->readplane]);
+        da2_log("da2_Rb: %05x=%02x\n", addr, da2->gdcla[da2->readplane]);
         if (da2->gdcreg[LG_MODE] & 0x08) { /* compare data across planes if the read mode bit (3EB 05, bit 3) is 1 */
             uint8_t ret = 0;
             for (int i = 0; i < 8; i++) {
@@ -2474,7 +2488,9 @@ static uint16_t
 da2_mmio_readw(uint32_t addr, void *p)
 {
     da2_t *da2 = (da2_t *) p;
-    //da2_log("da2_readW: %x %x %x %x %x\n", da2->ioctl[LS_MMIO], da2->fctl[LF_MMIO_SEL], da2->fctl[LF_MMIO_MODE], da2->fctl[LF_MMIO_ADDR], addr);
+    // da2_log("da2_readW: %x %x %x %x %x\n", da2->ioctl[LS_MMIO], da2->fctl[LF_MMIO_SEL], da2->fctl[LF_MMIO_MODE], da2->fctl[LF_MMIO_ADDR], addr);
+    da2_log("da2_readW: %x %x %x %x %x CS:PC=%4x:%4x\n", da2->ioctl[LS_MMIO], da2->fctl[LF_MMIO_SEL], da2->fctl[LF_MMIO_MODE], da2->fctl[LF_MMIO_ADDR], addr, CS, cpu_state.pc);
+
     if (da2->ioctl[LS_MMIO] & 0x10) {
         return (uint16_t) da2_mmio_read(addr, da2) | (uint16_t) (da2_mmio_read(addr + 1, da2) << 8);
     } else if (!(da2->ioctl[LS_MODE] & 1)) {/* 16 color or 256 color mode */
@@ -2566,11 +2582,14 @@ da2_mmio_write(uint32_t addr, uint8_t val, void *p)
         }
     } else if (!(da2->ioctl[LS_MODE] & 1)) { /* 16 color or 256 color mode */
         uint8_t bitmask;
-        /* align bitmask to even address */
-        if (addr & 1) bitmask = da2->gdcreg[LG_BIT_MASK_HIGH];
+        /* align bitmask with even address for OS/2 J2.0 (need to verify the condition with Win 3.x) */
+        if ((addr & 1)  && !(da2->gdcreg[LG_COMMAND] & 0x08)) bitmask = da2->gdcreg[LG_BIT_MASK_HIGH];
         else  bitmask = da2->gdcreg[LG_BIT_MASK_LOW];
+        
 #ifdef ENABLE_DA2_DEBUGVRAM
-        da2_log("da2_wB %x %02x\n", addr, val);
+        // da2_log("da2_wB %x %02x\n", addr, val);
+        da2_log("da2_wB %x %d %d %02x\n", addr,
+            addr % (da2->rowoffset * 2) * 8, addr / (da2->rowoffset * 2), val);
         // if (!(da2->gdcreg[LG_COMMAND] & 0x08))
         //{
         if (((int) addr - (int) da2->mmdbg_vidaddr) > 2 || (((int) da2->mmdbg_vidaddr - (int) addr) > 2) || da2->mmdbg_vidaddr == addr) {
@@ -2581,7 +2600,7 @@ da2_mmio_write(uint32_t addr, uint8_t val, void *p)
         for (int i = 0; i < 8; i++) {
             int pixeldata = 0;
             if (val & (1 << (7 - i)))
-                pixeldata = (da2->writemask & 0xf);
+                pixeldata = (da2->planemask & 0xf);
             fprintf(da2->mmdbg_fp, "%X", pixeldata);
         }
         da2->mmdbg_vidaddr = addr;
@@ -2614,7 +2633,7 @@ da2_mmio_write(uint32_t addr, uint8_t val, void *p)
         switch (da2->writemode) {
             case 2: /* equiv to vga write mode 1 */
                 for (int i = 0; i < 8; i++)
-                    if (da2->writemask & (1 << i))
+                    if (da2->planemask & (1 << i))
                         DA2_vram_w(addr | i, da2->gdcsrc[i], da2);
                 break;
             case 0:/* equiv to vga write mode 0 */
@@ -2622,7 +2641,7 @@ da2_mmio_write(uint32_t addr, uint8_t val, void *p)
                     val = svga_rotate[da2->gdcreg[LG_DATA_ROTATION] & 7][val];
                 if (bitmask == 0xff && !(da2->gdcreg[LG_COMMAND] & 0x03) && (!da2->gdcreg[LG_ENABLE_SRJ])) {
                     for (int i = 0; i < 8; i++)
-                        if (da2->writemask & (1 << i))
+                        if (da2->planemask & (1 << i))
                             DA2_vram_w(addr | i, val, da2);
                 } else {
                     for (int i = 0; i < 8; i++)
@@ -2631,8 +2650,8 @@ da2_mmio_write(uint32_t addr, uint8_t val, void *p)
                         else
                             da2->gdcinput[i] = val;
 
-                    for (int i = 0; i < 8; i++)
-                        da2->debug_vramold[i] = da2->vram[addr | i]; /* use latch */
+                    // for (int i = 0; i < 8; i++)
+                    //     da2->debug_vramold[i] = da2->vram[addr | i]; /* use latch */
                     da2_gdcropB(addr, bitmask, da2);
                     // for (int i = 0; i < 8; i++)
                     //     da2_log("\tsrc %02x in %02x bitm %02x mod %x rop %x: %02x -> %02x\n", da2->gdcsrc[i], da2->gdcinput[i], bitmask, da2->gdcreg[5],da2->gdcreg[LG_COMMAND], da2->debug_vramold[i],  da2->vram[addr | i]);//use latch
@@ -2640,17 +2659,17 @@ da2_mmio_write(uint32_t addr, uint8_t val, void *p)
                 }
                 break;
             case 1:/* equiv to vga write mode 2 */
-                if (!(da2->gdcreg[LG_COMMAND] & 0x03) && (!da2->gdcreg[LG_ENABLE_SRJ])) {
-                    for (int i = 0; i < 8; i++)
-                        if (da2->writemask & (1 << i))
-                            DA2_vram_w(addr | i, (((val & (1 << i)) ? 0xff : 0) & bitmask) | (da2->gdcsrc[i] & ~bitmask), da2);
-                    //fprintf(da2->mmdbg_fp, "m1-1");
-                } else {
+                // if (!(da2->gdcreg[LG_COMMAND] & 0x03) && (!da2->gdcreg[LG_ENABLE_SRJ])) {
+                //     for (int i = 0; i < 8; i++)
+                //         if (da2->planemask & (1 << i))
+                //             DA2_vram_w(addr | i, (((val & (1 << i)) ? 0xff : 0) & bitmask) | (da2->gdcsrc[i] & ~bitmask), da2);
+                //     //fprintf(da2->mmdbg_fp, "m1-1");
+                // } else {
                     for (int i = 0; i < 8; i++)
                         da2->gdcinput[i] = ((val & (1 << i)) ? 0xff : 0);
                     da2_gdcropB(addr, bitmask, da2);
                     //fprintf(da2->mmdbg_fp, "m1-2");
-                }
+                // }
                 break;
             case 3:/* equiv to vga write mode 3 */
                 if (da2->gdcreg[LG_DATA_ROTATION] & 7)
@@ -2658,7 +2677,7 @@ da2_mmio_write(uint32_t addr, uint8_t val, void *p)
                 bitmask &= val;
 
                 for (int i = 0; i < 8; i++)
-                    da2->gdcinput[i] = (da2->gdcreg[LG_SET_RESETJ] & (1 << i)) ? 0xff : 0;
+                        da2->gdcinput[i] = (da2->gdcreg[LG_SET_RESETJ] & (1 << i)) ? 0xff : 0;
                 da2_gdcropB(addr, bitmask, da2);
                 break;
         }
@@ -2678,22 +2697,25 @@ static void
 da2_mmio_gc_writeW(uint32_t addr, uint16_t val, void *p)
 {
     da2_t   *da2     = (da2_t *) p;
-    uint16_t bitmask = da2->gdcreg[LG_BIT_MASK_HIGH];
+    uint16_t bitmask;
     addr &= DA2_MASK_MMIO;
+    bitmask = da2->gdcreg[LG_BIT_MASK_HIGH];
     bitmask <<= 8;
     bitmask |= (uint16_t) da2->gdcreg[LG_BIT_MASK_LOW];
 #ifdef ENABLE_DA2_DEBUGVRAM
+    da2_log("da2_wW %x %d %d %02x\n", addr,
+    addr % (da2->rowoffset * 2) * 8, addr / (da2->rowoffset * 2), val);
     // if (!(da2->gdcreg[LG_COMMAND] & 0x08))
     //{
     if (((int) addr - (int) da2->mmdbg_vidaddr) > 2 || (((int) da2->mmdbg_vidaddr - (int) addr) > 2) || da2->mmdbg_vidaddr == addr) {
-        fprintf(da2->mmdbg_fp, "\nW %x ", addr);
+        fprintf(da2->mmdbg_fp, "\nW %x %x ", addr, val);
         for (int i = 0; i <= 0xb; i++)
             fprintf(da2->mmdbg_fp, "%02x ", da2->gdcreg[i]);
     }
     for (int i = 0; i < 16; i++) {
         int pixeldata = 0;
         if (val & (1 << (15 - i)))
-            pixeldata = (da2->writemask & 0xf);
+            pixeldata = (da2->planemask & 0xf);
         fprintf(da2->mmdbg_fp, "%X", pixeldata);
     }
     da2->mmdbg_vidaddr = addr;
@@ -2727,7 +2749,7 @@ da2_mmio_gc_writeW(uint32_t addr, uint16_t val, void *p)
     switch (da2->writemode) {
         case 2:
             for (int i = 0; i < 8; i++)
-                if (da2->writemask & (1 << i)) {
+                if (da2->planemask & (1 << i)) {
                     DA2_vram_w(addr | i, da2->gdcsrc[i] & 0xff, da2);
                     DA2_vram_w((addr + 8) | i, da2->gdcsrc[i] >> 8, da2);
                 }
@@ -2737,7 +2759,7 @@ da2_mmio_gc_writeW(uint32_t addr, uint16_t val, void *p)
                 val = rightRotate(val, da2->gdcreg[LG_DATA_ROTATION] & 15); // val = svga_rotate[da2->gdcreg[LG_DATA_ROTATION] & 7][val]; TODO this wont work
             if (bitmask == 0xffff && !(da2->gdcreg[LG_COMMAND] & 0x03) && (!da2->gdcreg[LG_ENABLE_SRJ])) {
                 for (int i = 0; i < 8; i++)
-                    if (da2->writemask & (1 << i)) {
+                    if (da2->planemask & (1 << i)) {
                         DA2_vram_w(addr | i, val & 0xff, da2);
                         DA2_vram_w((addr + 8) | i, val >> 8, da2);
                     }
@@ -2752,18 +2774,18 @@ da2_mmio_gc_writeW(uint32_t addr, uint16_t val, void *p)
             }
             break;
         case 1:
-            if (!(da2->gdcreg[LG_COMMAND] & 0x03) && (!da2->gdcreg[LG_ENABLE_SRJ])) {
-                for (int i = 0; i < 8; i++)
-                    if (da2->writemask & (1 << i)) {
-                        uint16_t wdata = (((val & (1 << i)) ? 0xffff : 0) & bitmask) | (da2->gdcsrc[i] & ~bitmask);
-                        DA2_vram_w(addr | i, wdata & 0xff, da2);
-                        DA2_vram_w((addr + 8) | i, wdata >> 8, da2);
-                    }
-            } else {
+            // if (!(da2->gdcreg[LG_COMMAND] & 0x03) && (!da2->gdcreg[LG_ENABLE_SRJ])) {
+            //     for (int i = 0; i < 8; i++)
+            //         if (da2->planemask & (1 << i)) {
+            //             uint16_t wdata = (((val & (1 << i)) ? 0xffff : 0) & bitmask) | (da2->gdcsrc[i] & ~bitmask);
+            //             DA2_vram_w(addr | i, wdata & 0xff, da2);
+            //             DA2_vram_w((addr + 8) | i, wdata >> 8, da2);
+            //         }
+            // } else {
                 for (int i = 0; i < 8; i++)
                     da2->gdcinput[i] = ((val & (1 << i)) ? 0xffff : 0);
                 da2_gdcropW(addr, bitmask, da2);
-            }
+            // }
             break;
         case 3:
             if (da2->gdcreg[LG_DATA_ROTATION] & 15)
@@ -2771,7 +2793,7 @@ da2_mmio_gc_writeW(uint32_t addr, uint16_t val, void *p)
             bitmask &= val;
 
             for (int i = 0; i < 8; i++)
-                da2->gdcinput[i] = (da2->gdcreg[LG_SET_RESETJ] & (1 << i)) ? 0xffff : 0;
+                    da2->gdcinput[i] = (da2->gdcreg[LG_SET_RESETJ] & (1 << i)) ? 0xffff : 0;
             da2_gdcropW(addr, bitmask, da2);
             break;
     }
@@ -2782,13 +2804,14 @@ static void
 da2_mmio_writew(uint32_t addr, uint16_t val, void *p)
 {
     da2_t *da2 = (da2_t *) p;
+    // if (da2->bitblt.exec != DA2_BLT_CIDLE) /* Bitblt is in operation. */
+    //     return;
     // if ((addr & ~0x1ffff) != 0xA0000) return;
     if (da2->ioctl[LS_MMIO] & 0x10) {
         // da2_log("da2_mmio_writeW %x %x %x %x %x %x\n", da2->ioctl[LS_MMIO], da2->fctl[LF_MMIO_SEL], da2->fctl[LF_MMIO_MODE], da2->fctl[LF_MMIO_ADDR], addr, val);
         da2_mmio_write(addr, val & 0xff, da2);
         da2_mmio_write(addr + 1, val >> 8, da2);
     } else if (!(da2->ioctl[LS_MODE] & 1)) { /* 16 color or 256 color mode */
-        // return;
         // da2_log("da2_mmio_writeGW %x %x %x %x %x %x\n", da2->ioctl[LS_MMIO], da2->fctl[LF_MMIO_SEL], da2->fctl[LF_MMIO_MODE], da2->fctl[LF_MMIO_ADDR], addr, val);
         da2_mmio_gc_writeW(addr, val, da2);
     } else { /* mode 3h text */
@@ -3181,7 +3204,7 @@ da2_init(UNUSED(const device_t *info))
     mem_mapping_disable(&da2->cmapping);
 
     timer_add(&da2->timer, da2_poll, da2, 0);
-    da2->bitblt.timerspeed = 8ull * TIMER_USEC; /* Bitblt execution speed */
+    da2->bitblt.timerspeed = 5ull * TIMER_USEC; /* Bitblt execution speed */
     timer_add(&da2->bitblt.timer, da2_bitblt_dopayload, da2, 0);
 
     return da2;

From e798030022d9ac1d0adfdbeada832a26aa9d33ed Mon Sep 17 00:00:00 2001
From: Akamaki <97360908+akmed772@users.noreply.github.com>
Date: Tue, 4 Mar 2025 11:38:31 +0900
Subject: [PATCH 32/58] modify bit mask alignment, added delay for vidsel

for Win 3.1.
And added delay for switching video selector
---
 src/video/vid_ps55da2.c | 72 +++++++++++++++++++++++------------------
 1 file changed, 41 insertions(+), 31 deletions(-)

diff --git a/src/video/vid_ps55da2.c b/src/video/vid_ps55da2.c
index 1b9ec473e..0f5851c6f 100644
--- a/src/video/vid_ps55da2.c
+++ b/src/video/vid_ps55da2.c
@@ -431,6 +431,7 @@ typedef struct da2_t {
     uint8_t pos_regs[8];
     svga_t *mb_vga;
     uint8_t monitorid;
+    pc_timer_t timer_vidupd;
 
     int old_pos2;
 } da2_t;
@@ -1108,11 +1109,13 @@ da2_bitblt_dopayload(void *priv)
 {
     da2_t *da2 = (da2_t *) priv;
     timer_set_delay_u64(&da2->bitblt.timer, da2->bitblt.timerspeed);
-    if (da2->bitblt.exec != DA2_BLT_CIDLE)
-        da2_bitblt_exec(da2);
-    else if (da2->bitblt.indata && !(da2->ioctl[LS_MMIO] & 0x10) && (da2->bitblt.exec == DA2_BLT_CIDLE)) {
-            da2->bitblt.exec = DA2_BLT_CLOAD;
+    /* do async operation but it causes the scrolling text glitch in OS/2 J1.3 Command Prompt (TODO) */
+    if (da2->bitblt.exec != DA2_BLT_CIDLE) {
+        while (da2->bitblt.exec != DA2_BLT_CIDLE) 
             da2_bitblt_exec(da2);
+    } else if (da2->bitblt.indata && !(da2->ioctl[LS_MMIO] & 0x10) && (da2->bitblt.exec == DA2_BLT_CIDLE)) {
+        da2->bitblt.exec = DA2_BLT_CLOAD;
+        da2_bitblt_exec(da2);
     } else {
         // timer_disable(&da2->bitblt.timer);
     }
@@ -1202,8 +1205,7 @@ da2_out(uint16_t addr, uint16_t val, void *p)
             oldval                   = da2->fctl[da2->fctladdr];
             da2->fctl[da2->fctladdr] = val;
             if (da2->fctladdr == 0 && oldval != val) {
-                da2->fullchange = changeframecount;
-                da2_recalctimings(da2);
+                da2_log("DA2 Out FCTL addr %03X idx %02X val %02X %04X:%04X\n", addr, da2->fctladdr, val, cs >> 4, cpu_state.pc);
             }
             break;
         case LC_INDEX:
@@ -1338,7 +1340,7 @@ da2_out(uint16_t addr, uint16_t val, void *p)
                     break;
                 case LG_MODE:
                     da2->writemode = val & 3;
-                     /* Resettting masks here gliches the screen in IBM Multitool Chart K3.1 */
+                     /* Resettting masks here gliches chart drawing in IBM Multitool Chart K3.1 */
                     // da2->gdcreg[LG_BIT_MASK_LOW] = 0xff;
                     // da2->gdcreg[LG_BIT_MASK_HIGH] = 0xff;
                     // da2->planemask = 0xff;
@@ -1558,11 +1560,13 @@ da2_outw(uint16_t addr, uint16_t val, void *p)
             // val = rightRotate(val, 8);
             // da2_out(LG_DATA, val, da2);
             da2_out(LG_DATA, val >> 8, da2);
-            /* reset masks */
-            da2->gdcreg[LG_BIT_MASK_LOW] = 0xff;
-            da2->gdcreg[LG_BIT_MASK_HIGH] = 0xff;
-            da2->planemask = 0xff;
-            da2->gdcreg[LG_MAP_MASKJ] = 0xff;
+            /* reset masks for compatibility with Win 3.1 solitaire */
+            if (da2->gdcaddr == LG_MODE) {
+                da2->gdcreg[LG_BIT_MASK_LOW]  = 0xff;
+                da2->gdcreg[LG_BIT_MASK_HIGH] = 0xff;
+                da2->planemask                = 0xff;
+                da2->gdcreg[LG_MAP_MASKJ]     = 0xff;
+            }
             break;
         case 0x3ED:
             da2->gdcaddr = LG_MODE;
@@ -2151,8 +2155,9 @@ da2_render_color_8bpp(da2_t *da2)
 }
 
 void
-da2_updatevidselector(da2_t *da2)
+da2_updatevidselector_tick(void *priv)
 {
+    da2_t *da2 = (da2_t *) priv;
     if (da2->ioctl[LS_MODE] & 0x02) {
         /* VGA passthrough mode */
         da2->override = 1;
@@ -2165,6 +2170,12 @@ da2_updatevidselector(da2_t *da2)
     }
 }
 
+void
+da2_updatevidselector(da2_t *da2)
+{
+    timer_set_delay_u64(&da2->timer_vidupd, 100000ull * TIMER_USEC);
+}
+
 void
 da2_recalctimings(da2_t *da2)
 {
@@ -2386,6 +2397,7 @@ static void
 da2_gdcropW(uint32_t addr, uint16_t bitmask, da2_t *da2)
 {
     if((addr & 8) && !(da2->gdcreg[LG_COMMAND] & 0x08)) bitmask = rightRotate(bitmask, 8);
+    // if((addr & 8)) bitmask = rightRotate(bitmask, 8);
     uint8_t bitmask_l = bitmask & 0xff;
     uint8_t bitmask_h = bitmask >> 8;
     for (int i = 0; i < 8; i++) {
@@ -2582,9 +2594,15 @@ da2_mmio_write(uint32_t addr, uint8_t val, void *p)
         }
     } else if (!(da2->ioctl[LS_MODE] & 1)) { /* 16 color or 256 color mode */
         uint8_t bitmask;
-        /* align bitmask with even address for OS/2 J2.0 (need to verify the condition with Win 3.x) */
-        if ((addr & 1)  && !(da2->gdcreg[LG_COMMAND] & 0x08)) bitmask = da2->gdcreg[LG_BIT_MASK_HIGH];
-        else  bitmask = da2->gdcreg[LG_BIT_MASK_LOW];
+        /* align bitmask with even address (need to verify the condition with OS/2 J2.x, Win 3.x and A-Train IV Win) */
+        /* With byte align: Win 3.1 (Window Title) - ok, Solitaire 3.1 - ok, A-Train IV (splash): bad,  OS/2 J2.0(cmd) - ok */
+        // if ((addr & 1)  && !(da2->gdcreg[LG_COMMAND] & 0x08)) bitmask = da2->gdcreg[LG_BIT_MASK_HIGH];
+        /* Without byte align: Win 3.1 (Window Title) - bad, Solitaire 3.1 - ok, A-Train IV (splash): ok,  OS/2 J2.0(cmd) - ok */
+        /* With byte align: Win 3.1 (Window Title) - ok, Solitaire 3.1 - ok, A-Train IV (splash): ok,  OS/2 J2.0(cmd) - ok, DOS J4.0 MC - ok */
+        if ((addr & 1)) bitmask = da2->gdcreg[LG_BIT_MASK_HIGH];
+        else  
+        /* No align: Win 3.1 (Window Title) - ok, Solitaire 3.1 - ok, A-Train IV (splash): bad,  OS/2 J2.0(cmd) - bad */
+            bitmask = da2->gdcreg[LG_BIT_MASK_LOW];
         
 #ifdef ENABLE_DA2_DEBUGVRAM
         // da2_log("da2_wB %x %02x\n", addr, val);
@@ -2756,7 +2774,7 @@ da2_mmio_gc_writeW(uint32_t addr, uint16_t val, void *p)
             break;
         case 0:
             if (da2->gdcreg[LG_DATA_ROTATION] & 15)
-                val = rightRotate(val, da2->gdcreg[LG_DATA_ROTATION] & 15); // val = svga_rotate[da2->gdcreg[LG_DATA_ROTATION] & 7][val]; TODO this wont work
+                val = rightRotate(val, da2->gdcreg[LG_DATA_ROTATION] & 15);
             if (bitmask == 0xffff && !(da2->gdcreg[LG_COMMAND] & 0x03) && (!da2->gdcreg[LG_ENABLE_SRJ])) {
                 for (int i = 0; i < 8; i++)
                     if (da2->planemask & (1 << i)) {
@@ -2774,22 +2792,13 @@ da2_mmio_gc_writeW(uint32_t addr, uint16_t val, void *p)
             }
             break;
         case 1:
-            // if (!(da2->gdcreg[LG_COMMAND] & 0x03) && (!da2->gdcreg[LG_ENABLE_SRJ])) {
-            //     for (int i = 0; i < 8; i++)
-            //         if (da2->planemask & (1 << i)) {
-            //             uint16_t wdata = (((val & (1 << i)) ? 0xffff : 0) & bitmask) | (da2->gdcsrc[i] & ~bitmask);
-            //             DA2_vram_w(addr | i, wdata & 0xff, da2);
-            //             DA2_vram_w((addr + 8) | i, wdata >> 8, da2);
-            //         }
-            // } else {
-                for (int i = 0; i < 8; i++)
-                    da2->gdcinput[i] = ((val & (1 << i)) ? 0xffff : 0);
-                da2_gdcropW(addr, bitmask, da2);
-            // }
+            for (int i = 0; i < 8; i++)
+                da2->gdcinput[i] = ((val & (1 << i)) ? 0xffff : 0);
+            da2_gdcropW(addr, bitmask, da2);
             break;
         case 3:
             if (da2->gdcreg[LG_DATA_ROTATION] & 15)
-                val = rightRotate(val, da2->gdcreg[LG_DATA_ROTATION] & 15); // val = svga_rotate[da2->gdcreg[LG_DATA_ROTATION] & 7][val];; TODO this wont work
+                val = rightRotate(val, da2->gdcreg[LG_DATA_ROTATION] & 15);
             bitmask &= val;
 
             for (int i = 0; i < 8; i++)
@@ -3117,7 +3126,6 @@ da2_reset(void *priv)
 
     /* Initialize drawing */
     da2->bitblt.exec = DA2_BLT_CIDLE;
-    da2->render      = da2_render_blank;
     da2_reset_ioctl(da2);
 
     da2->pos_regs[0]       = DA2_POSID_L;                    /* Adapter Identification Byte (Low byte) */
@@ -3192,6 +3200,8 @@ da2_init(UNUSED(const device_t *info))
     da2->mmrdbg_fp           = fopen("da2_mmiordat.txt", "w");
 #endif
     da2->bitblt.payload_addr = 0;
+
+    timer_add(&da2->timer_vidupd, da2_updatevidselector_tick, da2, 0);/* Init timer before executing reset */
     da2_reset(da2);
 
     mem_mapping_add(&da2->mmio.mapping, 0xA0000, 0x20000, da2_mmio_read, da2_mmio_readw, NULL, da2_mmio_write, da2_mmio_writew, NULL, NULL, MEM_MAPPING_EXTERNAL, da2);

From 5a77092a717164bd67c3a8f786753f30e8c67e04 Mon Sep 17 00:00:00 2001
From: Akamaki <97360908+akmed772@users.noreply.github.com>
Date: Wed, 5 Mar 2025 00:38:39 +0900
Subject: [PATCH 33/58] add support for 2nd sbcs in videomode 03

---
 src/video/vid_ps55da2.c | 22 +++++++++++++++++++---
 1 file changed, 19 insertions(+), 3 deletions(-)

diff --git a/src/video/vid_ps55da2.c b/src/video/vid_ps55da2.c
index 0f5851c6f..1338304f1 100644
--- a/src/video/vid_ps55da2.c
+++ b/src/video/vid_ps55da2.c
@@ -1111,7 +1111,7 @@ da2_bitblt_dopayload(void *priv)
     timer_set_delay_u64(&da2->bitblt.timer, da2->bitblt.timerspeed);
     /* do async operation but it causes the scrolling text glitch in OS/2 J1.3 Command Prompt (TODO) */
     if (da2->bitblt.exec != DA2_BLT_CIDLE) {
-        while (da2->bitblt.exec != DA2_BLT_CIDLE) 
+        while (da2->bitblt.exec != DA2_BLT_CIDLE) /* this disables async operation */
             da2_bitblt_exec(da2);
     } else if (da2->bitblt.indata && !(da2->ioctl[LS_MMIO] & 0x10) && (da2->bitblt.exec == DA2_BLT_CIDLE)) {
         da2->bitblt.exec = DA2_BLT_CLOAD;
@@ -2015,9 +2015,14 @@ da2_render_textm3(da2_t *da2)
                     }
                 } else {
                     /* the char code is SBCS (ANK) */
-                    uint16_t font = da2->mmio.ram[DA2_GAIJIRAM_SBCS + chr * 0x40 + da2->sc * 2]; /* w13xh29 font */
+                    uint32_t fontbase;
+                    if (extattr & 0x80) /* second map of SBCS font */
+                        fontbase = DA2_GAIJIRAM_SBEX;
+                    else
+                        fontbase = DA2_GAIJIRAM_SBCS;
+                    uint16_t font = da2->mmio.ram[fontbase+ chr * 0x40 + da2->sc * 2]; /* w13xh29 font */
                     font <<= 8;
-                    font |= da2->mmio.ram[DA2_GAIJIRAM_SBCS+ chr * 0x40 + da2->sc * 2 + 1]; /* w13xh29 font */
+                    font |= da2->mmio.ram[fontbase + chr * 0x40 + da2->sc * 2 + 1]; /* w13xh29 font */
                     // if(chr!=0x20) da2_log("ma: %x, sc: %x, chr: %x, font: %x    ", da2->ma, da2->sc, chr, font);
                     for (uint32_t n = 0; n < 13; n++) {
                         p[n] = da2->pallook[da2->egapal[(font & 0x8000) ? fg : bg]];
@@ -2176,6 +2181,17 @@ da2_updatevidselector(da2_t *da2)
     timer_set_delay_u64(&da2->timer_vidupd, 100000ull * TIMER_USEC);
 }
 
+/*
+    Video modes supported by DOS J4.0 for the DA-2 (The DA-2 doesn't have a video BIOS on its card.)
+    Mode (hex)  Type    Colors  Format  Base Address    PELs
+    3           A/N/K   16      80 x 25 B8000h          1040 x 725
+    8           A/N/K   1       80 x 25 E0000h          1040 x 725
+    A           APA     1       78 x 25 A0000h          1024 x 768
+    D           APA     16      78 x 25 A0000h          1024 x 768
+    E           A/N/K   16      80 x 25 E0000h          1040 x 725
+    F           APA     256             A0000h          1024 x 768
+    45 (undoc)  APA     16              A0000h          1040 x 768
+*/
 void
 da2_recalctimings(da2_t *da2)
 {

From 73576bb61ed1fb90be962f87028c3a6244b7723d Mon Sep 17 00:00:00 2001
From: Cacodemon345 <wahil1976@outlook.com>
Date: Wed, 5 Mar 2025 01:23:37 +0600
Subject: [PATCH 34/58] Revert "S3 ViRGE: Make IRQs happen in main thread"

This reverts commit fae26729f103010ddc7e621a37083bc4ae6d5736.
---
 src/video/vid_s3_virge.c | 57 +++++-----------------------------------
 1 file changed, 6 insertions(+), 51 deletions(-)

diff --git a/src/video/vid_s3_virge.c b/src/video/vid_s3_virge.c
index 62ffcff0c..af33003cb 100644
--- a/src/video/vid_s3_virge.c
+++ b/src/video/vid_s3_virge.c
@@ -338,7 +338,6 @@ typedef struct virge_t {
     event_t *    fifo_not_full_event;
 
     atomic_int   virge_busy;
-    atomic_uint  virge_irq_req;
 
     uint8_t      subsys_stat;
     uint8_t      subsys_cntl;
@@ -374,9 +373,6 @@ typedef struct virge_t {
     uint32_t     dma_mmio_addr;
     uint32_t     dma_data_type;
 
-    pc_timer_t   wake_timer;
-    pc_timer_t   irq_timer;
-
     int          pci;
     int          is_agp;
 } virge_t;
@@ -387,28 +383,6 @@ wake_fifo_thread(virge_t *virge) {
     thread_set_event(virge->wake_fifo_thread);
 }
 
-void
-s3_virge_wake_fifo_timer(void* priv)
-{
-    virge_t *virge = (virge_t *) priv;
-
-    thread_set_event(virge->wake_fifo_thread); /*Wake up FIFO thread if moving from idle*/
-}
-
-void
-s3_virge_wake_fifo_thread(virge_t *virge)
-{
-    if (!timer_is_enabled(&virge->wake_timer)) {
-        /*Don't wake FIFO thread immediately - if we do that it will probably
-          process one word and go back to sleep, requiring it to be woken on
-          almost every write. Instead, wait a short while so that the CPU
-          emulation writes more data so we have more batched-up work.*/
-        timer_on_auto(&virge->wake_timer, 100.0);
-    }
-    if (!timer_is_enabled(&virge->irq_timer))
-        timer_on_auto(&virge->irq_timer, 100.0);
-}
-
 static virge_t         *reset_state = NULL;
 
 static video_timings_t timing_diamond_stealth3d_2000_pci = { .type = VIDEO_PCI, .write_b = 2, .write_w = 2, .write_l = 3, .read_b = 28, .read_w = 28, .read_l = 45 };
@@ -497,19 +471,6 @@ s3_virge_update_irqs(virge_t *virge)
         pci_clear_irq(virge->pci_slot, PCI_INTA, &virge->irq_state);
 }
 
-void
-s3_virge_update_irq_timer(void* priv)
-{
-    virge_t *virge = (virge_t *) priv;
-
-    if (virge->virge_irq_req) {
-        virge->virge_irq_req--;
-        s3_virge_update_irqs(virge);
-    }
-
-    timer_on_auto(&virge->irq_timer, 100.);
-}
-
 static void
 s3_virge_out(uint16_t addr, uint8_t val, void *priv)
 {
@@ -1140,9 +1101,6 @@ static void
 s3_virge_vblank_start(svga_t *svga) {
     virge_t *virge = (virge_t *) svga->priv;
 
-    if (virge->virge_irq_req)
-        virge->virge_irq_req--;
-
     virge->subsys_stat |= INT_VSY;
     s3_virge_update_irqs(virge);
 }
@@ -1837,8 +1795,7 @@ fifo_thread(void *param)
          if (virge->cmd_dma)
             virge->subsys_stat |= (INT_HOST_DONE | INT_CMD_DONE);
 
-         //s3_virge_update_irqs(virge);
-         virge->virge_irq_req++;
+         s3_virge_update_irqs(virge);
     }
 }
 
@@ -1878,8 +1835,10 @@ s3_virge_queue(virge_t *virge, uint32_t addr, uint32_t val, uint32_t type)
 
     virge->fifo_write_idx++;
 
-    if (FIFO_ENTRIES > 0xe000 || FIFO_ENTRIES < 16)
-        s3_virge_wake_fifo_thread(virge);
+    if (FIFO_ENTRIES > 0xe000)
+        wake_fifo_thread(virge);
+    if (FIFO_ENTRIES > 0xe000 || FIFO_ENTRIES < 8)
+        wake_fifo_thread(virge);
 }
 
 static void
@@ -4445,8 +4404,7 @@ render_thread(void *param)
         }
         virge->s3d_busy = 0;
         virge->subsys_stat |= INT_S3D_DONE;
-        //s3_virge_update_irqs(virge);
-        virge->virge_irq_req++;
+        s3_virge_update_irqs(virge);
     }
 }
 
@@ -5375,9 +5333,6 @@ s3_virge_init(const device_t *info)
     virge->fifo_not_full_event = thread_create_event();
     virge->fifo_thread = thread_create(fifo_thread, virge);
 
-    timer_add(&virge->wake_timer, s3_virge_wake_fifo_timer, virge, 0);
-    timer_add(&virge->irq_timer, s3_virge_update_irq_timer, virge, 0);
-
     virge->local = info->local;
 
     *reset_state = *virge;

From 92d69475f4eea136226842784ffdcf2d862dbb77 Mon Sep 17 00:00:00 2001
From: Cacodemon345 <wahil1976@outlook.com>
Date: Wed, 5 Mar 2025 02:17:51 +0600
Subject: [PATCH 35/58] Only retain the newer IRQ updating code

---
 src/video/vid_s3_virge.c | 43 ++++++++++++++++++++++++++++++----------
 1 file changed, 32 insertions(+), 11 deletions(-)

diff --git a/src/video/vid_s3_virge.c b/src/video/vid_s3_virge.c
index af33003cb..904d4a4cf 100644
--- a/src/video/vid_s3_virge.c
+++ b/src/video/vid_s3_virge.c
@@ -338,6 +338,7 @@ typedef struct virge_t {
     event_t *    fifo_not_full_event;
 
     atomic_int   virge_busy;
+    atomic_uint  irq_pending;
 
     uint8_t      subsys_stat;
     uint8_t      subsys_cntl;
@@ -375,6 +376,8 @@ typedef struct virge_t {
 
     int          pci;
     int          is_agp;
+
+    pc_timer_t   irq_timer;
 } virge_t;
 
 static __inline void
@@ -471,6 +474,19 @@ s3_virge_update_irqs(virge_t *virge)
         pci_clear_irq(virge->pci_slot, PCI_INTA, &virge->irq_state);
 }
 
+static void
+s3_virge_update_irq_timer(void* priv)
+{
+    virge_t *virge = (virge_t *) priv;
+
+    if (virge->irq_pending) {
+        virge->irq_pending--;
+        s3_virge_update_irqs(virge);
+    }
+
+    timer_on_auto(&virge->irq_timer, 100.);
+}
+
 static void
 s3_virge_out(uint16_t addr, uint8_t val, void *priv)
 {
@@ -1101,6 +1117,8 @@ static void
 s3_virge_vblank_start(svga_t *svga) {
     virge_t *virge = (virge_t *) svga->priv;
 
+    if (virge->irq_pending)
+        virge->irq_pending--;
     virge->subsys_stat |= INT_VSY;
     s3_virge_update_irqs(virge);
 }
@@ -1784,18 +1802,18 @@ fifo_thread(void *param)
             virge->fifo_read_idx++;
             fifo->addr_type = FIFO_INVALID;
 
-             if (FIFO_ENTRIES > 0xe000)
-                 thread_set_event(virge->fifo_not_full_event);
+            if (FIFO_ENTRIES > 0xe000)
+                thread_set_event(virge->fifo_not_full_event);
 
-             end_time = plat_timer_read();
-             virge_time += end_time - start_time;
-         }
-         virge->virge_busy = 0;
-         virge->subsys_stat |= (INT_FIFO_EMP | INT_3DF_EMP);
-         if (virge->cmd_dma)
-            virge->subsys_stat |= (INT_HOST_DONE | INT_CMD_DONE);
+            end_time = plat_timer_read();
+            virge_time += end_time - start_time;
+        }
+        virge->virge_busy = 0;
+        virge->subsys_stat |= (INT_FIFO_EMP | INT_3DF_EMP);
+        if (virge->cmd_dma)
+           virge->subsys_stat |= (INT_HOST_DONE | INT_CMD_DONE);
 
-         s3_virge_update_irqs(virge);
+        virge->irq_pending++;
     }
 }
 
@@ -4404,7 +4422,7 @@ render_thread(void *param)
         }
         virge->s3d_busy = 0;
         virge->subsys_stat |= INT_S3D_DONE;
-        s3_virge_update_irqs(virge);
+        virge->irq_pending++;
     }
 }
 
@@ -5062,6 +5080,7 @@ s3_virge_disable_handlers(virge_t *dev)
 
     reset_state->svga.timer       = dev->svga.timer;
     reset_state->svga.timer8514   = dev->svga.timer8514;
+    reset_state->irq_timer        = dev->irq_timer;
 }
 
 static void
@@ -5333,6 +5352,8 @@ s3_virge_init(const device_t *info)
     virge->fifo_not_full_event = thread_create_event();
     virge->fifo_thread = thread_create(fifo_thread, virge);
 
+    timer_add(&virge->irq_timer, s3_virge_update_irq_timer, virge, 1);
+
     virge->local = info->local;
 
     *reset_state = *virge;

From 6c92eb61bf4e8c0d225b9f6354324ff766cc13ab Mon Sep 17 00:00:00 2001
From: Akamaki <97360908+akmed772@users.noreply.github.com>
Date: Wed, 5 Mar 2025 07:12:15 +0900
Subject: [PATCH 36/58] fix a screen glitch in OS/2 DOS MODE 1

fix a glitch in OS/2 DOS MODE 1 (monochrome graphics)
---
 src/video/vid_ps55da2.c | 22 ++++++++++++----------
 1 file changed, 12 insertions(+), 10 deletions(-)

diff --git a/src/video/vid_ps55da2.c b/src/video/vid_ps55da2.c
index 1338304f1..6568d9b45 100644
--- a/src/video/vid_ps55da2.c
+++ b/src/video/vid_ps55da2.c
@@ -259,7 +259,7 @@
 #ifdef ENABLE_DA2_LOG
 #    define ENABLE_DA2_DEBUGBLT 1
 #    define ENABLE_DA2_DEBUGVRAM 1
-#    define ENABLE_DA2_DEBUGFULLSCREEN 1
+// #    define ENABLE_DA2_DEBUGFULLSCREEN 1
 // #    define ENABLE_DA2_DEBUGMONWAIT 1
 int da2_do_log = ENABLE_DA2_LOG;
 
@@ -1227,7 +1227,9 @@ da2_out(uint16_t addr, uint16_t val, void *p)
                         val = 0;
                     break;
                 case LC_START_ADDRESS_HIGH:
-                    // if (da2->crtc[0x1c] & 0x40) return;
+                case LC_START_ADDRESS_LOW:
+                /* OS/2 DOS in MODE 1 read this to set the base line, but it does not work correctly. */
+                    val = 0;
                     break;
                 case LC_VERTICAL_TOTALJ:      /* Vertical Total */
                 case LC_VERTICAL_SYNC_START:  /* Vertical Retrace Start Register */
@@ -1248,8 +1250,8 @@ da2_out(uint16_t addr, uint16_t val, void *p)
                 case LC_H_DISPLAY_ENABLE_END:
                 case LC_VERTICAL_TOTALJ:
                 case LC_MAXIMUM_SCAN_LINE:
-                case LC_START_ADDRESS_HIGH:
-                case LC_START_ADDRESS_LOW:
+                // case LC_START_ADDRESS_HIGH:
+                // case LC_START_ADDRESS_LOW:
                 case LC_VERTICAL_SYNC_START:
                 case LC_V_DISPLAY_ENABLE_END:
                 case LC_START_VERTICAL_BLANK:
@@ -2235,7 +2237,7 @@ da2_recalctimings(da2_t *da2)
 
     // da2->interlace = 0;
 
-    // da2->ma_latch = ((da2->crtc[0xc] & 0x3ff) << 8) | da2->crtc[0xd];//w + b
+    // da2->ma_latch = ((da2->crtc[LC_START_ADDRESS_HIGH] & 0x3ff) << 8) | da2->crtc[LC_START_ADDRESS_LOW];//w + b
     da2->ca_adj = 0;
 
     da2->rowcount = da2->crtc[LC_MAXIMUM_SCAN_LINE];
@@ -2517,7 +2519,7 @@ da2_mmio_readw(uint32_t addr, void *p)
 {
     da2_t *da2 = (da2_t *) p;
     // da2_log("da2_readW: %x %x %x %x %x\n", da2->ioctl[LS_MMIO], da2->fctl[LF_MMIO_SEL], da2->fctl[LF_MMIO_MODE], da2->fctl[LF_MMIO_ADDR], addr);
-    da2_log("da2_readW: %x %x %x %x %x CS:PC=%4x:%4x\n", da2->ioctl[LS_MMIO], da2->fctl[LF_MMIO_SEL], da2->fctl[LF_MMIO_MODE], da2->fctl[LF_MMIO_ADDR], addr, CS, cpu_state.pc);
+    // da2_log("da2_readW: %x %x %x %x %x CS:PC=%4x:%4x\n", da2->ioctl[LS_MMIO], da2->fctl[LF_MMIO_SEL], da2->fctl[LF_MMIO_MODE], da2->fctl[LF_MMIO_ADDR], addr, CS, cpu_state.pc);
 
     if (da2->ioctl[LS_MMIO] & 0x10) {
         return (uint16_t) da2_mmio_read(addr, da2) | (uint16_t) (da2_mmio_read(addr + 1, da2) << 8);
@@ -2553,7 +2555,7 @@ da2_mmio_readw(uint32_t addr, void *p)
             }
             return ~ret;
         } else {
-            // da2_log("da2_Rw: %05x=%04x\n", addr, da2->gdcla[da2->readplane]);
+            da2_log("da2_Rw: %05x(%d) = %04x\n", addr, da2->readplane, da2->gdcla[da2->readplane]);
             return da2->gdcla[da2->readplane];
         }
     } else {
@@ -3050,11 +3052,11 @@ da2_poll(void *priv)
                 = da2->maback = da2->ma_latch;
             da2->ca           = ((da2->crtc[LC_CURSOR_LOC_HIGH] << 8) | da2->crtc[LC_CURSOR_LOC_LOWJ]) + da2->ca_adj;
 
-            da2->ma <<= 1;
-            da2->maback <<= 1;
+            // da2->ma <<= 1;
+            // da2->maback <<= 1;
             da2->ca <<= 1;
 
-            // da2_log("Addr %08X vson %03X vsoff %01X  %02X %02X %02X %i %i\n",ma,da2_vsyncstart,crtc[0x11]&0xF,crtc[0xD],crtc[0xC],crtc[0x33], da2_interlace, oddeven);
+            // da2_log("Addr %08X vson %03X vsoff %01X\n",da2->ma,da2->vsyncstart,da2->crtc[0x11]&0xF);
         }
         if (da2->vc == da2->vtotal) {
             // da2_log("VC vtotal\n");

From 5b7766bb7ab571aee906dac14f8548e53905586c Mon Sep 17 00:00:00 2001
From: Akamaki <97360908+akmed772@users.noreply.github.com>
Date: Wed, 5 Mar 2025 12:35:56 +0900
Subject: [PATCH 37/58] disable debug logging, update comments

---
 src/video/vid_ps55da2.c | 45 ++++++++++++++++++++++++-----------------
 1 file changed, 26 insertions(+), 19 deletions(-)

diff --git a/src/video/vid_ps55da2.c b/src/video/vid_ps55da2.c
index 6568d9b45..335cbf86d 100644
--- a/src/video/vid_ps55da2.c
+++ b/src/video/vid_ps55da2.c
@@ -10,7 +10,7 @@
  * 
  *   Notes: There are some known issues that should be corrected.
  *            - Incorrect foreground text color appears on an active window in OS/2 J1.3. 
- *            - BitBlt's text drawing function does not work correctly.
+ *            - Glitches some part of graphics on the Control Panel in OS/2 J2.1 beta. 
  *            - The screen resolution and blanking interval time maybe not correct.
  * 
  *          The code should be tested with following cases.
@@ -253,11 +253,11 @@
 #define LG_SET_RESET_2          0x10
 
 #ifndef RELEASE_BUILD
-#define ENABLE_DA2_LOG 1
+// #define ENABLE_DA2_LOG 1
 #endif
 
 #ifdef ENABLE_DA2_LOG
-#    define ENABLE_DA2_DEBUGBLT 1
+// #    define ENABLE_DA2_DEBUGBLT 1
 #    define ENABLE_DA2_DEBUGVRAM 1
 // #    define ENABLE_DA2_DEBUGFULLSCREEN 1
 // #    define ENABLE_DA2_DEBUGMONWAIT 1
@@ -1228,7 +1228,8 @@ da2_out(uint16_t addr, uint16_t val, void *p)
                     break;
                 case LC_START_ADDRESS_HIGH:
                 case LC_START_ADDRESS_LOW:
-                /* OS/2 DOS in MODE 1 read this to set the base line, but it does not work correctly. */
+                /* The DOS J4.0 MODE 4 command and OS/2 driver write 0xFF00.
+                   OS/2 DOS MODE 1 setup reads this to set the base line, but it causes the screen glitch. */
                     val = 0;
                     break;
                 case LC_VERTICAL_TOTALJ:      /* Vertical Total */
@@ -1625,11 +1626,13 @@ da2_inb(uint16_t addr, void *p)
 uint16_t
 da2_inw(uint16_t addr, void *p)
 {
-    // uint16_t temp;
+    uint16_t temp;
     da2_t *da2       = (da2_t *) p;
     da2->inflipflop  = 0;
     da2->outflipflop = 0;
-    return da2_in(addr, da2);
+    temp = da2_in(addr, da2);
+    da2_log("DA2 Inw addr %03X val %04X\n", addr, temp);
+    return temp;
 }
 /* IO 03DAh : Input Status Register 2 for DOSSHELL used by DOS J4.0 */
 uint8_t
@@ -2184,15 +2187,15 @@ da2_updatevidselector(da2_t *da2)
 }
 
 /*
-    Video modes supported by DOS J4.0 for the DA-2 (The DA-2 doesn't have a video BIOS on its card.)
-    Mode (hex)  Type    Colors  Format  Base Address    PELs
-    3           A/N/K   16      80 x 25 B8000h          1040 x 725
-    8           A/N/K   1       80 x 25 E0000h          1040 x 725
-    A           APA     1       78 x 25 A0000h          1024 x 768
-    D           APA     16      78 x 25 A0000h          1024 x 768
-    E           A/N/K   16      80 x 25 E0000h          1040 x 725
-    F           APA     256             A0000h          1024 x 768
-    45 (undoc)  APA     16              A0000h          1040 x 768
+    INT 10h video modes supported in DOS J4.0 (The DA-2 doesn't have a video BIOS on its card.)
+    Mode        Type    Colors  Text    Base Address    PELs        Render
+    3           A/N/K   16      80 x 25 B0000h/B8000h   1040 x 725  textm3
+    8           A/N/K   2       80 x 25 E0000h          1040 x 725  text
+    Ah          APA     1       78 x 25 A0000h          1024 x 768  color_4bpp
+    Dh          APA     16      78 x 25 A0000h          1024 x 768  color_4bpp
+    Eh          A/N/K   16      80 x 25 E0000h          1040 x 725  text
+    Fh          APA     256     NA      A0000h          1024 x 768  color_8bpp
+    45h(undoc)  APA     16      NA      A0000h          1040 x 768  color_4bpp
 */
 void
 da2_recalctimings(da2_t *da2)
@@ -2499,7 +2502,9 @@ da2_mmio_read(uint32_t addr, void *p)
         cycles -= video_timing_read_b;
         for (int i = 0; i < 8; i++)
             da2->gdcla[i] = da2->vram[(addr << 3) | i]; /* read in byte */
+#ifdef ENABLE_DA2_DEBUGVRAM
         da2_log("da2_Rb: %05x=%02x\n", addr, da2->gdcla[da2->readplane]);
+#endif
         if (da2->gdcreg[LG_MODE] & 0x08) { /* compare data across planes if the read mode bit (3EB 05, bit 3) is 1 */
             uint8_t ret = 0;
             for (int i = 0; i < 8; i++) {
@@ -2555,7 +2560,9 @@ da2_mmio_readw(uint32_t addr, void *p)
             }
             return ~ret;
         } else {
+#ifdef ENABLE_DA2_DEBUGVRAM
             da2_log("da2_Rw: %05x(%d) = %04x\n", addr, da2->readplane, da2->gdcla[da2->readplane]);
+#endif
             return da2->gdcla[da2->readplane];
         }
     } else {
@@ -2612,11 +2619,11 @@ da2_mmio_write(uint32_t addr, uint8_t val, void *p)
         }
     } else if (!(da2->ioctl[LS_MODE] & 1)) { /* 16 color or 256 color mode */
         uint8_t bitmask;
-        /* align bitmask with even address (need to verify the condition with OS/2 J2.x, Win 3.x and A-Train IV Win) */
-        /* With byte align: Win 3.1 (Window Title) - ok, Solitaire 3.1 - ok, A-Train IV (splash): bad,  OS/2 J2.0(cmd) - ok */
+        /* Align bitmask with even address */
+        /* With byte align: Win 3.1 (Window) - ok, Solitaire 3.1 - ok, A-Train IV (splash): bad,  OS/2 J2.0(cmd) - ok */
         // if ((addr & 1)  && !(da2->gdcreg[LG_COMMAND] & 0x08)) bitmask = da2->gdcreg[LG_BIT_MASK_HIGH];
-        /* Without byte align: Win 3.1 (Window Title) - bad, Solitaire 3.1 - ok, A-Train IV (splash): ok,  OS/2 J2.0(cmd) - ok */
-        /* With byte align: Win 3.1 (Window Title) - ok, Solitaire 3.1 - ok, A-Train IV (splash): ok,  OS/2 J2.0(cmd) - ok, DOS J4.0 MC - ok */
+        /* Without byte align: Win 3.1 (Window) - bad, Solitaire 3.1 - ok, A-Train IV (splash): ok,  OS/2 J2.0(cmd) - ok */
+        /* With byte align: Win 3.1 (Window) - ok, Solitaire 3.1 - ok, A-Train IV (splash): ok, OS/2 J2.0(cmd) - ok */
         if ((addr & 1)) bitmask = da2->gdcreg[LG_BIT_MASK_HIGH];
         else  
         /* No align: Win 3.1 (Window Title) - ok, Solitaire 3.1 - ok, A-Train IV (splash): bad,  OS/2 J2.0(cmd) - bad */

From 1a60f5aac4b7874cd0e16f78fa513b7b72a78949 Mon Sep 17 00:00:00 2001
From: Akamaki <97360908+akmed772@users.noreply.github.com>
Date: Wed, 5 Mar 2025 13:01:58 +0900
Subject: [PATCH 38/58] remove video timing modification

---
 src/video/vid_ps55da2.c | 9 +--------
 1 file changed, 1 insertion(+), 8 deletions(-)

diff --git a/src/video/vid_ps55da2.c b/src/video/vid_ps55da2.c
index 335cbf86d..823ef4d2f 100644
--- a/src/video/vid_ps55da2.c
+++ b/src/video/vid_ps55da2.c
@@ -2220,14 +2220,7 @@ da2_recalctimings(da2_t *da2)
     if (da2->crtc[LC_START_H_DISPLAY_ENAB] & 1) {
         da2->hdisp--;
         da2->dispend -= 29;
-    } else {
-        // da2->vtotal += 2;
-        da2->dispend--;
-        // da2->vsyncstart++;
-        // da2->split++;
-        // da2->vblankstart++;
-        // da2->hdisp--;
-    }
+    } 
 
     da2->htotal = da2->crtc[LC_HORIZONTAL_TOTAL];
     da2->htotal += 1;

From 8c2db2892d98b8b599f002c944667c47b10dc39e Mon Sep 17 00:00:00 2001
From: OBattler <oubattler@gmail.com>
Date: Wed, 5 Mar 2025 21:52:17 +0100
Subject: [PATCH 39/58] CPU: Fix Cyrix SMM instructions.

---
 src/cpu/386_ops.h   | 104 +++++++++++++++++++++++++++++++++++++++++---
 src/cpu/cpu.c       |  10 +----
 src/cpu/cpu.h       |  50 +++++++++++----------
 src/cpu/cpu_table.c |  87 ++++++++++++++++++++++++++++++++++++
 src/cpu/x86_ops.h   |   2 +
 5 files changed, 215 insertions(+), 38 deletions(-)

diff --git a/src/cpu/386_ops.h b/src/cpu/386_ops.h
index b8ef9edda..80a856482 100644
--- a/src/cpu/386_ops.h
+++ b/src/cpu/386_ops.h
@@ -859,12 +859,12 @@ const OpFn OP_TABLE(c486_0f)[1024] = {
 /*00*/  op0F00_a16,     op0F01_l_a16,   opLAR_l_a16,    opLSL_l_a16,    ILLEGAL,        ILLEGAL,        opCLTS,         opLOADALL386,   opINVD,         opWBINVD,       ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
 /*10*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
 /*20*/  opMOV_r_CRx_a16,opMOV_r_DRx_a16,opMOV_CRx_r_a16,opMOV_DRx_r_a16,opMOV_r_TRx_a16,ILLEGAL,        opMOV_TRx_r_a16,ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
-/*30*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        opRDSHR_a16,    opWRSHR_a16,    opSVDC_a16,     opRSDC_a16,     opSVLDT_a16,    opRSLDT_a16,    opSVTS_a16,     opRSTS_a16,     opSMINT,        ILLEGAL,
+/*30*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        opRDSHR_a16,    opWRSHR_a16,    ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
 
 /*40*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
 /*50*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
 /*60*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
-/*70*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*70*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        opSVDC_a16,     opRSDC_a16,     opSVLDT_a16,    opRSLDT_a16,    opSVTS_a16,     opRSTS_a16,     opSMINT,        ILLEGAL,
 
 /*80*/  opJO_l,         opJNO_l,        opJB_l,         opJNB_l,        opJE_l,         opJNE_l,        opJBE_l,        opJNBE_l,       opJS_l,         opJNS_l,        opJP_l,         opJNP_l,        opJL_l,         opJNL_l,        opJLE_l,        opJNLE_l,
 /*90*/  opSETO_a16,     opSETNO_a16,    opSETB_a16,     opSETNB_a16,    opSETE_a16,     opSETNE_a16,    opSETBE_a16,    opSETNBE_a16,   opSETS_a16,     opSETNS_a16,    opSETP_a16,     opSETNP_a16,    opSETL_a16,     opSETNL_a16,    opSETLE_a16,    opSETNLE_a16,
@@ -1392,7 +1392,7 @@ const OpFn OP_TABLE(c6x86_0f)[1024] = {
 /*00*/  op0F00_a16,     op0F01_w_a16,   opLAR_w_a16,    opLSL_w_a16,    ILLEGAL,        ILLEGAL,        opCLTS,         ILLEGAL,        opINVD,         opWBINVD,       ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
 /*10*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
 /*20*/  opMOV_r_CRx_a16,opMOV_r_DRx_a16,opMOV_CRx_r_a16,opMOV_DRx_r_a16,opMOV_r_TRx_a16,ILLEGAL,        opMOV_TRx_r_a16,ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
-/*30*/  opWRMSR,        opRDTSC,        opRDMSR,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*30*/  opWRMSR,        opRDTSC,        opRDMSR,        ILLEGAL,        ILLEGAL,        ILLEGAL,        opRDSHR_a16,    opWRSHR_a16,    ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
 
 /*40*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
 /*50*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
@@ -1414,7 +1414,7 @@ const OpFn OP_TABLE(c6x86_0f)[1024] = {
 /*00*/  op0F00_a16,     op0F01_l_a16,   opLAR_l_a16,    opLSL_l_a16,    ILLEGAL,        ILLEGAL,        opCLTS,         ILLEGAL,        opINVD,         opWBINVD,       ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
 /*10*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
 /*20*/  opMOV_r_CRx_a16,opMOV_r_DRx_a16,opMOV_CRx_r_a16,opMOV_DRx_r_a16,opMOV_r_TRx_a16,ILLEGAL,        opMOV_TRx_r_a16,ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
-/*30*/  opWRMSR,        opRDTSC,        opRDMSR,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*30*/  opWRMSR,        opRDTSC,        opRDMSR,        ILLEGAL,        ILLEGAL,        ILLEGAL,        opRDSHR_a16,    opWRSHR_a16,    ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
 
 /*40*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
 /*50*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
@@ -1436,7 +1436,7 @@ const OpFn OP_TABLE(c6x86_0f)[1024] = {
 /*00*/  op0F00_a32,     op0F01_w_a32,   opLAR_w_a32,    opLSL_w_a32,    ILLEGAL,        ILLEGAL,        opCLTS,         ILLEGAL,        opINVD,         opWBINVD,       ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
 /*10*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
 /*20*/  opMOV_r_CRx_a32,opMOV_r_DRx_a32,opMOV_CRx_r_a32,opMOV_DRx_r_a32,opMOV_r_TRx_a32,ILLEGAL,        opMOV_TRx_r_a32,ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
-/*30*/  opWRMSR,        opRDTSC,        opRDMSR,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*30*/  opWRMSR,        opRDTSC,        opRDMSR,        ILLEGAL,        ILLEGAL,        ILLEGAL,        opRDSHR_a32,    opWRSHR_a32,    ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
 
 /*40*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
 /*50*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
@@ -1458,7 +1458,99 @@ const OpFn OP_TABLE(c6x86_0f)[1024] = {
 /*00*/  op0F00_a32,     op0F01_l_a32,   opLAR_l_a32,    opLSL_l_a32,    ILLEGAL,        ILLEGAL,        opCLTS,         ILLEGAL,        opINVD,         opWBINVD,       ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
 /*10*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
 /*20*/  opMOV_r_CRx_a32,opMOV_r_DRx_a32,opMOV_CRx_r_a32,opMOV_DRx_r_a32,opMOV_r_TRx_a32,ILLEGAL,        opMOV_TRx_r_a32,ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
-/*30*/  opWRMSR,        opRDTSC,        opRDMSR,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*30*/  opWRMSR,        opRDTSC,        opRDMSR,        ILLEGAL,        ILLEGAL,        ILLEGAL,        opRDSHR_a32,    opWRSHR_a32,    ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+
+/*40*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*50*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*60*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*70*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        opSVDC_a32,     opRSDC_a32,     opSVLDT_a32,    opRSLDT_a32,    opSVTS_a32,     opRSTS_a32,     opSMINT,        ILLEGAL,
+
+/*80*/  opJO_l,         opJNO_l,        opJB_l,         opJNB_l,        opJE_l,         opJNE_l,        opJBE_l,        opJNBE_l,       opJS_l,         opJNS_l,        opJP_l,         opJNP_l,        opJL_l,         opJNL_l,        opJLE_l,        opJNLE_l,
+/*90*/  opSETO_a32,     opSETNO_a32,    opSETB_a32,     opSETNB_a32,    opSETE_a32,     opSETNE_a32,    opSETBE_a32,    opSETNBE_a32,   opSETS_a32,     opSETNS_a32,    opSETP_a32,     opSETNP_a32,    opSETL_a32,     opSETNL_a32,    opSETLE_a32,    opSETNLE_a32,
+/*a0*/  opPUSH_FS_l,    opPOP_FS_l,     opCPUID,        opBT_l_r_a32,   opSHLD_l_i_a32, opSHLD_l_CL_a32,ILLEGAL,        ILLEGAL,        opPUSH_GS_l,    opPOP_GS_l,     opRSM,          opBTS_l_r_a32,  opSHRD_l_i_a32, opSHRD_l_CL_a32,ILLEGAL,        opIMUL_l_l_a32,
+/*b0*/  opCMPXCHG_b_a32,opCMPXCHG_l_a32,opLSS_l_a32,    opBTR_l_r_a32,  opLFS_l_a32,    opLGS_l_a32,    opMOVZX_l_b_a32,opMOVZX_l_w_a32,ILLEGAL,        ILLEGAL,        opBA_l_a32,     opBTC_l_r_a32,  opBSF_l_a32,    opBSR_l_a32,    opMOVSX_l_b_a32,opMOVSX_l_w_a32,
+
+/*c0*/  opXADD_b_a32,   opXADD_l_a32,   ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        opCMPXCHG8B_a32,opBSWAP_EAX,    opBSWAP_ECX,    opBSWAP_EDX,    opBSWAP_EBX,    opBSWAP_ESP,    opBSWAP_EBP,    opBSWAP_ESI,    opBSWAP_EDI,
+/*d0*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*e0*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*f0*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+    // clang-format on
+};
+
+const OpFn OP_TABLE(c6x86l_0f)[1024] = {
+    // clang-format off
+        /*16-bit data, 16-bit addr*/
+/*      00              01              02              03              04              05              06              07              08              09              0a              0b              0c              0d              0e              0f*/
+/*00*/  op0F00_a16,     op0F01_w_a16,   opLAR_w_a16,    opLSL_w_a16,    ILLEGAL,        ILLEGAL,        opCLTS,         ILLEGAL,        opINVD,         opWBINVD,       ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*10*/  opMOV_b_r_a16,  opMOV_w_r_a16,  opMOV_r_b_a16,  opMOV_r_w_a16,  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*20*/  opMOV_r_CRx_a16,opMOV_r_DRx_a16,opMOV_CRx_r_a16,opMOV_DRx_r_a16,opMOV_r_TRx_a16,ILLEGAL,        opMOV_TRx_r_a16,ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*30*/  opWRMSR,        opRDTSC,        opRDMSR,        ILLEGAL,        ILLEGAL,        ILLEGAL,        opRDSHR_a16,    opWRSHR_a16,    ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+
+/*40*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*50*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*60*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*70*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        opSVDC_a16,     opRSDC_a16,     opSVLDT_a16,    opRSLDT_a16,    opSVTS_a16,     opRSTS_a16,     opSMINT,        ILLEGAL,
+
+/*80*/  opJO_w,         opJNO_w,        opJB_w,         opJNB_w,        opJE_w,         opJNE_w,        opJBE_w,        opJNBE_w,       opJS_w,         opJNS_w,        opJP_w,         opJNP_w,        opJL_w,         opJNL_w,        opJLE_w,        opJNLE_w,
+/*90*/  opSETO_a16,     opSETNO_a16,    opSETB_a16,     opSETNB_a16,    opSETE_a16,     opSETNE_a16,    opSETBE_a16,    opSETNBE_a16,   opSETS_a16,     opSETNS_a16,    opSETP_a16,     opSETNP_a16,    opSETL_a16,     opSETNL_a16,    opSETLE_a16,    opSETNLE_a16,
+/*a0*/  opPUSH_FS_w,    opPOP_FS_w,     opCPUID,        opBT_w_r_a16,   opSHLD_w_i_a16, opSHLD_w_CL_a16,ILLEGAL,        ILLEGAL,        opPUSH_GS_w,    opPOP_GS_w,     opRSM,          opBTS_w_r_a16,  opSHRD_w_i_a16, opSHRD_w_CL_a16,ILLEGAL,        opIMUL_w_w_a16,
+/*b0*/  opCMPXCHG_b_a16,opCMPXCHG_w_a16,opLSS_w_a16,    opBTR_w_r_a16,  opLFS_w_a16,    opLGS_w_a16,    opMOVZX_w_b_a16,opMOVZX_w_w_a16,ILLEGAL,        ILLEGAL,        opBA_w_a16,     opBTC_w_r_a16,  opBSF_w_a16,    opBSR_w_a16,    opMOVSX_w_b_a16,ILLEGAL,
+
+/*c0*/  opXADD_b_a16,   opXADD_w_a16,   ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        opCMPXCHG8B_a16,opBSWAP_EAX,    opBSWAP_ECX,    opBSWAP_EDX,    opBSWAP_EBX,    opBSWAP_ESP,    opBSWAP_EBP,    opBSWAP_ESI,    opBSWAP_EDI,
+/*d0*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*e0*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*f0*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        opSVDC_a16,     opRSDC_a16,     opSVLDT_a16,    opRSLDT_a16,    opSVTS_a16,     opRSTS_a16,     opSMINT,        ILLEGAL,
+
+        /*32-bit data, 16-bit addr*/
+/*      00              01              02              03              04              05              06              07              08              09              0a              0b              0c              0d              0e              0f*/
+/*00*/  op0F00_a16,     op0F01_l_a16,   opLAR_l_a16,    opLSL_l_a16,    ILLEGAL,        ILLEGAL,        opCLTS,         ILLEGAL,        opINVD,         opWBINVD,       ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*10*/  opMOV_b_r_a16,  opMOV_l_r_a16,  opMOV_r_b_a16,  opMOV_r_l_a16,  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*20*/  opMOV_r_CRx_a16,opMOV_r_DRx_a16,opMOV_CRx_r_a16,opMOV_DRx_r_a16,opMOV_r_TRx_a16,ILLEGAL,        opMOV_TRx_r_a16,ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*30*/  opWRMSR,        opRDTSC,        opRDMSR,        ILLEGAL,        ILLEGAL,        ILLEGAL,        opRDSHR_a16,    opWRSHR_a16,    ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+
+/*40*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*50*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*60*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*70*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+
+/*80*/  opJO_l,         opJNO_l,        opJB_l,         opJNB_l,        opJE_l,         opJNE_l,        opJBE_l,        opJNBE_l,       opJS_l,         opJNS_l,        opJP_l,         opJNP_l,        opJL_l,         opJNL_l,        opJLE_l,        opJNLE_l,
+/*90*/  opSETO_a16,     opSETNO_a16,    opSETB_a16,     opSETNB_a16,    opSETE_a16,     opSETNE_a16,    opSETBE_a16,    opSETNBE_a16,   opSETS_a16,     opSETNS_a16,    opSETP_a16,     opSETNP_a16,    opSETL_a16,     opSETNL_a16,    opSETLE_a16,    opSETNLE_a16,
+/*a0*/  opPUSH_FS_l,    opPOP_FS_l,     opCPUID,        opBT_l_r_a16,   opSHLD_l_i_a16, opSHLD_l_CL_a16,ILLEGAL,        ILLEGAL,        opPUSH_GS_l,    opPOP_GS_l,     opRSM,          opBTS_l_r_a16,  opSHRD_l_i_a16, opSHRD_l_CL_a16,ILLEGAL,        opIMUL_l_l_a16,
+/*b0*/  opCMPXCHG_b_a16,opCMPXCHG_l_a16,opLSS_l_a16,    opBTR_l_r_a16,  opLFS_l_a16,    opLGS_l_a16,    opMOVZX_l_b_a16,opMOVZX_l_w_a16,ILLEGAL,        ILLEGAL,        opBA_l_a16,     opBTC_l_r_a16,  opBSF_l_a16,    opBSR_l_a16,    opMOVSX_l_b_a16,opMOVSX_l_w_a16,
+
+/*c0*/  opXADD_b_a16,   opXADD_l_a16,   ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        opCMPXCHG8B_a16,opBSWAP_EAX,    opBSWAP_ECX,    opBSWAP_EDX,    opBSWAP_EBX,    opBSWAP_ESP,    opBSWAP_EBP,    opBSWAP_ESI,    opBSWAP_EDI,
+/*d0*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*e0*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*f0*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+
+        /*16-bit data, 32-bit addr*/
+/*      00              01              02              03              04              05              06              07              08              09              0a              0b              0c              0d              0e              0f*/
+/*00*/  op0F00_a32,     op0F01_w_a32,   opLAR_w_a32,    opLSL_w_a32,    ILLEGAL,        ILLEGAL,        opCLTS,         ILLEGAL,        opINVD,         opWBINVD,       ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*10*/  opMOV_b_r_a32,  opMOV_w_r_a32,  opMOV_r_b_a32,  opMOV_r_w_a32,  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*20*/  opMOV_r_CRx_a32,opMOV_r_DRx_a32,opMOV_CRx_r_a32,opMOV_DRx_r_a32,opMOV_r_TRx_a32,ILLEGAL,        opMOV_TRx_r_a32,ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*30*/  opWRMSR,        opRDTSC,        opRDMSR,        ILLEGAL,        ILLEGAL,        ILLEGAL,        opRDSHR_a32,    opWRSHR_a32,    ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+
+/*40*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*50*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*60*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*70*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        opSVDC_a32,     opRSDC_a32,     opSVLDT_a32,    opRSLDT_a32,    opSVTS_a32,     opRSTS_a32,     opSMINT,        ILLEGAL,
+
+/*80*/  opJO_w,         opJNO_w,        opJB_w,         opJNB_w,        opJE_w,         opJNE_w,        opJBE_w,        opJNBE_w,       opJS_w,         opJNS_w,        opJP_w,         opJNP_w,        opJL_w,         opJNL_w,        opJLE_w,        opJNLE_w,
+/*90*/  opSETO_a32,     opSETNO_a32,    opSETB_a32,     opSETNB_a32,    opSETE_a32,     opSETNE_a32,    opSETBE_a32,    opSETNBE_a32,   opSETS_a32,     opSETNS_a32,    opSETP_a32,     opSETNP_a32,    opSETL_a32,     opSETNL_a32,    opSETLE_a32,    opSETNLE_a32,
+/*a0*/  opPUSH_FS_w,    opPOP_FS_w,     opCPUID,        opBT_w_r_a32,   opSHLD_w_i_a32, opSHLD_w_CL_a32,ILLEGAL,        ILLEGAL,        opPUSH_GS_w,    opPOP_GS_w,     opRSM,          opBTS_w_r_a32,  opSHRD_w_i_a32, opSHRD_w_CL_a32,ILLEGAL,        opIMUL_w_w_a32,
+/*b0*/  opCMPXCHG_b_a32,opCMPXCHG_w_a32,opLSS_w_a32,    opBTR_w_r_a32,  opLFS_w_a32,    opLGS_w_a32,    opMOVZX_w_b_a32,opMOVZX_w_w_a32,ILLEGAL,        ILLEGAL,        opBA_w_a32,     opBTC_w_r_a32,  opBSF_w_a32,    opBSR_w_a32,    opMOVSX_w_b_a32,ILLEGAL,
+
+/*c0*/  opXADD_b_a32,   opXADD_w_a32,   ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        opCMPXCHG8B_a32,opBSWAP_EAX,    opBSWAP_ECX,    opBSWAP_EDX,    opBSWAP_EBX,    opBSWAP_ESP,    opBSWAP_EBP,    opBSWAP_ESI,    opBSWAP_EDI,
+/*d0*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*e0*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*f0*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+
+        /*32-bit data, 32-bit addr*/
+/*      00              01              02              03              04              05              06              07              08              09              0a              0b              0c              0d              0e              0f*/
+/*00*/  op0F00_a32,     op0F01_l_a32,   opLAR_l_a32,    opLSL_l_a32,    ILLEGAL,        ILLEGAL,        opCLTS,         ILLEGAL,        opINVD,         opWBINVD,       ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*10*/  opMOV_b_r_a32,  opMOV_l_r_a32,  opMOV_r_b_a32,  opMOV_r_l_a32,  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*20*/  opMOV_r_CRx_a32,opMOV_r_DRx_a32,opMOV_CRx_r_a32,opMOV_DRx_r_a32,opMOV_r_TRx_a32,ILLEGAL,        opMOV_TRx_r_a32,ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*30*/  opWRMSR,        opRDTSC,        opRDMSR,        ILLEGAL,        ILLEGAL,        ILLEGAL,        opRDSHR_a32,    opWRSHR_a32,    ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
 
 /*40*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
 /*50*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
diff --git a/src/cpu/cpu.c b/src/cpu/cpu.c
index 0802fa3ac..2d22ed796 100644
--- a/src/cpu/cpu.c
+++ b/src/cpu/cpu.c
@@ -1462,22 +1462,16 @@ cpu_set(void)
             if (cpu_s->cpu_type == CPU_Cx6x86MX)
                 x86_setopcodes(ops_386, ops_c6x86mx_0f, dynarec_ops_386, dynarec_ops_c6x86mx_0f);
             else if (cpu_s->cpu_type == CPU_Cx6x86L)
-                x86_setopcodes(ops_386, ops_pentium_0f, dynarec_ops_386, dynarec_ops_pentium_0f);
+                x86_setopcodes(ops_386, ops_c6x86l_0f, dynarec_ops_386, dynarec_ops_c6x86l_0f);
             else
-                x86_setopcodes(ops_386, ops_c6x86mx_0f, dynarec_ops_386, dynarec_ops_c6x86mx_0f);
-#        if 0
                 x86_setopcodes(ops_386, ops_c6x86_0f, dynarec_ops_386, dynarec_ops_c6x86_0f);
-#        endif
 #    else
             if (cpu_s->cpu_type == CPU_Cx6x86MX)
                 x86_setopcodes(ops_386, ops_c6x86mx_0f);
             else if (cpu_s->cpu_type == CPU_Cx6x86L)
-                x86_setopcodes(ops_386, ops_pentium_0f);
+                x86_setopcodes(ops_386, ops_c6x86l_0f);
             else
-                x86_setopcodes(ops_386, ops_c6x86mx_0f);
-#        if 0
                 x86_setopcodes(ops_386, ops_c6x86_0f);
-#        endif
 #    endif /* USE_DYNAREC */
 
             timing_rr  = 1; /* register dest - register src */
diff --git a/src/cpu/cpu.h b/src/cpu/cpu.h
index 1d003ddc9..3e72db005 100644
--- a/src/cpu/cpu.h
+++ b/src/cpu/cpu.h
@@ -87,30 +87,32 @@ enum {
 enum {
     CPU_PKG_8088             = (1 << 0),
     CPU_PKG_8088_EUROPC      = (1 << 1),
-    CPU_PKG_8086             = (1 << 2),
-    CPU_PKG_8086_MAZOVIA     = (1 << 3),
-    CPU_PKG_188              = (1 << 4),
-    CPU_PKG_186              = (1 << 5),
-    CPU_PKG_286              = (1 << 6),
-    CPU_PKG_386SX            = (1 << 7),
-    CPU_PKG_386DX            = (1 << 8),
-    CPU_PKG_386DX_DESKPRO386 = (1 << 9),
-    CPU_PKG_M6117            = (1 << 10),
-    CPU_PKG_386SLC_IBM       = (1 << 11),
-    CPU_PKG_486SLC           = (1 << 12),
-    CPU_PKG_486SLC_IBM       = (1 << 13),
-    CPU_PKG_486BL            = (1 << 14),
-    CPU_PKG_486DLC           = (1 << 15),
-    CPU_PKG_SOCKET1          = (1 << 16),
-    CPU_PKG_SOCKET3          = (1 << 17),
-    CPU_PKG_SOCKET3_PC330    = (1 << 18),
-    CPU_PKG_STPC             = (1 << 19),
-    CPU_PKG_SOCKET4          = (1 << 20),
-    CPU_PKG_SOCKET5_7        = (1 << 21),
-    CPU_PKG_SOCKET8          = (1 << 22),
-    CPU_PKG_SLOT1            = (1 << 23),
-    CPU_PKG_SLOT2            = (1 << 24),
-    CPU_PKG_SOCKET370        = (1 << 25)
+    CPU_PKG_8088_VTECH       = (1 << 2),
+    CPU_PKG_8086             = (1 << 3),
+    CPU_PKG_8086_MAZOVIA     = (1 << 4),
+    CPU_PKG_8086_VTECH       = (1 << 5),
+    CPU_PKG_188              = (1 << 6),
+    CPU_PKG_186              = (1 << 7),
+    CPU_PKG_286              = (1 << 8),
+    CPU_PKG_386SX            = (1 << 9),
+    CPU_PKG_386DX            = (1 << 10),
+    CPU_PKG_386DX_DESKPRO386 = (1 << 11),
+    CPU_PKG_M6117            = (1 << 12),
+    CPU_PKG_386SLC_IBM       = (1 << 13),
+    CPU_PKG_486SLC           = (1 << 14),
+    CPU_PKG_486SLC_IBM       = (1 << 15),
+    CPU_PKG_486BL            = (1 << 16),
+    CPU_PKG_486DLC           = (1 << 17),
+    CPU_PKG_SOCKET1          = (1 << 18),
+    CPU_PKG_SOCKET3          = (1 << 19),
+    CPU_PKG_SOCKET3_PC330    = (1 << 20),
+    CPU_PKG_STPC             = (1 << 21),
+    CPU_PKG_SOCKET4          = (1 << 22),
+    CPU_PKG_SOCKET5_7        = (1 << 23),
+    CPU_PKG_SOCKET8          = (1 << 24),
+    CPU_PKG_SLOT1            = (1 << 25),
+    CPU_PKG_SLOT2            = (1 << 26),
+    CPU_PKG_SOCKET370        = (1 << 27)
 };
 
 #define CPU_SUPPORTS_DYNAREC 1
diff --git a/src/cpu/cpu_table.c b/src/cpu/cpu_table.c
index b2ef71075..104d56d05 100644
--- a/src/cpu/cpu_table.c
+++ b/src/cpu/cpu_table.c
@@ -262,6 +262,50 @@ const cpu_family_t cpu_families[] = {
             { .name = "", 0 }
         }
     },
+    {
+        .package       = CPU_PKG_8088_VTECH,
+        .manufacturer  = "Intel",
+        .name          = "8088",
+        .internal_name = "8088_vtech",
+        .cpus          = (const CPU[]) {
+            {
+                .name               = "4.77",
+                .cpu_type           = CPU_8088,
+                .fpus               = fpus_8088,
+                .rspeed             = 4772728,
+                .multi              = 1,
+                .voltage            = 5000,
+                .edx_reset          = 0,
+                .cpuid_model        = 0,
+                .cyrix_id           = 0,
+                .cpu_flags          = 0,
+                .mem_read_cycles    = 0,
+                .mem_write_cycles   = 0,
+                .cache_read_cycles  = 0,
+                .cache_write_cycles = 0,
+                .atclk_div          = 1
+            },
+
+            {
+                .name               = "10",
+                .cpu_type           = CPU_8088,
+                .fpus               = fpus_8088,
+                .rspeed             = 10000000,
+                .multi              = 1,
+                .voltage            = 5000,
+                .edx_reset          = 0,
+                .cpuid_model        = 0,
+                .cyrix_id           = 0,
+                .cpu_flags          = 0,
+                .mem_read_cycles    = 0,
+                .mem_write_cycles   = 0,
+                .cache_read_cycles  = 0,
+                .cache_write_cycles = 0,
+                .atclk_div          = 1
+            },
+            { .name = "", 0 }
+        }
+    },
     {
         .package       = CPU_PKG_8086,
         .manufacturer  = "Intel",
@@ -399,6 +443,49 @@ const cpu_family_t cpu_families[] = {
             { .name = "", 0 }
         }
     },
+    {
+        .package       = CPU_PKG_8086_VTECH,
+        .manufacturer  = "Intel",
+        .name          = "8086",
+        .internal_name = "8086_vtech",
+        .cpus          = (const CPU[]) {
+            {
+                .name               = "4.77",
+                .cpu_type           = CPU_8086,
+                .fpus               = fpus_8088,
+                .rspeed             = 4772728,
+                .multi              = 1,
+                .voltage            = 5000,
+                .edx_reset          = 0,
+                .cpuid_model        = 0,
+                .cyrix_id           = 0,
+                .cpu_flags          = 0,
+                .mem_read_cycles    = 0,
+                .mem_write_cycles   = 0,
+                .cache_read_cycles  = 0,
+                .cache_write_cycles = 0,
+                .atclk_div          = 1
+            },
+            {
+                .name               = "10",
+                .cpu_type           = CPU_8086,
+                .fpus               = fpus_8088,
+                .rspeed             = 10000000,
+                .multi              = 1,
+                .voltage            = 5000,
+                .edx_reset          = 0,
+                .cpuid_model        = 0,
+                .cyrix_id           = 0,
+                .cpu_flags          = 0,
+                .mem_read_cycles    = 0,
+                .mem_write_cycles   = 0,
+                .cache_read_cycles  = 0,
+                .cache_write_cycles = 0,
+                .atclk_div          = 1
+            },
+            { .name = "", 0 }
+        }
+    },
     {
         .package       = CPU_PKG_188,
         .manufacturer  = "Intel",
diff --git a/src/cpu/x86_ops.h b/src/cpu/x86_ops.h
index 3dfb3f917..7aec16855 100644
--- a/src/cpu/x86_ops.h
+++ b/src/cpu/x86_ops.h
@@ -92,6 +92,7 @@ extern const OpFn dynarec_ops_pentiummmx_0f[1024];
 
 #    ifdef USE_CYRIX_6X86
 extern const OpFn dynarec_ops_c6x86_0f[1024];
+extern const OpFn dynarec_ops_c6x86l_0f[1024];
 extern const OpFn dynarec_ops_c6x86mx_0f[1024];
 #    endif /* USE_CYRIX_6X86 */
 
@@ -234,6 +235,7 @@ extern const OpFn ops_pentiummmx_0f[1024];
 
 #ifdef USE_CYRIX_6X86
 extern const OpFn ops_c6x86_0f[1024];
+extern const OpFn ops_c6x86l_0f[1024];
 extern const OpFn ops_c6x86mx_0f[1024];
 #endif /* USE_CYRIX_6X86 */
 

From 5f3641ecbd7faf5caa77016392207a42ff9a2377 Mon Sep 17 00:00:00 2001
From: Cacodemon345 <wahil1976@outlook.com>
Date: Thu, 6 Mar 2025 00:54:28 +0600
Subject: [PATCH 40/58] Implement Cyrix EMMI extensions and 4 FPU instructions

PADDSIW, PSUBSIW, PMULHRW (named PMULHRWC in the code as recognized by some assemblers), PMULHRIW, PDISTIB, PMACHRIW, PAVEB, PMAGW, PMVZB, PMVNZB, PMVLZB, PMVGEZB, FTSTP, FRINT2, FRINEAR, FRICHOP are implemented for Cyrix 6x86MX. Cyrix 6x86(L) only has the last 4 instructions.
---
 src/cpu/386_ops.h            |   9 +-
 src/cpu/cpu.c                |  47 +-
 src/cpu/cpu.h                |  10 +
 src/cpu/x86.c                |   4 +
 src/cpu/x86_ops.h            |  18 +
 src/cpu/x86_ops_mmx_emmi.h   | 807 +++++++++++++++++++++++++++++++++++
 src/cpu/x87_ops.h            | 664 ++++++++++++++++++++++++++++
 src/cpu/x87_ops_misc.h       |  77 ++++
 src/cpu/x87_ops_sf_arith.h   |  81 ++++
 src/cpu/x87_ops_sf_compare.h |  26 ++
 10 files changed, 1723 insertions(+), 20 deletions(-)
 create mode 100644 src/cpu/x86_ops_mmx_emmi.h

diff --git a/src/cpu/386_ops.h b/src/cpu/386_ops.h
index 80a856482..e8300f134 100644
--- a/src/cpu/386_ops.h
+++ b/src/cpu/386_ops.h
@@ -206,6 +206,7 @@ extern void x386_dynarec_log(const char *fmt, ...);
 #    include "x86_ops_mmx_mov.h"
 #    include "x86_ops_mmx_pack.h"
 #    include "x86_ops_mmx_shift.h"
+#    include "x86_ops_mmx_emmi.h"
 #endif
 #include "x86_ops_mov.h"
 #ifdef OPS_286_386
@@ -1857,7 +1858,7 @@ const OpFn OP_TABLE(c6x86mx_0f)[1024] = {
 /*30*/  opWRMSR,        opRDTSC,        opRDMSR,        opRDPMC,        ILLEGAL,        ILLEGAL,        opRDSHR_a16,    opWRSHR_a16,    ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
 
 /*40*/  opCMOVO_w_a16,  opCMOVNO_w_a16, opCMOVB_w_a16,  opCMOVNB_w_a16, opCMOVE_w_a16,  opCMOVNE_w_a16, opCMOVBE_w_a16, opCMOVNBE_w_a16,opCMOVS_w_a16,  opCMOVNS_w_a16, opCMOVP_w_a16,  opCMOVNP_w_a16, opCMOVL_w_a16,  opCMOVNL_w_a16, opCMOVLE_w_a16, opCMOVNLE_w_a16,
-/*50*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*50*/  opPAVEB_a16,    opPADDSIW_a16,  opPMAGW_a16,    ILLEGAL,        opPDISTIB_a16,  opPSUBSIW_a16,  ILLEGAL,        ILLEGAL,        opPMVZB_a16,    opPMULHRWC_a16, opPMVNZB_a16,   opPMVLZB_a16,   opPMVGEZB_a16,  opPMULHRIW_a16, opPMACHRIW_a16, ILLEGAL,
 /*60*/  opPUNPCKLBW_a16,opPUNPCKLWD_a16,opPUNPCKLDQ_a16,opPACKSSWB_a16, opPCMPGTB_a16,  opPCMPGTW_a16,  opPCMPGTD_a16,  opPACKUSWB_a16, opPUNPCKHBW_a16,opPUNPCKHWD_a16,opPUNPCKHDQ_a16,opPACKSSDW_a16, ILLEGAL,        ILLEGAL,        opMOVD_l_mm_a16,opMOVQ_q_mm_a16,
 /*70*/  ILLEGAL,        opPSxxW_imm,    opPSxxD_imm,    opPSxxQ_imm,    opPCMPEQB_a16,  opPCMPEQW_a16,  opPCMPEQD_a16,  opEMMS,         opSVDC_a16,     opRSDC_a16,     opSVLDT_a16,    opRSLDT_a16,    opSVTS_a16,     opRSTS_a16,     opMOVD_mm_l_a16_cx,opMOVQ_mm_q_a16,
 
@@ -1879,7 +1880,7 @@ const OpFn OP_TABLE(c6x86mx_0f)[1024] = {
 /*30*/  opWRMSR,        opRDTSC,        opRDMSR,        opRDPMC,        ILLEGAL,        ILLEGAL,        opRDSHR_a16,    opWRSHR_a16,    ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
 
 /*40*/  opCMOVO_l_a16,  opCMOVNO_l_a16, opCMOVB_l_a16,  opCMOVNB_l_a16, opCMOVE_l_a16,  opCMOVNE_l_a16, opCMOVBE_l_a16, opCMOVNBE_l_a16,opCMOVS_l_a16,  opCMOVNS_l_a16, opCMOVP_l_a16,  opCMOVNP_l_a16, opCMOVL_l_a16,  opCMOVNL_l_a16, opCMOVLE_l_a16, opCMOVNLE_l_a16,
-/*50*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*50*/  opPAVEB_a16,    opPADDSIW_a16,  opPMAGW_a16,    ILLEGAL,        opPDISTIB_a16,  opPSUBSIW_a16,  ILLEGAL,        ILLEGAL,        opPMVZB_a16,    opPMULHRWC_a16, opPMVNZB_a16,   opPMVLZB_a16,   opPMVGEZB_a16,  opPMULHRIW_a16, opPMACHRIW_a16, ILLEGAL,
 /*60*/  opPUNPCKLBW_a16,opPUNPCKLWD_a16,opPUNPCKLDQ_a16,opPACKSSWB_a16, opPCMPGTB_a16,  opPCMPGTW_a16,  opPCMPGTD_a16,  opPACKUSWB_a16, opPUNPCKHBW_a16,opPUNPCKHWD_a16,opPUNPCKHDQ_a16,opPACKSSDW_a16, ILLEGAL,        ILLEGAL,        opMOVD_l_mm_a16,opMOVQ_q_mm_a16,
 /*70*/  ILLEGAL,        opPSxxW_imm,    opPSxxD_imm,    opPSxxQ_imm,    opPCMPEQB_a16,  opPCMPEQW_a16,  opPCMPEQD_a16,  opEMMS,         opSVDC_a16,     opRSDC_a16,     opSVLDT_a16,    opRSLDT_a16,    opSVTS_a16,     opRSTS_a16,     opMOVD_mm_l_a16_cx,opMOVQ_mm_q_a16,
 
@@ -1901,7 +1902,7 @@ const OpFn OP_TABLE(c6x86mx_0f)[1024] = {
 /*30*/  opWRMSR,        opRDTSC,        opRDMSR,        opRDPMC,        ILLEGAL,        ILLEGAL,        opRDSHR_a32,    opWRSHR_a32,    ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
 
 /*40*/  opCMOVO_w_a32,  opCMOVNO_w_a32, opCMOVB_w_a32,  opCMOVNB_w_a32, opCMOVE_w_a32,  opCMOVNE_w_a32, opCMOVBE_w_a32, opCMOVNBE_w_a32,opCMOVS_w_a32,  opCMOVNS_w_a32, opCMOVP_w_a32,  opCMOVNP_w_a32, opCMOVL_w_a32,  opCMOVNL_w_a32, opCMOVLE_w_a32, opCMOVNLE_w_a32,
-/*50*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*50*/  opPAVEB_a32,    opPADDSIW_a32,  opPMAGW_a32,    ILLEGAL,        opPDISTIB_a32,  opPSUBSIW_a32,  ILLEGAL,        ILLEGAL,        opPMVZB_a32,    opPMULHRWC_a32, opPMVNZB_a32,   opPMVLZB_a32,   opPMVGEZB_a32,  opPMULHRIW_a32, opPMACHRIW_a32, ILLEGAL,
 /*60*/  opPUNPCKLBW_a32,opPUNPCKLWD_a32,opPUNPCKLDQ_a32,opPACKSSWB_a32, opPCMPGTB_a32,  opPCMPGTW_a32,  opPCMPGTD_a32,  opPACKUSWB_a32, opPUNPCKHBW_a32,opPUNPCKHWD_a32,opPUNPCKHDQ_a32,opPACKSSDW_a32, ILLEGAL,        ILLEGAL,        opMOVD_l_mm_a32,opMOVQ_q_mm_a32,
 /*70*/  ILLEGAL,        opPSxxW_imm,    opPSxxD_imm,    opPSxxQ_imm,    opPCMPEQB_a32,  opPCMPEQW_a32,  opPCMPEQD_a32,  opEMMS,         opSVDC_a32,     opRSDC_a32,     opSVLDT_a32,    opRSLDT_a32,    opSVTS_a32,     opRSTS_a32,     opMOVD_mm_l_a32_cx,opMOVQ_mm_q_a32,
 
@@ -1923,7 +1924,7 @@ const OpFn OP_TABLE(c6x86mx_0f)[1024] = {
 /*30*/  opWRMSR,        opRDTSC,        opRDMSR,        opRDPMC,        ILLEGAL,        ILLEGAL,        opRDSHR_a32,    opWRSHR_a32,    ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
 
 /*40*/  opCMOVO_l_a32,  opCMOVNO_l_a32, opCMOVB_l_a32,  opCMOVNB_l_a32, opCMOVE_l_a32,  opCMOVNE_l_a32, opCMOVBE_l_a32, opCMOVNBE_l_a32,opCMOVS_l_a32,  opCMOVNS_l_a32, opCMOVP_l_a32,  opCMOVNP_l_a32, opCMOVL_l_a32,  opCMOVNL_l_a32, opCMOVLE_l_a32, opCMOVNLE_l_a32,
-/*50*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
+/*50*/  opPAVEB_a32,    opPADDSIW_a32,  opPMAGW_a32,    ILLEGAL,        opPDISTIB_a32,  opPSUBSIW_a32,  ILLEGAL,        ILLEGAL,        opPMVZB_a32,    opPMULHRWC_a32, opPMVNZB_a32,   opPMVLZB_a32,   opPMVGEZB_a32,  opPMULHRIW_a32, opPMACHRIW_a32, ILLEGAL,
 /*60*/  opPUNPCKLBW_a32,opPUNPCKLWD_a32,opPUNPCKLDQ_a32,opPACKSSWB_a32, opPCMPGTB_a32,  opPCMPGTW_a32,  opPCMPGTD_a32,  opPACKUSWB_a32, opPUNPCKHBW_a32,opPUNPCKHWD_a32,opPUNPCKHDQ_a32,opPACKSSDW_a32, ILLEGAL,        ILLEGAL,        opMOVD_l_mm_a32,opMOVQ_q_mm_a32,
 /*70*/  ILLEGAL,        opPSxxW_imm,    opPSxxD_imm,    opPSxxQ_imm,    opPCMPEQB_a32,  opPCMPEQW_a32,  opPCMPEQD_a32,  opEMMS,         opSVDC_a16,     opRSDC_a16,     opSVLDT_a16,    opRSLDT_a16,    opSVTS_a16,     opRSTS_a16,     opMOVD_mm_l_a32_cx,opMOVQ_mm_q_a32,
 
diff --git a/src/cpu/cpu.c b/src/cpu/cpu.c
index 2d22ed796..13a109d38 100644
--- a/src/cpu/cpu.c
+++ b/src/cpu/cpu.c
@@ -279,13 +279,14 @@ uint8_t do_translate2 = 0;
 
 void (*cpu_exec)(int32_t cycs);
 
-static uint8_t ccr0;
-static uint8_t ccr1;
-static uint8_t ccr2;
-static uint8_t ccr3;
-static uint8_t ccr4;
-static uint8_t ccr5;
-static uint8_t ccr6;
+uint8_t ccr0;
+uint8_t ccr1;
+uint8_t ccr2;
+uint8_t ccr3;
+uint8_t ccr4;
+uint8_t ccr5;
+uint8_t ccr6;
+uint8_t ccr7;
 
 static int cyrix_addr;
 
@@ -557,7 +558,8 @@ cpu_set(void)
         cpu_busspeed = cpu_s->rspeed;
     cpu_multi  = (int) ceil(cpu_s->multi);
     cpu_dmulti = cpu_s->multi;
-    ccr0 = ccr1 = ccr2 = ccr3 = ccr4 = ccr5 = ccr6 = 0;
+    ccr0 = ccr1 = ccr2 = ccr3 = ccr4 = ccr5 = ccr6 = ccr7 = 0;
+    ccr4 = 0x85;
 
     cpu_update_waitstates();
 
@@ -1442,19 +1444,27 @@ cpu_set(void)
                 }
 #    endif /* USE_DYNAREC */
                 if (fpu_softfloat) {
+                    x86_opcodes_d9_a16 = ops_sf_fpu_cyrix_d9_a16;
+                    x86_opcodes_d9_a32 = ops_sf_fpu_cyrix_d9_a32;
                     x86_opcodes_da_a16 = ops_sf_fpu_686_da_a16;
                     x86_opcodes_da_a32 = ops_sf_fpu_686_da_a32;
-                    x86_opcodes_db_a16 = ops_sf_fpu_686_db_a16;
-                    x86_opcodes_db_a32 = ops_sf_fpu_686_db_a32;
-                    x86_opcodes_df_a16 = ops_sf_fpu_686_df_a16;
-                    x86_opcodes_df_a32 = ops_sf_fpu_686_df_a32;
+                    x86_opcodes_db_a16 = ops_sf_fpu_cyrix_686_db_a16;
+                    x86_opcodes_db_a32 = ops_sf_fpu_cyrix_686_db_a32;
+                    x86_opcodes_dd_a16 = ops_sf_fpu_cyrix_dd_a16;
+                    x86_opcodes_dd_a32 = ops_sf_fpu_cyrix_dd_a32;
+                    x86_opcodes_df_a16 = ops_sf_fpu_cyrix_686_df_a16;
+                    x86_opcodes_df_a32 = ops_sf_fpu_cyrix_686_df_a32;
                 } else {
+                    x86_opcodes_d9_a16 = ops_fpu_cyrix_d9_a16;
+                    x86_opcodes_d9_a32 = ops_fpu_cyrix_d9_a32;
                     x86_opcodes_da_a16 = ops_fpu_686_da_a16;
                     x86_opcodes_da_a32 = ops_fpu_686_da_a32;
-                    x86_opcodes_db_a16 = ops_fpu_686_db_a16;
-                    x86_opcodes_db_a32 = ops_fpu_686_db_a32;
-                    x86_opcodes_df_a16 = ops_fpu_686_df_a16;
-                    x86_opcodes_df_a32 = ops_fpu_686_df_a32;
+                    x86_opcodes_db_a16 = ops_fpu_cyrix_686_db_a16;
+                    x86_opcodes_db_a32 = ops_fpu_cyrix_686_db_a32;
+                    x86_opcodes_dd_a16 = ops_fpu_cyrix_dd_a16;
+                    x86_opcodes_dd_a32 = ops_fpu_cyrix_dd_a32;
+                    x86_opcodes_df_a16 = ops_fpu_cyrix_686_df_a16;
+                    x86_opcodes_df_a32 = ops_fpu_cyrix_686_df_a32;
                 }
             }
 
@@ -4291,6 +4301,9 @@ cpu_write(uint16_t addr, uint8_t val, UNUSED(void *priv))
                 if ((ccr3 & 0xf0) == 0x10)
                     ccr6 = val;
                 break;
+            case 0xeb: /* CCR7 */
+                ccr7 = val & 5;
+                break;
         }
 }
 
@@ -4319,6 +4332,8 @@ cpu_read(uint16_t addr, UNUSED(void *priv))
                 return ((ccr3 & 0xf0) == 0x10) ? ccr5 : 0xff;
             case 0xea:
                 return ((ccr3 & 0xf0) == 0x10) ? ccr6 : 0xff;
+            case 0xeb:
+                return ccr7;
             case 0xfe:
                 return cpu_s->cyrix_id & 0xff;
             case 0xff:
diff --git a/src/cpu/cpu.h b/src/cpu/cpu.h
index 3e72db005..8324c543e 100644
--- a/src/cpu/cpu.h
+++ b/src/cpu/cpu.h
@@ -589,6 +589,16 @@ extern uint32_t _tr[8];
 extern uint32_t cache_index;
 extern uint8_t  _cache[2048];
 
+/* For the Cyrix 6x86(MX) */
+extern uint8_t ccr0;
+extern uint8_t ccr1;
+extern uint8_t ccr2;
+extern uint8_t ccr3;
+extern uint8_t ccr4;
+extern uint8_t ccr5;
+extern uint8_t ccr6;
+extern uint8_t ccr7;
+
 /*Segments -
   _cs,_ds,_es,_ss are the segment structures
   CS,DS,ES,SS is the 16-bit data
diff --git a/src/cpu/x86.c b/src/cpu/x86.c
index 6f15560f3..f1f5004d9 100644
--- a/src/cpu/x86.c
+++ b/src/cpu/x86.c
@@ -272,6 +272,10 @@ reset_common(int hard)
     msr.fcr        = (1 << 8) | (1 << 9) | (1 << 12) | (1 << 16) | (1 << 19) | (1 << 21);
     msw            = 0;
     new_ne         = 0;
+
+    ccr0 = ccr1 = ccr2 = ccr3 = ccr4 = ccr5 = ccr6 = ccr7 = 0;
+    ccr4 = 0x85;
+
     if (hascache)
         cr0 = 1 << 30;
     else
diff --git a/src/cpu/x86_ops.h b/src/cpu/x86_ops.h
index 7aec16855..6d93fb435 100644
--- a/src/cpu/x86_ops.h
+++ b/src/cpu/x86_ops.h
@@ -265,14 +265,20 @@ extern const OpFn ops_sf_fpu_d8_a16[32];
 extern const OpFn ops_sf_fpu_d8_a32[32];
 extern const OpFn ops_sf_fpu_d9_a16[256];
 extern const OpFn ops_sf_fpu_d9_a32[256];
+extern const OpFn ops_sf_fpu_cyrix_d9_a16[256];
+extern const OpFn ops_sf_fpu_cyrix_d9_a32[256];
 extern const OpFn ops_sf_fpu_da_a16[256];
 extern const OpFn ops_sf_fpu_da_a32[256];
 extern const OpFn ops_sf_fpu_db_a16[256];
 extern const OpFn ops_sf_fpu_db_a32[256];
+extern const OpFn ops_sf_fpu_cyrix_686_db_a16[256];
+extern const OpFn ops_sf_fpu_cyrix_686_db_a32[256];
 extern const OpFn ops_sf_fpu_dc_a16[32];
 extern const OpFn ops_sf_fpu_dc_a32[32];
 extern const OpFn ops_sf_fpu_dd_a16[256];
 extern const OpFn ops_sf_fpu_dd_a32[256];
+extern const OpFn ops_sf_fpu_cyrix_dd_a16[256];
+extern const OpFn ops_sf_fpu_cyrix_dd_a32[256];
 extern const OpFn ops_sf_fpu_de_a16[256];
 extern const OpFn ops_sf_fpu_de_a32[256];
 extern const OpFn ops_sf_fpu_df_a16[256];
@@ -297,6 +303,8 @@ extern const OpFn ops_fpu_d8_a16[32];
 extern const OpFn ops_fpu_d8_a32[32];
 extern const OpFn ops_fpu_d9_a16[256];
 extern const OpFn ops_fpu_d9_a32[256];
+extern const OpFn ops_fpu_cyrix_d9_a16[256];
+extern const OpFn ops_fpu_cyrix_d9_a32[256];
 extern const OpFn ops_fpu_da_a16[256];
 extern const OpFn ops_fpu_da_a32[256];
 extern const OpFn ops_fpu_db_a16[256];
@@ -305,6 +313,8 @@ extern const OpFn ops_fpu_dc_a16[32];
 extern const OpFn ops_fpu_dc_a32[32];
 extern const OpFn ops_fpu_dd_a16[256];
 extern const OpFn ops_fpu_dd_a32[256];
+extern const OpFn ops_fpu_cyrix_dd_a16[256];
+extern const OpFn ops_fpu_cyrix_dd_a32[256];
 extern const OpFn ops_fpu_de_a16[256];
 extern const OpFn ops_fpu_de_a32[256];
 extern const OpFn ops_fpu_df_a16[256];
@@ -316,15 +326,23 @@ extern const OpFn ops_sf_fpu_686_da_a16[256];
 extern const OpFn ops_sf_fpu_686_da_a32[256];
 extern const OpFn ops_sf_fpu_686_db_a16[256];
 extern const OpFn ops_sf_fpu_686_db_a32[256];
+extern const OpFn ops_sf_fpu_cyrix_686_db_a16[256];
+extern const OpFn ops_sf_fpu_cyrix_686_db_a32[256];
 extern const OpFn ops_sf_fpu_686_df_a16[256];
 extern const OpFn ops_sf_fpu_686_df_a32[256];
+extern const OpFn ops_sf_fpu_cyrix_686_df_a16[256];
+extern const OpFn ops_sf_fpu_cyrix_686_df_a32[256];
 
 extern const OpFn ops_fpu_686_da_a16[256];
 extern const OpFn ops_fpu_686_da_a32[256];
 extern const OpFn ops_fpu_686_db_a16[256];
 extern const OpFn ops_fpu_686_db_a32[256];
+extern const OpFn ops_fpu_cyrix_686_db_a16[256];
+extern const OpFn ops_fpu_cyrix_686_db_a32[256];
 extern const OpFn ops_fpu_686_df_a16[256];
 extern const OpFn ops_fpu_686_df_a32[256];
+extern const OpFn ops_fpu_cyrix_686_df_a16[256];
+extern const OpFn ops_fpu_cyrix_686_df_a32[256];
 
 extern const OpFn ops_REPE[1024];
 extern const OpFn ops_REPNE[1024];
diff --git a/src/cpu/x86_ops_mmx_emmi.h b/src/cpu/x86_ops_mmx_emmi.h
new file mode 100644
index 000000000..4c994be0f
--- /dev/null
+++ b/src/cpu/x86_ops_mmx_emmi.h
@@ -0,0 +1,807 @@
+/* Notes: "src2" means "first operand" */
+
+static int
+opPMULHRWC_a16(UNUSED(uint32_t fetchdat))
+{
+    MMX_REG  src;
+    MMX_REG *dst;
+
+    if (!(ccr7 & 1)) {
+        x86illegal();
+        return 1;
+    }
+
+    MMX_ENTER();
+
+    fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
+    MMX_GETSRC();
+
+    dst->w[0] = ((int32_t) (dst->sw[0] * (int32_t) src.sw[0]) + 0x4000) >> 15;
+    dst->w[1] = ((int32_t) (dst->sw[1] * (int32_t) src.sw[1]) + 0x4000) >> 15;
+    dst->w[2] = ((int32_t) (dst->sw[2] * (int32_t) src.sw[2]) + 0x4000) >> 15;
+    dst->w[3] = ((int32_t) (dst->sw[3] * (int32_t) src.sw[3]) + 0x4000) >> 15;
+
+    MMX_SETEXP(cpu_reg);
+
+    return 0;
+}
+
+static int
+opPMULHRWC_a32(UNUSED(uint32_t fetchdat))
+{
+    MMX_REG  src;
+    MMX_REG *dst;
+
+    if (!(ccr7 & 1)) {
+        x86illegal();
+        return 1;
+    }
+
+    MMX_ENTER();
+
+    fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
+    MMX_GETSRC();
+
+    dst->w[0] = ((int32_t) (dst->sw[0] * (int32_t) src.sw[0]) + 0x4000) >> 15;
+    dst->w[1] = ((int32_t) (dst->sw[1] * (int32_t) src.sw[1]) + 0x4000) >> 15;
+    dst->w[2] = ((int32_t) (dst->sw[2] * (int32_t) src.sw[2]) + 0x4000) >> 15;
+    dst->w[3] = ((int32_t) (dst->sw[3] * (int32_t) src.sw[3]) + 0x4000) >> 15;
+
+    MMX_SETEXP(cpu_reg);
+
+    return 0;
+}
+
+static int
+opPMULHRIW_a16(uint32_t fetchdat)
+{
+    MMX_REG  src, src2;
+    MMX_REG *dst;
+
+    if (!(ccr7 & 1)) {
+        x86illegal();
+        return 1;
+    }
+
+    MMX_ENTER();
+
+    fetch_ea_16(fetchdat);
+
+    src2 = MMX_GETREG(cpu_reg);
+    dst  = MMX_GETREGP(cpu_reg ^ 1);
+
+    MMX_GETSRC();
+
+    dst->w[0] = (((int32_t) src2.sw[0] * (int32_t) src.sw[0]) + 0x4000) >> 15;
+    dst->w[1] = (((int32_t) src2.sw[1] * (int32_t) src.sw[1]) + 0x4000) >> 15;
+    dst->w[2] = (((int32_t) src2.sw[2] * (int32_t) src.sw[2]) + 0x4000) >> 15;
+    dst->w[3] = (((int32_t) src2.sw[3] * (int32_t) src.sw[3]) + 0x4000) >> 15;
+    
+    MMX_SETEXP(cpu_reg ^ 1);
+    return 0;
+}
+
+static int
+opPMULHRIW_a32(uint32_t fetchdat)
+{
+    MMX_REG  src, src2;
+    MMX_REG *dst;
+
+    if (!(ccr7 & 1)) {
+        x86illegal();
+        return 1;
+    }
+
+    MMX_ENTER();
+
+    fetch_ea_32(fetchdat);
+
+    src2 = MMX_GETREG(cpu_reg);
+    dst  = MMX_GETREGP(cpu_reg ^ 1);
+
+    MMX_GETSRC();
+
+    dst->w[0] = (((int32_t) src2.sw[0] * (int32_t) src.sw[0]) + 0x4000) >> 15;
+    dst->w[1] = (((int32_t) src2.sw[1] * (int32_t) src.sw[1]) + 0x4000) >> 15;
+    dst->w[2] = (((int32_t) src2.sw[2] * (int32_t) src.sw[2]) + 0x4000) >> 15;
+    dst->w[3] = (((int32_t) src2.sw[3] * (int32_t) src.sw[3]) + 0x4000) >> 15;
+    
+    MMX_SETEXP(cpu_reg ^ 1);
+    return 0;
+}
+
+static int
+opPDISTIB_a16(uint32_t fetchdat)
+{
+    MMX_REG  src, src2;
+    MMX_REG *dst;
+
+    if (!(ccr7 & 1)) {
+        x86illegal();
+        return 1;
+    }
+
+    MMX_ENTER();
+
+    fetch_ea_16(fetchdat);
+
+    if (cpu_mod == 3) {
+        x86illegal();
+        return 1;
+    }
+
+    src2 = MMX_GETREG(cpu_reg);
+    dst  = MMX_GETREGP(cpu_reg ^ 1);
+
+    MMX_GETSRC();
+
+    dst->b[0] = USATB(dst->b[0] + abs(src2.b[0] - src.b[0]));
+    dst->b[1] = USATB(dst->b[1] + abs(src2.b[1] - src.b[1]));
+    dst->b[2] = USATB(dst->b[2] + abs(src2.b[2] - src.b[2]));
+    dst->b[3] = USATB(dst->b[3] + abs(src2.b[3] - src.b[3]));
+    dst->b[4] = USATB(dst->b[4] + abs(src2.b[4] - src.b[4]));
+    dst->b[5] = USATB(dst->b[5] + abs(src2.b[5] - src.b[5]));
+    dst->b[6] = USATB(dst->b[6] + abs(src2.b[6] - src.b[6]));
+    dst->b[7] = USATB(dst->b[7] + abs(src2.b[7] - src.b[7]));
+
+    MMX_SETEXP(cpu_reg ^ 1);
+    return 0;
+}
+
+static int
+opPDISTIB_a32(uint32_t fetchdat)
+{
+    MMX_REG  src, src2;
+    MMX_REG *dst;
+
+    if (!(ccr7 & 1)) {
+        x86illegal();
+        return 1;
+    }
+
+    MMX_ENTER();
+
+    fetch_ea_32(fetchdat);
+
+    if (cpu_mod == 3) {
+        x86illegal();
+        return 1;
+    }
+
+    src2 = MMX_GETREG(cpu_reg);
+    dst  = MMX_GETREGP(cpu_reg ^ 1);
+
+    MMX_GETSRC();
+
+    dst->b[0] = USATB(dst->b[0] + abs(src2.b[0] - src.b[0]));
+    dst->b[1] = USATB(dst->b[1] + abs(src2.b[1] - src.b[1]));
+    dst->b[2] = USATB(dst->b[2] + abs(src2.b[2] - src.b[2]));
+    dst->b[3] = USATB(dst->b[3] + abs(src2.b[3] - src.b[3]));
+    dst->b[4] = USATB(dst->b[4] + abs(src2.b[4] - src.b[4]));
+    dst->b[5] = USATB(dst->b[5] + abs(src2.b[5] - src.b[5]));
+    dst->b[6] = USATB(dst->b[6] + abs(src2.b[6] - src.b[6]));
+    dst->b[7] = USATB(dst->b[7] + abs(src2.b[7] - src.b[7]));
+
+    MMX_SETEXP(cpu_reg ^ 1);
+    return 0;
+}
+
+static int
+opPMACHRIW_a16(uint32_t fetchdat)
+{
+    MMX_REG  src, src2;
+    MMX_REG *dst;
+
+    if (!(ccr7 & 1)) {
+        x86illegal();
+        return 1;
+    }
+
+    MMX_ENTER();
+
+    fetch_ea_16(fetchdat);
+
+    if (cpu_mod == 3) {
+        x86illegal();
+        return 1;
+    }
+
+    src2 = MMX_GETREG(cpu_reg);
+    dst  = MMX_GETREGP(cpu_reg ^ 1);
+
+    MMX_GETSRC();
+
+    dst->w[0] += (((int32_t) src2.sw[0] * (int32_t) src.sw[0]) + 0x4000) >> 15;
+    dst->w[1] += (((int32_t) src2.sw[1] * (int32_t) src.sw[1]) + 0x4000) >> 15;
+    dst->w[2] += (((int32_t) src2.sw[2] * (int32_t) src.sw[2]) + 0x4000) >> 15;
+    dst->w[3] += (((int32_t) src2.sw[3] * (int32_t) src.sw[3]) + 0x4000) >> 15;
+    
+    MMX_SETEXP(cpu_reg ^ 1);
+    return 0;
+}
+
+static int
+opPMACHRIW_a32(uint32_t fetchdat)
+{
+    MMX_REG  src, src2;
+    MMX_REG *dst;
+
+    if (!(ccr7 & 1)) {
+        x86illegal();
+        return 1;
+    }
+
+    MMX_ENTER();
+
+    fetch_ea_32(fetchdat);
+
+    if (cpu_mod == 3) {
+        x86illegal();
+        return 1;
+    }
+
+    src2 = MMX_GETREG(cpu_reg);
+    dst  = MMX_GETREGP(cpu_reg ^ 1);
+
+    MMX_GETSRC();
+
+    dst->w[0] += (((int32_t) src2.sw[0] * (int32_t) src.sw[0]) + 0x4000) >> 15;
+    dst->w[1] += (((int32_t) src2.sw[1] * (int32_t) src.sw[1]) + 0x4000) >> 15;
+    dst->w[2] += (((int32_t) src2.sw[2] * (int32_t) src.sw[2]) + 0x4000) >> 15;
+    dst->w[3] += (((int32_t) src2.sw[3] * (int32_t) src.sw[3]) + 0x4000) >> 15;
+    
+    MMX_SETEXP(cpu_reg ^ 1);
+    return 0;
+}
+
+static int
+opPADDSIW_a16(uint32_t fetchdat)
+{
+    MMX_REG  src, src2;
+    MMX_REG *dst;
+
+    if (!(ccr7 & 1)) {
+        x86illegal();
+        return 1;
+    }
+
+    MMX_ENTER();
+
+    fetch_ea_16(fetchdat);
+
+    src2 = MMX_GETREG(cpu_reg);
+    dst  = MMX_GETREGP(cpu_reg ^ 1);
+
+    MMX_GETSRC();
+
+    dst->sw[0] = SSATW(src2.sw[0] + src.sw[0]);
+    dst->sw[1] = SSATW(src2.sw[1] + src.sw[1]);
+    dst->sw[2] = SSATW(src2.sw[2] + src.sw[2]);
+    dst->sw[3] = SSATW(src2.sw[3] + src.sw[3]);
+
+    MMX_SETEXP(cpu_reg ^ 1);
+
+    return 0;
+}
+
+static int
+opPADDSIW_a32(uint32_t fetchdat)
+{
+    MMX_REG  src, src2;
+    MMX_REG *dst;
+
+    if (!(ccr7 & 1)) {
+        x86illegal();
+        return 1;
+    }
+
+    MMX_ENTER();
+
+    fetch_ea_32(fetchdat);
+
+    src2 = MMX_GETREG(cpu_reg);
+    dst  = MMX_GETREGP(cpu_reg ^ 1);
+
+    MMX_GETSRC();
+
+    dst->sw[0] = SSATW(src2.sw[0] + src.sw[0]);
+    dst->sw[1] = SSATW(src2.sw[1] + src.sw[1]);
+    dst->sw[2] = SSATW(src2.sw[2] + src.sw[2]);
+    dst->sw[3] = SSATW(src2.sw[3] + src.sw[3]);
+
+    MMX_SETEXP(cpu_reg ^ 1);
+
+    return 0;
+}
+
+static int
+opPSUBSIW_a16(uint32_t fetchdat)
+{
+    MMX_REG  src, src2;
+    MMX_REG *dst;
+
+    if (!(ccr7 & 1)) {
+        x86illegal();
+        return 1;
+    }
+
+    MMX_ENTER();
+
+    fetch_ea_16(fetchdat);
+
+    src2 = MMX_GETREG(cpu_reg);
+    dst  = MMX_GETREGP(cpu_reg ^ 1);
+
+    MMX_GETSRC();
+
+    dst->sw[0] = SSATW(src2.sw[0] - src.sw[0]);
+    dst->sw[1] = SSATW(src2.sw[1] - src.sw[1]);
+    dst->sw[2] = SSATW(src2.sw[2] - src.sw[2]);
+    dst->sw[3] = SSATW(src2.sw[3] - src.sw[3]);
+
+    MMX_SETEXP(cpu_reg);
+
+    return 0;
+}
+
+static int
+opPSUBSIW_a32(uint32_t fetchdat)
+{
+    MMX_REG  src, src2;
+    MMX_REG *dst;
+
+    if (!(ccr7 & 1)) {
+        x86illegal();
+        return 1;
+    }
+
+    MMX_ENTER();
+
+    fetch_ea_32(fetchdat);
+
+    src2 = MMX_GETREG(cpu_reg);
+    dst  = MMX_GETREGP(cpu_reg ^ 1);
+
+    MMX_GETSRC();
+
+    dst->sw[0] = SSATW(src2.sw[0] - src.sw[0]);
+    dst->sw[1] = SSATW(src2.sw[1] - src.sw[1]);
+    dst->sw[2] = SSATW(src2.sw[2] - src.sw[2]);
+    dst->sw[3] = SSATW(src2.sw[3] - src.sw[3]);
+
+    MMX_SETEXP(cpu_reg);
+
+    return 0;
+}
+
+static int
+opPAVEB_a16(uint32_t fetchdat)
+{
+    MMX_REG  src;
+    MMX_REG *dst;
+
+    if (!(ccr7 & 1)) {
+        x86illegal();
+        return 1;
+    }
+
+    MMX_ENTER();
+
+    fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
+    MMX_GETSRC();
+
+    dst->b[0] = (dst->b[0] + src.b[0]) >> 1;
+    dst->b[1] = (dst->b[1] + src.b[1]) >> 1;
+    dst->b[2] = (dst->b[2] + src.b[2]) >> 1;
+    dst->b[3] = (dst->b[3] + src.b[3]) >> 1;
+    dst->b[4] = (dst->b[4] + src.b[4]) >> 1;
+    dst->b[5] = (dst->b[5] + src.b[5]) >> 1;
+    dst->b[6] = (dst->b[6] + src.b[6]) >> 1;
+    dst->b[7] = (dst->b[7] + src.b[7]) >> 1;
+
+    MMX_SETEXP(cpu_reg);
+    return 0;
+}
+
+static int
+opPAVEB_a32(uint32_t fetchdat)
+{
+    MMX_REG  src;
+    MMX_REG *dst;
+
+    if (!(ccr7 & 1)) {
+        x86illegal();
+        return 1;
+    }
+
+    MMX_ENTER();
+
+    fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
+    MMX_GETSRC();
+
+    dst->b[0] = (dst->b[0] + src.b[0]) >> 1;
+    dst->b[1] = (dst->b[1] + src.b[1]) >> 1;
+    dst->b[2] = (dst->b[2] + src.b[2]) >> 1;
+    dst->b[3] = (dst->b[3] + src.b[3]) >> 1;
+    dst->b[4] = (dst->b[4] + src.b[4]) >> 1;
+    dst->b[5] = (dst->b[5] + src.b[5]) >> 1;
+    dst->b[6] = (dst->b[6] + src.b[6]) >> 1;
+    dst->b[7] = (dst->b[7] + src.b[7]) >> 1;
+
+    MMX_SETEXP(cpu_reg);
+    return 0;
+}
+
+static int
+opPMAGW_a16(uint32_t fetchdat)
+{
+    MMX_REG  src;
+    MMX_REG *dst;
+
+    if (!(ccr7 & 1)) {
+        x86illegal();
+        return 1;
+    }
+
+    MMX_ENTER();
+
+    fetch_ea_16(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
+    MMX_GETSRC();
+
+    if (abs(src.sw[0]) > abs(dst->sw[0])) dst->sw[0] = src.sw[0];
+    if (abs(src.sw[1]) > abs(dst->sw[1])) dst->sw[1] = src.sw[1];
+    if (abs(src.sw[2]) > abs(dst->sw[2])) dst->sw[2] = src.sw[2];
+    if (abs(src.sw[3]) > abs(dst->sw[3])) dst->sw[3] = src.sw[3];
+
+    MMX_SETEXP(cpu_reg);
+
+    return 0;
+}
+
+static int
+opPMAGW_a32(uint32_t fetchdat)
+{
+    MMX_REG  src;
+    MMX_REG *dst;
+
+    if (!(ccr7 & 1)) {
+        x86illegal();
+        return 1;
+    }
+
+    MMX_ENTER();
+
+    fetch_ea_32(fetchdat);
+
+    dst = MMX_GETREGP(cpu_reg);
+
+    MMX_GETSRC();
+
+    if (abs(src.sw[0]) > abs(dst->sw[0])) dst->sw[0] = src.sw[0];
+    if (abs(src.sw[1]) > abs(dst->sw[1])) dst->sw[1] = src.sw[1];
+    if (abs(src.sw[2]) > abs(dst->sw[2])) dst->sw[2] = src.sw[2];
+    if (abs(src.sw[3]) > abs(dst->sw[3])) dst->sw[3] = src.sw[3];
+
+    MMX_SETEXP(cpu_reg);
+
+    return 0;
+}
+
+static int
+opPMVZB_a16(uint32_t fetchdat)
+{
+    MMX_REG  src, src2;
+    MMX_REG *dst;
+
+    if (!(ccr7 & 1)) {
+        x86illegal();
+        return 1;
+    }
+
+    MMX_ENTER();
+
+    fetch_ea_16(fetchdat);
+
+    if (cpu_mod == 3) {
+        x86illegal();
+        return 1;
+    }
+
+    dst = MMX_GETREGP(cpu_reg);
+    src2 = MMX_GETREG(cpu_reg ^ 1);
+
+    MMX_GETSRC();
+
+    if (src2.b[0] == 0) dst->b[0] = src.b[0];
+    if (src2.b[1] == 0) dst->b[1] = src.b[1];
+    if (src2.b[2] == 0) dst->b[2] = src.b[2];
+    if (src2.b[3] == 0) dst->b[3] = src.b[3];
+    if (src2.b[4] == 0) dst->b[4] = src.b[4];
+    if (src2.b[5] == 0) dst->b[5] = src.b[5];
+    if (src2.b[6] == 0) dst->b[6] = src.b[6];
+    if (src2.b[7] == 0) dst->b[7] = src.b[7];
+
+    MMX_SETEXP(cpu_reg);
+    return 0;
+}
+
+static int
+opPMVZB_a32(uint32_t fetchdat)
+{
+    MMX_REG  src, src2;
+    MMX_REG *dst;
+
+    if (!(ccr7 & 1)) {
+        x86illegal();
+        return 1;
+    }
+
+    MMX_ENTER();
+
+    fetch_ea_32(fetchdat);
+
+    if (cpu_mod == 3) {
+        x86illegal();
+        return 1;
+    }
+
+    dst = MMX_GETREGP(cpu_reg);
+    src2 = MMX_GETREG(cpu_reg ^ 1);
+
+    MMX_GETSRC();
+
+    if (src2.b[0] == 0) dst->b[0] = src.b[0];
+    if (src2.b[1] == 0) dst->b[1] = src.b[1];
+    if (src2.b[2] == 0) dst->b[2] = src.b[2];
+    if (src2.b[3] == 0) dst->b[3] = src.b[3];
+    if (src2.b[4] == 0) dst->b[4] = src.b[4];
+    if (src2.b[5] == 0) dst->b[5] = src.b[5];
+    if (src2.b[6] == 0) dst->b[6] = src.b[6];
+    if (src2.b[7] == 0) dst->b[7] = src.b[7];
+
+    MMX_SETEXP(cpu_reg);
+    return 0;
+}
+
+static int
+opPMVNZB_a16(uint32_t fetchdat)
+{
+    MMX_REG  src, src2;
+    MMX_REG *dst;
+
+    if (!(ccr7 & 1)) {
+        x86illegal();
+        return 1;
+    }
+
+    MMX_ENTER();
+
+    fetch_ea_16(fetchdat);
+
+    if (cpu_mod == 3) {
+        x86illegal();
+        return 1;
+    }
+
+    dst = MMX_GETREGP(cpu_reg);
+    src2 = MMX_GETREG(cpu_reg ^ 1);
+
+    MMX_GETSRC();
+
+    if (src2.b[0] != 0) dst->b[0] = src.b[0];
+    if (src2.b[1] != 0) dst->b[1] = src.b[1];
+    if (src2.b[2] != 0) dst->b[2] = src.b[2];
+    if (src2.b[3] != 0) dst->b[3] = src.b[3];
+    if (src2.b[4] != 0) dst->b[4] = src.b[4];
+    if (src2.b[5] != 0) dst->b[5] = src.b[5];
+    if (src2.b[6] != 0) dst->b[6] = src.b[6];
+    if (src2.b[7] != 0) dst->b[7] = src.b[7];
+
+    MMX_SETEXP(cpu_reg);
+    return 0;
+}
+
+static int
+opPMVNZB_a32(uint32_t fetchdat)
+{
+    MMX_REG  src, src2;
+    MMX_REG *dst;
+
+    if (!(ccr7 & 1)) {
+        x86illegal();
+        return 1;
+    }
+
+    MMX_ENTER();
+
+    fetch_ea_32(fetchdat);
+
+    if (cpu_mod == 3) {
+        x86illegal();
+        return 1;
+    }
+
+    dst = MMX_GETREGP(cpu_reg);
+    src2 = MMX_GETREG(cpu_reg ^ 1);
+
+    MMX_GETSRC();
+
+    if (src2.b[0] != 0) dst->b[0] = src.b[0];
+    if (src2.b[1] != 0) dst->b[1] = src.b[1];
+    if (src2.b[2] != 0) dst->b[2] = src.b[2];
+    if (src2.b[3] != 0) dst->b[3] = src.b[3];
+    if (src2.b[4] != 0) dst->b[4] = src.b[4];
+    if (src2.b[5] != 0) dst->b[5] = src.b[5];
+    if (src2.b[6] != 0) dst->b[6] = src.b[6];
+    if (src2.b[7] != 0) dst->b[7] = src.b[7];
+
+    MMX_SETEXP(cpu_reg);
+    return 0;
+}
+
+static int
+opPMVLZB_a16(uint32_t fetchdat)
+{
+    MMX_REG  src, src2;
+    MMX_REG *dst;
+
+    if (!(ccr7 & 1)) {
+        x86illegal();
+        return 1;
+    }
+
+    MMX_ENTER();
+
+    fetch_ea_16(fetchdat);
+
+    if (cpu_mod == 3) {
+        x86illegal();
+        return 1;
+    }
+
+    dst = MMX_GETREGP(cpu_reg);
+    src2 = MMX_GETREG(cpu_reg ^ 1);
+
+    MMX_GETSRC();
+
+    if (src2.sb[0] < 0) dst->b[0] = src.b[0];
+    if (src2.sb[1] < 0) dst->b[1] = src.b[1];
+    if (src2.sb[2] < 0) dst->b[2] = src.b[2];
+    if (src2.sb[3] < 0) dst->b[3] = src.b[3];
+    if (src2.sb[4] < 0) dst->b[4] = src.b[4];
+    if (src2.sb[5] < 0) dst->b[5] = src.b[5];
+    if (src2.sb[6] < 0) dst->b[6] = src.b[6];
+    if (src2.sb[7] < 0) dst->b[7] = src.b[7];
+
+    MMX_SETEXP(cpu_reg);
+    return 0;
+}
+
+static int
+opPMVLZB_a32(uint32_t fetchdat)
+{
+    MMX_REG  src, src2;
+    MMX_REG *dst;
+
+    if (!(ccr7 & 1)) {
+        x86illegal();
+        return 1;
+    }
+
+    MMX_ENTER();
+
+    fetch_ea_32(fetchdat);
+
+    if (cpu_mod == 3) {
+        x86illegal();
+        return 1;
+    }
+
+    dst = MMX_GETREGP(cpu_reg);
+    src2 = MMX_GETREG(cpu_reg ^ 1);
+
+    MMX_GETSRC();
+
+    if (src2.sb[0] < 0) dst->b[0] = src.b[0];
+    if (src2.sb[1] < 0) dst->b[1] = src.b[1];
+    if (src2.sb[2] < 0) dst->b[2] = src.b[2];
+    if (src2.sb[3] < 0) dst->b[3] = src.b[3];
+    if (src2.sb[4] < 0) dst->b[4] = src.b[4];
+    if (src2.sb[5] < 0) dst->b[5] = src.b[5];
+    if (src2.sb[6] < 0) dst->b[6] = src.b[6];
+    if (src2.sb[7] < 0) dst->b[7] = src.b[7];
+
+    MMX_SETEXP(cpu_reg);
+    return 0;
+}
+
+static int
+opPMVGEZB_a16(uint32_t fetchdat)
+{
+    MMX_REG  src, src2;
+    MMX_REG *dst;
+
+    if (!(ccr7 & 1)) {
+        x86illegal();
+        return 1;
+    }
+
+    MMX_ENTER();
+
+    fetch_ea_16(fetchdat);
+
+    if (cpu_mod == 3) {
+        x86illegal();
+        return 1;
+    }
+
+    dst = MMX_GETREGP(cpu_reg);
+    src2 = MMX_GETREG(cpu_reg ^ 1);
+
+    MMX_GETSRC();
+
+    if (src2.sb[0] >= 0) dst->b[0] = src.b[0];
+    if (src2.sb[1] >= 0) dst->b[1] = src.b[1];
+    if (src2.sb[2] >= 0) dst->b[2] = src.b[2];
+    if (src2.sb[3] >= 0) dst->b[3] = src.b[3];
+    if (src2.sb[4] >= 0) dst->b[4] = src.b[4];
+    if (src2.sb[5] >= 0) dst->b[5] = src.b[5];
+    if (src2.sb[6] >= 0) dst->b[6] = src.b[6];
+    if (src2.sb[7] >= 0) dst->b[7] = src.b[7];
+
+    MMX_SETEXP(cpu_reg);
+    return 0;
+}
+
+static int
+opPMVGEZB_a32(uint32_t fetchdat)
+{
+    MMX_REG  src, src2;
+    MMX_REG *dst;
+
+    if (!(ccr7 & 1)) {
+        x86illegal();
+        return 1;
+    }
+
+    MMX_ENTER();
+
+    fetch_ea_32(fetchdat);
+
+    if (cpu_mod == 3) {
+        x86illegal();
+        return 1;
+    }
+
+    dst = MMX_GETREGP(cpu_reg);
+    src2 = MMX_GETREG(cpu_reg ^ 1);
+
+    MMX_GETSRC();
+
+    if (src2.sb[0] >= 0) dst->b[0] = src.b[0];
+    if (src2.sb[1] >= 0) dst->b[1] = src.b[1];
+    if (src2.sb[2] >= 0) dst->b[2] = src.b[2];
+    if (src2.sb[3] >= 0) dst->b[3] = src.b[3];
+    if (src2.sb[4] >= 0) dst->b[4] = src.b[4];
+    if (src2.sb[5] >= 0) dst->b[5] = src.b[5];
+    if (src2.sb[6] >= 0) dst->b[6] = src.b[6];
+    if (src2.sb[7] >= 0) dst->b[7] = src.b[7];
+
+    MMX_SETEXP(cpu_reg);
+    return 0;
+}
\ No newline at end of file
diff --git a/src/cpu/x87_ops.h b/src/cpu/x87_ops.h
index f1362bf76..97f5415b9 100644
--- a/src/cpu/x87_ops.h
+++ b/src/cpu/x87_ops.h
@@ -241,6 +241,24 @@ x87_fround32_64(double b)
     return (int64_t) x87_fround32(b);
 }
 
+static __inline int64_t
+x87_fround_nearest(double b)
+{
+    double da, dc;
+    int64_t a, c;
+    
+    da = floor(b);
+    dc = floor(b + 1.0);
+    a = (int64_t) da;
+    c = (int64_t) dc;
+    if ((b - a) < (c - b))
+        return a;
+    else if ((b - a) > (c - b))
+        return c;
+    else
+        return (a & 1) ? c : a;
+}
+
 static __inline int64_t
 x87_fround(double b)
 {
@@ -1267,6 +1285,86 @@ const OpFn OP_TABLE(sf_fpu_d9_a32)[256] = {
     // clang-format on
 };
 
+const OpFn OP_TABLE(sf_fpu_cyrix_d9_a16)[256] = {
+    // clang-format off
+    /*0x00*/    sf_FLDs_a16,   sf_FLDs_a16,   sf_FLDs_a16,   sf_FLDs_a16,   sf_FLDs_a16,   sf_FLDs_a16,   sf_FLDs_a16,   sf_FLDs_a16,
+    /*0x08*/    ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+    /*0x10*/    sf_FSTs_a16,   sf_FSTs_a16,   sf_FSTs_a16,   sf_FSTs_a16,   sf_FSTs_a16,   sf_FSTs_a16,   sf_FSTs_a16,   sf_FSTs_a16,
+    /*0x18*/    sf_FSTPs_a16,  sf_FSTPs_a16,  sf_FSTPs_a16,  sf_FSTPs_a16,  sf_FSTPs_a16,  sf_FSTPs_a16,  sf_FSTPs_a16,  sf_FSTPs_a16,
+    /*0x20*/    sf_FLDENV_a16, sf_FLDENV_a16, sf_FLDENV_a16, sf_FLDENV_a16, sf_FLDENV_a16, sf_FLDENV_a16, sf_FLDENV_a16, sf_FLDENV_a16,
+    /*0x28*/    sf_FLDCW_a16,  sf_FLDCW_a16,  sf_FLDCW_a16,  sf_FLDCW_a16,  sf_FLDCW_a16,  sf_FLDCW_a16,  sf_FLDCW_a16,  sf_FLDCW_a16,
+    /*0x30*/    sf_FNSTENV_a16, sf_FNSTENV_a16, sf_FNSTENV_a16, sf_FNSTENV_a16, sf_FNSTENV_a16, sf_FNSTENV_a16, sf_FNSTENV_a16, sf_FNSTENV_a16,
+    /*0x38*/    sf_FNSTCW_a16,  sf_FNSTCW_a16,  sf_FNSTCW_a16,  sf_FNSTCW_a16,  sf_FNSTCW_a16,  sf_FNSTCW_a16,  sf_FNSTCW_a16,  sf_FNSTCW_a16,
+
+    /*0x40*/    sf_FLDs_a16,   sf_FLDs_a16,   sf_FLDs_a16,   sf_FLDs_a16,   sf_FLDs_a16,   sf_FLDs_a16,   sf_FLDs_a16,   sf_FLDs_a16,
+    /*0x48*/    ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+    /*0x50*/    sf_FSTs_a16,   sf_FSTs_a16,   sf_FSTs_a16,   sf_FSTs_a16,   sf_FSTs_a16,   sf_FSTs_a16,   sf_FSTs_a16,   sf_FSTs_a16,
+    /*0x58*/    sf_FSTPs_a16,  sf_FSTPs_a16,  sf_FSTPs_a16,  sf_FSTPs_a16,  sf_FSTPs_a16,  sf_FSTPs_a16,  sf_FSTPs_a16,  sf_FSTPs_a16,
+    /*0x60*/    sf_FLDENV_a16, sf_FLDENV_a16, sf_FLDENV_a16, sf_FLDENV_a16, sf_FLDENV_a16, sf_FLDENV_a16, sf_FLDENV_a16, sf_FLDENV_a16,
+    /*0x68*/    sf_FLDCW_a16,  sf_FLDCW_a16,  sf_FLDCW_a16,  sf_FLDCW_a16,  sf_FLDCW_a16,  sf_FLDCW_a16,  sf_FLDCW_a16,  sf_FLDCW_a16,
+    /*0x70*/    sf_FNSTENV_a16, sf_FNSTENV_a16, sf_FNSTENV_a16, sf_FNSTENV_a16, sf_FNSTENV_a16, sf_FNSTENV_a16, sf_FNSTENV_a16, sf_FNSTENV_a16,
+    /*0x78*/    sf_FNSTCW_a16,  sf_FNSTCW_a16,  sf_FNSTCW_a16,  sf_FNSTCW_a16,  sf_FNSTCW_a16,  sf_FNSTCW_a16,  sf_FNSTCW_a16,  sf_FNSTCW_a16,
+
+    /*0x80*/    sf_FLDs_a16,   sf_FLDs_a16,   sf_FLDs_a16,   sf_FLDs_a16,   sf_FLDs_a16,   sf_FLDs_a16,   sf_FLDs_a16,   sf_FLDs_a16,
+    /*0x88*/    ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+    /*0x90*/    sf_FSTs_a16,   sf_FSTs_a16,   sf_FSTs_a16,   sf_FSTs_a16,   sf_FSTs_a16,   sf_FSTs_a16,   sf_FSTs_a16,   sf_FSTs_a16,
+    /*0x98*/    sf_FSTPs_a16,  sf_FSTPs_a16,  sf_FSTPs_a16,  sf_FSTPs_a16,  sf_FSTPs_a16,  sf_FSTPs_a16,  sf_FSTPs_a16,  sf_FSTPs_a16,
+    /*0xa0*/    sf_FLDENV_a16, sf_FLDENV_a16, sf_FLDENV_a16, sf_FLDENV_a16, sf_FLDENV_a16, sf_FLDENV_a16, sf_FLDENV_a16, sf_FLDENV_a16,
+    /*0xa8*/    sf_FLDCW_a16,  sf_FLDCW_a16,  sf_FLDCW_a16,  sf_FLDCW_a16,  sf_FLDCW_a16,  sf_FLDCW_a16,  sf_FLDCW_a16,  sf_FLDCW_a16,
+    /*0xb0*/    sf_FNSTENV_a16, sf_FNSTENV_a16, sf_FNSTENV_a16, sf_FNSTENV_a16, sf_FNSTENV_a16, sf_FNSTENV_a16, sf_FNSTENV_a16, sf_FNSTENV_a16,
+    /*0xb8*/    sf_FNSTCW_a16,  sf_FNSTCW_a16,  sf_FNSTCW_a16,  sf_FNSTCW_a16,  sf_FNSTCW_a16,  sf_FNSTCW_a16,  sf_FNSTCW_a16,  sf_FNSTCW_a16,
+
+    /*0xc0*/    sf_FLD_sti,    sf_FLD_sti,    sf_FLD_sti,   sf_FLD_sti,    sf_FLD_sti,   sf_FLD_sti,   sf_FLD_sti,  sf_FLD_sti,
+    /*0xc8*/    sf_FXCH_sti,   sf_FXCH_sti,   sf_FXCH_sti,  sf_FXCH_sti,   sf_FXCH_sti,  sf_FXCH_sti,  sf_FXCH_sti, sf_FXCH_sti,
+    /*0xd0*/    sf_FNOP,       ILLEGAL_a16,   ILLEGAL_a16,  ILLEGAL_a16,   ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16, ILLEGAL_a16,
+    /*0xd8*/    sf_FSTP_sti,   sf_FSTP_sti,   sf_FSTP_sti,  sf_FSTP_sti,   sf_FSTP_sti,  sf_FSTP_sti,  sf_FSTP_sti, sf_FSTP_sti,  /*Invalid*/
+    /*0xe0*/    sf_FCHS,       sf_FABS,       ILLEGAL_a16,  ILLEGAL_a16,   sf_FTST,      sf_FXAM,      sf_FTSTP,    ILLEGAL_a16,
+    /*0xe8*/    sf_FLD1,       sf_FLDL2T,     sf_FLDL2E,    sf_FLDPI,      sf_FLDEG2,    sf_FLDLN2,    sf_FLDZ,     ILLEGAL_a16,
+    /*0xf0*/    sf_F2XM1,      sf_FYL2X,      sf_FPTAN,     sf_FPATAN,     sf_FXTRACT,   sf_FPREM1,    sf_FDECSTP,  sf_FINCSTP,
+    /*0xf8*/    sf_FPREM,      sf_FYL2XP1,    sf_FSQRT,     sf_FSINCOS,    sf_FRNDINT,   sf_FSCALE,    sf_FSIN,     sf_FCOS,
+    // clang-format on
+};
+
+const OpFn OP_TABLE(sf_fpu_cyrix_d9_a32)[256] = {
+    // clang-format off
+        sf_FLDs_a32,   sf_FLDs_a32,   sf_FLDs_a32,   sf_FLDs_a32,   sf_FLDs_a32,   sf_FLDs_a32,   sf_FLDs_a32,   sf_FLDs_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        sf_FSTs_a32,   sf_FSTs_a32,   sf_FSTs_a32,   sf_FSTs_a32,   sf_FSTs_a32,   sf_FSTs_a32,   sf_FSTs_a32,   sf_FSTs_a32,
+        sf_FSTPs_a32,  sf_FSTPs_a32,  sf_FSTPs_a32,  sf_FSTPs_a32,  sf_FSTPs_a32,  sf_FSTPs_a32,  sf_FSTPs_a32,  sf_FSTPs_a32,
+        sf_FLDENV_a32, sf_FLDENV_a32, sf_FLDENV_a32, sf_FLDENV_a32, sf_FLDENV_a32, sf_FLDENV_a32, sf_FLDENV_a32, sf_FLDENV_a32,
+        sf_FLDCW_a32,  sf_FLDCW_a32,  sf_FLDCW_a32,  sf_FLDCW_a32,  sf_FLDCW_a32,  sf_FLDCW_a32,  sf_FLDCW_a32,  sf_FLDCW_a32,
+        sf_FNSTENV_a32, sf_FNSTENV_a32, sf_FNSTENV_a32, sf_FNSTENV_a32, sf_FNSTENV_a32, sf_FNSTENV_a32, sf_FNSTENV_a32, sf_FNSTENV_a32,
+        sf_FNSTCW_a32,  sf_FNSTCW_a32,  sf_FNSTCW_a32,  sf_FNSTCW_a32,  sf_FNSTCW_a32,  sf_FNSTCW_a32,  sf_FNSTCW_a32,  sf_FNSTCW_a32,
+
+        sf_FLDs_a32,   sf_FLDs_a32,   sf_FLDs_a32,   sf_FLDs_a32,   sf_FLDs_a32,   sf_FLDs_a32,   sf_FLDs_a32,   sf_FLDs_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        sf_FSTs_a32,   sf_FSTs_a32,   sf_FSTs_a32,   sf_FSTs_a32,   sf_FSTs_a32,   sf_FSTs_a32,   sf_FSTs_a32,   sf_FSTs_a32,
+        sf_FSTPs_a32,  sf_FSTPs_a32,  sf_FSTPs_a32,  sf_FSTPs_a32,  sf_FSTPs_a32,  sf_FSTPs_a32,  sf_FSTPs_a32,  sf_FSTPs_a32,
+        sf_FLDENV_a32, sf_FLDENV_a32, sf_FLDENV_a32, sf_FLDENV_a32, sf_FLDENV_a32, sf_FLDENV_a32, sf_FLDENV_a32, sf_FLDENV_a32,
+        sf_FLDCW_a32,  sf_FLDCW_a32,  sf_FLDCW_a32,  sf_FLDCW_a32,  sf_FLDCW_a32,  sf_FLDCW_a32,  sf_FLDCW_a32,  sf_FLDCW_a32,
+        sf_FNSTENV_a32, sf_FNSTENV_a32, sf_FNSTENV_a32, sf_FNSTENV_a32, sf_FNSTENV_a32, sf_FNSTENV_a32, sf_FNSTENV_a32, sf_FNSTENV_a32,
+        sf_FNSTCW_a32,  sf_FNSTCW_a32,  sf_FNSTCW_a32,  sf_FNSTCW_a32,  sf_FNSTCW_a32,  sf_FNSTCW_a32,  sf_FNSTCW_a32,  sf_FNSTCW_a32,
+
+        sf_FLDs_a32,   sf_FLDs_a32,   sf_FLDs_a32,   sf_FLDs_a32,   sf_FLDs_a32,   sf_FLDs_a32,   sf_FLDs_a32,   sf_FLDs_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        sf_FSTs_a32,   sf_FSTs_a32,   sf_FSTs_a32,   sf_FSTs_a32,   sf_FSTs_a32,   sf_FSTs_a32,   sf_FSTs_a32,   sf_FSTs_a32,
+        sf_FSTPs_a32,  sf_FSTPs_a32,  sf_FSTPs_a32,  sf_FSTPs_a32,  sf_FSTPs_a32,  sf_FSTPs_a32,  sf_FSTPs_a32,  sf_FSTPs_a32,
+        sf_FLDENV_a32, sf_FLDENV_a32, sf_FLDENV_a32, sf_FLDENV_a32, sf_FLDENV_a32, sf_FLDENV_a32, sf_FLDENV_a32, sf_FLDENV_a32,
+        sf_FLDCW_a32,  sf_FLDCW_a32,  sf_FLDCW_a32,  sf_FLDCW_a32,  sf_FLDCW_a32,  sf_FLDCW_a32,  sf_FLDCW_a32,  sf_FLDCW_a32,
+        sf_FNSTENV_a32, sf_FNSTENV_a32, sf_FNSTENV_a32, sf_FNSTENV_a32, sf_FNSTENV_a32, sf_FNSTENV_a32, sf_FNSTENV_a32, sf_FNSTENV_a32,
+        sf_FNSTCW_a32,  sf_FNSTCW_a32,  sf_FNSTCW_a32,  sf_FNSTCW_a32,  sf_FNSTCW_a32,  sf_FNSTCW_a32,  sf_FNSTCW_a32,  sf_FNSTCW_a32,
+
+        sf_FLD_sti,        sf_FLD_sti,        sf_FLD_sti,        sf_FLD_sti,        sf_FLD_sti,        sf_FLD_sti,        sf_FLD_sti,        sf_FLD_sti,
+        sf_FXCH_sti,       sf_FXCH_sti,       sf_FXCH_sti,       sf_FXCH_sti,       sf_FXCH_sti,       sf_FXCH_sti,       sf_FXCH_sti,       sf_FXCH_sti,
+        sf_FNOP,       ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        sf_FSTP_sti,       sf_FSTP_sti,       sf_FSTP_sti,       sf_FSTP_sti,       sf_FSTP_sti,       sf_FSTP_sti,       sf_FSTP_sti,       sf_FSTP_sti,  /*Invalid*/
+        sf_FCHS,       sf_FABS,       ILLEGAL_a32,  ILLEGAL_a32,  sf_FTST,       sf_FXAM,       sf_FTSTP,        ILLEGAL_a32,
+        sf_FLD1,       sf_FLDL2T,     sf_FLDL2E,     sf_FLDPI,      sf_FLDEG2,     sf_FLDLN2,     sf_FLDZ,       ILLEGAL_a32,
+        sf_F2XM1,      sf_FYL2X,      sf_FPTAN,      sf_FPATAN,     sf_FXTRACT,  sf_FPREM1,     sf_FDECSTP,    sf_FINCSTP,
+        sf_FPREM,      sf_FYL2XP1,    sf_FSQRT,      sf_FSINCOS,    sf_FRNDINT,    sf_FSCALE,     sf_FSIN,       sf_FCOS,
+    // clang-format on
+};
+
 const OpFn OP_TABLE(sf_fpu_287_da_a16)[256] = {
     // clang-format off
         sf_FADDil_a16,  sf_FADDil_a16,  sf_FADDil_a16,  sf_FADDil_a16,  sf_FADDil_a16,  sf_FADDil_a16,  sf_FADDil_a16,  sf_FADDil_a16,
@@ -1750,6 +1848,87 @@ const OpFn OP_TABLE(sf_fpu_686_db_a32)[256] = {
 };
 #    endif
 
+#    ifndef OPS_286_386
+const OpFn OP_TABLE(sf_fpu_cyrix_686_db_a16)[256] = {
+    // clang-format off
+        sf_FILDil_a16,  sf_FILDil_a16,  sf_FILDil_a16,  sf_FILDil_a16,  sf_FILDil_a16,  sf_FILDil_a16,  sf_FILDil_a16,  sf_FILDil_a16,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        sf_FISTil_a16,  sf_FISTil_a16,  sf_FISTil_a16,  sf_FISTil_a16,  sf_FISTil_a16,  sf_FISTil_a16,  sf_FISTil_a16,  sf_FISTil_a16,
+        sf_FISTPil_a16, sf_FISTPil_a16, sf_FISTPil_a16, sf_FISTPil_a16, sf_FISTPil_a16, sf_FISTPil_a16, sf_FISTPil_a16, sf_FISTPil_a16,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        sf_FLDe_a16,    sf_FLDe_a16,    sf_FLDe_a16,    sf_FLDe_a16,    sf_FLDe_a16,    sf_FLDe_a16,    sf_FLDe_a16,    sf_FLDe_a16,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        sf_FSTPe_a16,   sf_FSTPe_a16,   sf_FSTPe_a16,   sf_FSTPe_a16,   sf_FSTPe_a16,   sf_FSTPe_a16,   sf_FSTPe_a16,   sf_FSTPe_a16,
+
+        sf_FILDil_a16,  sf_FILDil_a16,  sf_FILDil_a16,  sf_FILDil_a16,  sf_FILDil_a16,  sf_FILDil_a16,  sf_FILDil_a16,  sf_FILDil_a16,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        sf_FISTil_a16,  sf_FISTil_a16,  sf_FISTil_a16,  sf_FISTil_a16,  sf_FISTil_a16,  sf_FISTil_a16,  sf_FISTil_a16,  sf_FISTil_a16,
+        sf_FISTPil_a16, sf_FISTPil_a16, sf_FISTPil_a16, sf_FISTPil_a16, sf_FISTPil_a16, sf_FISTPil_a16, sf_FISTPil_a16, sf_FISTPil_a16,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        sf_FLDe_a16,    sf_FLDe_a16,    sf_FLDe_a16,    sf_FLDe_a16,    sf_FLDe_a16,    sf_FLDe_a16,    sf_FLDe_a16,    sf_FLDe_a16,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        sf_FSTPe_a16,   sf_FSTPe_a16,   sf_FSTPe_a16,   sf_FSTPe_a16,   sf_FSTPe_a16,   sf_FSTPe_a16,   sf_FSTPe_a16,   sf_FSTPe_a16,
+
+        sf_FILDil_a16,  sf_FILDil_a16,  sf_FILDil_a16,  sf_FILDil_a16,  sf_FILDil_a16,  sf_FILDil_a16,  sf_FILDil_a16,  sf_FILDil_a16,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        sf_FISTil_a16,  sf_FISTil_a16,  sf_FISTil_a16,  sf_FISTil_a16,  sf_FISTil_a16,  sf_FISTil_a16,  sf_FISTil_a16,  sf_FISTil_a16,
+        sf_FISTPil_a16, sf_FISTPil_a16, sf_FISTPil_a16, sf_FISTPil_a16, sf_FISTPil_a16, sf_FISTPil_a16, sf_FISTPil_a16, sf_FISTPil_a16,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        sf_FLDe_a16,    sf_FLDe_a16,    sf_FLDe_a16,    sf_FLDe_a16,    sf_FLDe_a16,    sf_FLDe_a16,    sf_FLDe_a16,    sf_FLDe_a16,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        sf_FSTPe_a16,   sf_FSTPe_a16,   sf_FSTPe_a16,   sf_FSTPe_a16,   sf_FSTPe_a16,   sf_FSTPe_a16,   sf_FSTPe_a16,   sf_FSTPe_a16,
+
+        sf_FCMOVNB,     sf_FCMOVNB,     sf_FCMOVNB,     sf_FCMOVNB,     sf_FCMOVNB,     sf_FCMOVNB,     sf_FCMOVNB,     sf_FCMOVNB,
+        sf_FCMOVNE,     sf_FCMOVNE,     sf_FCMOVNE,     sf_FCMOVNE,     sf_FCMOVNE,     sf_FCMOVNE,     sf_FCMOVNE,     sf_FCMOVNE,
+        sf_FCMOVNBE,    sf_FCMOVNBE,    sf_FCMOVNBE,    sf_FCMOVNBE,    sf_FCMOVNBE,    sf_FCMOVNBE,    sf_FCMOVNBE,    sf_FCMOVNBE,
+        sf_FCMOVNU,     sf_FCMOVNU,     sf_FCMOVNU,     sf_FCMOVNU,     sf_FCMOVNU,     sf_FCMOVNU,     sf_FCMOVNU,     sf_FCMOVNU,
+        sf_FNOP,        sf_FNOP,        sf_FNCLEX,       sf_FNINIT,       sf_FNOP,        sf_FNOP,        ILLEGAL_a16,       ILLEGAL_a16,
+        sf_FUCOMI_st0_stj,      sf_FUCOMI_st0_stj,      sf_FUCOMI_st0_stj,      sf_FUCOMI_st0_stj,      sf_FUCOMI_st0_stj,      sf_FUCOMI_st0_stj,      sf_FUCOMI_st0_stj,      sf_FUCOMI_st0_stj,
+        sf_FCOMI_st0_stj,       sf_FCOMI_st0_stj,       sf_FCOMI_st0_stj,       sf_FCOMI_st0_stj,       sf_FCOMI_st0_stj,       sf_FCOMI_st0_stj,       sf_FCOMI_st0_stj,       sf_FCOMI_st0_stj,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  sf_FRINT2,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+    // clang-format on
+};
+const OpFn OP_TABLE(sf_fpu_cyrix_686_db_a32)[256] = {
+    // clang-format off
+        sf_FILDil_a32,  sf_FILDil_a32,  sf_FILDil_a32,  sf_FILDil_a32,  sf_FILDil_a32,  sf_FILDil_a32,  sf_FILDil_a32,  sf_FILDil_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        sf_FISTil_a32,  sf_FISTil_a32,  sf_FISTil_a32,  sf_FISTil_a32,  sf_FISTil_a32,  sf_FISTil_a32,  sf_FISTil_a32,  sf_FISTil_a32,
+        sf_FISTPil_a32, sf_FISTPil_a32, sf_FISTPil_a32, sf_FISTPil_a32, sf_FISTPil_a32, sf_FISTPil_a32, sf_FISTPil_a32, sf_FISTPil_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        sf_FLDe_a32,    sf_FLDe_a32,    sf_FLDe_a32,    sf_FLDe_a32,    sf_FLDe_a32,    sf_FLDe_a32,    sf_FLDe_a32,    sf_FLDe_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        sf_FSTPe_a32,   sf_FSTPe_a32,   sf_FSTPe_a32,   sf_FSTPe_a32,   sf_FSTPe_a32,   sf_FSTPe_a32,   sf_FSTPe_a32,   sf_FSTPe_a32,
+
+        sf_FILDil_a32,  sf_FILDil_a32,  sf_FILDil_a32,  sf_FILDil_a32,  sf_FILDil_a32,  sf_FILDil_a32,  sf_FILDil_a32,  sf_FILDil_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        sf_FISTil_a32,  sf_FISTil_a32,  sf_FISTil_a32,  sf_FISTil_a32,  sf_FISTil_a32,  sf_FISTil_a32,  sf_FISTil_a32,  sf_FISTil_a32,
+        sf_FISTPil_a32, sf_FISTPil_a32, sf_FISTPil_a32, sf_FISTPil_a32, sf_FISTPil_a32, sf_FISTPil_a32, sf_FISTPil_a32, sf_FISTPil_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        sf_FLDe_a32,    sf_FLDe_a32,    sf_FLDe_a32,    sf_FLDe_a32,    sf_FLDe_a32,    sf_FLDe_a32,    sf_FLDe_a32,    sf_FLDe_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        sf_FSTPe_a32,   sf_FSTPe_a32,   sf_FSTPe_a32,   sf_FSTPe_a32,   sf_FSTPe_a32,   sf_FSTPe_a32,   sf_FSTPe_a32,   sf_FSTPe_a32,
+
+        sf_FILDil_a32,  sf_FILDil_a32,  sf_FILDil_a32,  sf_FILDil_a32,  sf_FILDil_a32,  sf_FILDil_a32,  sf_FILDil_a32,  sf_FILDil_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        sf_FISTil_a32,  sf_FISTil_a32,  sf_FISTil_a32,  sf_FISTil_a32,  sf_FISTil_a32,  sf_FISTil_a32,  sf_FISTil_a32,  sf_FISTil_a32,
+        sf_FISTPil_a32, sf_FISTPil_a32, sf_FISTPil_a32, sf_FISTPil_a32, sf_FISTPil_a32, sf_FISTPil_a32, sf_FISTPil_a32, sf_FISTPil_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        sf_FLDe_a32,    sf_FLDe_a32,    sf_FLDe_a32,    sf_FLDe_a32,    sf_FLDe_a32,    sf_FLDe_a32,    sf_FLDe_a32,    sf_FLDe_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        sf_FSTPe_a32,   sf_FSTPe_a32,   sf_FSTPe_a32,   sf_FSTPe_a32,   sf_FSTPe_a32,   sf_FSTPe_a32,   sf_FSTPe_a32,   sf_FSTPe_a32,
+
+        sf_FCMOVNB,     sf_FCMOVNB,     sf_FCMOVNB,     sf_FCMOVNB,     sf_FCMOVNB,     sf_FCMOVNB,     sf_FCMOVNB,     sf_FCMOVNB,
+        sf_FCMOVNE,     sf_FCMOVNE,     sf_FCMOVNE,     sf_FCMOVNE,     sf_FCMOVNE,     sf_FCMOVNE,     sf_FCMOVNE,     sf_FCMOVNE,
+        sf_FCMOVNBE,    sf_FCMOVNBE,    sf_FCMOVNBE,    sf_FCMOVNBE,    sf_FCMOVNBE,    sf_FCMOVNBE,    sf_FCMOVNBE,    sf_FCMOVNBE,
+        sf_FCMOVNU,     sf_FCMOVNU,     sf_FCMOVNU,     sf_FCMOVNU,     sf_FCMOVNU,     sf_FCMOVNU,     sf_FCMOVNU,     sf_FCMOVNU,
+        sf_FNOP,        sf_FNOP,        sf_FNCLEX,       sf_FNINIT,       sf_FNOP,        sf_FNOP,        ILLEGAL_a32,       ILLEGAL_a32,
+        sf_FUCOMI_st0_stj,      sf_FUCOMI_st0_stj,      sf_FUCOMI_st0_stj,      sf_FUCOMI_st0_stj,      sf_FUCOMI_st0_stj,      sf_FUCOMI_st0_stj,      sf_FUCOMI_st0_stj,      sf_FUCOMI_st0_stj,
+        sf_FCOMI_st0_stj,       sf_FCOMI_st0_stj,       sf_FCOMI_st0_stj,       sf_FCOMI_st0_stj,       sf_FCOMI_st0_stj,       sf_FCOMI_st0_stj,       sf_FCOMI_st0_stj,       sf_FCOMI_st0_stj,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  sf_FRINT2,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+    // clang-format on
+};
+#    endif
+
 const OpFn OP_TABLE(sf_fpu_287_dc_a16)[32] = {
     // clang-format off
         sf_FADDd_a16, sf_FMULd_a16, sf_FCOMd_a16, sf_FCOMPd_a16, sf_FSUBd_a16, sf_FSUBRd_a16, sf_FDIVd_a16, sf_FDIVRd_a16,
@@ -1946,6 +2125,86 @@ const OpFn OP_TABLE(sf_fpu_dd_a32)[256] = {
     // clang-format on
 };
 
+const OpFn OP_TABLE(sf_fpu_cyrix_dd_a16)[256] = {
+    // clang-format off
+    /*0x00*/    sf_FLDd_a16,    sf_FLDd_a16,    sf_FLDd_a16,    sf_FLDd_a16,    sf_FLDd_a16,    sf_FLDd_a16,    sf_FLDd_a16,    sf_FLDd_a16,
+    /*0x08*/    ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+    /*0x10*/    sf_FSTd_a16,    sf_FSTd_a16,    sf_FSTd_a16,    sf_FSTd_a16,    sf_FSTd_a16,    sf_FSTd_a16,    sf_FSTd_a16,    sf_FSTd_a16,
+    /*0x18*/    sf_FSTPd_a16,   sf_FSTPd_a16,   sf_FSTPd_a16,   sf_FSTPd_a16,   sf_FSTPd_a16,   sf_FSTPd_a16,   sf_FSTPd_a16,   sf_FSTPd_a16,
+    /*0x20*/    sf_FRSTOR_a16,   sf_FRSTOR_a16,   sf_FRSTOR_a16,   sf_FRSTOR_a16,   sf_FRSTOR_a16,   sf_FRSTOR_a16,   sf_FRSTOR_a16,   sf_FRSTOR_a16,
+    /*0x28*/    ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+    /*0x30*/    sf_FNSAVE_a16,   sf_FNSAVE_a16,   sf_FNSAVE_a16,   sf_FNSAVE_a16,   sf_FNSAVE_a16,   sf_FNSAVE_a16,   sf_FNSAVE_a16,   sf_FNSAVE_a16,
+    /*0x38*/    sf_FNSTSW_a16,   sf_FNSTSW_a16,   sf_FNSTSW_a16,   sf_FNSTSW_a16,   sf_FNSTSW_a16,   sf_FNSTSW_a16,   sf_FNSTSW_a16,   sf_FNSTSW_a16,
+
+    /*0x40*/    sf_FLDd_a16,    sf_FLDd_a16,    sf_FLDd_a16,    sf_FLDd_a16,    sf_FLDd_a16,    sf_FLDd_a16,    sf_FLDd_a16,    sf_FLDd_a16,
+    /*0x48*/    ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+    /*0x50*/    sf_FSTd_a16,    sf_FSTd_a16,    sf_FSTd_a16,    sf_FSTd_a16,    sf_FSTd_a16,    sf_FSTd_a16,    sf_FSTd_a16,    sf_FSTd_a16,
+    /*0x58*/    sf_FSTPd_a16,   sf_FSTPd_a16,   sf_FSTPd_a16,   sf_FSTPd_a16,   sf_FSTPd_a16,   sf_FSTPd_a16,   sf_FSTPd_a16,   sf_FSTPd_a16,
+    /*0x60*/    sf_FRSTOR_a16,   sf_FRSTOR_a16,   sf_FRSTOR_a16,   sf_FRSTOR_a16,   sf_FRSTOR_a16,   sf_FRSTOR_a16,   sf_FRSTOR_a16,   sf_FRSTOR_a16,
+    /*0x68*/    ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+    /*0x70*/    sf_FNSAVE_a16,   sf_FNSAVE_a16,   sf_FNSAVE_a16,   sf_FNSAVE_a16,   sf_FNSAVE_a16,   sf_FNSAVE_a16,   sf_FNSAVE_a16,   sf_FNSAVE_a16,
+    /*0x78*/    sf_FNSTSW_a16,   sf_FNSTSW_a16,   sf_FNSTSW_a16,   sf_FNSTSW_a16,   sf_FNSTSW_a16,   sf_FNSTSW_a16,   sf_FNSTSW_a16,   sf_FNSTSW_a16,
+
+    /*0x80*/    sf_FLDd_a16,    sf_FLDd_a16,    sf_FLDd_a16,    sf_FLDd_a16,    sf_FLDd_a16,    sf_FLDd_a16,    sf_FLDd_a16,    sf_FLDd_a16,
+    /*0x88*/    ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+    /*0x90*/    sf_FSTd_a16,    sf_FSTd_a16,    sf_FSTd_a16,    sf_FSTd_a16,    sf_FSTd_a16,    sf_FSTd_a16,    sf_FSTd_a16,    sf_FSTd_a16,
+    /*0x98*/    sf_FSTPd_a16,   sf_FSTPd_a16,   sf_FSTPd_a16,   sf_FSTPd_a16,   sf_FSTPd_a16,   sf_FSTPd_a16,   sf_FSTPd_a16,   sf_FSTPd_a16,
+    /*0xa0*/    sf_FRSTOR_a16,   sf_FRSTOR_a16,   sf_FRSTOR_a16,   sf_FRSTOR_a16,   sf_FRSTOR_a16,   sf_FRSTOR_a16,   sf_FRSTOR_a16,   sf_FRSTOR_a16,
+    /*0xa8*/    ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+    /*0xb0*/    sf_FNSAVE_a16,   sf_FNSAVE_a16,   sf_FNSAVE_a16,   sf_FNSAVE_a16,   sf_FNSAVE_a16,   sf_FNSAVE_a16,   sf_FNSAVE_a16,   sf_FNSAVE_a16,
+    /*0xb8*/    sf_FNSTSW_a16,   sf_FNSTSW_a16,   sf_FNSTSW_a16,   sf_FNSTSW_a16,   sf_FNSTSW_a16,   sf_FNSTSW_a16,   sf_FNSTSW_a16,   sf_FNSTSW_a16,
+
+    /*0xc0*/    sf_FFREE_sti,       sf_FFREE_sti,       sf_FFREE_sti,       sf_FFREE_sti,       sf_FFREE_sti,       sf_FFREE_sti,       sf_FFREE_sti,       sf_FFREE_sti,
+    /*0xc8*/    sf_FXCH_sti,        sf_FXCH_sti,        sf_FXCH_sti,        sf_FXCH_sti,        sf_FXCH_sti,        sf_FXCH_sti,        sf_FXCH_sti,        sf_FXCH_sti,
+    /*0xd0*/    sf_FST_sti,         sf_FST_sti,         sf_FST_sti,         sf_FST_sti,         sf_FST_sti,         sf_FST_sti,         sf_FST_sti,         sf_FST_sti,
+    /*0xd8*/    sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,
+    /*0xe0*/    sf_FUCOM_sti,       sf_FUCOM_sti,       sf_FUCOM_sti,       sf_FUCOM_sti,       sf_FUCOM_sti,       sf_FUCOM_sti,       sf_FUCOM_sti,       sf_FUCOM_sti,
+    /*0xe8*/    sf_FUCOMP_sti,      sf_FUCOMP_sti,      sf_FUCOMP_sti,      sf_FUCOMP_sti,      sf_FUCOMP_sti,      sf_FUCOMP_sti,      sf_FUCOMP_sti,      sf_FUCOMP_sti,
+    /*0xf0*/    ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+    /*0xf8*/    ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  sf_FRICHOP,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+    // clang-format on
+};
+
+const OpFn OP_TABLE(sf_fpu_cyrix_dd_a32)[256] = {
+    // clang-format off
+        sf_FLDd_a32,    sf_FLDd_a32,    sf_FLDd_a32,    sf_FLDd_a32,    sf_FLDd_a32,    sf_FLDd_a32,    sf_FLDd_a32,    sf_FLDd_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        sf_FSTd_a32,    sf_FSTd_a32,    sf_FSTd_a32,    sf_FSTd_a32,    sf_FSTd_a32,    sf_FSTd_a32,    sf_FSTd_a32,    sf_FSTd_a32,
+        sf_FSTPd_a32,   sf_FSTPd_a32,   sf_FSTPd_a32,   sf_FSTPd_a32,   sf_FSTPd_a32,   sf_FSTPd_a32,   sf_FSTPd_a32,   sf_FSTPd_a32,
+        sf_FRSTOR_a32,   sf_FRSTOR_a32,   sf_FRSTOR_a32,   sf_FRSTOR_a32,   sf_FRSTOR_a32,   sf_FRSTOR_a32,   sf_FRSTOR_a32,   sf_FRSTOR_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        sf_FNSAVE_a32,   sf_FNSAVE_a32,   sf_FNSAVE_a32,   sf_FNSAVE_a32,   sf_FNSAVE_a32,   sf_FNSAVE_a32,   sf_FNSAVE_a32,   sf_FNSAVE_a32,
+        sf_FNSTSW_a32,   sf_FNSTSW_a32,   sf_FNSTSW_a32,   sf_FNSTSW_a32,   sf_FNSTSW_a32,   sf_FNSTSW_a32,   sf_FNSTSW_a32,   sf_FNSTSW_a32,
+
+        sf_FLDd_a32,    sf_FLDd_a32,    sf_FLDd_a32,    sf_FLDd_a32,    sf_FLDd_a32,    sf_FLDd_a32,    sf_FLDd_a32,    sf_FLDd_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        sf_FSTd_a32,    sf_FSTd_a32,    sf_FSTd_a32,    sf_FSTd_a32,    sf_FSTd_a32,    sf_FSTd_a32,    sf_FSTd_a32,    sf_FSTd_a32,
+        sf_FSTPd_a32,   sf_FSTPd_a32,   sf_FSTPd_a32,   sf_FSTPd_a32,   sf_FSTPd_a32,   sf_FSTPd_a32,   sf_FSTPd_a32,   sf_FSTPd_a32,
+        sf_FRSTOR_a32,   sf_FRSTOR_a32,   sf_FRSTOR_a32,   sf_FRSTOR_a32,   sf_FRSTOR_a32,   sf_FRSTOR_a32,   sf_FRSTOR_a32,   sf_FRSTOR_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        sf_FNSAVE_a32,   sf_FNSAVE_a32,   sf_FNSAVE_a32,   sf_FNSAVE_a32,   sf_FNSAVE_a32,   sf_FNSAVE_a32,   sf_FNSAVE_a32,   sf_FNSAVE_a32,
+        sf_FNSTSW_a32,   sf_FNSTSW_a32,   sf_FNSTSW_a32,   sf_FNSTSW_a32,   sf_FNSTSW_a32,   sf_FNSTSW_a32,   sf_FNSTSW_a32,   sf_FNSTSW_a32,
+
+        sf_FLDd_a32,    sf_FLDd_a32,    sf_FLDd_a32,    sf_FLDd_a32,    sf_FLDd_a32,    sf_FLDd_a32,    sf_FLDd_a32,    sf_FLDd_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        sf_FSTd_a32,    sf_FSTd_a32,    sf_FSTd_a32,    sf_FSTd_a32,    sf_FSTd_a32,    sf_FSTd_a32,    sf_FSTd_a32,    sf_FSTd_a32,
+        sf_FSTPd_a32,   sf_FSTPd_a32,   sf_FSTPd_a32,   sf_FSTPd_a32,   sf_FSTPd_a32,   sf_FSTPd_a32,   sf_FSTPd_a32,   sf_FSTPd_a32,
+        sf_FRSTOR_a32,   sf_FRSTOR_a32,   sf_FRSTOR_a32,   sf_FRSTOR_a32,   sf_FRSTOR_a32,   sf_FRSTOR_a32,   sf_FRSTOR_a32,   sf_FRSTOR_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        sf_FNSAVE_a32,   sf_FNSAVE_a32,   sf_FNSAVE_a32,   sf_FNSAVE_a32,   sf_FNSAVE_a32,   sf_FNSAVE_a32,   sf_FNSAVE_a32,   sf_FNSAVE_a32,
+        sf_FNSTSW_a32,   sf_FNSTSW_a32,   sf_FNSTSW_a32,   sf_FNSTSW_a32,   sf_FNSTSW_a32,   sf_FNSTSW_a32,   sf_FNSTSW_a32,   sf_FNSTSW_a32,
+
+        sf_FFREE_sti,       sf_FFREE_sti,       sf_FFREE_sti,       sf_FFREE_sti,       sf_FFREE_sti,       sf_FFREE_sti,       sf_FFREE_sti,       sf_FFREE_sti,
+        sf_FXCH_sti,        sf_FXCH_sti,        sf_FXCH_sti,        sf_FXCH_sti,        sf_FXCH_sti,        sf_FXCH_sti,        sf_FXCH_sti,        sf_FXCH_sti,
+        sf_FST_sti,         sf_FST_sti,         sf_FST_sti,         sf_FST_sti,         sf_FST_sti,         sf_FST_sti,         sf_FST_sti,         sf_FST_sti,
+        sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,
+        sf_FUCOM_sti,       sf_FUCOM_sti,       sf_FUCOM_sti,       sf_FUCOM_sti,       sf_FUCOM_sti,       sf_FUCOM_sti,       sf_FUCOM_sti,       sf_FUCOM_sti,
+        sf_FUCOMP_sti,      sf_FUCOMP_sti,      sf_FUCOMP_sti,      sf_FUCOMP_sti,      sf_FUCOMP_sti,      sf_FUCOMP_sti,      sf_FUCOMP_sti,      sf_FUCOMP_sti,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  sf_FRICHOP,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+    // clang-format on
+};
+
 const OpFn OP_TABLE(sf_fpu_287_de_a16)[256] = {
     // clang-format off
         sf_FADDiw_a16,  sf_FADDiw_a16,  sf_FADDiw_a16,  sf_FADDiw_a16,  sf_FADDiw_a16,  sf_FADDiw_a16,  sf_FADDiw_a16,  sf_FADDiw_a16,
@@ -2348,6 +2607,88 @@ const OpFn OP_TABLE(sf_fpu_686_df_a32)[256] = {
 };
 #    endif
 
+#    ifndef OPS_286_386
+const OpFn OP_TABLE(sf_fpu_cyrix_686_df_a16)[256] = {
+    // clang-format off
+        sf_FILDiw_a16,  sf_FILDiw_a16,  sf_FILDiw_a16,  sf_FILDiw_a16,  sf_FILDiw_a16,  sf_FILDiw_a16,  sf_FILDiw_a16,  sf_FILDiw_a16,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        sf_FISTiw_a16,  sf_FISTiw_a16,  sf_FISTiw_a16,  sf_FISTiw_a16,  sf_FISTiw_a16,  sf_FISTiw_a16,  sf_FISTiw_a16,  sf_FISTiw_a16,
+        sf_FISTPiw_a16, sf_FISTPiw_a16, sf_FISTPiw_a16, sf_FISTPiw_a16, sf_FISTPiw_a16, sf_FISTPiw_a16, sf_FISTPiw_a16, sf_FISTPiw_a16,
+        sf_FBLD_PACKED_BCD_a16,     sf_FBLD_PACKED_BCD_a16,     sf_FBLD_PACKED_BCD_a16,     sf_FBLD_PACKED_BCD_a16,     sf_FBLD_PACKED_BCD_a16,     sf_FBLD_PACKED_BCD_a16,     sf_FBLD_PACKED_BCD_a16,     sf_FBLD_PACKED_BCD_a16,
+        sf_FILDiq_a16,  sf_FILDiq_a16,  sf_FILDiq_a16,  sf_FILDiq_a16,  sf_FILDiq_a16,  sf_FILDiq_a16,  sf_FILDiq_a16,  sf_FILDiq_a16,
+        sf_FBSTP_PACKED_BCD_a16,     sf_FBSTP_PACKED_BCD_a16,     sf_FBSTP_PACKED_BCD_a16,     sf_FBSTP_PACKED_BCD_a16,     sf_FBSTP_PACKED_BCD_a16,     sf_FBSTP_PACKED_BCD_a16,     sf_FBSTP_PACKED_BCD_a16,     sf_FBSTP_PACKED_BCD_a16,
+        sf_FISTPiq_a16,   sf_FISTPiq_a16,   sf_FISTPiq_a16,   sf_FISTPiq_a16,   sf_FISTPiq_a16,   sf_FISTPiq_a16,   sf_FISTPiq_a16,   sf_FISTPiq_a16,
+
+        sf_FILDiw_a16,  sf_FILDiw_a16,  sf_FILDiw_a16,  sf_FILDiw_a16,  sf_FILDiw_a16,  sf_FILDiw_a16,  sf_FILDiw_a16,  sf_FILDiw_a16,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        sf_FISTiw_a16,  sf_FISTiw_a16,  sf_FISTiw_a16,  sf_FISTiw_a16,  sf_FISTiw_a16,  sf_FISTiw_a16,  sf_FISTiw_a16,  sf_FISTiw_a16,
+        sf_FISTPiw_a16, sf_FISTPiw_a16, sf_FISTPiw_a16, sf_FISTPiw_a16, sf_FISTPiw_a16, sf_FISTPiw_a16, sf_FISTPiw_a16, sf_FISTPiw_a16,
+        sf_FBLD_PACKED_BCD_a16,     sf_FBLD_PACKED_BCD_a16,     sf_FBLD_PACKED_BCD_a16,     sf_FBLD_PACKED_BCD_a16,     sf_FBLD_PACKED_BCD_a16,     sf_FBLD_PACKED_BCD_a16,     sf_FBLD_PACKED_BCD_a16,     sf_FBLD_PACKED_BCD_a16,
+        sf_FILDiq_a16,  sf_FILDiq_a16,  sf_FILDiq_a16,  sf_FILDiq_a16,  sf_FILDiq_a16,  sf_FILDiq_a16,  sf_FILDiq_a16,  sf_FILDiq_a16,
+        sf_FBSTP_PACKED_BCD_a16,     sf_FBSTP_PACKED_BCD_a16,     sf_FBSTP_PACKED_BCD_a16,     sf_FBSTP_PACKED_BCD_a16,     sf_FBSTP_PACKED_BCD_a16,     sf_FBSTP_PACKED_BCD_a16,     sf_FBSTP_PACKED_BCD_a16,     sf_FBSTP_PACKED_BCD_a16,
+        sf_FISTPiq_a16,   sf_FISTPiq_a16,   sf_FISTPiq_a16,   sf_FISTPiq_a16,   sf_FISTPiq_a16,   sf_FISTPiq_a16,   sf_FISTPiq_a16,   sf_FISTPiq_a16,
+
+        sf_FILDiw_a16,  sf_FILDiw_a16,  sf_FILDiw_a16,  sf_FILDiw_a16,  sf_FILDiw_a16,  sf_FILDiw_a16,  sf_FILDiw_a16,  sf_FILDiw_a16,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        sf_FISTiw_a16,  sf_FISTiw_a16,  sf_FISTiw_a16,  sf_FISTiw_a16,  sf_FISTiw_a16,  sf_FISTiw_a16,  sf_FISTiw_a16,  sf_FISTiw_a16,
+        sf_FISTPiw_a16, sf_FISTPiw_a16, sf_FISTPiw_a16, sf_FISTPiw_a16, sf_FISTPiw_a16, sf_FISTPiw_a16, sf_FISTPiw_a16, sf_FISTPiw_a16,
+        sf_FBLD_PACKED_BCD_a16,     sf_FBLD_PACKED_BCD_a16,     sf_FBLD_PACKED_BCD_a16,     sf_FBLD_PACKED_BCD_a16,     sf_FBLD_PACKED_BCD_a16,     sf_FBLD_PACKED_BCD_a16,     sf_FBLD_PACKED_BCD_a16,     sf_FBLD_PACKED_BCD_a16,
+        sf_FILDiq_a16,  sf_FILDiq_a16,  sf_FILDiq_a16,  sf_FILDiq_a16,  sf_FILDiq_a16,  sf_FILDiq_a16,  sf_FILDiq_a16,  sf_FILDiq_a16,
+        sf_FBSTP_PACKED_BCD_a16,     sf_FBSTP_PACKED_BCD_a16,     sf_FBSTP_PACKED_BCD_a16,     sf_FBSTP_PACKED_BCD_a16,     sf_FBSTP_PACKED_BCD_a16,     sf_FBSTP_PACKED_BCD_a16,     sf_FBSTP_PACKED_BCD_a16,     sf_FBSTP_PACKED_BCD_a16,
+        sf_FISTPiq_a16,   sf_FISTPiq_a16,   sf_FISTPiq_a16,   sf_FISTPiq_a16,   sf_FISTPiq_a16,   sf_FISTPiq_a16,   sf_FISTPiq_a16,   sf_FISTPiq_a16,
+
+        sf_FFREEP_sti,      sf_FFREEP_sti,      sf_FFREEP_sti,      sf_FFREEP_sti,      sf_FFREEP_sti,      sf_FFREEP_sti,      sf_FFREEP_sti,      sf_FFREEP_sti,
+        sf_FXCH_sti,        sf_FXCH_sti,        sf_FXCH_sti,        sf_FXCH_sti,        sf_FXCH_sti,        sf_FXCH_sti,        sf_FXCH_sti,        sf_FXCH_sti,
+        sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,
+        sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,
+        sf_FNSTSW_AX,   ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        sf_FUCOMIP_st0_stj,     sf_FUCOMIP_st0_stj,     sf_FUCOMIP_st0_stj,     sf_FUCOMIP_st0_stj,     sf_FUCOMIP_st0_stj,     sf_FUCOMIP_st0_stj,     sf_FUCOMIP_st0_stj,     sf_FUCOMIP_st0_stj,
+        sf_FCOMIP_st0_stj,      sf_FCOMIP_st0_stj,      sf_FCOMIP_st0_stj,      sf_FCOMIP_st0_stj,      sf_FCOMIP_st0_stj,      sf_FCOMIP_st0_stj,      sf_FCOMIP_st0_stj,      sf_FCOMIP_st0_stj,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  sf_FRINEAR,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+    // clang-format on
+};
+
+const OpFn OP_TABLE(sf_fpu_cyrix_686_df_a32)[256] = {
+    // clang-format off
+        sf_FILDiw_a32,  sf_FILDiw_a32,  sf_FILDiw_a32,  sf_FILDiw_a32,  sf_FILDiw_a32,  sf_FILDiw_a32,  sf_FILDiw_a32,  sf_FILDiw_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        sf_FISTiw_a32,  sf_FISTiw_a32,  sf_FISTiw_a32,  sf_FISTiw_a32,  sf_FISTiw_a32,  sf_FISTiw_a32,  sf_FISTiw_a32,  sf_FISTiw_a32,
+        sf_FISTPiw_a32, sf_FISTPiw_a32, sf_FISTPiw_a32, sf_FISTPiw_a32, sf_FISTPiw_a32, sf_FISTPiw_a32, sf_FISTPiw_a32, sf_FISTPiw_a32,
+        sf_FBLD_PACKED_BCD_a32,     sf_FBLD_PACKED_BCD_a32,     sf_FBLD_PACKED_BCD_a32,     sf_FBLD_PACKED_BCD_a32,     sf_FBLD_PACKED_BCD_a32,     sf_FBLD_PACKED_BCD_a32,     sf_FBLD_PACKED_BCD_a32,     sf_FBLD_PACKED_BCD_a32,
+        sf_FILDiq_a32,  sf_FILDiq_a32,  sf_FILDiq_a32,  sf_FILDiq_a32,  sf_FILDiq_a32,  sf_FILDiq_a32,  sf_FILDiq_a32,  sf_FILDiq_a32,
+        sf_FBSTP_PACKED_BCD_a32,     sf_FBSTP_PACKED_BCD_a32,     sf_FBSTP_PACKED_BCD_a32,     sf_FBSTP_PACKED_BCD_a32,     sf_FBSTP_PACKED_BCD_a32,     sf_FBSTP_PACKED_BCD_a32,     sf_FBSTP_PACKED_BCD_a32,     sf_FBSTP_PACKED_BCD_a32,
+        sf_FISTPiq_a32,   sf_FISTPiq_a32,   sf_FISTPiq_a32,   sf_FISTPiq_a32,   sf_FISTPiq_a32,   sf_FISTPiq_a32,   sf_FISTPiq_a32,   sf_FISTPiq_a32,
+
+        sf_FILDiw_a32,  sf_FILDiw_a32,  sf_FILDiw_a32,  sf_FILDiw_a32,  sf_FILDiw_a32,  sf_FILDiw_a32,  sf_FILDiw_a32,  sf_FILDiw_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        sf_FISTiw_a32,  sf_FISTiw_a32,  sf_FISTiw_a32,  sf_FISTiw_a32,  sf_FISTiw_a32,  sf_FISTiw_a32,  sf_FISTiw_a32,  sf_FISTiw_a32,
+        sf_FISTPiw_a32, sf_FISTPiw_a32, sf_FISTPiw_a32, sf_FISTPiw_a32, sf_FISTPiw_a32, sf_FISTPiw_a32, sf_FISTPiw_a32, sf_FISTPiw_a32,
+        sf_FBLD_PACKED_BCD_a32,     sf_FBLD_PACKED_BCD_a32,     sf_FBLD_PACKED_BCD_a32,     sf_FBLD_PACKED_BCD_a32,     sf_FBLD_PACKED_BCD_a32,     sf_FBLD_PACKED_BCD_a32,     sf_FBLD_PACKED_BCD_a32,     sf_FBLD_PACKED_BCD_a32,
+        sf_FILDiq_a32,  sf_FILDiq_a32,  sf_FILDiq_a32,  sf_FILDiq_a32,  sf_FILDiq_a32,  sf_FILDiq_a32,  sf_FILDiq_a32,  sf_FILDiq_a32,
+        sf_FBSTP_PACKED_BCD_a32,     sf_FBSTP_PACKED_BCD_a32,     sf_FBSTP_PACKED_BCD_a32,     sf_FBSTP_PACKED_BCD_a32,     sf_FBSTP_PACKED_BCD_a32,     sf_FBSTP_PACKED_BCD_a32,     sf_FBSTP_PACKED_BCD_a32,     sf_FBSTP_PACKED_BCD_a32,
+        sf_FISTPiq_a32,   sf_FISTPiq_a32,   sf_FISTPiq_a32,   sf_FISTPiq_a32,   sf_FISTPiq_a32,   sf_FISTPiq_a32,   sf_FISTPiq_a32,   sf_FISTPiq_a32,
+
+        sf_FILDiw_a32,  sf_FILDiw_a32,  sf_FILDiw_a32,  sf_FILDiw_a32,  sf_FILDiw_a32,  sf_FILDiw_a32,  sf_FILDiw_a32,  sf_FILDiw_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        sf_FISTiw_a32,  sf_FISTiw_a32,  sf_FISTiw_a32,  sf_FISTiw_a32,  sf_FISTiw_a32,  sf_FISTiw_a32,  sf_FISTiw_a32,  sf_FISTiw_a32,
+        sf_FISTPiw_a32, sf_FISTPiw_a32, sf_FISTPiw_a32, sf_FISTPiw_a32, sf_FISTPiw_a32, sf_FISTPiw_a32, sf_FISTPiw_a32, sf_FISTPiw_a32,
+        sf_FBLD_PACKED_BCD_a32,     sf_FBLD_PACKED_BCD_a32,     sf_FBLD_PACKED_BCD_a32,     sf_FBLD_PACKED_BCD_a32,     sf_FBLD_PACKED_BCD_a32,     sf_FBLD_PACKED_BCD_a32,     sf_FBLD_PACKED_BCD_a32,     sf_FBLD_PACKED_BCD_a32,
+        sf_FILDiq_a32,  sf_FILDiq_a32,  sf_FILDiq_a32,  sf_FILDiq_a32,  sf_FILDiq_a32,  sf_FILDiq_a32,  sf_FILDiq_a32,  sf_FILDiq_a32,
+        sf_FBSTP_PACKED_BCD_a32,     sf_FBSTP_PACKED_BCD_a32,     sf_FBSTP_PACKED_BCD_a32,     sf_FBSTP_PACKED_BCD_a32,     sf_FBSTP_PACKED_BCD_a32,     sf_FBSTP_PACKED_BCD_a32,     sf_FBSTP_PACKED_BCD_a32,     sf_FBSTP_PACKED_BCD_a32,
+        sf_FISTPiq_a32,   sf_FISTPiq_a32,   sf_FISTPiq_a32,   sf_FISTPiq_a32,   sf_FISTPiq_a32,   sf_FISTPiq_a32,   sf_FISTPiq_a32,   sf_FISTPiq_a32,
+
+        sf_FFREEP_sti,      sf_FFREEP_sti,      sf_FFREEP_sti,      sf_FFREEP_sti,      sf_FFREEP_sti,      sf_FFREEP_sti,      sf_FFREEP_sti,      sf_FFREEP_sti,
+        sf_FXCH_sti,        sf_FXCH_sti,        sf_FXCH_sti,        sf_FXCH_sti,        sf_FXCH_sti,        sf_FXCH_sti,        sf_FXCH_sti,        sf_FXCH_sti,
+        sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,
+        sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,        sf_FSTP_sti,
+        sf_FNSTSW_AX,   ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        sf_FUCOMIP_st0_stj,     sf_FUCOMIP_st0_stj,     sf_FUCOMIP_st0_stj,     sf_FUCOMIP_st0_stj,     sf_FUCOMIP_st0_stj,     sf_FUCOMIP_st0_stj,     sf_FUCOMIP_st0_stj,     sf_FUCOMIP_st0_stj,
+        sf_FCOMIP_st0_stj,      sf_FCOMIP_st0_stj,      sf_FCOMIP_st0_stj,      sf_FCOMIP_st0_stj,      sf_FCOMIP_st0_stj,      sf_FCOMIP_st0_stj,      sf_FCOMIP_st0_stj,      sf_FCOMIP_st0_stj,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  sf_FRINEAR,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+    // clang-format on
+};
+#    endif
+
 const OpFn OP_TABLE(fpu_d8_a16)[32] = {
     // clang-format off
         opFADDs_a16, opFMULs_a16, opFCOMs_a16, opFCOMPs_a16, opFSUBs_a16, opFSUBRs_a16, opFDIVs_a16, opFDIVRs_a16,
@@ -2526,6 +2867,86 @@ const OpFn OP_TABLE(fpu_d9_a32)[256] = {
     // clang-format on
 };
 
+const OpFn OP_TABLE(fpu_cyrix_d9_a16)[256] = {
+    // clang-format off
+        opFLDs_a16,   opFLDs_a16,   opFLDs_a16,   opFLDs_a16,   opFLDs_a16,   opFLDs_a16,   opFLDs_a16,   opFLDs_a16,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        opFSTs_a16,   opFSTs_a16,   opFSTs_a16,   opFSTs_a16,   opFSTs_a16,   opFSTs_a16,   opFSTs_a16,   opFSTs_a16,
+        opFSTPs_a16,  opFSTPs_a16,  opFSTPs_a16,  opFSTPs_a16,  opFSTPs_a16,  opFSTPs_a16,  opFSTPs_a16,  opFSTPs_a16,
+        opFLDENV_a16, opFLDENV_a16, opFLDENV_a16, opFLDENV_a16, opFLDENV_a16, opFLDENV_a16, opFLDENV_a16, opFLDENV_a16,
+        opFLDCW_a16,  opFLDCW_a16,  opFLDCW_a16,  opFLDCW_a16,  opFLDCW_a16,  opFLDCW_a16,  opFLDCW_a16,  opFLDCW_a16,
+        opFSTENV_a16, opFSTENV_a16, opFSTENV_a16, opFSTENV_a16, opFSTENV_a16, opFSTENV_a16, opFSTENV_a16, opFSTENV_a16,
+        opFSTCW_a16,  opFSTCW_a16,  opFSTCW_a16,  opFSTCW_a16,  opFSTCW_a16,  opFSTCW_a16,  opFSTCW_a16,  opFSTCW_a16,
+
+        opFLDs_a16,   opFLDs_a16,   opFLDs_a16,   opFLDs_a16,   opFLDs_a16,   opFLDs_a16,   opFLDs_a16,   opFLDs_a16,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        opFSTs_a16,   opFSTs_a16,   opFSTs_a16,   opFSTs_a16,   opFSTs_a16,   opFSTs_a16,   opFSTs_a16,   opFSTs_a16,
+        opFSTPs_a16,  opFSTPs_a16,  opFSTPs_a16,  opFSTPs_a16,  opFSTPs_a16,  opFSTPs_a16,  opFSTPs_a16,  opFSTPs_a16,
+        opFLDENV_a16, opFLDENV_a16, opFLDENV_a16, opFLDENV_a16, opFLDENV_a16, opFLDENV_a16, opFLDENV_a16, opFLDENV_a16,
+        opFLDCW_a16,  opFLDCW_a16,  opFLDCW_a16,  opFLDCW_a16,  opFLDCW_a16,  opFLDCW_a16,  opFLDCW_a16,  opFLDCW_a16,
+        opFSTENV_a16, opFSTENV_a16, opFSTENV_a16, opFSTENV_a16, opFSTENV_a16, opFSTENV_a16, opFSTENV_a16, opFSTENV_a16,
+        opFSTCW_a16,  opFSTCW_a16,  opFSTCW_a16,  opFSTCW_a16,  opFSTCW_a16,  opFSTCW_a16,  opFSTCW_a16,  opFSTCW_a16,
+
+        opFLDs_a16,   opFLDs_a16,   opFLDs_a16,   opFLDs_a16,   opFLDs_a16,   opFLDs_a16,   opFLDs_a16,   opFLDs_a16,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        opFSTs_a16,   opFSTs_a16,   opFSTs_a16,   opFSTs_a16,   opFSTs_a16,   opFSTs_a16,   opFSTs_a16,   opFSTs_a16,
+        opFSTPs_a16,  opFSTPs_a16,  opFSTPs_a16,  opFSTPs_a16,  opFSTPs_a16,  opFSTPs_a16,  opFSTPs_a16,  opFSTPs_a16,
+        opFLDENV_a16, opFLDENV_a16, opFLDENV_a16, opFLDENV_a16, opFLDENV_a16, opFLDENV_a16, opFLDENV_a16, opFLDENV_a16,
+        opFLDCW_a16,  opFLDCW_a16,  opFLDCW_a16,  opFLDCW_a16,  opFLDCW_a16,  opFLDCW_a16,  opFLDCW_a16,  opFLDCW_a16,
+        opFSTENV_a16, opFSTENV_a16, opFSTENV_a16, opFSTENV_a16, opFSTENV_a16, opFSTENV_a16, opFSTENV_a16, opFSTENV_a16,
+        opFSTCW_a16,  opFSTCW_a16,  opFSTCW_a16,  opFSTCW_a16,  opFSTCW_a16,  opFSTCW_a16,  opFSTCW_a16,  opFSTCW_a16,
+
+        opFLD,        opFLD,        opFLD,        opFLD,        opFLD,        opFLD,        opFLD,        opFLD,
+        opFXCH,       opFXCH,       opFXCH,       opFXCH,       opFXCH,       opFXCH,       opFXCH,       opFXCH,
+        opFNOP,       ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        opFSTP,       opFSTP,       opFSTP,       opFSTP,       opFSTP,       opFSTP,       opFSTP,       opFSTP,  /*Invalid*/
+        opFCHS,       opFABS,       ILLEGAL_a16,  ILLEGAL_a16,  opFTST,       opFXAM,       opFTSTP,      ILLEGAL_a16,
+        opFLD1,       opFLDL2T,     opFLDL2E,     opFLDPI,      opFLDEG2,     opFLDLN2,     opFLDZ,       ILLEGAL_a16,
+        opF2XM1,      opFYL2X,      opFPTAN,      opFPATAN,     opFXTRACT,  opFPREM1,     opFDECSTP,    opFINCSTP,
+        opFPREM,      opFYL2XP1,    opFSQRT,      opFSINCOS,    opFRNDINT,    opFSCALE,     opFSIN,       opFCOS
+    // clang-format on
+};
+
+const OpFn OP_TABLE(fpu_cyrix_d9_a32)[256] = {
+    // clang-format off
+        opFLDs_a32,   opFLDs_a32,   opFLDs_a32,   opFLDs_a32,   opFLDs_a32,   opFLDs_a32,   opFLDs_a32,   opFLDs_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        opFSTs_a32,   opFSTs_a32,   opFSTs_a32,   opFSTs_a32,   opFSTs_a32,   opFSTs_a32,   opFSTs_a32,   opFSTs_a32,
+        opFSTPs_a32,  opFSTPs_a32,  opFSTPs_a32,  opFSTPs_a32,  opFSTPs_a32,  opFSTPs_a32,  opFSTPs_a32,  opFSTPs_a32,
+        opFLDENV_a32, opFLDENV_a32, opFLDENV_a32, opFLDENV_a32, opFLDENV_a32, opFLDENV_a32, opFLDENV_a32, opFLDENV_a32,
+        opFLDCW_a32,  opFLDCW_a32,  opFLDCW_a32,  opFLDCW_a32,  opFLDCW_a32,  opFLDCW_a32,  opFLDCW_a32,  opFLDCW_a32,
+        opFSTENV_a32, opFSTENV_a32, opFSTENV_a32, opFSTENV_a32, opFSTENV_a32, opFSTENV_a32, opFSTENV_a32, opFSTENV_a32,
+        opFSTCW_a32,  opFSTCW_a32,  opFSTCW_a32,  opFSTCW_a32,  opFSTCW_a32,  opFSTCW_a32,  opFSTCW_a32,  opFSTCW_a32,
+
+        opFLDs_a32,   opFLDs_a32,   opFLDs_a32,   opFLDs_a32,   opFLDs_a32,   opFLDs_a32,   opFLDs_a32,   opFLDs_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        opFSTs_a32,   opFSTs_a32,   opFSTs_a32,   opFSTs_a32,   opFSTs_a32,   opFSTs_a32,   opFSTs_a32,   opFSTs_a32,
+        opFSTPs_a32,  opFSTPs_a32,  opFSTPs_a32,  opFSTPs_a32,  opFSTPs_a32,  opFSTPs_a32,  opFSTPs_a32,  opFSTPs_a32,
+        opFLDENV_a32, opFLDENV_a32, opFLDENV_a32, opFLDENV_a32, opFLDENV_a32, opFLDENV_a32, opFLDENV_a32, opFLDENV_a32,
+        opFLDCW_a32,  opFLDCW_a32,  opFLDCW_a32,  opFLDCW_a32,  opFLDCW_a32,  opFLDCW_a32,  opFLDCW_a32,  opFLDCW_a32,
+        opFSTENV_a32, opFSTENV_a32, opFSTENV_a32, opFSTENV_a32, opFSTENV_a32, opFSTENV_a32, opFSTENV_a32, opFSTENV_a32,
+        opFSTCW_a32,  opFSTCW_a32,  opFSTCW_a32,  opFSTCW_a32,  opFSTCW_a32,  opFSTCW_a32,  opFSTCW_a32,  opFSTCW_a32,
+
+        opFLDs_a32,   opFLDs_a32,   opFLDs_a32,   opFLDs_a32,   opFLDs_a32,   opFLDs_a32,   opFLDs_a32,   opFLDs_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        opFSTs_a32,   opFSTs_a32,   opFSTs_a32,   opFSTs_a32,   opFSTs_a32,   opFSTs_a32,   opFSTs_a32,   opFSTs_a32,
+        opFSTPs_a32,  opFSTPs_a32,  opFSTPs_a32,  opFSTPs_a32,  opFSTPs_a32,  opFSTPs_a32,  opFSTPs_a32,  opFSTPs_a32,
+        opFLDENV_a32, opFLDENV_a32, opFLDENV_a32, opFLDENV_a32, opFLDENV_a32, opFLDENV_a32, opFLDENV_a32, opFLDENV_a32,
+        opFLDCW_a32,  opFLDCW_a32,  opFLDCW_a32,  opFLDCW_a32,  opFLDCW_a32,  opFLDCW_a32,  opFLDCW_a32,  opFLDCW_a32,
+        opFSTENV_a32, opFSTENV_a32, opFSTENV_a32, opFSTENV_a32, opFSTENV_a32, opFSTENV_a32, opFSTENV_a32, opFSTENV_a32,
+        opFSTCW_a32,  opFSTCW_a32,  opFSTCW_a32,  opFSTCW_a32,  opFSTCW_a32,  opFSTCW_a32,  opFSTCW_a32,  opFSTCW_a32,
+
+        opFLD,        opFLD,        opFLD,        opFLD,        opFLD,        opFLD,        opFLD,        opFLD,
+        opFXCH,       opFXCH,       opFXCH,       opFXCH,       opFXCH,       opFXCH,       opFXCH,       opFXCH,
+        opFNOP,       ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        opFSTP,       opFSTP,       opFSTP,       opFSTP,       opFSTP,       opFSTP,       opFSTP,       opFSTP,  /*Invalid*/
+        opFCHS,       opFABS,       ILLEGAL_a32,  ILLEGAL_a32,  opFTST,       opFXAM,       opFTSTP,      ILLEGAL_a32,
+        opFLD1,       opFLDL2T,     opFLDL2E,     opFLDPI,      opFLDEG2,     opFLDLN2,     opFLDZ,       ILLEGAL_a32,
+        opF2XM1,      opFYL2X,      opFPTAN,      opFPATAN,     opFXTRACT,    opFPREM1,     opFDECSTP,    opFINCSTP,
+        opFPREM,      opFYL2XP1,    opFSQRT,      opFSINCOS,    opFRNDINT,    opFSCALE,     opFSIN,       opFCOS
+    // clang-format on
+};
+
 const OpFn OP_TABLE(fpu_287_da_a16)[256] = {
     // clang-format off
         opFADDil_a16,  opFADDil_a16,  opFADDil_a16,  opFADDil_a16,  opFADDil_a16,  opFADDil_a16,  opFADDil_a16,  opFADDil_a16,
@@ -3009,6 +3430,87 @@ const OpFn OP_TABLE(fpu_686_db_a32)[256] = {
 };
 #    endif
 
+#    ifndef OPS_286_386
+const OpFn OP_TABLE(fpu_cyrix_686_db_a16)[256] = {
+    // clang-format off
+        opFILDil_a16,  opFILDil_a16,  opFILDil_a16,  opFILDil_a16,  opFILDil_a16,  opFILDil_a16,  opFILDil_a16,  opFILDil_a16,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        opFISTil_a16,  opFISTil_a16,  opFISTil_a16,  opFISTil_a16,  opFISTil_a16,  opFISTil_a16,  opFISTil_a16,  opFISTil_a16,
+        opFISTPil_a16, opFISTPil_a16, opFISTPil_a16, opFISTPil_a16, opFISTPil_a16, opFISTPil_a16, opFISTPil_a16, opFISTPil_a16,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        opFLDe_a16,    opFLDe_a16,    opFLDe_a16,    opFLDe_a16,    opFLDe_a16,    opFLDe_a16,    opFLDe_a16,    opFLDe_a16,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        opFSTPe_a16,   opFSTPe_a16,   opFSTPe_a16,   opFSTPe_a16,   opFSTPe_a16,   opFSTPe_a16,   opFSTPe_a16,   opFSTPe_a16,
+
+        opFILDil_a16,  opFILDil_a16,  opFILDil_a16,  opFILDil_a16,  opFILDil_a16,  opFILDil_a16,  opFILDil_a16,  opFILDil_a16,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        opFISTil_a16,  opFISTil_a16,  opFISTil_a16,  opFISTil_a16,  opFISTil_a16,  opFISTil_a16,  opFISTil_a16,  opFISTil_a16,
+        opFISTPil_a16, opFISTPil_a16, opFISTPil_a16, opFISTPil_a16, opFISTPil_a16, opFISTPil_a16, opFISTPil_a16, opFISTPil_a16,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        opFLDe_a16,    opFLDe_a16,    opFLDe_a16,    opFLDe_a16,    opFLDe_a16,    opFLDe_a16,    opFLDe_a16,    opFLDe_a16,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        opFSTPe_a16,   opFSTPe_a16,   opFSTPe_a16,   opFSTPe_a16,   opFSTPe_a16,   opFSTPe_a16,   opFSTPe_a16,   opFSTPe_a16,
+
+        opFILDil_a16,  opFILDil_a16,  opFILDil_a16,  opFILDil_a16,  opFILDil_a16,  opFILDil_a16,  opFILDil_a16,  opFILDil_a16,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        opFISTil_a16,  opFISTil_a16,  opFISTil_a16,  opFISTil_a16,  opFISTil_a16,  opFISTil_a16,  opFISTil_a16,  opFISTil_a16,
+        opFISTPil_a16, opFISTPil_a16, opFISTPil_a16, opFISTPil_a16, opFISTPil_a16, opFISTPil_a16, opFISTPil_a16, opFISTPil_a16,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        opFLDe_a16,    opFLDe_a16,    opFLDe_a16,    opFLDe_a16,    opFLDe_a16,    opFLDe_a16,    opFLDe_a16,    opFLDe_a16,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        opFSTPe_a16,   opFSTPe_a16,   opFSTPe_a16,   opFSTPe_a16,   opFSTPe_a16,   opFSTPe_a16,   opFSTPe_a16,   opFSTPe_a16,
+
+        opFCMOVNB,     opFCMOVNB,     opFCMOVNB,     opFCMOVNB,     opFCMOVNB,     opFCMOVNB,     opFCMOVNB,     opFCMOVNB,
+        opFCMOVNE,     opFCMOVNE,     opFCMOVNE,     opFCMOVNE,     opFCMOVNE,     opFCMOVNE,     opFCMOVNE,     opFCMOVNE,
+        opFCMOVNBE,    opFCMOVNBE,    opFCMOVNBE,    opFCMOVNBE,    opFCMOVNBE,    opFCMOVNBE,    opFCMOVNBE,    opFCMOVNBE,
+        opFCMOVNU,     opFCMOVNU,     opFCMOVNU,     opFCMOVNU,     opFCMOVNU,     opFCMOVNU,     opFCMOVNU,     opFCMOVNU,
+        opFNOP,        opFNOP,        opFCLEX,       opFINIT,       opFNOP,        opFNOP,        ILLEGAL_a16,       ILLEGAL_a16,
+        opFUCOMI,      opFUCOMI,      opFUCOMI,      opFUCOMI,      opFUCOMI,      opFUCOMI,      opFUCOMI,      opFUCOMI,
+        opFCOMI,       opFCOMI,       opFCOMI,       opFCOMI,       opFCOMI,       opFCOMI,       opFCOMI,       opFCOMI,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,      opFRINT2,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+    // clang-format on
+};
+const OpFn OP_TABLE(fpu_cyrix_686_db_a32)[256] = {
+    // clang-format off
+        opFILDil_a32,  opFILDil_a32,  opFILDil_a32,  opFILDil_a32,  opFILDil_a32,  opFILDil_a32,  opFILDil_a32,  opFILDil_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        opFISTil_a32,  opFISTil_a32,  opFISTil_a32,  opFISTil_a32,  opFISTil_a32,  opFISTil_a32,  opFISTil_a32,  opFISTil_a32,
+        opFISTPil_a32, opFISTPil_a32, opFISTPil_a32, opFISTPil_a32, opFISTPil_a32, opFISTPil_a32, opFISTPil_a32, opFISTPil_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        opFLDe_a32,    opFLDe_a32,    opFLDe_a32,    opFLDe_a32,    opFLDe_a32,    opFLDe_a32,    opFLDe_a32,    opFLDe_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        opFSTPe_a32,   opFSTPe_a32,   opFSTPe_a32,   opFSTPe_a32,   opFSTPe_a32,   opFSTPe_a32,   opFSTPe_a32,   opFSTPe_a32,
+
+        opFILDil_a32,  opFILDil_a32,  opFILDil_a32,  opFILDil_a32,  opFILDil_a32,  opFILDil_a32,  opFILDil_a32,  opFILDil_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        opFISTil_a32,  opFISTil_a32,  opFISTil_a32,  opFISTil_a32,  opFISTil_a32,  opFISTil_a32,  opFISTil_a32,  opFISTil_a32,
+        opFISTPil_a32, opFISTPil_a32, opFISTPil_a32, opFISTPil_a32, opFISTPil_a32, opFISTPil_a32, opFISTPil_a32, opFISTPil_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        opFLDe_a32,    opFLDe_a32,    opFLDe_a32,    opFLDe_a32,    opFLDe_a32,    opFLDe_a32,    opFLDe_a32,    opFLDe_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        opFSTPe_a32,   opFSTPe_a32,   opFSTPe_a32,   opFSTPe_a32,   opFSTPe_a32,   opFSTPe_a32,   opFSTPe_a32,   opFSTPe_a32,
+
+        opFILDil_a32,  opFILDil_a32,  opFILDil_a32,  opFILDil_a32,  opFILDil_a32,  opFILDil_a32,  opFILDil_a32,  opFILDil_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        opFISTil_a32,  opFISTil_a32,  opFISTil_a32,  opFISTil_a32,  opFISTil_a32,  opFISTil_a32,  opFISTil_a32,  opFISTil_a32,
+        opFISTPil_a32, opFISTPil_a32, opFISTPil_a32, opFISTPil_a32, opFISTPil_a32, opFISTPil_a32, opFISTPil_a32, opFISTPil_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        opFLDe_a32,    opFLDe_a32,    opFLDe_a32,    opFLDe_a32,    opFLDe_a32,    opFLDe_a32,    opFLDe_a32,    opFLDe_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        opFSTPe_a32,   opFSTPe_a32,   opFSTPe_a32,   opFSTPe_a32,   opFSTPe_a32,   opFSTPe_a32,   opFSTPe_a32,   opFSTPe_a32,
+
+        opFCMOVNB,     opFCMOVNB,     opFCMOVNB,     opFCMOVNB,     opFCMOVNB,     opFCMOVNB,     opFCMOVNB,     opFCMOVNB,
+        opFCMOVNE,     opFCMOVNE,     opFCMOVNE,     opFCMOVNE,     opFCMOVNE,     opFCMOVNE,     opFCMOVNE,     opFCMOVNE,
+        opFCMOVNBE,    opFCMOVNBE,    opFCMOVNBE,    opFCMOVNBE,    opFCMOVNBE,    opFCMOVNBE,    opFCMOVNBE,    opFCMOVNBE,
+        opFCMOVNU,     opFCMOVNU,     opFCMOVNU,     opFCMOVNU,     opFCMOVNU,     opFCMOVNU,     opFCMOVNU,     opFCMOVNU,
+        opFNOP,        opFNOP,        opFCLEX,       opFINIT,       opFNOP,        opFNOP,        ILLEGAL_a32,       ILLEGAL_a32,
+        opFUCOMI,      opFUCOMI,      opFUCOMI,      opFUCOMI,      opFUCOMI,      opFUCOMI,      opFUCOMI,      opFUCOMI,
+        opFCOMI,       opFCOMI,       opFCOMI,       opFCOMI,       opFCOMI,       opFCOMI,       opFCOMI,       opFCOMI,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,      opFRINT2,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+    // clang-format on
+};
+#    endif
+
 const OpFn OP_TABLE(fpu_287_dc_a16)[32] = {
     // clang-format off
         opFADDd_a16, opFMULd_a16, opFCOMd_a16, opFCOMPd_a16, opFSUBd_a16, opFSUBRd_a16, opFDIVd_a16, opFDIVRd_a16,
@@ -3205,6 +3707,86 @@ const OpFn OP_TABLE(fpu_dd_a32)[256] = {
     // clang-format on
 };
 
+const OpFn OP_TABLE(fpu_cyrix_dd_a16)[256] = {
+    // clang-format off
+        opFLDd_a16,    opFLDd_a16,    opFLDd_a16,    opFLDd_a16,    opFLDd_a16,    opFLDd_a16,    opFLDd_a16,    opFLDd_a16,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        opFSTd_a16,    opFSTd_a16,    opFSTd_a16,    opFSTd_a16,    opFSTd_a16,    opFSTd_a16,    opFSTd_a16,    opFSTd_a16,
+        opFSTPd_a16,   opFSTPd_a16,   opFSTPd_a16,   opFSTPd_a16,   opFSTPd_a16,   opFSTPd_a16,   opFSTPd_a16,   opFSTPd_a16,
+        opFSTOR_a16,   opFSTOR_a16,   opFSTOR_a16,   opFSTOR_a16,   opFSTOR_a16,   opFSTOR_a16,   opFSTOR_a16,   opFSTOR_a16,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        opFSAVE_a16,   opFSAVE_a16,   opFSAVE_a16,   opFSAVE_a16,   opFSAVE_a16,   opFSAVE_a16,   opFSAVE_a16,   opFSAVE_a16,
+        opFSTSW_a16,   opFSTSW_a16,   opFSTSW_a16,   opFSTSW_a16,   opFSTSW_a16,   opFSTSW_a16,   opFSTSW_a16,   opFSTSW_a16,
+
+        opFLDd_a16,    opFLDd_a16,    opFLDd_a16,    opFLDd_a16,    opFLDd_a16,    opFLDd_a16,    opFLDd_a16,    opFLDd_a16,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        opFSTd_a16,    opFSTd_a16,    opFSTd_a16,    opFSTd_a16,    opFSTd_a16,    opFSTd_a16,    opFSTd_a16,    opFSTd_a16,
+        opFSTPd_a16,   opFSTPd_a16,   opFSTPd_a16,   opFSTPd_a16,   opFSTPd_a16,   opFSTPd_a16,   opFSTPd_a16,   opFSTPd_a16,
+        opFSTOR_a16,   opFSTOR_a16,   opFSTOR_a16,   opFSTOR_a16,   opFSTOR_a16,   opFSTOR_a16,   opFSTOR_a16,   opFSTOR_a16,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        opFSAVE_a16,   opFSAVE_a16,   opFSAVE_a16,   opFSAVE_a16,   opFSAVE_a16,   opFSAVE_a16,   opFSAVE_a16,   opFSAVE_a16,
+        opFSTSW_a16,   opFSTSW_a16,   opFSTSW_a16,   opFSTSW_a16,   opFSTSW_a16,   opFSTSW_a16,   opFSTSW_a16,   opFSTSW_a16,
+
+        opFLDd_a16,    opFLDd_a16,    opFLDd_a16,    opFLDd_a16,    opFLDd_a16,    opFLDd_a16,    opFLDd_a16,    opFLDd_a16,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        opFSTd_a16,    opFSTd_a16,    opFSTd_a16,    opFSTd_a16,    opFSTd_a16,    opFSTd_a16,    opFSTd_a16,    opFSTd_a16,
+        opFSTPd_a16,   opFSTPd_a16,   opFSTPd_a16,   opFSTPd_a16,   opFSTPd_a16,   opFSTPd_a16,   opFSTPd_a16,   opFSTPd_a16,
+        opFSTOR_a16,   opFSTOR_a16,   opFSTOR_a16,   opFSTOR_a16,   opFSTOR_a16,   opFSTOR_a16,   opFSTOR_a16,   opFSTOR_a16,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        opFSAVE_a16,   opFSAVE_a16,   opFSAVE_a16,   opFSAVE_a16,   opFSAVE_a16,   opFSAVE_a16,   opFSAVE_a16,   opFSAVE_a16,
+        opFSTSW_a16,   opFSTSW_a16,   opFSTSW_a16,   opFSTSW_a16,   opFSTSW_a16,   opFSTSW_a16,   opFSTSW_a16,   opFSTSW_a16,
+
+        opFFREE,       opFFREE,       opFFREE,       opFFREE,       opFFREE,       opFFREE,       opFFREE,       opFFREE,
+        opFXCH,        opFXCH,        opFXCH,        opFXCH,        opFXCH,        opFXCH,        opFXCH,        opFXCH,
+        opFST,         opFST,         opFST,         opFST,         opFST,         opFST,         opFST,         opFST,
+        opFSTP,        opFSTP,        opFSTP,        opFSTP,        opFSTP,        opFSTP,        opFSTP,        opFSTP,
+        opFUCOM,       opFUCOM,       opFUCOM,       opFUCOM,       opFUCOM,       opFUCOM,       opFUCOM,       opFUCOM,
+        opFUCOMP,      opFUCOMP,      opFUCOMP,      opFUCOMP,      opFUCOMP,      opFUCOMP,      opFUCOMP,      opFUCOMP,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,    opFRICHOP,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+    // clang-format on
+};
+
+const OpFn OP_TABLE(fpu_cyrix_dd_a32)[256] = {
+    // clang-format off
+        opFLDd_a32,    opFLDd_a32,    opFLDd_a32,    opFLDd_a32,    opFLDd_a32,    opFLDd_a32,    opFLDd_a32,    opFLDd_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        opFSTd_a32,    opFSTd_a32,    opFSTd_a32,    opFSTd_a32,    opFSTd_a32,    opFSTd_a32,    opFSTd_a32,    opFSTd_a32,
+        opFSTPd_a32,   opFSTPd_a32,   opFSTPd_a32,   opFSTPd_a32,   opFSTPd_a32,   opFSTPd_a32,   opFSTPd_a32,   opFSTPd_a32,
+        opFSTOR_a32,   opFSTOR_a32,   opFSTOR_a32,   opFSTOR_a32,   opFSTOR_a32,   opFSTOR_a32,   opFSTOR_a32,   opFSTOR_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        opFSAVE_a32,   opFSAVE_a32,   opFSAVE_a32,   opFSAVE_a32,   opFSAVE_a32,   opFSAVE_a32,   opFSAVE_a32,   opFSAVE_a32,
+        opFSTSW_a32,   opFSTSW_a32,   opFSTSW_a32,   opFSTSW_a32,   opFSTSW_a32,   opFSTSW_a32,   opFSTSW_a32,   opFSTSW_a32,
+
+        opFLDd_a32,    opFLDd_a32,    opFLDd_a32,    opFLDd_a32,    opFLDd_a32,    opFLDd_a32,    opFLDd_a32,    opFLDd_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        opFSTd_a32,    opFSTd_a32,    opFSTd_a32,    opFSTd_a32,    opFSTd_a32,    opFSTd_a32,    opFSTd_a32,    opFSTd_a32,
+        opFSTPd_a32,   opFSTPd_a32,   opFSTPd_a32,   opFSTPd_a32,   opFSTPd_a32,   opFSTPd_a32,   opFSTPd_a32,   opFSTPd_a32,
+        opFSTOR_a32,   opFSTOR_a32,   opFSTOR_a32,   opFSTOR_a32,   opFSTOR_a32,   opFSTOR_a32,   opFSTOR_a32,   opFSTOR_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        opFSAVE_a32,   opFSAVE_a32,   opFSAVE_a32,   opFSAVE_a32,   opFSAVE_a32,   opFSAVE_a32,   opFSAVE_a32,   opFSAVE_a32,
+        opFSTSW_a32,   opFSTSW_a32,   opFSTSW_a32,   opFSTSW_a32,   opFSTSW_a32,   opFSTSW_a32,   opFSTSW_a32,   opFSTSW_a32,
+
+        opFLDd_a32,    opFLDd_a32,    opFLDd_a32,    opFLDd_a32,    opFLDd_a32,    opFLDd_a32,    opFLDd_a32,    opFLDd_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        opFSTd_a32,    opFSTd_a32,    opFSTd_a32,    opFSTd_a32,    opFSTd_a32,    opFSTd_a32,    opFSTd_a32,    opFSTd_a32,
+        opFSTPd_a32,   opFSTPd_a32,   opFSTPd_a32,   opFSTPd_a32,   opFSTPd_a32,   opFSTPd_a32,   opFSTPd_a32,   opFSTPd_a32,
+        opFSTOR_a32,   opFSTOR_a32,   opFSTOR_a32,   opFSTOR_a32,   opFSTOR_a32,   opFSTOR_a32,   opFSTOR_a32,   opFSTOR_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        opFSAVE_a32,   opFSAVE_a32,   opFSAVE_a32,   opFSAVE_a32,   opFSAVE_a32,   opFSAVE_a32,   opFSAVE_a32,   opFSAVE_a32,
+        opFSTSW_a32,   opFSTSW_a32,   opFSTSW_a32,   opFSTSW_a32,   opFSTSW_a32,   opFSTSW_a32,   opFSTSW_a32,   opFSTSW_a32,
+
+        opFFREE,       opFFREE,       opFFREE,       opFFREE,       opFFREE,       opFFREE,       opFFREE,       opFFREE,
+        opFXCH,        opFXCH,        opFXCH,        opFXCH,        opFXCH,        opFXCH,        opFXCH,        opFXCH,
+        opFST,         opFST,         opFST,         opFST,         opFST,         opFST,         opFST,         opFST,
+        opFSTP,        opFSTP,        opFSTP,        opFSTP,        opFSTP,        opFSTP,        opFSTP,        opFSTP,
+        opFUCOM,       opFUCOM,       opFUCOM,       opFUCOM,       opFUCOM,       opFUCOM,       opFUCOM,       opFUCOM,
+        opFUCOMP,      opFUCOMP,      opFUCOMP,      opFUCOMP,      opFUCOMP,      opFUCOMP,      opFUCOMP,      opFUCOMP,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,    opFRICHOP,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+    // clang-format on
+};
+
 const OpFn OP_TABLE(fpu_287_de_a16)[256] = {
     // clang-format off
         opFADDiw_a16,  opFADDiw_a16,  opFADDiw_a16,  opFADDiw_a16,  opFADDiw_a16,  opFADDiw_a16,  opFADDiw_a16,  opFADDiw_a16,
@@ -3607,6 +4189,88 @@ const OpFn OP_TABLE(fpu_686_df_a32)[256] = {
 };
 #    endif
 
+#    ifndef OPS_286_386
+const OpFn OP_TABLE(fpu_cyrix_686_df_a16)[256] = {
+    // clang-format off
+        opFILDiw_a16,  opFILDiw_a16,  opFILDiw_a16,  opFILDiw_a16,  opFILDiw_a16,  opFILDiw_a16,  opFILDiw_a16,  opFILDiw_a16,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        opFISTiw_a16,  opFISTiw_a16,  opFISTiw_a16,  opFISTiw_a16,  opFISTiw_a16,  opFISTiw_a16,  opFISTiw_a16,  opFISTiw_a16,
+        opFISTPiw_a16, opFISTPiw_a16, opFISTPiw_a16, opFISTPiw_a16, opFISTPiw_a16, opFISTPiw_a16, opFISTPiw_a16, opFISTPiw_a16,
+        FBLD_a16,     FBLD_a16,     FBLD_a16,     FBLD_a16,     FBLD_a16,     FBLD_a16,     FBLD_a16,     FBLD_a16,
+        opFILDiq_a16,  opFILDiq_a16,  opFILDiq_a16,  opFILDiq_a16,  opFILDiq_a16,  opFILDiq_a16,  opFILDiq_a16,  opFILDiq_a16,
+        FBSTP_a16,     FBSTP_a16,     FBSTP_a16,     FBSTP_a16,     FBSTP_a16,     FBSTP_a16,     FBSTP_a16,     FBSTP_a16,
+        FISTPiq_a16,   FISTPiq_a16,   FISTPiq_a16,   FISTPiq_a16,   FISTPiq_a16,   FISTPiq_a16,   FISTPiq_a16,   FISTPiq_a16,
+
+        opFILDiw_a16,  opFILDiw_a16,  opFILDiw_a16,  opFILDiw_a16,  opFILDiw_a16,  opFILDiw_a16,  opFILDiw_a16,  opFILDiw_a16,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        opFISTiw_a16,  opFISTiw_a16,  opFISTiw_a16,  opFISTiw_a16,  opFISTiw_a16,  opFISTiw_a16,  opFISTiw_a16,  opFISTiw_a16,
+        opFISTPiw_a16, opFISTPiw_a16, opFISTPiw_a16, opFISTPiw_a16, opFISTPiw_a16, opFISTPiw_a16, opFISTPiw_a16, opFISTPiw_a16,
+        FBLD_a16,     FBLD_a16,     FBLD_a16,     FBLD_a16,     FBLD_a16,     FBLD_a16,     FBLD_a16,     FBLD_a16,
+        opFILDiq_a16,  opFILDiq_a16,  opFILDiq_a16,  opFILDiq_a16,  opFILDiq_a16,  opFILDiq_a16,  opFILDiq_a16,  opFILDiq_a16,
+        FBSTP_a16,     FBSTP_a16,     FBSTP_a16,     FBSTP_a16,     FBSTP_a16,     FBSTP_a16,     FBSTP_a16,     FBSTP_a16,
+        FISTPiq_a16,   FISTPiq_a16,   FISTPiq_a16,   FISTPiq_a16,   FISTPiq_a16,   FISTPiq_a16,   FISTPiq_a16,   FISTPiq_a16,
+
+        opFILDiw_a16,  opFILDiw_a16,  opFILDiw_a16,  opFILDiw_a16,  opFILDiw_a16,  opFILDiw_a16,  opFILDiw_a16,  opFILDiw_a16,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        opFISTiw_a16,  opFISTiw_a16,  opFISTiw_a16,  opFISTiw_a16,  opFISTiw_a16,  opFISTiw_a16,  opFISTiw_a16,  opFISTiw_a16,
+        opFISTPiw_a16, opFISTPiw_a16, opFISTPiw_a16, opFISTPiw_a16, opFISTPiw_a16, opFISTPiw_a16, opFISTPiw_a16, opFISTPiw_a16,
+        FBLD_a16,     FBLD_a16,     FBLD_a16,     FBLD_a16,     FBLD_a16,     FBLD_a16,     FBLD_a16,     FBLD_a16,
+        opFILDiq_a16,  opFILDiq_a16,  opFILDiq_a16,  opFILDiq_a16,  opFILDiq_a16,  opFILDiq_a16,  opFILDiq_a16,  opFILDiq_a16,
+        FBSTP_a16,     FBSTP_a16,     FBSTP_a16,     FBSTP_a16,     FBSTP_a16,     FBSTP_a16,     FBSTP_a16,     FBSTP_a16,
+        FISTPiq_a16,   FISTPiq_a16,   FISTPiq_a16,   FISTPiq_a16,   FISTPiq_a16,   FISTPiq_a16,   FISTPiq_a16,   FISTPiq_a16,
+
+        opFFREEP,      opFFREEP,      opFFREEP,      opFFREEP,      opFFREEP,      opFFREEP,      opFFREEP,      opFFREEP,
+        opFXCH,        opFXCH,        opFXCH,        opFXCH,        opFXCH,        opFXCH,        opFXCH,        opFXCH,
+        opFSTP,        opFSTP,        opFSTP,        opFSTP,        opFSTP,        opFSTP,        opFSTP,        opFSTP,
+        opFSTP,        opFSTP,        opFSTP,        opFSTP,        opFSTP,        opFSTP,        opFSTP,        opFSTP,
+        opFSTSW_AX,   ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+        opFUCOMIP,     opFUCOMIP,     opFUCOMIP,     opFUCOMIP,     opFUCOMIP,     opFUCOMIP,     opFUCOMIP,     opFUCOMIP,
+        opFCOMIP,      opFCOMIP,      opFCOMIP,      opFCOMIP,      opFCOMIP,      opFCOMIP,      opFCOMIP,      opFCOMIP,
+        ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,      opFRINEAR,  ILLEGAL_a16,  ILLEGAL_a16,  ILLEGAL_a16,
+    // clang-format on
+};
+
+const OpFn OP_TABLE(fpu_cyrix_686_df_a32)[256] = {
+    // clang-format off
+        opFILDiw_a32,  opFILDiw_a32,  opFILDiw_a32,  opFILDiw_a32,  opFILDiw_a32,  opFILDiw_a32,  opFILDiw_a32,  opFILDiw_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        opFISTiw_a32,  opFISTiw_a32,  opFISTiw_a32,  opFISTiw_a32,  opFISTiw_a32,  opFISTiw_a32,  opFISTiw_a32,  opFISTiw_a32,
+        opFISTPiw_a32, opFISTPiw_a32, opFISTPiw_a32, opFISTPiw_a32, opFISTPiw_a32, opFISTPiw_a32, opFISTPiw_a32, opFISTPiw_a32,
+        FBLD_a32,     FBLD_a32,     FBLD_a32,     FBLD_a32,     FBLD_a32,     FBLD_a32,     FBLD_a32,     FBLD_a32,
+        opFILDiq_a32,  opFILDiq_a32,  opFILDiq_a32,  opFILDiq_a32,  opFILDiq_a32,  opFILDiq_a32,  opFILDiq_a32,  opFILDiq_a32,
+        FBSTP_a32,     FBSTP_a32,     FBSTP_a32,     FBSTP_a32,     FBSTP_a32,     FBSTP_a32,     FBSTP_a32,     FBSTP_a32,
+        FISTPiq_a32,   FISTPiq_a32,   FISTPiq_a32,   FISTPiq_a32,   FISTPiq_a32,   FISTPiq_a32,   FISTPiq_a32,   FISTPiq_a32,
+
+        opFILDiw_a32,  opFILDiw_a32,  opFILDiw_a32,  opFILDiw_a32,  opFILDiw_a32,  opFILDiw_a32,  opFILDiw_a32,  opFILDiw_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        opFISTiw_a32,  opFISTiw_a32,  opFISTiw_a32,  opFISTiw_a32,  opFISTiw_a32,  opFISTiw_a32,  opFISTiw_a32,  opFISTiw_a32,
+        opFISTPiw_a32, opFISTPiw_a32, opFISTPiw_a32, opFISTPiw_a32, opFISTPiw_a32, opFISTPiw_a32, opFISTPiw_a32, opFISTPiw_a32,
+        FBLD_a32,     FBLD_a32,     FBLD_a32,     FBLD_a32,     FBLD_a32,     FBLD_a32,     FBLD_a32,     FBLD_a32,
+        opFILDiq_a32,  opFILDiq_a32,  opFILDiq_a32,  opFILDiq_a32,  opFILDiq_a32,  opFILDiq_a32,  opFILDiq_a32,  opFILDiq_a32,
+        FBSTP_a32,     FBSTP_a32,     FBSTP_a32,     FBSTP_a32,     FBSTP_a32,     FBSTP_a32,     FBSTP_a32,     FBSTP_a32,
+        FISTPiq_a32,   FISTPiq_a32,   FISTPiq_a32,   FISTPiq_a32,   FISTPiq_a32,   FISTPiq_a32,   FISTPiq_a32,   FISTPiq_a32,
+
+        opFILDiw_a32,  opFILDiw_a32,  opFILDiw_a32,  opFILDiw_a32,  opFILDiw_a32,  opFILDiw_a32,  opFILDiw_a32,  opFILDiw_a32,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        opFISTiw_a32,  opFISTiw_a32,  opFISTiw_a32,  opFISTiw_a32,  opFISTiw_a32,  opFISTiw_a32,  opFISTiw_a32,  opFISTiw_a32,
+        opFISTPiw_a32, opFISTPiw_a32, opFISTPiw_a32, opFISTPiw_a32, opFISTPiw_a32, opFISTPiw_a32, opFISTPiw_a32, opFISTPiw_a32,
+        FBLD_a32,     FBLD_a32,     FBLD_a32,     FBLD_a32,     FBLD_a32,     FBLD_a32,     FBLD_a32,     FBLD_a32,
+        opFILDiq_a32,  opFILDiq_a32,  opFILDiq_a32,  opFILDiq_a32,  opFILDiq_a32,  opFILDiq_a32,  opFILDiq_a32,  opFILDiq_a32,
+        FBSTP_a32,     FBSTP_a32,     FBSTP_a32,     FBSTP_a32,     FBSTP_a32,     FBSTP_a32,     FBSTP_a32,     FBSTP_a32,
+        FISTPiq_a32,   FISTPiq_a32,   FISTPiq_a32,   FISTPiq_a32,   FISTPiq_a32,   FISTPiq_a32,   FISTPiq_a32,   FISTPiq_a32,
+
+        opFFREEP,      opFFREEP,      opFFREEP,      opFFREEP,      opFFREEP,      opFFREEP,      opFFREEP,      opFFREEP,
+        opFXCH,        opFXCH,        opFXCH,        opFXCH,        opFXCH,        opFXCH,        opFXCH,        opFXCH,
+        opFSTP,        opFSTP,        opFSTP,        opFSTP,        opFSTP,        opFSTP,        opFSTP,        opFSTP,
+        opFSTP,        opFSTP,        opFSTP,        opFSTP,        opFSTP,        opFSTP,        opFSTP,        opFSTP,
+        opFSTSW_AX,   ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+        opFUCOMIP,     opFUCOMIP,     opFUCOMIP,     opFUCOMIP,     opFUCOMIP,     opFUCOMIP,     opFUCOMIP,     opFUCOMIP,
+        opFCOMIP,      opFCOMIP,      opFCOMIP,      opFCOMIP,      opFCOMIP,      opFCOMIP,      opFCOMIP,      opFCOMIP,
+        ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,      opFRINEAR,  ILLEGAL_a32,  ILLEGAL_a32,  ILLEGAL_a32,
+    // clang-format on
+};
+#    endif
+
 const OpFn OP_TABLE(nofpu_a16)[256] = {
     // clang-format off
         op_nofpu_a16, op_nofpu_a16, op_nofpu_a16, op_nofpu_a16, op_nofpu_a16, op_nofpu_a16, op_nofpu_a16, op_nofpu_a16,
diff --git a/src/cpu/x87_ops_misc.h b/src/cpu/x87_ops_misc.h
index c1f09fda9..cbad12e56 100644
--- a/src/cpu/x87_ops_misc.h
+++ b/src/cpu/x87_ops_misc.h
@@ -554,6 +554,22 @@ opFTST(UNUSED(uint32_t fetchdat))
     return 0;
 }
 
+static int
+opFTSTP(UNUSED(uint32_t fetchdat))
+{
+    FP_ENTER();
+    cpu_state.pc++;
+    cpu_state.npxs &= ~(FPU_SW_C0 | FPU_SW_C2 | FPU_SW_C3);
+    if (ST(0) == 0.0)
+        cpu_state.npxs |= FPU_SW_C3;
+    else if (ST(0) < 0.0)
+        cpu_state.npxs |= FPU_SW_C0;
+    x87_pop();
+    CLOCK_CYCLES_FPU((fpu_type >= FPU_487SX) ? (x87_timings.ftst) : (x87_timings.ftst * cpu_multi));
+    CONCURRENCY_CYCLES((fpu_type >= FPU_487SX) ? (x87_concurrency.ftst) : (x87_concurrency.ftst * cpu_multi));
+    return 0;
+}
+
 static int
 opFXAM(UNUSED(uint32_t fetchdat))
 {
@@ -838,6 +854,67 @@ opFRNDINT(UNUSED(uint32_t fetchdat))
     return 0;
 }
 
+static int
+opFRINT2(UNUSED(uint32_t fetchdat))
+{
+    double dst0, st0, integral, frac;
+    int prevRound;
+
+    FP_ENTER();
+    cpu_state.pc++;
+    prevRound = fegetround();
+    fesetround(FE_TONEAREST);
+    st0 = ST(0);
+    frac = modf(st0, &integral);
+    if (frac == 0.5 || frac == -0.5) {
+        dst0 = (st0 < 0) ? floor(st0) : ceil(st0);
+    } else {
+        dst0 = round(st0);
+    }
+    fesetround(prevRound);
+    ST(0) = (double) dst0;
+    FP_TAG_VALID;
+    CLOCK_CYCLES_FPU((fpu_type >= FPU_487SX) ? (x87_timings.frndint) : (x87_timings.frndint * cpu_multi));
+    CONCURRENCY_CYCLES((fpu_type >= FPU_487SX) ? (x87_concurrency.frndint) : (x87_concurrency.frndint * cpu_multi));
+    return 0;
+}
+
+static int
+opFRINEAR(UNUSED(uint32_t fetchdat))
+{
+    double dst0, st0, integral, frac;
+    int prevRound;
+
+    FP_ENTER();
+    cpu_state.pc++;
+    prevRound = fegetround();
+    fesetround(FE_TONEAREST);
+    ST(0) = (double) x87_fround_nearest(ST(0));
+    fesetround(prevRound);
+    FP_TAG_VALID;
+    CLOCK_CYCLES_FPU((fpu_type >= FPU_487SX) ? (x87_timings.frndint) : (x87_timings.frndint * cpu_multi));
+    CONCURRENCY_CYCLES((fpu_type >= FPU_487SX) ? (x87_concurrency.frndint) : (x87_concurrency.frndint * cpu_multi));
+    return 0;
+}
+
+static int
+opFRICHOP(UNUSED(uint32_t fetchdat))
+{
+    double dst0, st0, integral, frac;
+    int prevRound;
+
+    FP_ENTER();
+    cpu_state.pc++;
+    prevRound = fegetround();
+    fesetround(FE_TONEAREST);
+    ST(0) = (double) ((int64_t)(ST(0)));
+    fesetround(prevRound);
+    FP_TAG_VALID;
+    CLOCK_CYCLES_FPU((fpu_type >= FPU_487SX) ? (x87_timings.frndint) : (x87_timings.frndint * cpu_multi));
+    CONCURRENCY_CYCLES((fpu_type >= FPU_487SX) ? (x87_concurrency.frndint) : (x87_concurrency.frndint * cpu_multi));
+    return 0;
+}
+
 static int
 opFSCALE(UNUSED(uint32_t fetchdat))
 {
diff --git a/src/cpu/x87_ops_sf_arith.h b/src/cpu/x87_ops_sf_arith.h
index c9e25e97c..620c44691 100644
--- a/src/cpu/x87_ops_sf_arith.h
+++ b/src/cpu/x87_ops_sf_arith.h
@@ -800,3 +800,84 @@ next_ins:
     CONCURRENCY_CYCLES((fpu_type >= FPU_487SX) ? (x87_concurrency.frndint) : (x87_concurrency.frndint * cpu_multi));
     return 0;
 }
+
+static int
+sf_FRINT2(uint32_t fetchdat)
+{
+    floatx80              result;
+    struct softfloat_status_t status;
+
+    FP_ENTER();
+    FPU_check_pending_exceptions();
+    cpu_state.pc++;
+    clear_C1();
+    if (IS_TAG_EMPTY(0)) {
+        FPU_stack_underflow(fetchdat, 0, 0);
+        goto next_ins;
+    }
+    status = i387cw_to_softfloat_status_word(i387_get_control_word());
+    result = extF80_roundToInt(FPU_read_regi(0), softfloat_round_near_maxMag, true, &status);
+
+    if (!FPU_exception(fetchdat, status.softfloat_exceptionFlags, 0)) {
+        FPU_save_regi(result, 0);
+    }
+
+next_ins:
+    CLOCK_CYCLES_FPU((fpu_type >= FPU_487SX) ? (x87_timings.frndint) : (x87_timings.frndint * cpu_multi));
+    CONCURRENCY_CYCLES((fpu_type >= FPU_487SX) ? (x87_concurrency.frndint) : (x87_concurrency.frndint * cpu_multi));
+    return 0;
+}
+
+static int
+sf_FRINEAR(uint32_t fetchdat)
+{
+    floatx80              result;
+    struct softfloat_status_t status;
+
+    FP_ENTER();
+    FPU_check_pending_exceptions();
+    cpu_state.pc++;
+    clear_C1();
+    if (IS_TAG_EMPTY(0)) {
+        FPU_stack_underflow(fetchdat, 0, 0);
+        goto next_ins;
+    }
+    status = i387cw_to_softfloat_status_word(i387_get_control_word());
+    result = extF80_roundToInt(FPU_read_regi(0), softfloat_round_near_even, true, &status);
+
+    if (!FPU_exception(fetchdat, status.softfloat_exceptionFlags, 0)) {
+        FPU_save_regi(result, 0);
+    }
+
+next_ins:
+    CLOCK_CYCLES_FPU((fpu_type >= FPU_487SX) ? (x87_timings.frndint) : (x87_timings.frndint * cpu_multi));
+    CONCURRENCY_CYCLES((fpu_type >= FPU_487SX) ? (x87_concurrency.frndint) : (x87_concurrency.frndint * cpu_multi));
+    return 0;
+}
+
+static int
+sf_FRICHOP(uint32_t fetchdat)
+{
+    floatx80              result;
+    struct softfloat_status_t status;
+
+    FP_ENTER();
+    FPU_check_pending_exceptions();
+    cpu_state.pc++;
+    clear_C1();
+    if (IS_TAG_EMPTY(0)) {
+        FPU_stack_underflow(fetchdat, 0, 0);
+        goto next_ins;
+    }
+    status = i387cw_to_softfloat_status_word(i387_get_control_word());
+    result = extF80_roundToInt(FPU_read_regi(0), softfloat_round_to_zero, true, &status);
+
+    if (!FPU_exception(fetchdat, status.softfloat_exceptionFlags, 0)) {
+        FPU_save_regi(result, 0);
+    }
+
+next_ins:
+    CLOCK_CYCLES_FPU((fpu_type >= FPU_487SX) ? (x87_timings.frndint) : (x87_timings.frndint * cpu_multi));
+    CONCURRENCY_CYCLES((fpu_type >= FPU_487SX) ? (x87_concurrency.frndint) : (x87_concurrency.frndint * cpu_multi));
+    return 0;
+}
diff --git a/src/cpu/x87_ops_sf_compare.h b/src/cpu/x87_ops_sf_compare.h
index 945ed99d3..d1d9bc1dd 100644
--- a/src/cpu/x87_ops_sf_compare.h
+++ b/src/cpu/x87_ops_sf_compare.h
@@ -459,6 +459,32 @@ sf_FTST(uint32_t fetchdat)
     return 0;
 }
 
+static int
+sf_FTSTP(uint32_t fetchdat)
+{
+    const floatx80            Const_Z = packFloatx80(0, 0x0000, 0);
+    struct softfloat_status_t status;
+    int                       rc;
+
+    FP_ENTER();
+    FPU_check_pending_exceptions();
+    cpu_state.pc++;
+    clear_C1();
+    if (IS_TAG_EMPTY(0)) {
+        FPU_exception(fetchdat, FPU_EX_Stack_Underflow, 0);
+        setcc(FPU_SW_C0 | FPU_SW_C2 | FPU_SW_C3);
+    } else {
+        status = i387cw_to_softfloat_status_word(i387_get_control_word());
+        rc     = extF80_compare_normal(FPU_read_regi(0), Const_Z, &status);
+        setcc(FPU_status_word_flags_fpu_compare(rc));
+        FPU_exception(fetchdat, status.softfloat_exceptionFlags, 0);
+        FPU_pop();
+    }
+    CLOCK_CYCLES_FPU((fpu_type >= FPU_487SX) ? (x87_timings.ftst) : (x87_timings.ftst * cpu_multi));
+    CONCURRENCY_CYCLES((fpu_type >= FPU_487SX) ? (x87_concurrency.ftst) : (x87_concurrency.ftst * cpu_multi));
+    return 0;
+}
+
 static int
 sf_FXAM(UNUSED(uint32_t fetchdat))
 {

From c40aa61be419c6c6ed2365a2c69607964878b95a Mon Sep 17 00:00:00 2001
From: Cacodemon345 <wahil1976@outlook.com>
Date: Thu, 6 Mar 2025 02:09:23 +0600
Subject: [PATCH 41/58] Cyrix 6x86: Correctly initalize ARR3 on reset to avoid
 some SMM problems

---
 src/cpu/x86.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/src/cpu/x86.c b/src/cpu/x86.c
index f1f5004d9..97ae19f3a 100644
--- a/src/cpu/x86.c
+++ b/src/cpu/x86.c
@@ -275,6 +275,8 @@ reset_common(int hard)
 
     ccr0 = ccr1 = ccr2 = ccr3 = ccr4 = ccr5 = ccr6 = ccr7 = 0;
     ccr4 = 0x85;
+    cyrix.arr[3].base = 0x30000;
+    cyrix.arr[3].size = 65536;
 
     if (hascache)
         cr0 = 1 << 30;

From 4c3ea296c3dbe7ab10c9660e2225059eebb1b676 Mon Sep 17 00:00:00 2001
From: Jasmine Iwanek <jasmine@iwanek.co.uk>
Date: Wed, 5 Mar 2025 17:58:01 -0500
Subject: [PATCH 42/58] Fix ES1370 detection

---
 src/sound/snd_audiopci.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/src/sound/snd_audiopci.c b/src/sound/snd_audiopci.c
index f8a6e4153..e3e2b9f1a 100644
--- a/src/sound/snd_audiopci.c
+++ b/src/sound/snd_audiopci.c
@@ -489,7 +489,10 @@ es137x_reset(void *priv)
 
     /* Serial Interface Control Register, Address 20H
        Addressable as byte, word, longword */
-    dev->si_cr = 0xff800000;
+    if (dev->type == AUDIOPCI_ES1370)
+        dev->si_cr = 0x00000000;
+    else
+        dev->si_cr = 0xff800000;
 
     /* DAC1 Channel Sample Count Register, Address 24H
        Addressable as word, longword */

From 2300339588f42744c8f5485a60a91b884fe6f8b3 Mon Sep 17 00:00:00 2001
From: OBattler <oubattler@gmail.com>
Date: Thu, 6 Mar 2025 00:05:16 +0100
Subject: [PATCH 43/58] Included stdlib.h.

---
 src/cpu/386_dynarec_ops.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/src/cpu/386_dynarec_ops.c b/src/cpu/386_dynarec_ops.c
index c31b725b2..3d9a7e080 100644
--- a/src/cpu/386_dynarec_ops.c
+++ b/src/cpu/386_dynarec_ops.c
@@ -1,6 +1,7 @@
 #include <stdarg.h>
-#include <stdio.h>
 #include <stdint.h>
+#include <stdio.h>
+#include <stdlib.h>
 #include <string.h>
 #include <wchar.h>
 #include <math.h>

From 9b47522f43116dfaa04a03d18ced128ecbf31299 Mon Sep 17 00:00:00 2001
From: OBattler <oubattler@gmail.com>
Date: Thu, 6 Mar 2025 00:12:45 +0100
Subject: [PATCH 44/58] FPU: Fix newly-introduced x87-related warnings.

---
 src/cpu/x87_ops_misc.h       | 8 ++++++--
 src/cpu/x87_ops_sf_arith.h   | 4 ++++
 src/cpu/x87_ops_sf_compare.h | 2 ++
 3 files changed, 12 insertions(+), 2 deletions(-)

diff --git a/src/cpu/x87_ops_misc.h b/src/cpu/x87_ops_misc.h
index cbad12e56..5821a5cd5 100644
--- a/src/cpu/x87_ops_misc.h
+++ b/src/cpu/x87_ops_misc.h
@@ -554,6 +554,7 @@ opFTST(UNUSED(uint32_t fetchdat))
     return 0;
 }
 
+#ifndef FPU_8087
 static int
 opFTSTP(UNUSED(uint32_t fetchdat))
 {
@@ -569,6 +570,7 @@ opFTSTP(UNUSED(uint32_t fetchdat))
     CONCURRENCY_CYCLES((fpu_type >= FPU_487SX) ? (x87_concurrency.ftst) : (x87_concurrency.ftst * cpu_multi));
     return 0;
 }
+#endif
 
 static int
 opFXAM(UNUSED(uint32_t fetchdat))
@@ -854,6 +856,8 @@ opFRNDINT(UNUSED(uint32_t fetchdat))
     return 0;
 }
 
+#ifndef FPU_8087
+#ifndef OPS_286_386
 static int
 opFRINT2(UNUSED(uint32_t fetchdat))
 {
@@ -882,7 +886,6 @@ opFRINT2(UNUSED(uint32_t fetchdat))
 static int
 opFRINEAR(UNUSED(uint32_t fetchdat))
 {
-    double dst0, st0, integral, frac;
     int prevRound;
 
     FP_ENTER();
@@ -896,11 +899,11 @@ opFRINEAR(UNUSED(uint32_t fetchdat))
     CONCURRENCY_CYCLES((fpu_type >= FPU_487SX) ? (x87_concurrency.frndint) : (x87_concurrency.frndint * cpu_multi));
     return 0;
 }
+#endif
 
 static int
 opFRICHOP(UNUSED(uint32_t fetchdat))
 {
-    double dst0, st0, integral, frac;
     int prevRound;
 
     FP_ENTER();
@@ -914,6 +917,7 @@ opFRICHOP(UNUSED(uint32_t fetchdat))
     CONCURRENCY_CYCLES((fpu_type >= FPU_487SX) ? (x87_concurrency.frndint) : (x87_concurrency.frndint * cpu_multi));
     return 0;
 }
+#endif
 
 static int
 opFSCALE(UNUSED(uint32_t fetchdat))
diff --git a/src/cpu/x87_ops_sf_arith.h b/src/cpu/x87_ops_sf_arith.h
index 620c44691..7abde322b 100644
--- a/src/cpu/x87_ops_sf_arith.h
+++ b/src/cpu/x87_ops_sf_arith.h
@@ -801,6 +801,8 @@ next_ins:
     return 0;
 }
 
+#ifndef FPU_8087
+#ifndef OPS_286_386
 static int
 sf_FRINT2(uint32_t fetchdat)
 {
@@ -854,6 +856,7 @@ next_ins:
     CONCURRENCY_CYCLES((fpu_type >= FPU_487SX) ? (x87_concurrency.frndint) : (x87_concurrency.frndint * cpu_multi));
     return 0;
 }
+#endif
 
 static int
 sf_FRICHOP(uint32_t fetchdat)
@@ -881,3 +884,4 @@ next_ins:
     CONCURRENCY_CYCLES((fpu_type >= FPU_487SX) ? (x87_concurrency.frndint) : (x87_concurrency.frndint * cpu_multi));
     return 0;
 }
+#endif
diff --git a/src/cpu/x87_ops_sf_compare.h b/src/cpu/x87_ops_sf_compare.h
index d1d9bc1dd..d90526b82 100644
--- a/src/cpu/x87_ops_sf_compare.h
+++ b/src/cpu/x87_ops_sf_compare.h
@@ -459,6 +459,7 @@ sf_FTST(uint32_t fetchdat)
     return 0;
 }
 
+#ifndef FPU_8087
 static int
 sf_FTSTP(uint32_t fetchdat)
 {
@@ -484,6 +485,7 @@ sf_FTSTP(uint32_t fetchdat)
     CONCURRENCY_CYCLES((fpu_type >= FPU_487SX) ? (x87_concurrency.ftst) : (x87_concurrency.ftst * cpu_multi));
     return 0;
 }
+#endif
 
 static int
 sf_FXAM(UNUSED(uint32_t fetchdat))

From 78f50c5b042ebd69c4fec875d47e10cf434e5806 Mon Sep 17 00:00:00 2001
From: OBattler <oubattler@gmail.com>
Date: Thu, 6 Mar 2025 00:17:16 +0100
Subject: [PATCH 45/58] Move the Cyrix 6x86 out of the Dev branch.

---
 CMakeLists.txt            |  1 -
 src/cpu/386_ops.h         |  4 ----
 src/cpu/CMakeLists.txt    | 10 +---------
 src/cpu/cpu.c             | 10 ----------
 src/cpu/cpu_table.c       |  2 --
 src/cpu/x86_ops.h         |  4 ----
 src/cpu/x86_ops_mmx_mov.h |  2 --
 7 files changed, 1 insertion(+), 32 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 5c33bb3e9..fb71553e2 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -163,7 +163,6 @@ endif()
 cmake_dependent_option(AMD_K5         "AMD K5"                                   ON    "DEV_BRANCH"  OFF)
 cmake_dependent_option(AN430TX        "Intel AN430TX"                            ON    "DEV_BRANCH"  OFF)
 cmake_dependent_option(CDROM_MITSUMI  "Mitsumi CDROM"                            ON    "DEV_BRANCH"  OFF)
-cmake_dependent_option(CYRIX_6X86     "Cyrix 6x86"                               ON    "DEV_BRANCH"  OFF)
 cmake_dependent_option(G100           "Matrox Productiva G100"                   ON    "DEV_BRANCH"  OFF)
 cmake_dependent_option(GUSMAX         "Gravis UltraSound MAX"                    ON    "DEV_BRANCH"  OFF)
 cmake_dependent_option(ISAMEM_RAMPAGE "AST Rampage"                              ON    "DEV_BRANCH"  OFF)
diff --git a/src/cpu/386_ops.h b/src/cpu/386_ops.h
index e8300f134..5113ca817 100644
--- a/src/cpu/386_ops.h
+++ b/src/cpu/386_ops.h
@@ -1385,7 +1385,6 @@ const OpFn OP_TABLE(pentium_0f)[1024] = {
     // clang-format on
 };
 
-#    ifdef USE_CYRIX_6X86
 const OpFn OP_TABLE(c6x86_0f)[1024] = {
     // clang-format off
         /*16-bit data, 16-bit addr*/
@@ -1569,7 +1568,6 @@ const OpFn OP_TABLE(c6x86l_0f)[1024] = {
 /*f0*/  ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,        ILLEGAL,
     // clang-format on
 };
-#    endif /* USE_CYRIX_6X86 */
 
 const OpFn OP_TABLE(pentiummmx_0f)[1024] = {
     // clang-format off
@@ -1847,7 +1845,6 @@ const OpFn OP_TABLE(k62_0f)[1024] = {
     // clang-format on
 };
 
-#    ifdef USE_CYRIX_6X86
 const OpFn OP_TABLE(c6x86mx_0f)[1024] = {
     // clang-format off
         /*16-bit data, 16-bit addr*/
@@ -1939,7 +1936,6 @@ const OpFn OP_TABLE(c6x86mx_0f)[1024] = {
 /*f0*/  ILLEGAL,        opPSLLW_a32,    opPSLLD_a32,    opPSLLQ_a32,    ILLEGAL,        opPMADDWD_a32,  ILLEGAL,        ILLEGAL,        opPSUBB_a32,    opPSUBW_a32,    opPSUBD_a32,    ILLEGAL,        opPADDB_a32,    opPADDW_a32,    opPADDD_a32,    ILLEGAL,
     // clang-format on
 };
-#    endif /* USE_CYRIX_6X86 */
 
 const OpFn OP_TABLE(pentiumpro_0f)[1024] = {
     // clang-format off
diff --git a/src/cpu/CMakeLists.txt b/src/cpu/CMakeLists.txt
index 890d02e3e..9c1385d4f 100644
--- a/src/cpu/CMakeLists.txt
+++ b/src/cpu/CMakeLists.txt
@@ -42,20 +42,12 @@ endif()
 
 endif()
 
-if(CYRIX_6X86)
-    target_compile_definitions(cpu PRIVATE USE_CYRIX_6X86)
-
-if(DYNAREC)
-    add_library(ct686 OBJECT codegen_timing_686.c)
-    target_link_libraries(86Box ct686)
-endif()
-endif()
-
 if(DYNAREC)
     target_sources(cpu PRIVATE 386_dynarec_ops.c)
 
     add_library(cgt OBJECT
         codegen_timing_486.c
+        codegen_timing_686.c
         codegen_timing_common.c
         codegen_timing_k6.c
         codegen_timing_pentium.c
diff --git a/src/cpu/cpu.c b/src/cpu/cpu.c
index 13a109d38..e6a22394a 100644
--- a/src/cpu/cpu.c
+++ b/src/cpu/cpu.c
@@ -1420,7 +1420,6 @@ cpu_set(void)
 #endif /* USE_DYNAREC */
             break;
 
-#ifdef USE_CYRIX_6X86
         case CPU_Cx6x86:
         case CPU_Cx6x86L:
         case CPU_CxGX1:
@@ -1541,7 +1540,6 @@ cpu_set(void)
             else if (CPU_Cx6x86)
                 CPUID = 0; /* Disabled on powerup by default */
             break;
-#endif /* USE_CYRIX_6X86 */
 
 #ifdef USE_AMD_K5
         case CPU_K5:
@@ -2386,7 +2384,6 @@ cpu_CPUID(void)
                 EAX = EBX = ECX = EDX = 0;
             break;
 
-#ifdef USE_CYRIX_6X86
         case CPU_Cx6x86:
             if (!EAX) {
                 EAX = 0x00000001;
@@ -2447,7 +2444,6 @@ cpu_CPUID(void)
             } else
                 EAX = EBX = ECX = EDX = 0;
             break;
-#endif /* USE_CYRIX_6X86 */
 
         case CPU_PENTIUMPRO:
             if (!EAX) {
@@ -3124,7 +3120,6 @@ pentium_invalid_rdmsr:
             cpu_log("RDMSR: ECX = %08X, val = %08X%08X\n", ECX, EDX, EAX);
             break;
 
-#ifdef USE_CYRIX_6X86
         case CPU_Cx6x86:
         case CPU_Cx6x86L:
         case CPU_CxGX1:
@@ -3164,7 +3159,6 @@ pentium_invalid_rdmsr:
             }
             cpu_log("RDMSR: ECX = %08X, val = %08X%08X\n", ECX, EDX, EAX);
             break;
-#endif /* USE_CYRIX_6X86 */
 
         case CPU_PENTIUMPRO:
         case CPU_PENTIUM2:
@@ -3945,7 +3939,6 @@ pentium_invalid_wrmsr:
             }
             break;
 
-#ifdef USE_CYRIX_6X86
         case CPU_Cx6x86:
         case CPU_Cx6x86L:
         case CPU_CxGX1:
@@ -3979,7 +3972,6 @@ pentium_invalid_wrmsr:
                     break;
             }
             break;
-#endif /* USE_CYRIX_6X86 */
 
         case CPU_PENTIUMPRO:
         case CPU_PENTIUM2:
@@ -4283,14 +4275,12 @@ cpu_write(uint16_t addr, uint8_t val, UNUSED(void *priv))
             case 0xe8: /* CCR4 */
                 if ((ccr3 & 0xf0) == 0x10) {
                     ccr4 = val;
-#ifdef USE_CYRIX_6X86
                     if (cpu_s->cpu_type >= CPU_Cx6x86) {
                         if (val & 0x80)
                             CPUID = cpu_s->cpuid_model;
                         else
                             CPUID = 0;
                     }
-#endif /* USE_CYRIX_6X86 */
                 }
                 break;
             case 0xe9: /* CCR5 */
diff --git a/src/cpu/cpu_table.c b/src/cpu/cpu_table.c
index 104d56d05..5a156853e 100644
--- a/src/cpu/cpu_table.c
+++ b/src/cpu/cpu_table.c
@@ -6095,7 +6095,6 @@ const cpu_family_t cpu_families[] = {
             { .name = "", 0 }
         }
     },
-#ifdef USE_CYRIX_6X86
     {
         .package       = CPU_PKG_SOCKET5_7,
         .manufacturer  = "Cyrix",
@@ -6455,7 +6454,6 @@ const cpu_family_t cpu_families[] = {
             { .name = "", 0 }
         }
     },
-#endif /* USE_CYRIX_6X86 */
     {
         .package       = CPU_PKG_SOCKET8,
         .manufacturer  = "Intel",
diff --git a/src/cpu/x86_ops.h b/src/cpu/x86_ops.h
index 6d93fb435..ed95d04d9 100644
--- a/src/cpu/x86_ops.h
+++ b/src/cpu/x86_ops.h
@@ -90,11 +90,9 @@ extern const OpFn dynarec_ops_winchip2_0f[1024];
 extern const OpFn dynarec_ops_pentium_0f[1024];
 extern const OpFn dynarec_ops_pentiummmx_0f[1024];
 
-#    ifdef USE_CYRIX_6X86
 extern const OpFn dynarec_ops_c6x86_0f[1024];
 extern const OpFn dynarec_ops_c6x86l_0f[1024];
 extern const OpFn dynarec_ops_c6x86mx_0f[1024];
-#    endif /* USE_CYRIX_6X86 */
 
 extern const OpFn dynarec_ops_k6_0f[1024];
 extern const OpFn dynarec_ops_k62_0f[1024];
@@ -233,11 +231,9 @@ extern const OpFn ops_winchip2_0f[1024];
 extern const OpFn ops_pentium_0f[1024];
 extern const OpFn ops_pentiummmx_0f[1024];
 
-#ifdef USE_CYRIX_6X86
 extern const OpFn ops_c6x86_0f[1024];
 extern const OpFn ops_c6x86l_0f[1024];
 extern const OpFn ops_c6x86mx_0f[1024];
-#endif /* USE_CYRIX_6X86 */
 
 extern const OpFn ops_k6_0f[1024];
 extern const OpFn ops_k62_0f[1024];
diff --git a/src/cpu/x86_ops_mmx_mov.h b/src/cpu/x86_ops_mmx_mov.h
index 8855f8ccd..0bf928198 100644
--- a/src/cpu/x86_ops_mmx_mov.h
+++ b/src/cpu/x86_ops_mmx_mov.h
@@ -110,7 +110,6 @@ opMOVD_mm_l_a32(uint32_t fetchdat)
     return 0;
 }
 
-#ifdef USE_CYRIX_6X86
 /*Cyrix maps both MOVD and SMINT to the same opcode*/
 static int
 opMOVD_mm_l_a16_cx(uint32_t fetchdat)
@@ -170,7 +169,6 @@ opMOVD_mm_l_a32_cx(uint32_t fetchdat)
 
     return 0;
 }
-#endif /* USE_CYRIX_6X86 */
 
 static int
 opMOVQ_q_mm_a16(uint32_t fetchdat)

From 74368bfe7ab523f92d3533f0bdaf49c7c0611b4b Mon Sep 17 00:00:00 2001
From: Cacodemon345 <wahil1976@outlook.com>
Date: Thu, 6 Mar 2025 19:26:11 +0600
Subject: [PATCH 46/58] Don't include `cpu.h` in `timer.h`

Change the name of the define used to indicate SVGA multi-monitor-capable cards
---
 src/chipset/ali1531.c         | 1 +
 src/chipset/ali1541.c         | 1 +
 src/chipset/ali1543.c         | 1 +
 src/chipset/ali6117.c         | 1 +
 src/chipset/gc100.c           | 1 +
 src/chipset/sis_5511_h2p.c    | 1 +
 src/chipset/sis_5513_p2i.c    | 1 +
 src/chipset/sis_5571_h2p.c    | 1 +
 src/chipset/sis_5581_h2p.c    | 1 +
 src/chipset/sis_5591_h2p.c    | 1 +
 src/chipset/sis_5600_h2p.c    | 1 +
 src/chipset/sis_85c50x.c      | 1 +
 src/device/cartridge.c        | 1 +
 src/device/keyboard_xt.c      | 1 +
 src/device/serial.c           | 1 +
 src/include/86box/timer.h     | 2 +-
 src/include/86box/video.h     | 2 ++
 src/machine/m_at_socket5.c    | 1 +
 src/machine/m_xt_olivetti.c   | 1 +
 src/machine/m_xt_philips.c    | 1 +
 src/mem/sst_flash.c           | 1 +
 src/qt/qt_main.cpp            | 1 +
 src/qt/qt_settingsdisplay.cpp | 6 +++---
 src/scsi/scsi_t128.c          | 1 +
 src/sound/snd_audiopci.c      | 1 +
 src/sound/snd_gus.c           | 1 +
 src/sound/snd_opl_esfm.c      | 1 +
 src/sound/snd_opl_nuked.c     | 1 +
 src/sound/snd_sb.c            | 1 +
 src/timer.c                   | 1 +
 src/video/vid_ati_mach64.c    | 1 +
 src/video/vid_ati_mach8.c     | 1 +
 src/video/vid_chips_69000.c   | 1 +
 src/video/vid_mga.c           | 1 +
 src/video/vid_nga.c           | 1 +
 src/video/vid_ogc.c           | 1 +
 src/video/vid_table.c         | 6 +++---
 37 files changed, 42 insertions(+), 7 deletions(-)

diff --git a/src/chipset/ali1531.c b/src/chipset/ali1531.c
index e765e6d45..53324f8e6 100644
--- a/src/chipset/ali1531.c
+++ b/src/chipset/ali1531.c
@@ -22,6 +22,7 @@
 #include <wchar.h>
 #define HAVE_STDARG_H
 #include <86box/86box.h>
+#include "cpu.h"
 #include <86box/timer.h>
 
 #include <86box/device.h>
diff --git a/src/chipset/ali1541.c b/src/chipset/ali1541.c
index eed6ddbc5..ebbcd7e63 100644
--- a/src/chipset/ali1541.c
+++ b/src/chipset/ali1541.c
@@ -22,6 +22,7 @@
 #include <wchar.h>
 #define HAVE_STDARG_H
 #include <86box/86box.h>
+#include "cpu.h"
 #include <86box/timer.h>
 
 #include <86box/device.h>
diff --git a/src/chipset/ali1543.c b/src/chipset/ali1543.c
index 69a8990e9..2f9273736 100644
--- a/src/chipset/ali1543.c
+++ b/src/chipset/ali1543.c
@@ -22,6 +22,7 @@
 #include <wchar.h>
 #define HAVE_STDARG_H
 #include <86box/86box.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/device.h>
 #include <86box/io.h>
diff --git a/src/chipset/ali6117.c b/src/chipset/ali6117.c
index 707b528b2..88bb4b572 100644
--- a/src/chipset/ali6117.c
+++ b/src/chipset/ali6117.c
@@ -26,6 +26,7 @@
 #include <86box/io.h>
 #include <86box/pci.h>
 #include <86box/pic.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/pit.h>
 #include <86box/device.h>
diff --git a/src/chipset/gc100.c b/src/chipset/gc100.c
index 23455b952..e9eb05ecf 100644
--- a/src/chipset/gc100.c
+++ b/src/chipset/gc100.c
@@ -27,6 +27,7 @@
 #define HAVE_STDARG_H
 #include <86box/86box.h>
 #include <86box/nmi.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/pit.h>
 #include <86box/mem.h>
diff --git a/src/chipset/sis_5511_h2p.c b/src/chipset/sis_5511_h2p.c
index 543fcacc5..b94e69f26 100644
--- a/src/chipset/sis_5511_h2p.c
+++ b/src/chipset/sis_5511_h2p.c
@@ -22,6 +22,7 @@
 #include <86box/86box.h>
 #include <86box/device.h>
 #include <86box/io.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/dma.h>
 #include <86box/mem.h>
diff --git a/src/chipset/sis_5513_p2i.c b/src/chipset/sis_5513_p2i.c
index 95a890fcf..b2346bbc4 100644
--- a/src/chipset/sis_5513_p2i.c
+++ b/src/chipset/sis_5513_p2i.c
@@ -22,6 +22,7 @@
 #include <86box/86box.h>
 #include <86box/device.h>
 #include <86box/io.h>
+#include "cpu.h"
 #include <86box/timer.h>
 
 #include <86box/dma.h>
diff --git a/src/chipset/sis_5571_h2p.c b/src/chipset/sis_5571_h2p.c
index d04964581..a28553487 100644
--- a/src/chipset/sis_5571_h2p.c
+++ b/src/chipset/sis_5571_h2p.c
@@ -22,6 +22,7 @@
 #include <86box/86box.h>
 #include <86box/device.h>
 #include <86box/io.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/dma.h>
 #include <86box/mem.h>
diff --git a/src/chipset/sis_5581_h2p.c b/src/chipset/sis_5581_h2p.c
index d01e9dd28..83983376d 100644
--- a/src/chipset/sis_5581_h2p.c
+++ b/src/chipset/sis_5581_h2p.c
@@ -22,6 +22,7 @@
 #include <86box/86box.h>
 #include <86box/device.h>
 #include <86box/io.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/dma.h>
 #include <86box/mem.h>
diff --git a/src/chipset/sis_5591_h2p.c b/src/chipset/sis_5591_h2p.c
index 048e7deea..74dd6dfbd 100644
--- a/src/chipset/sis_5591_h2p.c
+++ b/src/chipset/sis_5591_h2p.c
@@ -22,6 +22,7 @@
 #include <86box/86box.h>
 #include <86box/device.h>
 #include <86box/io.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/dma.h>
 #include <86box/mem.h>
diff --git a/src/chipset/sis_5600_h2p.c b/src/chipset/sis_5600_h2p.c
index a15c6fff5..c23309de4 100644
--- a/src/chipset/sis_5600_h2p.c
+++ b/src/chipset/sis_5600_h2p.c
@@ -22,6 +22,7 @@
 #include <86box/86box.h>
 #include <86box/device.h>
 #include <86box/io.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/dma.h>
 #include <86box/mem.h>
diff --git a/src/chipset/sis_85c50x.c b/src/chipset/sis_85c50x.c
index 192ae3767..e932ff6ca 100644
--- a/src/chipset/sis_85c50x.c
+++ b/src/chipset/sis_85c50x.c
@@ -24,6 +24,7 @@
 #include <86box/86box.h>
 #include <86box/device.h>
 #include <86box/io.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/apm.h>
 #include <86box/machine.h>
diff --git a/src/device/cartridge.c b/src/device/cartridge.c
index d2e91ecb3..00464026a 100644
--- a/src/device/cartridge.c
+++ b/src/device/cartridge.c
@@ -22,6 +22,7 @@
 #include <wchar.h>
 #define HAVE_STDARG_H
 #include <86box/86box.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/plat.h>
 #include <86box/ui.h>
diff --git a/src/device/keyboard_xt.c b/src/device/keyboard_xt.c
index 644168004..ddbcae61b 100644
--- a/src/device/keyboard_xt.c
+++ b/src/device/keyboard_xt.c
@@ -29,6 +29,7 @@
 #include <wchar.h>
 #include <86box/86box.h>
 #include <86box/device.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/fdd.h>
 #include <86box/machine.h>
diff --git a/src/device/serial.c b/src/device/serial.c
index c25da0070..deb97225a 100644
--- a/src/device/serial.c
+++ b/src/device/serial.c
@@ -27,6 +27,7 @@
 #define HAVE_STDARG_H
 #include <86box/86box.h>
 #include <86box/device.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/machine.h>
 #include <86box/io.h>
diff --git a/src/include/86box/timer.h b/src/include/86box/timer.h
index 25aff6b2f..37a03d9ca 100644
--- a/src/include/86box/timer.h
+++ b/src/include/86box/timer.h
@@ -1,7 +1,7 @@
 #ifndef _TIMER_H_
 #define _TIMER_H_
 
-#include "cpu.h"
+extern uint64_t tsc;
 
 /* Maximum period, currently 1 second. */
 #define MAX_USEC64    1000000ULL
diff --git a/src/include/86box/video.h b/src/include/86box/video.h
index 7ad29838c..2985559e1 100644
--- a/src/include/86box/video.h
+++ b/src/include/86box/video.h
@@ -69,6 +69,8 @@ enum {
 #define VIDEO_FLAG_TYPE_NONE    5
 #define VIDEO_FLAG_TYPE_MASK    7
 
+#define VIDEO_FLAG_TYPE_SECONDARY VIDEO_FLAG_TYPE_SPECIAL
+
 typedef struct video_timings_t {
     int type;
     int write_b;
diff --git a/src/machine/m_at_socket5.c b/src/machine/m_at_socket5.c
index b6e82301b..02922b425 100644
--- a/src/machine/m_at_socket5.c
+++ b/src/machine/m_at_socket5.c
@@ -29,6 +29,7 @@
 #include <86box/fdc_ext.h>
 #include <86box/hdc.h>
 #include <86box/hdc_ide.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/fdd.h>
 #include <86box/fdc.h>
diff --git a/src/machine/m_xt_olivetti.c b/src/machine/m_xt_olivetti.c
index b6e4fb94a..34ca441ec 100644
--- a/src/machine/m_xt_olivetti.c
+++ b/src/machine/m_xt_olivetti.c
@@ -30,6 +30,7 @@
 #include <wchar.h>
 #define HAVE_STDARG_H
 #include <86box/86box.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/io.h>
 #include <86box/pic.h>
diff --git a/src/machine/m_xt_philips.c b/src/machine/m_xt_philips.c
index 604ccebaf..eed54e07f 100644
--- a/src/machine/m_xt_philips.c
+++ b/src/machine/m_xt_philips.c
@@ -23,6 +23,7 @@
 #define HAVE_STDARG_H
 #include <86box/86box.h>
 #include <86box/nmi.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/pit.h>
 #include <86box/mem.h>
diff --git a/src/mem/sst_flash.c b/src/mem/sst_flash.c
index b41e0f2e2..34bfae83a 100644
--- a/src/mem/sst_flash.c
+++ b/src/mem/sst_flash.c
@@ -25,6 +25,7 @@
 #include <86box/device.h>
 #include <86box/mem.h>
 #include <86box/machine.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/nvr.h>
 #include <86box/plat.h>
diff --git a/src/qt/qt_main.cpp b/src/qt/qt_main.cpp
index 7723261b4..719208c18 100644
--- a/src/qt/qt_main.cpp
+++ b/src/qt/qt_main.cpp
@@ -86,6 +86,7 @@ extern MainWindow   *main_window;
 
 extern "C" {
 #include <86box/keyboard.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/nvr.h>
 extern int qt_nvr_save(void);
diff --git a/src/qt/qt_settingsdisplay.cpp b/src/qt/qt_settingsdisplay.cpp
index 75f8d376b..b7a930711 100644
--- a/src/qt/qt_settingsdisplay.cpp
+++ b/src/qt/qt_settingsdisplay.cpp
@@ -220,9 +220,9 @@ SettingsDisplay::on_comboBoxVideo_currentIndexChanged(int index)
         int secondaryFlags = video_card_get_flags(c);
         if (video_card_available(c)
             && device_is_valid(video_dev, machineId)
-            && !((secondaryFlags == primaryFlags) && (secondaryFlags != VIDEO_FLAG_TYPE_SPECIAL))
-            && !(((primaryFlags == VIDEO_FLAG_TYPE_8514) || (primaryFlags == VIDEO_FLAG_TYPE_XGA)) && (secondaryFlags != VIDEO_FLAG_TYPE_MDA) && (secondaryFlags != VIDEO_FLAG_TYPE_SPECIAL))
-            && !((primaryFlags != VIDEO_FLAG_TYPE_MDA) && (primaryFlags != VIDEO_FLAG_TYPE_SPECIAL) && ((secondaryFlags == VIDEO_FLAG_TYPE_8514) || (secondaryFlags == VIDEO_FLAG_TYPE_XGA)))) {
+            && !((secondaryFlags == primaryFlags) && (secondaryFlags != VIDEO_FLAG_TYPE_SECONDARY))
+            && !(((primaryFlags == VIDEO_FLAG_TYPE_8514) || (primaryFlags == VIDEO_FLAG_TYPE_XGA)) && (secondaryFlags != VIDEO_FLAG_TYPE_MDA) && (secondaryFlags != VIDEO_FLAG_TYPE_SECONDARY))
+            && !((primaryFlags != VIDEO_FLAG_TYPE_MDA) && (primaryFlags != VIDEO_FLAG_TYPE_SECONDARY) && ((secondaryFlags == VIDEO_FLAG_TYPE_8514) || (secondaryFlags == VIDEO_FLAG_TYPE_XGA)))) {
             ui->comboBoxVideoSecondary->addItem(name, c);
             if (c == curVideoCard_2)
                 ui->comboBoxVideoSecondary->setCurrentIndex(ui->comboBoxVideoSecondary->count() - 1);
diff --git a/src/scsi/scsi_t128.c b/src/scsi/scsi_t128.c
index a40d803f1..5ac3b5d67 100644
--- a/src/scsi/scsi_t128.c
+++ b/src/scsi/scsi_t128.c
@@ -29,6 +29,7 @@
 #define HAVE_STDARG_H
 #include <86box/86box.h>
 #include <86box/io.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/dma.h>
 #include <86box/pic.h>
diff --git a/src/sound/snd_audiopci.c b/src/sound/snd_audiopci.c
index e3e2b9f1a..8d1de442b 100644
--- a/src/sound/snd_audiopci.c
+++ b/src/sound/snd_audiopci.c
@@ -39,6 +39,7 @@
 #include <86box/pci.h>
 #include <86box/snd_ac97.h>
 #include <86box/sound.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/plat_unused.h>
 #include <86box/snd_akm4531.h>
diff --git a/src/sound/snd_gus.c b/src/sound/snd_gus.c
index 43638473c..1885581a5 100644
--- a/src/sound/snd_gus.c
+++ b/src/sound/snd_gus.c
@@ -15,6 +15,7 @@
 #include <86box/nmi.h>
 #include <86box/pic.h>
 #include <86box/sound.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #ifdef USE_GUSMAX
 #    include <86box/snd_ad1848.h>
diff --git a/src/sound/snd_opl_esfm.c b/src/sound/snd_opl_esfm.c
index bcd2a56e9..b80d264d5 100644
--- a/src/sound/snd_opl_esfm.c
+++ b/src/sound/snd_opl_esfm.c
@@ -32,6 +32,7 @@
 #include <86box/86box.h>
 #include <86box/sound.h>
 #include <86box/device.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/snd_opl.h>
 #include <86box/plat_unused.h>
diff --git a/src/sound/snd_opl_nuked.c b/src/sound/snd_opl_nuked.c
index 03851a589..60f5ed2a6 100644
--- a/src/sound/snd_opl_nuked.c
+++ b/src/sound/snd_opl_nuked.c
@@ -46,6 +46,7 @@
 #define HAVE_STDARG_H
 #include <86box/86box.h>
 #include <86box/sound.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/device.h>
 #include <86box/snd_opl.h>
diff --git a/src/sound/snd_sb.c b/src/sound/snd_sb.c
index 4f8b8a0bd..689d0b91e 100644
--- a/src/sound/snd_sb.c
+++ b/src/sound/snd_sb.c
@@ -39,6 +39,7 @@
 #include <86box/pic.h>
 #include <86box/rom.h>
 #include <86box/sound.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/snd_sb.h>
 #include <86box/plat_unused.h>
diff --git a/src/timer.c b/src/timer.c
index 9fed37511..03908890f 100644
--- a/src/timer.c
+++ b/src/timer.c
@@ -3,6 +3,7 @@
 #include <string.h>
 #include <wchar.h>
 #include <86box/86box.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/nv/vid_nv_rivatimer.h>
 
diff --git a/src/video/vid_ati_mach64.c b/src/video/vid_ati_mach64.c
index 2df7782ff..24935e3a2 100644
--- a/src/video/vid_ati_mach64.c
+++ b/src/video/vid_ati_mach64.c
@@ -28,6 +28,7 @@
 #include <86box/device.h>
 #include <86box/io.h>
 #include <86box/mem.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/pci.h>
 #include <86box/rom.h>
diff --git a/src/video/vid_ati_mach8.c b/src/video/vid_ati_mach8.c
index 71dd973da..2038f1b1d 100644
--- a/src/video/vid_ati_mach8.c
+++ b/src/video/vid_ati_mach8.c
@@ -27,6 +27,7 @@
 #include <86box/device.h>
 #include <86box/io.h>
 #include <86box/mem.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/mca.h>
 #include <86box/pci.h>
diff --git a/src/video/vid_chips_69000.c b/src/video/vid_chips_69000.c
index 012a16348..a387e99fa 100644
--- a/src/video/vid_chips_69000.c
+++ b/src/video/vid_chips_69000.c
@@ -27,6 +27,7 @@
 #include <86box/mem.h>
 #include <86box/rom.h>
 #include <86box/device.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/video.h>
 #include <86box/vid_svga.h>
diff --git a/src/video/vid_mga.c b/src/video/vid_mga.c
index f11983a20..00570bd9f 100644
--- a/src/video/vid_mga.c
+++ b/src/video/vid_mga.c
@@ -23,6 +23,7 @@
 #include <stdatomic.h>
 #include <86box/86box.h>
 #include <86box/io.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/mem.h>
 #include <86box/pci.h>
diff --git a/src/video/vid_nga.c b/src/video/vid_nga.c
index 3640e5106..ef1c6cd40 100644
--- a/src/video/vid_nga.c
+++ b/src/video/vid_nga.c
@@ -30,6 +30,7 @@
 #include <86box/io.h>
 #include <86box/video.h>
 #include <86box/86box.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/mem.h>
 #include <86box/pit.h>
diff --git a/src/video/vid_ogc.c b/src/video/vid_ogc.c
index 7fe5fbbeb..a49cd8a22 100644
--- a/src/video/vid_ogc.c
+++ b/src/video/vid_ogc.c
@@ -31,6 +31,7 @@
 #include <86box/io.h>
 #include <86box/video.h>
 #include <86box/86box.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/mem.h>
 #include <86box/pit.h>
diff --git a/src/video/vid_table.c b/src/video/vid_table.c
index ff3073c73..d1df2ba20 100644
--- a/src/video/vid_table.c
+++ b/src/video/vid_table.c
@@ -141,10 +141,10 @@ video_cards[] = {
     { .device = &chips_69000_device,                            .flags = VIDEO_FLAG_TYPE_NONE },
     { .device = &gd5430_pci_device,                             .flags = VIDEO_FLAG_TYPE_NONE },
     { .device = &gd5434_pci_device,                             .flags = VIDEO_FLAG_TYPE_NONE },
-    { .device = &gd5436_pci_device,                             .flags = VIDEO_FLAG_TYPE_SPECIAL },
+    { .device = &gd5436_pci_device,                             .flags = VIDEO_FLAG_TYPE_SECONDARY },
     { .device = &gd5440_pci_device,                             .flags = VIDEO_FLAG_TYPE_NONE },
-    { .device = &gd5446_pci_device,                             .flags = VIDEO_FLAG_TYPE_SPECIAL },
-    { .device = &gd5446_stb_pci_device,                         .flags = VIDEO_FLAG_TYPE_SPECIAL },
+    { .device = &gd5446_pci_device,                             .flags = VIDEO_FLAG_TYPE_SECONDARY },
+    { .device = &gd5446_stb_pci_device,                         .flags = VIDEO_FLAG_TYPE_SECONDARY },
     { .device = &gd5480_pci_device,                             .flags = VIDEO_FLAG_TYPE_NONE },
     { .device = &et4000w32p_videomagic_revb_pci_device,         .flags = VIDEO_FLAG_TYPE_NONE },
     { .device = &et4000w32p_revc_pci_device,                    .flags = VIDEO_FLAG_TYPE_NONE },

From 0e42547cca39beb1d07fd007b4c2d8a1e0a92a8d Mon Sep 17 00:00:00 2001
From: Cacodemon345 <wahil1976@outlook.com>
Date: Thu, 6 Mar 2025 19:34:24 +0600
Subject: [PATCH 47/58] Remove some `timer.h`-related include hacks

---
 src/qt/qt_machinestatus.cpp | 3 ---
 src/qt/qt_mainwindow.cpp    | 9 ++++-----
 2 files changed, 4 insertions(+), 8 deletions(-)

diff --git a/src/qt/qt_machinestatus.cpp b/src/qt/qt_machinestatus.cpp
index d1335873c..dee90f11c 100644
--- a/src/qt/qt_machinestatus.cpp
+++ b/src/qt/qt_machinestatus.cpp
@@ -19,9 +19,6 @@
 #include "qt_machinestatus.hpp"
 
 extern "C" {
-#define EMU_CPU_H // superhack - don't want timer.h to include cpu.h here, and some combo is preventing a compile
-extern uint64_t tsc;
-
 #include <86box/hdd.h>
 #include <86box/timer.h>
 #include <86box/86box.h>
diff --git a/src/qt/qt_mainwindow.cpp b/src/qt/qt_mainwindow.cpp
index 89d17a07c..552fd0310 100644
--- a/src/qt/qt_mainwindow.cpp
+++ b/src/qt/qt_mainwindow.cpp
@@ -48,11 +48,10 @@ extern "C" {
 #include <86box/machine.h>
 #include <86box/vid_ega.h>
 #include <86box/version.h>
-#if 0
-#include <86box/acpi.h> /* Requires timer.h include, which conflicts with Qt headers */
-#endif
-extern atomic_int acpi_pwrbut_pressed;
-extern int acpi_enabled;
+#include <86box/timer.h>
+#include <86box/apm.h>
+#include <86box/nvr.h>
+#include <86box/acpi.h>
 
 #ifdef USE_VNC
 #    include <86box/vnc.h>

From 9badd32c9cc14efde3fc7352d4edc9e8c1e1ec71 Mon Sep 17 00:00:00 2001
From: Cacodemon345 <wahil1976@outlook.com>
Date: Thu, 6 Mar 2025 19:44:37 +0600
Subject: [PATCH 48/58] Remove more `nvr.h` hacks

---
 src/qt/qt_settingsmachine.cpp | 6 +-----
 1 file changed, 1 insertion(+), 5 deletions(-)

diff --git a/src/qt/qt_settingsmachine.cpp b/src/qt/qt_settingsmachine.cpp
index 771f1cbec..9b1f9849e 100644
--- a/src/qt/qt_settingsmachine.cpp
+++ b/src/qt/qt_settingsmachine.cpp
@@ -32,13 +32,9 @@ extern "C" {
 #include <86box/config.h>
 #include <86box/device.h>
 #include <86box/machine.h>
+#include <86box/nvr.h>
 }
 
-// from nvr.h, which we can't import into CPP code
-#define TIME_SYNC_DISABLED 0
-#define TIME_SYNC_ENABLED  1
-#define TIME_SYNC_UTC      2
-
 #include "qt_deviceconfig.hpp"
 #include "qt_models_common.hpp"
 

From 4090c5a62cf61bc4f4437bb7df0577aed37106ee Mon Sep 17 00:00:00 2001
From: Cacodemon345 <wahil1976@outlook.com>
Date: Thu, 6 Mar 2025 19:56:08 +0600
Subject: [PATCH 49/58] Fix SDL builds

---
 src/unix/unix.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/src/unix/unix.c b/src/unix/unix.c
index 7f653b9b6..33b78ddb1 100644
--- a/src/unix/unix.c
+++ b/src/unix/unix.c
@@ -38,6 +38,7 @@
 #include <86box/device.h>
 #include <86box/gameport.h>
 #include <86box/unix_sdl.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/nvr.h>
 #include <86box/version.h>

From 5c78fa52035892bbe695afc7eca72ec58d68de5a Mon Sep 17 00:00:00 2001
From: TC1995 <mariogplayer90@gmail.com>
Date: Thu, 6 Mar 2025 16:03:11 +0100
Subject: [PATCH 50/58] 8514/A compatible changes of the day (March 6th, 2025)

1. Follow the Mach32 manual more closely regarding vblank support.
2. The subsystem status now takes account of the other bits more accurately.
3. The Mach32 PCI, when used with the ATI 68860 ramdac, has its own bpp's when in accelerator mode, separate from the VGA compatible side, so fix this accordingly.
4. Reset the vram when a mapping change occurs, should clear the messups in the ATI Mach8/32 accel video mode tests.
---
 src/include/86box/vid_8514a.h   |  2 +-
 src/video/vid_8514a.c           | 71 +++++++++++++++-----------
 src/video/vid_ati68860_ramdac.c | 18 +++----
 src/video/vid_ati_mach8.c       | 88 ++++++++++++++++++++++++---------
 4 files changed, 115 insertions(+), 64 deletions(-)

diff --git a/src/include/86box/vid_8514a.h b/src/include/86box/vid_8514a.h
index 57e98cc44..bfde22d5e 100644
--- a/src/include/86box/vid_8514a.h
+++ b/src/include/86box/vid_8514a.h
@@ -213,7 +213,7 @@ typedef struct ibm8514_t {
     int     vdisp2;
     int     disp_cntl;
     int     interlace;
-    uint8_t subsys_cntl;
+    uint16_t subsys_cntl;
     uint8_t subsys_stat;
 
     atomic_int fifo_idx;
diff --git a/src/video/vid_8514a.c b/src/video/vid_8514a.c
index 379772820..765258237 100644
--- a/src/video/vid_8514a.c
+++ b/src/video/vid_8514a.c
@@ -715,27 +715,13 @@ ibm8514_accel_out(uint16_t port, uint32_t val, svga_t *svga, int len)
                 break;
 
             case 0x42e8:
-                if (val & 0x01)
-                    dev->subsys_stat &= ~0x01;
-                if (val & 0x02)
-                    dev->subsys_stat &= ~0x02;
-                if (val & 0x04)
-                    dev->subsys_stat &= ~0x04;
-                if (val & 0x08)
-                    dev->subsys_stat &= ~0x08;
+                ibm8514_log("VBLANK stat=%02x, val=%02x.\n", dev->subsys_stat, val);
+                dev->subsys_cntl = (dev->subsys_cntl & 0xff00) | val;
+                dev->subsys_stat &= ~val;
                 break;
             case 0x42e9:
-                dev->subsys_cntl = val;
-                if (val & 0x01)
-                    dev->subsys_stat |= 0x01;
-                if (val & 0x02)
-                    dev->subsys_stat |= 0x02;
-                if (val & 0x04)
-                    dev->subsys_stat |= 0x04;
-                if (val & 0x08)
-                    dev->subsys_stat |= 0x08;
-
-                if ((val & 0xc0) == 0xc0) {
+                dev->subsys_cntl = (dev->subsys_cntl & 0xff) | (val << 8);
+                if ((val & 0xc0) == 0x80) {
                     dev->fifo_idx = 0;
                     dev->force_busy = 0;
                     dev->force_busy2 = 0;
@@ -882,10 +868,10 @@ ibm8514_accel_in(uint16_t port, svga_t *svga)
 
     switch (port) {
         case 0x2e8:
-            if (dev->vc == dev->v_syncstart)
+            if (dev->vc == dev->dispend)
                 temp |= 0x02;
 
-            ibm8514_log("0x2E8 read: Display Status=%02x.\n", temp);
+            ibm8514_log("Read: Display Status1=%02x.\n", temp);
             break;
 
         case 0x6e8:
@@ -910,21 +896,25 @@ ibm8514_accel_in(uint16_t port, svga_t *svga)
 
         case 0x42e8:
         case 0x42e9:
-            if (dev->vc == dev->v_syncstart)
-                dev->subsys_stat |= 0x01;
+            if ((dev->subsys_cntl & 0x01) && !(dev->subsys_stat & 0x01) && (dev->vc == dev->dispend))
+                temp |= 0x01;
 
             if (cmd == 6) {
-                if ((dev->accel.dx >= clip_l) &&
+                if ((dev->subsys_cntl & 0x02) &&
+                    !(dev->subsys_stat & 0x02) &&
+                    (dev->accel.dx >= clip_l) &&
                     (dev->accel.dx <= clip_r_ibm) &&
                     (dev->accel.dy >= clip_t) &&
                     (dev->accel.dy <= clip_b_ibm))
-                    dev->subsys_stat |= 0x02;
+                    temp |= 0x02;
             } else {
-                if ((dev->accel.cx >= clip_l) &&
+                if ((dev->subsys_cntl & 0x02) &&
+                    !(dev->subsys_stat & 0x02) &&
+                    (dev->accel.cx >= clip_l) &&
                     (dev->accel.cx <= clip_r_ibm) &&
                     (dev->accel.cy >= clip_t) &&
                     (dev->accel.cy <= clip_b_ibm))
-                    dev->subsys_stat |= 0x02;
+                    temp |= 0x02;
             }
 
             if (!dev->fifo_idx) {
@@ -932,9 +922,10 @@ ibm8514_accel_in(uint16_t port, svga_t *svga)
                     temp |= 0x08;
             }
 
-            if (port & 1)
+            if (port & 1) {
                 temp = dev->vram_512k_8514 ? 0x00 : 0x80;
-            else {
+                temp |= (dev->subsys_cntl >> 8);
+            } else {
                 temp |= (dev->subsys_stat | (dev->vram_512k_8514 ? 0x00 : 0x80));
                 temp |= 0x20;
             }
@@ -1155,6 +1146,7 @@ ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat
                         (dev->accel.cx <= clip_r) &&
                         (dev->accel.cy >= clip_t) &&
                         (dev->accel.cy <= clip_b)) {
+                        dev->subsys_stat |= 0x02;
                         switch ((mix_dat & mix_mask) ? frgd_mix : bkgd_mix) {
                             case 0:
                                 src_dat = bkgd_color;
@@ -1250,6 +1242,7 @@ ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat
                         (dev->accel.cx <= clip_r) &&
                         (dev->accel.cy >= clip_t) &&
                         (dev->accel.cy <= clip_b)) {
+                        dev->subsys_stat |= 0x02;
                         switch ((mix_dat & mix_mask) ? frgd_mix : bkgd_mix) {
                             case 0:
                                 src_dat = bkgd_color;
@@ -1385,6 +1378,7 @@ ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat
                         (dev->accel.cx <= clip_r) &&
                         (dev->accel.cy >= clip_t) &&
                         (dev->accel.cy <= clip_b)) {
+                        dev->subsys_stat |= 0x02;
                         if (ibm8514_cpu_dest(svga) && (pixcntl == 0)) {
                             mix_dat = mix_mask; /* Mix data = forced to foreground register. */
                         } else if (ibm8514_cpu_dest(svga) && (pixcntl == 3)) {
@@ -1543,6 +1537,7 @@ ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat
                             (dev->accel.cx <= clip_r) &&
                             (dev->accel.cy >= clip_t) &&
                             (dev->accel.cy <= clip_b)) {
+                            dev->subsys_stat |= 0x02;
                             if (ibm8514_cpu_dest(svga)) {
                                 READ((dev->accel.cy * dev->pitch) + dev->accel.cx, src_dat);
                             } else
@@ -1634,6 +1629,7 @@ ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat
                             (dev->accel.cx <= clip_r) &&
                             (dev->accel.cy >= clip_t) &&
                             (dev->accel.cy <= clip_b)) {
+                            dev->subsys_stat |= 0x02;
                             if (ibm8514_cpu_dest(svga) && (pixcntl == 0)) {
                                 mix_dat = mix_mask; /* Mix data = forced to foreground register. */
                             } else if (ibm8514_cpu_dest(svga) && (pixcntl == 3)) {
@@ -1832,6 +1828,7 @@ ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat
                                 (dev->accel.cx <= clip_r) &&
                                 (dev->accel.cy >= clip_t) &&
                                 (dev->accel.cy <= clip_b)) {
+                                dev->subsys_stat |= 0x02;
                                 switch ((mix_dat & mix_mask) ? frgd_mix : bkgd_mix) {
                                     case 0:
                                         src_dat = bkgd_color;
@@ -1989,6 +1986,7 @@ skip_vector_rect_write:
                             (dev->accel.cx <= clip_r) &&
                             (dev->accel.cy >= clip_t) &&
                             (dev->accel.cy <= clip_b)) {
+                            dev->subsys_stat |= 0x02;
                             if (ibm8514_cpu_dest(svga) && (pixcntl == 0)) {
                                 mix_dat = mix_mask; /* Mix data = forced to foreground register. */
                             } else if (ibm8514_cpu_dest(svga) && (pixcntl == 3)) {
@@ -2150,6 +2148,7 @@ skip_nibble_rect_write:
                                     (dev->accel.cx <= clip_r) &&
                                     (dev->accel.cy >= clip_t) &&
                                     (dev->accel.cy <= clip_b)) {
+                                    dev->subsys_stat |= 0x02;
                                     switch ((mix_dat & mix_mask) ? frgd_mix : bkgd_mix) {
                                         case 0:
                                             src_dat = bkgd_color;
@@ -2232,6 +2231,7 @@ skip_nibble_rect_write:
                                     (dev->accel.cx <= clip_r) &&
                                     (dev->accel.cy >= clip_t) &&
                                     (dev->accel.cy <= clip_b)) {
+                                    dev->subsys_stat |= 0x02;
                                     switch ((mix_dat & 0x01) ? frgd_mix : bkgd_mix) {
                                         case 0:
                                             src_dat = bkgd_color;
@@ -2313,6 +2313,7 @@ skip_nibble_rect_write:
                                 (dev->accel.cx <= clip_r) &&
                                 (dev->accel.cy >= clip_t) &&
                                 (dev->accel.cy <= clip_b)) {
+                                dev->subsys_stat |= 0x02;
                                 switch ((mix_dat & mix_mask) ? frgd_mix : bkgd_mix) {
                                     case 0:
                                         src_dat = bkgd_color;
@@ -2424,6 +2425,7 @@ skip_nibble_rect_write:
                                 (dev->accel.cx <= clip_r) &&
                                 (dev->accel.cy >= clip_t) &&
                                 (dev->accel.cy <= clip_b)) {
+                                dev->subsys_stat |= 0x02;
                                 switch ((mix_dat & mix_mask) ? frgd_mix : bkgd_mix) {
                                     case 0:
                                         src_dat = bkgd_color;
@@ -2546,6 +2548,7 @@ skip_nibble_rect_write:
                         (dev->accel.cx <= clip_r) &&
                         (dev->accel.cy >= clip_t) &&
                         (dev->accel.cy <= clip_b)) {
+                        dev->subsys_stat |= 0x02;
                         switch ((mix_dat & mix_mask) ? frgd_mix : bkgd_mix) {
                             case 0:
                                 src_dat = bkgd_color;
@@ -2650,9 +2653,10 @@ skip_nibble_rect_write:
                     dev->accel.cx = CLAMP(dev->accel.cx, clip_l, clip_r);
 
                     if ((dev->accel.cx >= clip_l) &&
-                        (dev->accel.cx < clip_r) &&
+                        (dev->accel.cx <= clip_r) &&
                         (dev->accel.cy >= clip_t) &&
                         (dev->accel.cy <= clip_b)) {
+                        dev->subsys_stat |= 0x02;
                         switch ((mix_dat & mix_mask) ? frgd_mix : bkgd_mix) {
                             case 0:
                                 src_dat = bkgd_color;
@@ -2811,6 +2815,7 @@ skip_nibble_rect_write:
                         (dev->accel.dx <= clip_r) &&
                         (dev->accel.dy >= clip_t) &&
                         (dev->accel.dy <= clip_b)) {
+                        dev->subsys_stat |= 0x02;
                         if (pixcntl == 3) {
                             if (!(dev->accel.cmd & 0x10) && ((frgd_mix != 3) || (bkgd_mix != 3))) {
                                 READ(dev->accel.src + dev->accel.cx, mix_dat);
@@ -2978,6 +2983,7 @@ skip_nibble_bitblt_write:
                                 (dev->accel.dx <= clip_r) &&
                                 (dev->accel.dy >= clip_t) &&
                                 (dev->accel.dy <= clip_b)) {
+                                dev->subsys_stat |= 0x02;
                                 switch ((mix_dat & mix_mask) ? frgd_mix : bkgd_mix) {
                                     case 0:
                                         src_dat = bkgd_color;
@@ -3075,6 +3081,7 @@ skip_nibble_bitblt_write:
                                 (dev->accel.dx <= clip_r) &&
                                 (dev->accel.dy >= clip_t) &&
                                 (dev->accel.dy <= clip_b)) {
+                                dev->subsys_stat |= 0x02;
                                 switch ((mix_dat & 0x01) ? frgd_mix : bkgd_mix) {
                                     case 0:
                                         src_dat = bkgd_color;
@@ -3174,6 +3181,7 @@ skip_nibble_bitblt_write:
                                 (dx <= (((uint64_t)clip_r) * 3)) &&
                                 (dev->accel.dy >= (clip_t << 1)) &&
                                 (dev->accel.dy <= (clip_b << 1))) {
+                                dev->subsys_stat |= 0x02;
 
                                 READ(dev->accel.src + cx, src_dat);
                                 READ(dev->accel.dest + dx, dest_dat);
@@ -3196,6 +3204,7 @@ skip_nibble_bitblt_write:
                                 (dev->accel.dx <= clip_r) &&
                                 (dev->accel.dy >= clip_t) &&
                                 (dev->accel.dy <= clip_b)) {
+                                dev->subsys_stat |= 0x02;
                                 if (pixcntl == 3) {
                                     if (!(dev->accel.cmd & 0x10) && ((frgd_mix != 3) || (bkgd_mix != 3))) {
                                         READ(dev->accel.src + dev->accel.cx, mix_dat);
@@ -3704,6 +3713,8 @@ ibm8514_poll(void *priv)
             dev->vc &= 0xfff;
 
             if (dev->vc == dev->dispend) {
+                dev->subsys_stat |= 0x01;
+                ibm8514_log("VBLANK irq.\n");
                 dev->dispon = 0;
 
                 for (x = 0; x < ((dev->vram_mask + 1) >> 12); x++) {
diff --git a/src/video/vid_ati68860_ramdac.c b/src/video/vid_ati68860_ramdac.c
index bb5ccbf31..17b2a6d22 100644
--- a/src/video/vid_ati68860_ramdac.c
+++ b/src/video/vid_ati68860_ramdac.c
@@ -70,20 +70,19 @@ void
 ati68860_ramdac_out(uint16_t addr, uint8_t val, void *priv, svga_t *svga)
 {
     ati68860_ramdac_t *ramdac = (ati68860_ramdac_t *) priv;
-    const ibm8514_t   *dev    = (ibm8514_t *) svga->dev8514;
 
     switch (addr) {
         case 0:
-            svga_out((dev && dev->on) ? 0x2ec : 0x3c8, val, svga);
+            svga_out(0x3c8, val, svga);
             break;
         case 1:
-            svga_out((dev && dev->on) ? 0x2ed : 0x3c9, val, svga);
+            svga_out(0x3c9, val, svga);
             break;
         case 2:
-            svga_out((dev && dev->on) ? 0x2ea : 0x3c6, val, svga);
+            svga_out(0x3c6, val, svga);
             break;
         case 3:
-            svga_out((dev && dev->on) ? 0x2eb : 0x3c7, val, svga);
+            svga_out(0x3c7, val, svga);
             break;
         default:
             ramdac->regs[addr & 0xf] = val;
@@ -176,21 +175,20 @@ uint8_t
 ati68860_ramdac_in(uint16_t addr, void *priv, svga_t *svga)
 {
     const ati68860_ramdac_t *ramdac = (ati68860_ramdac_t *) priv;
-    const ibm8514_t         *dev    = (ibm8514_t *) svga->dev8514;
     uint8_t                  temp   = 0;
 
     switch (addr) {
         case 0:
-            temp = svga_in((dev && dev->on) ? 0x2ec : 0x3c8, svga);
+            temp = svga_in(0x3c8, svga);
             break;
         case 1:
-            temp = svga_in((dev && dev->on) ? 0x2ed : 0x3c9, svga);
+            temp = svga_in(0x3c9, svga);
             break;
         case 2:
-            temp = svga_in((dev && dev->on) ? 0x2ea : 0x3c6, svga);
+            temp = svga_in(0x3c6, svga);
             break;
         case 3:
-            temp = svga_in((dev && dev->on) ? 0x2eb : 0x3c7, svga);
+            temp = svga_in(0x3c7, svga);
             break;
         case 4:
         case 8:
diff --git a/src/video/vid_ati_mach8.c b/src/video/vid_ati_mach8.c
index 2038f1b1d..6e2798295 100644
--- a/src/video/vid_ati_mach8.c
+++ b/src/video/vid_ati_mach8.c
@@ -438,6 +438,7 @@ mach_accel_start(int cmd_type, int cpu_input, int count, uint32_t mix_dat, uint3
                         (dev->accel.dx <= clip_r) &&
                         (dev->accel.dy >= clip_t) &&
                         (dev->accel.dy <= clip_b)) {
+                        dev->subsys_stat |= 0x02;
                         switch (mix ? frgd_sel : bkgd_sel) {
                             case 0:
                                 src_dat = dev->accel.bkgd_color;
@@ -661,6 +662,7 @@ mach_accel_start(int cmd_type, int cpu_input, int count, uint32_t mix_dat, uint3
                         (dev->accel.dx <= clip_r) &&
                         (dev->accel.dy >= clip_t) &&
                         (dev->accel.dy <= clip_b)) {
+                        dev->subsys_stat |= 0x02;
                         switch (mix ? frgd_sel : bkgd_sel) {
                             case 0:
                                 src_dat = dev->accel.bkgd_color;
@@ -1067,6 +1069,7 @@ mach_accel_start(int cmd_type, int cpu_input, int count, uint32_t mix_dat, uint3
                     (dev->accel.dx <= clip_r) &&
                     (dev->accel.dy >= clip_t) &&
                     (dev->accel.dy <= clip_b)) {
+                    dev->subsys_stat |= 0x02;
                     if (mach->accel.dp_config & 0x02) {
                         READ(dev->accel.src + dev->accel.cx, poly_src);
                         poly_src = ((poly_src & rd_mask) == rd_mask);
@@ -1308,6 +1311,7 @@ mach_accel_start(int cmd_type, int cpu_input, int count, uint32_t mix_dat, uint3
                             (dev->accel.cx <= clip_r) &&
                             (dev->accel.cy >= clip_t) &&
                             (dev->accel.cy <= clip_b)) {
+                            dev->subsys_stat |= 0x02;
                             mach->accel.clip_overrun = 0;
                             switch (mix ? frgd_sel : bkgd_sel) {
                                 case 0:
@@ -1441,6 +1445,7 @@ mach_accel_start(int cmd_type, int cpu_input, int count, uint32_t mix_dat, uint3
                             (dev->accel.cx <= clip_r) &&
                             (dev->accel.cy >= clip_t) &&
                             (dev->accel.cy <= clip_b)) {
+                            dev->subsys_stat |= 0x02;
                             mach->accel.clip_overrun = 0;
                             if (mach->accel.linedraw_opt & 0x02) {
                                 if (dev->bpp) {
@@ -1599,6 +1604,7 @@ mach_accel_start(int cmd_type, int cpu_input, int count, uint32_t mix_dat, uint3
                             (dev->accel.cx <= clip_r) &&
                             (dev->accel.cy >= clip_t) &&
                             (dev->accel.cy <= clip_b)) {
+                            dev->subsys_stat |= 0x02;
                             mach->accel.clip_overrun = 0;
                             switch (mix ? frgd_sel : bkgd_sel) {
                                 case 0:
@@ -1733,6 +1739,7 @@ mach_accel_start(int cmd_type, int cpu_input, int count, uint32_t mix_dat, uint3
                             (dev->accel.cx <= clip_r) &&
                             (dev->accel.cy >= clip_t) &&
                             (dev->accel.cy <= clip_b)) {
+                            dev->subsys_stat |= 0x02;
                             mach->accel.clip_overrun = 0;
                             switch (mix ? frgd_sel : bkgd_sel) {
                                 case 0:
@@ -2028,6 +2035,7 @@ mach_accel_start(int cmd_type, int cpu_input, int count, uint32_t mix_dat, uint3
                     (dev->accel.dx <= clip_r) &&
                     (dev->accel.dy >= clip_t) &&
                     (dev->accel.dy <= clip_b)) {
+                    dev->subsys_stat |= 0x02;
                     switch (mix ? frgd_sel : bkgd_sel) {
                         case 0:
                             src_dat = dev->accel.bkgd_color;
@@ -2333,6 +2341,7 @@ mach_out(uint16_t addr, uint8_t val, void *priv)
         case 0x2ed:
             rs2 = !!(mach->accel.ext_ge_config & 0x1000);
             rs3 = !!(mach->accel.ext_ge_config & 0x2000);
+            mach_log("8514/A RS2=%d, RS3=%d, addr=%03x.\n", rs2, rs3, addr);
             if ((dev->local & 0xff) >= 0x02) {
                 if (mach->regs[0xb0] & 0x20) { /*ATI extended 8514/A mode.*/
                     mach_log("Extended 8514/A mode.\n");
@@ -2341,10 +2350,14 @@ mach_out(uint16_t addr, uint8_t val, void *priv)
                     svga_recalctimings(svga);
                     mach32_updatemapping(mach, svga);
                 }
-                if (mach->pci_bus && !mach->ramdac_type)
-                    ati68860_ramdac_out((addr & 0x03) | (rs2 << 2) | (rs3 << 3), val, svga->ramdac, svga);
-                else
-                    ati68875_ramdac_out(addr, rs2, rs3, val, svga->ramdac, svga);
+                if (dev->on)
+                    svga_out(addr, val, svga);
+                else {
+                    if (mach->pci_bus && !mach->ramdac_type)
+                        ati68860_ramdac_out((addr & 0x03) | (rs2 << 2) | (rs3 << 3), val, svga->ramdac, svga);
+                    else
+                        ati68875_ramdac_out(addr, rs2, rs3, val, svga->ramdac, svga);
+                }
             } else
                 svga_out(addr, val, svga);
             return;
@@ -2355,6 +2368,7 @@ mach_out(uint16_t addr, uint8_t val, void *priv)
         case 0x3C9:
             rs2 = !!(mach->regs[0xa0] & 0x20);
             rs3 = !!(mach->regs[0xa0] & 0x40);
+            mach_log("VGA RS2=%d, RS3=%d, addr=%03x.\n", rs2, rs3, addr);
             if ((dev->local & 0xff) >= 0x02) {
                 if (svga->attrregs[0x10] & 0x40) {
                     mach_log("VGA mode.\n");
@@ -2363,10 +2377,14 @@ mach_out(uint16_t addr, uint8_t val, void *priv)
                     svga_recalctimings(svga);
                     mach32_updatemapping(mach, svga);
                 }
-                if (mach->pci_bus && !mach->ramdac_type)
-                    ati68860_ramdac_out((addr & 0x03) | (rs2 << 2) | (rs3 << 3), val, svga->ramdac, svga);
-                else
-                    ati68875_ramdac_out(addr, rs2, rs3, val, svga->ramdac, svga);
+                if (dev->on)
+                    svga_out(addr, val, svga);
+                else {
+                    if (mach->pci_bus && !mach->ramdac_type)
+                        ati68860_ramdac_out((addr & 0x03) | (rs2 << 2) | (rs3 << 3), val, svga->ramdac, svga);
+                    else
+                        ati68875_ramdac_out(addr, rs2, rs3, val, svga->ramdac, svga);
+                }
             } else
                 svga_out(addr, val, svga);
             return;
@@ -2481,10 +2499,14 @@ mach_in(uint16_t addr, void *priv)
             rs2 = !!(mach->accel.ext_ge_config & 0x1000);
             rs3 = !!(mach->accel.ext_ge_config & 0x2000);
             if ((dev->local & 0xff) >= 0x02) {
-                if (mach->pci_bus && !mach->ramdac_type)
-                    temp = ati68860_ramdac_in((addr & 3) | (rs2 << 2) | (rs3 << 3), svga->ramdac, svga);
-                else
-                    temp = ati68875_ramdac_in(addr, rs2, rs3, svga->ramdac, svga);
+                if (dev->on)
+                    temp = svga_in(addr, svga);
+                else {
+                    if (mach->pci_bus && !mach->ramdac_type)
+                        temp = ati68860_ramdac_in((addr & 3) | (rs2 << 2) | (rs3 << 3), svga->ramdac, svga);
+                    else
+                        temp = ati68875_ramdac_in(addr, rs2, rs3, svga->ramdac, svga);
+                }
             } else
                 temp = svga_in(addr, svga);
             break;
@@ -3407,7 +3429,7 @@ mach_accel_out_fifo(mach_t *mach, svga_t *svga, ibm8514_t *dev, uint16_t port, u
 static void
 mach_accel_out_call(uint16_t port, uint8_t val, mach_t *mach, svga_t *svga, ibm8514_t *dev)
 {
-    if (port != 0x42e8 && port != 0x42e9)
+    if (port == 0x42e8 || port == 0x42e9)
         mach_log("[%04X:%08X]: Port CALL OUT=%04x, val=%02x.\n", CS, cpu_state.pc, port, val);
 
     switch (port) {
@@ -3422,7 +3444,7 @@ mach_accel_out_call(uint16_t port, uint8_t val, mach_t *mach, svga_t *svga, ibm8
             break;
         case 0x42e9:
             ibm8514_accel_out(port, val, svga, 2);
-            if ((val & 0xc0) == 0xc0) {
+            if ((val & 0xc0) == 0x80) {
                 dev->ext_fifo_idx = 0;
                 mach->force_busy = 0;
             }
@@ -3513,6 +3535,11 @@ mach_accel_out_call(uint16_t port, uint8_t val, mach_t *mach, svga_t *svga, ibm8
             svga_recalctimings(svga);
             break;
 
+        case 0x46e8:
+        case 0x46e9:
+            mach_log("0x%04x write: VGA subsystem enable add-on=%02x.\n", port, val);
+            break;
+
         case 0x4ae8:
             dev->accel.advfunc_cntl = val;
             dev->on = dev->accel.advfunc_cntl & 0x01;
@@ -4201,6 +4228,7 @@ mach_accel_in_call(uint16_t port, mach_t *mach, svga_t *svga, ibm8514_t *dev)
 
     switch (port) {
         case 0x2e8:
+        case 0x2e9:
         case 0x6e8:
         case 0x22e8:
         case 0x26e8:
@@ -4212,18 +4240,22 @@ mach_accel_in_call(uint16_t port, mach_t *mach, svga_t *svga, ibm8514_t *dev)
 
         case 0x42e8:
         case 0x42e9:
-            if (dev->vc == dev->v_syncstart)
+            if ((dev->subsys_cntl & 0x01) && !(dev->subsys_stat & 0x01) && (dev->vc == dev->dispend))
                 temp |= 0x01;
 
             if (mach->accel.cmd_type == -1) {
                 if (cmd == 6) {
-                    if ((dev->accel.dx >= clip_l) &&
+                    if ((dev->subsys_cntl & 0x02) &&
+                        !(dev->subsys_stat & 0x02) &&
+                        (dev->accel.dx >= clip_l) &&
                         (dev->accel.dx <= clip_r_ibm) &&
                         (dev->accel.dy >= clip_t) &&
                         (dev->accel.dy <= clip_b_ibm))
                         temp |= 0x02;
                 } else {
-                    if ((dev->accel.cx >= clip_l) &&
+                    if ((dev->subsys_cntl & 0x02) &&
+                        !(dev->subsys_stat & 0x02) &&
+                        (dev->accel.cx >= clip_l) &&
                         (dev->accel.cx <= clip_r_ibm) &&
                         (dev->accel.cy >= clip_t) &&
                         (dev->accel.cy <= clip_b_ibm))
@@ -4234,7 +4266,9 @@ mach_accel_in_call(uint16_t port, mach_t *mach, svga_t *svga, ibm8514_t *dev)
                     case 1:
                     case 2:
                     case 5:
-                        if ((dev->accel.dx >= clip_l) &&
+                        if ((dev->subsys_cntl & 0x02) &&
+                            !(dev->subsys_stat & 0x02) &&
+                            (dev->accel.dx >= clip_l) &&
                             (dev->accel.dx <= clip_r) &&
                             (dev->accel.dy >= clip_t) &&
                             (dev->accel.dy <= clip_b))
@@ -4242,7 +4276,9 @@ mach_accel_in_call(uint16_t port, mach_t *mach, svga_t *svga, ibm8514_t *dev)
                         break;
                     case 3:
                     case 4:
-                        if ((dev->accel.cx >= clip_l) &&
+                        if ((dev->subsys_cntl & 0x02) &&
+                            !(dev->subsys_stat & 0x02) &&
+                            (dev->accel.cx >= clip_l) &&
                             (dev->accel.cx <= clip_r) &&
                             (dev->accel.cy >= clip_t) &&
                             (dev->accel.cy <= clip_b))
@@ -4257,16 +4293,17 @@ mach_accel_in_call(uint16_t port, mach_t *mach, svga_t *svga, ibm8514_t *dev)
                 if ((!dev->force_busy && !dev->force_busy2) || !mach->force_busy)
                     temp |= 0x08;
             }
-            if (port & 1)
+            if (port & 1) {
                 temp = dev->vram_512k_8514 ? 0x00 : 0x80;
-            else {
+                temp |= (dev->subsys_cntl >> 8);
+            } else {
                 temp |= (dev->subsys_stat | (dev->vram_512k_8514 ? 0x00 : 0x80));
                 if (mach->accel.ext_ge_config & 0x08)
                     temp |= ((mach->accel.ext_ge_config & 0x07) << 4);
                 else
                     temp |= 0x20;
             }
-            mach_log("0x%04x read: Subsystem Status=%02x.\n", port, temp);
+            mach_log("0x%04x read: Subsystem Status=%02x, monitoralias=%02x.\n", port, temp, mach->accel.ext_ge_config & 0x07);
             break;
 
             /*ATI Mach8/32 specific registers*/
@@ -4403,7 +4440,8 @@ mach_accel_in_call(uint16_t port, mach_t *mach, svga_t *svga, ibm8514_t *dev)
         default:
             break;
     }
-    mach_log("[%04X:%08X]: Port NORMAL IN=%04x, temp=%04x.\n", CS, cpu_state.pc, port, temp);
+    if (port == 0x2ee8 || port == 0x2ee9 || port == 0x42e8 || port == 0x42e9)
+        mach_log("[%04X:%08X]: Port NORMAL IN=%04x, temp=%04x.\n", CS, cpu_state.pc, port, temp);
 
     return temp;
 }
@@ -5419,6 +5457,10 @@ mach32_updatemapping(mach_t *mach, svga_t *svga)
             mem_mapping_set_handler(&svga->mapping, mach32_read, mach32_readw, mach32_readl, mach32_write, mach32_writew, mach32_writel);
             mem_mapping_set_p(&svga->mapping, mach);
         } else {
+            if (!dev->on) {
+                memset(dev->vram, 0, dev->vram_size);
+                memset(dev->changedvram, 0, (dev->vram_size >> 12) + 1);
+            }
             mach_log("IBM compatible banked mapping.\n");
             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);

From d25aed2da98cfc4e2208bb46cd24776d5e6b6020 Mon Sep 17 00:00:00 2001
From: Cacodemon345 <wahil1976@outlook.com>
Date: Fri, 7 Mar 2025 00:36:09 +0600
Subject: [PATCH 51/58] Add global mute option

---
 src/86box.c                 |   1 +
 src/config.c                |   5 ++++
 src/include/86box/86box.h   |   1 +
 src/qt/icons/sound_mute.ico | Bin 0 -> 9622 bytes
 src/qt/qt_machinestatus.cpp |  45 +++++++++++++++++++++++++++++++-----
 src/qt/qt_machinestatus.hpp |   6 +++++
 src/qt/qt_mainwindow.cpp    |   1 +
 src/qt_resources.qrc        |   1 +
 src/sound/openal.c          |   2 +-
 src/sound/xaudio2.c         |   2 +-
 10 files changed, 56 insertions(+), 8 deletions(-)
 create mode 100644 src/qt/icons/sound_mute.ico

diff --git a/src/86box.c b/src/86box.c
index 6b64cd2ec..aa3afb085 100644
--- a/src/86box.c
+++ b/src/86box.c
@@ -213,6 +213,7 @@ int      do_auto_pause                          = 0;              /* (C) Auto-pa
 int      hook_enabled                           = 1;              /* (C) Keyboard hook is enabled */
 int      test_mode                              = 0;              /* (C) Test mode */
 char     uuid[MAX_UUID_LEN]                     = { '\0' };       /* (C) UUID or machine identifier */
+int      sound_muted                            = 0;              /* (C) Is sound muted? */
 
 int      other_ide_present = 0;                                   /* IDE controllers from non-IDE cards are
                                                                      present */
diff --git a/src/config.c b/src/config.c
index 6c34c4790..07e770822 100644
--- a/src/config.c
+++ b/src/config.c
@@ -168,6 +168,7 @@ load_general(void)
     kbd_req_capture = ini_section_get_int(cat, "kbd_req_capture", 0);
     hide_status_bar = ini_section_get_int(cat, "hide_status_bar", 0);
     hide_tool_bar   = ini_section_get_int(cat, "hide_tool_bar", 0);
+    sound_muted     = ini_section_get_int(cat, "sound_muted", 0);
 
     confirm_reset = ini_section_get_int(cat, "confirm_reset", 1);
     confirm_exit  = ini_section_get_int(cat, "confirm_exit", 1);
@@ -1846,6 +1847,10 @@ save_general(void)
 
     const char *va_name;
 
+    ini_section_set_int(cat, "sound_muted", sound_muted);
+    if (sound_muted == 0)
+        ini_section_delete_var(cat, "sound_muted");
+
     ini_section_set_int(cat, "vid_resize", vid_resize);
     if (vid_resize == 0)
         ini_section_delete_var(cat, "vid_resize");
diff --git a/src/include/86box/86box.h b/src/include/86box/86box.h
index e4e5ecae1..79bb84009 100644
--- a/src/include/86box/86box.h
+++ b/src/include/86box/86box.h
@@ -158,6 +158,7 @@ extern int      other_scsi_present;         /* SCSI controllers from non-SCSI ca
 extern int    hard_reset_pending;
 extern int    fixed_size_x;
 extern int    fixed_size_y;
+extern int    sound_muted;                  /* (C) Is sound muted? */
 extern int    do_auto_pause;                /* (C) Auto-pause the emulator on focus loss */
 extern int    auto_paused;
 extern double mouse_sensitivity;            /* (C) Mouse sensitivity scale */
diff --git a/src/qt/icons/sound_mute.ico b/src/qt/icons/sound_mute.ico
new file mode 100644
index 0000000000000000000000000000000000000000..89a8407354ce0ef805689d75763cd17fd58ce52b
GIT binary patch
literal 9622
zcmeI2F^?QY5QTg3frUWg0!ujC3km%XxJ{hmz$3y#LPCTj#2ygBWDxKJHUNPJ38zQ|
zcNRau(TdPGxd2PZiFL4Mx2Wo#Z>PIwW_NDxI1$iN@9VCvdhd15-pqPr=O%8??e9Cq
z1NY&cbGMyyH*SRaPp>%lqsDID4D&Bub?$@LoI5xO^Y69p56yi`Yjmg#A+F#KweGyF
zEK7-MY)p^By+dDHbLd#2?s<9((G6%GsHMc_-5hDH<x_8CUk4gT(x}y-UpqNDQEckN
z`o;_O9;y9iK2PQ;2eroOTaw#ds}B5iu$s`(GKSOBQ_n%Ih7#_e4@aYEBY4PxgE>fE
zj>d5jeXF+yFZvqtd+5_=-D+t$2Xi>cQ{;T2P#97}4z<K{Y!PUH95|wRS9)iR8n82I
zcd&P#l*;=e>rX^huZt|d7J0reQt7;yDgK~X1Y9Wo{<E$tE{`?;p4N$o-DmW6<L@_i
zzq0m~-lHz5ng6_xrB)vP_G0Y+E_4Zn>rqBraTdITIc;y+;IpxD#LdsMSiBy_d{0t;
z<i(u3`BE<9i2dc9+xg(qD!+&D_XM8^>j90cpO0Mb5zKSW^g5eX()^rSxKe*UU-PUs
z?bKNYV=(9H=JedHkDLQ8koyowejl0u_ho9}(k0HsBkNJz%B_(XX8MU9+&L#)<V75C
zfg4T+IVT)&MqHHia#@1a!$~b%aQ7hsT?i+&aCYQw^%v*oobg&V9?YqAJ{O-Ek4?|R
zMfsQ-YDc5|`L@WjsmL#1ihTP><n|+3fA@~8W^c%HK9$9tY5lx9QaqMLaU@H{zZBOG
zjzx|hix?_cv*(ungm$;I^W1t*e@kCtvvbP(Wzg50rDEPUdSi7WPVU7qqUUwaqv3h&
z-1SI&w{hNB3;P<Zp?ua#PnU6orTWtn!n*Ba5Pt`EtmApcYSi53qs^Ilb^HxyaX1nW
z?5(*sxB0Cg<MJF?EaqqIvElCakB5ve*GxFsW6Eu|TK<!|SR)?S9umLYtI<Gni-sA9
zeA{=(Uw1rgp0jX*wSIl=;yu7&JaKP-Kk)}=?w@(k3L4EBz!%Ws)=U##+Z)dwG{Drq
zb3rTk4NU20vln9x4IMb~%RKn;x&Ay^Z@A*~Y4U1%4`}53&_%9M!cBg@p3kE^#~8#v
zHTY&-<9!WJed`-7G~h&|x3F0ltf3{HXM^T0Y$V$wYSD=1?c_AF(?98Og2V^lU=8Iu
zPp(Om4-T;N-%OFer~13=y2x+W^dD;zk=YYjm(OIicSn|=-;%}4`%-;c>idF+Qt5he
zIFV|8U5dMJN>x>|c=t?}4_?ShUnHy*<!B{xyxiP4?049K_CxLb4lC`&<<&L2MtQ#q
z^;)x3%=^aAzvUy+OrH^Y6w#XZamgOi=iugx5DkI{e2n&B=iAXz{c!eanptViVU)ht
zZr_t)?R1<vF!j+M*msAAwq~ZYv602$+^qKu$n(hasyT<YzVT#Qo;r)c+anuF>~s4w
zvs>TlGQE7hOW&`fZFI&tbpMuXE%p$PZOJEee{9v><$-GhHoAlU{t|8W4;R{CupeIo
zG)m=(`_}J1XVIz6S)mEyDXa$%>0Z&aXlNbx2CoSh`^_=;Kt6Fi7=KseEU2q7*)RL*
z)25a@=BWprCFjmHuz0)Wzpry`=D8Lg+Kj^^aH0Vxb5I$RYeXLPV4g?356$V&{o6IQ
zd~F`_5JKZV@x?yjWDcT9K6%U|^~7UGJPL&&9J|gzvy**Omm_(^nS;bZjXTC5`v$3H
z-yr$Mv7_~!OREe{v_ax%gDrwf2%v?wQYL4QHhDu49`d(YC8y5=>)SXkx`qZGn6v*8
z{DZ#-aHFa7<QhbAQimcQhUU9S%hyKp!5kM&L=(Iy-^luNpWR5E|I4o61SEM8aWav!
zhfn0GzKB?S_K{TjBBHpzzpn0=t70}=Rwt81ad@q&<}Ve+-KoBY(ErC!6ahbetnVeN
eMRD)frN;HWM6p(sLY_+{FJz(Pr`pf7N9`Z+b=u_s

literal 0
HcmV?d00001

diff --git a/src/qt/qt_machinestatus.cpp b/src/qt/qt_machinestatus.cpp
index dee90f11c..3c3f4019f 100644
--- a/src/qt/qt_machinestatus.cpp
+++ b/src/qt/qt_machinestatus.cpp
@@ -39,6 +39,7 @@ extern "C" {
 #include <86box/network.h>
 #include <86box/ui.h>
 #include <86box/machine_status.h>
+#include <86box/config.h>
 };
 
 #include <QIcon>
@@ -89,7 +90,7 @@ struct Pixmaps {
     PixmapSetEmptyActive mo;
     PixmapSetActive      hd;
     PixmapSetEmptyActive net;
-    QPixmap              sound;
+    QPixmap              sound, soundMuted;
 };
 
 struct StateActive {
@@ -215,6 +216,7 @@ struct MachineStatus::States {
         pixmaps.hd.load("/hard_disk%1.ico");
         pixmaps.net.load("/network%1.ico");
         pixmaps.sound = ProgSettings::loadIcon("/sound.ico").pixmap(pixmap_size);
+        pixmaps.soundMuted = ProgSettings::loadIcon("/sound_mute.ico").pixmap(pixmap_size);
 
         cartridge[0].pixmaps = &pixmaps.cartridge;
         cartridge[1].pixmaps = &pixmaps.cartridge;
@@ -256,12 +258,19 @@ MachineStatus::MachineStatus(QObject *parent)
     , refreshTimer(new QTimer(this))
 {
     d = std::make_unique<MachineStatus::States>(this);
+    muteUnmuteAction = nullptr;
     connect(refreshTimer, &QTimer::timeout, this, &MachineStatus::refreshIcons);
     refreshTimer->start(75);
 }
 
 MachineStatus::~MachineStatus() = default;
 
+void
+MachineStatus::setSoundGainAction(QAction* action)
+{
+    soundGainAction = action;
+}
+
 bool
 MachineStatus::hasCassette()
 {
@@ -494,6 +503,28 @@ MachineStatus::refresh(QStatusBar *sbar)
     }
     sbar->removeWidget(d->sound.get());
 
+    if (!muteUnmuteAction) {
+        muteUnmuteAction = new QAction;
+        connect(muteUnmuteAction, &QAction::triggered, this, [this]() {
+            sound_muted ^= 1;
+            config_save();
+            if (d->sound)
+                d->sound->setPixmap(sound_muted ? d->pixmaps.soundMuted : d->pixmaps.sound);
+            
+            muteUnmuteAction->setText(sound_muted ? tr("&Unmute") : tr("&Mute"));
+        });
+    }
+
+    if (!soundMenu) {
+        soundMenu = new QMenu((QWidget*)parent());
+
+        soundMenu->addAction(muteUnmuteAction);
+        soundMenu->addSeparator();
+        soundMenu->addAction(soundGainAction);
+
+        muteUnmuteAction->setParent(soundMenu);
+    }
+
     if (cassette_enable) {
         d->cassette.label = std::make_unique<ClickableLabel>();
         d->cassette.setEmpty(QString(cassette_fname).isEmpty());
@@ -662,12 +693,14 @@ MachineStatus::refresh(QStatusBar *sbar)
     }
 
     d->sound = std::make_unique<ClickableLabel>();
-    d->sound->setPixmap(d->pixmaps.sound);
-
-    connect(d->sound.get(), &ClickableLabel::doubleClicked, d->sound.get(), [](QPoint pos) {
-        SoundGain gain(main_window);
-        gain.exec();
+    d->sound->setPixmap(sound_muted ? d->pixmaps.soundMuted : d->pixmaps.sound);
+    if (muteUnmuteAction)
+        muteUnmuteAction->setText(sound_muted ? tr("&Unmute") : tr("&Mute"));
+    
+    connect(d->sound.get(), &ClickableLabel::clicked, this, [this](QPoint pos) {
+        this->soundMenu->popup(pos - QPoint(0, this->soundMenu->sizeHint().height()));
     });
+
     d->sound->setToolTip(tr("Sound"));
     sbar->addWidget(d->sound.get());
     d->text = std::make_unique<QLabel>();
diff --git a/src/qt/qt_machinestatus.hpp b/src/qt/qt_machinestatus.hpp
index fc0e33e91..90b420763 100644
--- a/src/qt/qt_machinestatus.hpp
+++ b/src/qt/qt_machinestatus.hpp
@@ -1,6 +1,8 @@
 #ifndef QT_MACHINESTATUS_HPP
 #define QT_MACHINESTATUS_HPP
 
+#include <QAction>
+#include <QMenu>
 #include <QWidget>
 #include <QLabel>
 #include <QMouseEvent>
@@ -71,6 +73,7 @@ public:
 
     QString getMessage();
     void    clearActivity();
+    void    setSoundGainAction(QAction* action);
 public slots:
     void refresh(QStatusBar *sbar);
     void message(const QString &msg);
@@ -82,6 +85,9 @@ private:
     struct States;
     std::unique_ptr<States> d;
     QTimer                 *refreshTimer;
+    QAction                *soundGainAction;
+    QAction                *muteUnmuteAction;
+    QMenu                  *soundMenu;
 };
 
 #endif // QT_MACHINESTATUS_HPP
diff --git a/src/qt/qt_mainwindow.cpp b/src/qt/qt_mainwindow.cpp
index 552fd0310..61f5f6fe1 100644
--- a/src/qt/qt_mainwindow.cpp
+++ b/src/qt/qt_mainwindow.cpp
@@ -178,6 +178,7 @@ MainWindow::MainWindow(QWidget *parent)
     extern MainWindow *main_window;
     main_window = this;
     ui->setupUi(this);
+    status->setSoundGainAction(ui->actionSound_gain);
     ui->stackedWidget->setMouseTracking(true);
     statusBar()->setVisible(!hide_status_bar);
     statusBar()->setStyleSheet("QStatusBar::item {border: None; } QStatusBar QLabel { margin-right: 2px; margin-bottom: 1px; }");
diff --git a/src/qt_resources.qrc b/src/qt_resources.qrc
index 1f6341786..6ca323b89 100644
--- a/src/qt_resources.qrc
+++ b/src/qt_resources.qrc
@@ -43,6 +43,7 @@
         <file>qt/icons/other_removable_devices.ico</file>
         <file>qt/icons/ports.ico</file>
         <file>qt/icons/sound.ico</file>
+        <file>qt/icons/sound_mute.ico</file>
         <file>qt/icons/storage_controllers.ico</file>
         <file>qt/icons/zip.ico</file>
         <file>qt/icons/zip_active.ico</file>
diff --git a/src/sound/openal.c b/src/sound/openal.c
index a41199b23..90f626362 100644
--- a/src/sound/openal.c
+++ b/src/sound/openal.c
@@ -278,7 +278,7 @@ givealbuffer_common(const void *buf, const uint8_t src, const int size, const in
 
     alGetSourcei(source[src], AL_BUFFERS_PROCESSED, &processed);
     if (processed >= 1) {
-        const double gain = pow(10.0, (double) sound_gain / 20.0);
+        const double gain = sound_muted ? 0.0 : pow(10.0, (double) sound_gain / 20.0);
         alListenerf(AL_GAIN, (float) gain);
 
         alSourceUnqueueBuffers(source[src], 1, &buffer);
diff --git a/src/sound/xaudio2.c b/src/sound/xaudio2.c
index 9b341f574..2aee97efc 100644
--- a/src/sound/xaudio2.c
+++ b/src/sound/xaudio2.c
@@ -245,7 +245,7 @@ givealbuffer_common(const void *buf, IXAudio2SourceVoice *sourcevoice, const siz
     if (!initialized)
         return;
 
-    (void) IXAudio2MasteringVoice_SetVolume(mastervoice, pow(10.0, (double) sound_gain / 20.0),
+    (void) IXAudio2MasteringVoice_SetVolume(mastervoice, sound_muted ? 0.0 : pow(10.0, (double) sound_gain / 20.0),
                                             XAUDIO2_COMMIT_NOW);
     XAUDIO2_BUFFER buffer = { 0 };
     buffer.Flags          = 0;

From c1c4bcc889d59f5a4f714d6802e5592054f5cacf Mon Sep 17 00:00:00 2001
From: Alexander Babikov <lemondrops358@gmail.com>
Date: Sat, 1 Mar 2025 15:54:50 +0500
Subject: [PATCH 52/58] Disable Win11 rounded window corners when the status
 bar is hidden

---
 src/qt/qt_mainwindow.cpp | 21 +++++++++++++++++++++
 1 file changed, 21 insertions(+)

diff --git a/src/qt/qt_mainwindow.cpp b/src/qt/qt_mainwindow.cpp
index 61f5f6fe1..9802a2666 100644
--- a/src/qt/qt_mainwindow.cpp
+++ b/src/qt/qt_mainwindow.cpp
@@ -122,6 +122,19 @@ extern int qt_nvr_save(void);
 #include "x11_util.h"
 #endif
 
+#ifdef Q_OS_WINDOWS
+#include <dwmapi.h>
+#ifndef DWMWA_WINDOW_CORNER_PREFERENCE
+#define DWMWA_WINDOW_CORNER_PREFERENCE 33
+#endif
+#ifndef DWMWCP_DEFAULT
+#define DWMWCP_DEFAULT 0
+#endif
+#ifndef DWMWCP_DONOTROUND
+#define DWMWCP_DONOTROUND 1
+#endif
+#endif
+
 #ifdef Q_OS_MACOS
 #    include "cocoa_keyboard.hpp"
 // The namespace is required to avoid clashing typedefs; we only use this
@@ -181,6 +194,10 @@ MainWindow::MainWindow(QWidget *parent)
     status->setSoundGainAction(ui->actionSound_gain);
     ui->stackedWidget->setMouseTracking(true);
     statusBar()->setVisible(!hide_status_bar);
+#ifdef Q_OS_WINDOWS
+    auto cornerPreference = (hide_status_bar ? DWMWCP_DONOTROUND : DWMWCP_DEFAULT);
+    DwmSetWindowAttribute((HWND) this->winId(), DWMWA_WINDOW_CORNER_PREFERENCE, (LPCVOID) &cornerPreference, sizeof(cornerPreference));
+#endif
     statusBar()->setStyleSheet("QStatusBar::item {border: None; } QStatusBar QLabel { margin-right: 2px; margin-bottom: 1px; }");
     this->centralWidget()->setStyleSheet("background-color: black;");
     ui->toolBar->setVisible(!hide_tool_bar);
@@ -1833,6 +1850,10 @@ MainWindow::on_actionHide_status_bar_triggered()
     hide_status_bar ^= 1;
     ui->actionHide_status_bar->setChecked(hide_status_bar);
     statusBar()->setVisible(!hide_status_bar);
+#ifdef Q_OS_WINDOWS
+    auto cornerPreference = (hide_status_bar ? DWMWCP_DONOTROUND : DWMWCP_DEFAULT);
+    DwmSetWindowAttribute((HWND) main_window->winId(), DWMWA_WINDOW_CORNER_PREFERENCE, (LPCVOID) &cornerPreference, sizeof(cornerPreference));
+#endif
     if (vid_resize >= 2) {
         setFixedSize(fixed_size_x, fixed_size_y + menuBar()->height() + (hide_status_bar ? 0 : statusBar()->height()) + (hide_tool_bar ? 0 : ui->toolBar->height()));
     } else {

From d6fa4d4f98f15f6b21ebffcff0bfa3a2f6b8a76a Mon Sep 17 00:00:00 2001
From: Alexander Babikov <lemondrops358@gmail.com>
Date: Thu, 6 Mar 2025 22:20:13 +0500
Subject: [PATCH 53/58] Disable Win11 rounded corners on the secondary monitor
 windows

---
 src/qt/qt_mainwindow.cpp | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/src/qt/qt_mainwindow.cpp b/src/qt/qt_mainwindow.cpp
index 9802a2666..036f56375 100644
--- a/src/qt/qt_mainwindow.cpp
+++ b/src/qt/qt_mainwindow.cpp
@@ -816,6 +816,10 @@ MainWindow::initRendererMonitorSlot(int monitor_index)
         if (vid_resize == 2)
             secondaryRenderer->setFixedSize(fixed_size_x, fixed_size_y);
         secondaryRenderer->setWindowIcon(this->windowIcon());
+#ifdef Q_OS_WINDOWS
+        auto cornerPreference = DWMWCP_DONOTROUND;
+        DwmSetWindowAttribute((HWND) secondaryRenderer->winId(), DWMWA_WINDOW_CORNER_PREFERENCE, (LPCVOID) &cornerPreference, sizeof(cornerPreference));
+#endif
         if (show_second_monitors) {
             secondaryRenderer->show();
             if (window_remember) {

From e6f96352e083a9df48e87babed6e3afadf7a0e89 Mon Sep 17 00:00:00 2001
From: Alexander Babikov <lemondrops358@gmail.com>
Date: Sat, 1 Mar 2025 02:01:00 +0500
Subject: [PATCH 54/58] Fix serial passthrough configure buttons not
 re-enabling when the checkboxes are toggled

---
 src/qt/qt_settingsports.cpp | 43 ++++++++++++++++---------------------
 src/qt/qt_settingsports.hpp | 14 ++++++------
 2 files changed, 25 insertions(+), 32 deletions(-)

diff --git a/src/qt/qt_settingsports.cpp b/src/qt/qt_settingsports.cpp
index 9679c9e21..3855a8b68 100644
--- a/src/qt/qt_settingsports.cpp
+++ b/src/qt/qt_settingsports.cpp
@@ -68,21 +68,14 @@ SettingsPorts::SettingsPorts(QWidget *parent)
     for (int i = 0; i < SERIAL_MAX; i++) {
         auto *checkBox     = findChild<QCheckBox *>(QString("checkBoxSerial%1").arg(i + 1));
         auto *checkBoxPass = findChild<QCheckBox *>(QString("checkBoxSerialPassThru%1").arg(i + 1));
+        auto *buttonPass   = findChild<QPushButton *>(QString("pushButtonSerialPassThru%1").arg(i + 1));
         if (checkBox != NULL)
             checkBox->setChecked(com_ports[i].enabled > 0);
-        if (checkBoxPass != NULL)
+        if (checkBoxPass != NULL) {
             checkBoxPass->setChecked(serial_passthrough_enabled[i]);
+            buttonPass->setEnabled(serial_passthrough_enabled[i]);
+        }
     }
-
-    ui->pushButtonSerialPassThru1->setEnabled(serial_passthrough_enabled[0]);
-    ui->pushButtonSerialPassThru2->setEnabled(serial_passthrough_enabled[1]);
-    ui->pushButtonSerialPassThru3->setEnabled(serial_passthrough_enabled[2]);
-    ui->pushButtonSerialPassThru4->setEnabled(serial_passthrough_enabled[3]);
-#if 0
-    ui->pushButtonSerialPassThru5->setEnabled(serial_passthrough_enabled[4]);
-    ui->pushButtonSerialPassThru6->setEnabled(serial_passthrough_enabled[5]);
-    ui->pushButtonSerialPassThru7->setEnabled(serial_passthrough_enabled[6]);
-#endif
 }
 
 SettingsPorts::~SettingsPorts()
@@ -181,45 +174,45 @@ SettingsPorts::on_pushButtonSerialPassThru7_clicked()
 #endif
 
 void
-SettingsPorts::on_checkBoxSerialPassThru1_clicked(bool checked)
+SettingsPorts::on_checkBoxSerialPassThru1_stateChanged(int state)
 {
-    ui->pushButtonSerialPassThru1->setEnabled(checked);
+    ui->pushButtonSerialPassThru1->setEnabled(state == Qt::Checked);
 }
 
 void
-SettingsPorts::on_checkBoxSerialPassThru2_clicked(bool checked)
+SettingsPorts::on_checkBoxSerialPassThru2_stateChanged(int state)
 {
-    ui->pushButtonSerialPassThru2->setEnabled(checked);
+    ui->pushButtonSerialPassThru2->setEnabled(state == Qt::Checked);
 }
 
 void
-SettingsPorts::on_checkBoxSerialPassThru3_clicked(bool checked)
+SettingsPorts::on_checkBoxSerialPassThru3_stateChanged(int state)
 {
-    ui->pushButtonSerialPassThru3->setEnabled(checked);
+    ui->pushButtonSerialPassThru3->setEnabled(state == Qt::Checked);
 }
 
 void
-SettingsPorts::on_checkBoxSerialPassThru4_clicked(bool checked)
+SettingsPorts::on_checkBoxSerialPassThru4_stateChanged(int state)
 {
-    ui->pushButtonSerialPassThru4->setEnabled(checked);
+    ui->pushButtonSerialPassThru4->setEnabled(state == Qt::Checked);
 }
 
 #if 0
 void
-SettingsPorts::on_checkBoxSerialPassThru5_clicked(bool checked)
+SettingsPorts::on_checkBoxSerialPassThru5_stateChanged(int state)
 {
-    ui->pushButtonSerialPassThru5->setEnabled(checked);
+    ui->pushButtonSerialPassThru5->setEnabled(state == Qt::Checked);
 }
 
 void
-SettingsPorts::on_checkBoxSerialPassThru6_clicked(bool checked)
+SettingsPorts::on_checkBoxSerialPassThru6_stateChanged(int state)
 {
-    ui->pushButtonSerialPassThru6->setEnabled(checked);
+    ui->pushButtonSerialPassThru6->setEnabled(state == Qt::Checked);
 }
 
 void
-SettingsPorts::on_checkBoxSerialPassThru7_clicked(bool checked)
+SettingsPorts::on_checkBoxSerialPassThru7_stateChanged(int state)
 {
-    ui->pushButtonSerialPassThru7->setEnabled(checked);
+    ui->pushButtonSerialPassThru7->setEnabled(state == Qt::Checked);
 }
 #endif
diff --git a/src/qt/qt_settingsports.hpp b/src/qt/qt_settingsports.hpp
index fb9cdb343..cba5bceff 100644
--- a/src/qt/qt_settingsports.hpp
+++ b/src/qt/qt_settingsports.hpp
@@ -18,14 +18,14 @@ public:
 
 #if 0
 private slots:
-    void on_checkBoxSerialPassThru7_clicked(bool checked);
-    void on_checkBoxSerialPassThru6_clicked(bool checked);
-    void on_checkBoxSerialPassThru5_clicked(bool checked);
+    void on_checkBoxSerialPassThru7_stateChanged(int state);
+    void on_checkBoxSerialPassThru6_stateChanged(int state);
+    void on_checkBoxSerialPassThru5_stateChanged(int state);
 #endif
-    void on_checkBoxSerialPassThru4_clicked(bool checked);
-    void on_checkBoxSerialPassThru3_clicked(bool checked);
-    void on_checkBoxSerialPassThru2_clicked(bool checked);
-    void on_checkBoxSerialPassThru1_clicked(bool checked);
+    void on_checkBoxSerialPassThru4_stateChanged(int state);
+    void on_checkBoxSerialPassThru3_stateChanged(int state);
+    void on_checkBoxSerialPassThru2_stateChanged(int state);
+    void on_checkBoxSerialPassThru1_stateChanged(int state);
 
 private slots:
 #if 0

From b018017c55a5f2c8c80cc6816369ac56ef76811f Mon Sep 17 00:00:00 2001
From: Alexander Babikov <lemondrops358@gmail.com>
Date: Sat, 1 Mar 2025 02:05:30 +0500
Subject: [PATCH 55/58] Disable serial passthrough checkboxes when the
 corresponding port is disabled

---
 src/qt/qt_settingsports.cpp | 54 ++++++++++++++++++++++++++++++++++++-
 src/qt/qt_settingsports.hpp | 22 ++++++++++-----
 2 files changed, 68 insertions(+), 8 deletions(-)

diff --git a/src/qt/qt_settingsports.cpp b/src/qt/qt_settingsports.cpp
index 3855a8b68..e5a8039a4 100644
--- a/src/qt/qt_settingsports.cpp
+++ b/src/qt/qt_settingsports.cpp
@@ -72,8 +72,9 @@ SettingsPorts::SettingsPorts(QWidget *parent)
         if (checkBox != NULL)
             checkBox->setChecked(com_ports[i].enabled > 0);
         if (checkBoxPass != NULL) {
+            checkBoxPass->setEnabled(com_ports[i].enabled > 0);
             checkBoxPass->setChecked(serial_passthrough_enabled[i]);
-            buttonPass->setEnabled(serial_passthrough_enabled[i]);
+            buttonPass->setEnabled((com_ports[i].enabled > 0) && serial_passthrough_enabled[i]);
         }
     }
 }
@@ -173,6 +174,57 @@ SettingsPorts::on_pushButtonSerialPassThru7_clicked()
 }
 #endif
 
+void
+SettingsPorts::on_checkBoxSerial1_stateChanged(int state)
+{
+    ui->checkBoxSerialPassThru1->setEnabled(state == Qt::Checked);
+    ui->pushButtonSerialPassThru1->setEnabled((state == Qt::Checked) && ui->checkBoxSerialPassThru1->isChecked());
+}
+
+void
+SettingsPorts::on_checkBoxSerial2_stateChanged(int state)
+{
+    ui->checkBoxSerialPassThru2->setEnabled(state == Qt::Checked);
+    ui->pushButtonSerialPassThru2->setEnabled((state == Qt::Checked) && ui->checkBoxSerialPassThru2->isChecked());
+}
+
+void
+SettingsPorts::on_checkBoxSerial3_stateChanged(int state)
+{
+    ui->checkBoxSerialPassThru3->setEnabled(state == Qt::Checked);
+    ui->pushButtonSerialPassThru3->setEnabled((state == Qt::Checked) && ui->checkBoxSerialPassThru3->isChecked());
+}
+
+void
+SettingsPorts::on_checkBoxSerial4_stateChanged(int state)
+{
+    ui->checkBoxSerialPassThru4->setEnabled(state == Qt::Checked);
+    ui->pushButtonSerialPassThru4->setEnabled((state == Qt::Checked) && ui->checkBoxSerialPassThru4->isChecked());
+}
+
+#if 0
+void
+SettingsPorts::on_checkBoxSerial5_stateChanged(int state)
+{
+    ui->checkBoxSerialPassThru5->setEnabled(state == Qt::Checked);
+    ui->pushButtonSerialPassThru5->setEnabled((state == Qt::Checked) && ui->checkBoxSerialPassThru5->isChecked());
+}
+
+void
+SettingsPorts::on_checkBoxSerial6_stateChanged(int state)
+{
+    ui->checkBoxSerialPassThru6->setEnabled(state == Qt::Checked);
+    ui->pushButtonSerialPassThru6->setEnabled((state == Qt::Checked) && ui->checkBoxSerialPassThru6->isChecked());
+}
+
+void
+SettingsPorts::on_checkBoxSerial7_stateChanged(int state)
+{
+    ui->checkBoxSerialPassThru7->setEnabled(state == Qt::Checked);
+    ui->pushButtonSerialPassThru7->setEnabled((state == Qt::Checked) && ui->checkBoxSerialPassThru7->isChecked());
+}
+#endif
+
 void
 SettingsPorts::on_checkBoxSerialPassThru1_stateChanged(int state)
 {
diff --git a/src/qt/qt_settingsports.hpp b/src/qt/qt_settingsports.hpp
index cba5bceff..7d332509c 100644
--- a/src/qt/qt_settingsports.hpp
+++ b/src/qt/qt_settingsports.hpp
@@ -16,8 +16,8 @@ public:
 
     void save();
 
-#if 0
 private slots:
+#if 0
     void on_checkBoxSerialPassThru7_stateChanged(int state);
     void on_checkBoxSerialPassThru6_stateChanged(int state);
     void on_checkBoxSerialPassThru5_stateChanged(int state);
@@ -27,7 +27,16 @@ private slots:
     void on_checkBoxSerialPassThru2_stateChanged(int state);
     void on_checkBoxSerialPassThru1_stateChanged(int state);
 
-private slots:
+#if 0
+    void on_checkBoxSerial7_stateChanged(int state);
+    void on_checkBoxSerial6_stateChanged(int state);
+    void on_checkBoxSerial5_stateChanged(int state);
+#endif
+    void on_checkBoxSerial4_stateChanged(int state);
+    void on_checkBoxSerial3_stateChanged(int state);
+    void on_checkBoxSerial2_stateChanged(int state);
+    void on_checkBoxSerial1_stateChanged(int state);
+
 #if 0
     void on_pushButtonSerialPassThru7_clicked();
     void on_pushButtonSerialPassThru6_clicked();
@@ -38,11 +47,10 @@ private slots:
     void on_pushButtonSerialPassThru2_clicked();
     void on_pushButtonSerialPassThru1_clicked();
 
-private slots:
-    void on_checkBoxParallel4_stateChanged(int arg1);
-    void on_checkBoxParallel3_stateChanged(int arg1);
-    void on_checkBoxParallel2_stateChanged(int arg1);
-    void on_checkBoxParallel1_stateChanged(int arg1);
+    void on_checkBoxParallel4_stateChanged(int state);
+    void on_checkBoxParallel3_stateChanged(int state);
+    void on_checkBoxParallel2_stateChanged(int state);
+    void on_checkBoxParallel1_stateChanged(int state);
 
 private:
     Ui::SettingsPorts *ui;

From 3bbeae91d1c405ea979b948c3407f3cef3479327 Mon Sep 17 00:00:00 2001
From: Alexander Babikov <lemondrops358@gmail.com>
Date: Sat, 1 Mar 2025 15:22:36 +0500
Subject: [PATCH 56/58] Add plumbing for updates on machine change to the port
 settings page

---
 src/qt/qt_settings.cpp      |   2 +
 src/qt/qt_settingsports.cpp | 149 +++++++++++++++++++-----------------
 src/qt/qt_settingsports.hpp |  69 +++++++++--------
 3 files changed, 116 insertions(+), 104 deletions(-)

diff --git a/src/qt/qt_settings.cpp b/src/qt/qt_settings.cpp
index 67064b73a..471558e22 100644
--- a/src/qt/qt_settings.cpp
+++ b/src/qt/qt_settings.cpp
@@ -153,6 +153,8 @@ Settings::Settings(QWidget *parent)
             &SettingsSound::onCurrentMachineChanged);
     connect(machine, &SettingsMachine::currentMachineChanged, network,
             &SettingsNetwork::onCurrentMachineChanged);
+    connect(machine, &SettingsMachine::currentMachineChanged, ports,
+            &SettingsPorts::onCurrentMachineChanged);
     connect(machine, &SettingsMachine::currentMachineChanged, storageControllers,
             &SettingsStorageControllers::onCurrentMachineChanged);
     connect(machine, &SettingsMachine::currentMachineChanged, otherPeripherals,
diff --git a/src/qt/qt_settingsports.cpp b/src/qt/qt_settingsports.cpp
index e5a8039a4..630c843b9 100644
--- a/src/qt/qt_settingsports.cpp
+++ b/src/qt/qt_settingsports.cpp
@@ -38,6 +38,40 @@ SettingsPorts::SettingsPorts(QWidget *parent)
     , ui(new Ui::SettingsPorts)
 {
     ui->setupUi(this);
+    onCurrentMachineChanged(machine);
+}
+
+SettingsPorts::~SettingsPorts()
+{
+    delete ui;
+}
+
+void
+SettingsPorts::save()
+{
+    for (int i = 0; i < PARALLEL_MAX; i++) {
+        auto *cbox     = findChild<QComboBox *>(QString("comboBoxLpt%1").arg(i + 1));
+        auto *checkBox       = findChild<QCheckBox *>(QString("checkBoxParallel%1").arg(i + 1));
+        if (cbox != NULL)
+            lpt_ports[i].device  = cbox->currentData().toInt();
+        if (checkBox != NULL)
+            lpt_ports[i].enabled = checkBox->isChecked() ? 1 : 0;
+    }
+
+    for (int i = 0; i < SERIAL_MAX; i++) {
+        auto *checkBox     = findChild<QCheckBox *>(QString("checkBoxSerial%1").arg(i + 1));
+        auto *checkBoxPass = findChild<QCheckBox *>(QString("checkBoxSerialPassThru%1").arg(i + 1));
+        if (checkBox != NULL)
+            com_ports[i].enabled = checkBox->isChecked() ? 1 : 0;
+        if (checkBoxPass != NULL)
+            serial_passthrough_enabled[i] = checkBoxPass->isChecked();
+    }
+}
+
+void
+SettingsPorts::onCurrentMachineChanged(int machineId)
+{
+    this->machineId = machineId;
 
     for (int i = 0; i < PARALLEL_MAX; i++) {
         auto *cbox        = findChild<QComboBox *>(QString("comboBoxLpt%1").arg(i + 1));
@@ -79,33 +113,6 @@ SettingsPorts::SettingsPorts(QWidget *parent)
     }
 }
 
-SettingsPorts::~SettingsPorts()
-{
-    delete ui;
-}
-
-void
-SettingsPorts::save()
-{
-    for (int i = 0; i < PARALLEL_MAX; i++) {
-        auto *cbox     = findChild<QComboBox *>(QString("comboBoxLpt%1").arg(i + 1));
-        auto *checkBox       = findChild<QCheckBox *>(QString("checkBoxParallel%1").arg(i + 1));
-        if (cbox != NULL)
-            lpt_ports[i].device  = cbox->currentData().toInt();
-        if (checkBox != NULL)
-            lpt_ports[i].enabled = checkBox->isChecked() ? 1 : 0;
-    }
-
-    for (int i = 0; i < SERIAL_MAX; i++) {
-        auto *checkBox     = findChild<QCheckBox *>(QString("checkBoxSerial%1").arg(i + 1));
-        auto *checkBoxPass = findChild<QCheckBox *>(QString("checkBoxSerialPassThru%1").arg(i + 1));
-        if (checkBox != NULL)
-            com_ports[i].enabled = checkBox->isChecked() ? 1 : 0;
-        if (checkBoxPass != NULL)
-            serial_passthrough_enabled[i] = checkBoxPass->isChecked();
-    }
-}
-
 void
 SettingsPorts::on_checkBoxParallel1_stateChanged(int state)
 {
@@ -130,50 +137,6 @@ SettingsPorts::on_checkBoxParallel4_stateChanged(int state)
     ui->comboBoxLpt4->setEnabled(state == Qt::Checked);
 }
 
-void
-SettingsPorts::on_pushButtonSerialPassThru1_clicked()
-{
-    DeviceConfig::ConfigureDevice(&serial_passthrough_device, 1, qobject_cast<Settings *>(Settings::settings));
-}
-
-void
-SettingsPorts::on_pushButtonSerialPassThru2_clicked()
-{
-    DeviceConfig::ConfigureDevice(&serial_passthrough_device, 2, qobject_cast<Settings *>(Settings::settings));
-}
-
-void
-SettingsPorts::on_pushButtonSerialPassThru3_clicked()
-{
-    DeviceConfig::ConfigureDevice(&serial_passthrough_device, 3, qobject_cast<Settings *>(Settings::settings));
-}
-
-void
-SettingsPorts::on_pushButtonSerialPassThru4_clicked()
-{
-    DeviceConfig::ConfigureDevice(&serial_passthrough_device, 4, qobject_cast<Settings *>(Settings::settings));
-}
-
-#if 0
-void
-SettingsPorts::on_pushButtonSerialPassThru5_clicked()
-{
-    DeviceConfig::ConfigureDevice(&serial_passthrough_device, 5, qobject_cast<Settings *>(Settings::settings));
-}
-
-void
-SettingsPorts::on_pushButtonSerialPassThru6_clicked()
-{
-    DeviceConfig::ConfigureDevice(&serial_passthrough_device, 6, qobject_cast<Settings *>(Settings::settings));
-}
-
-void
-SettingsPorts::on_pushButtonSerialPassThru7_clicked()
-{
-    DeviceConfig::ConfigureDevice(&serial_passthrough_device, 7, qobject_cast<Settings *>(Settings::settings));
-}
-#endif
-
 void
 SettingsPorts::on_checkBoxSerial1_stateChanged(int state)
 {
@@ -268,3 +231,47 @@ SettingsPorts::on_checkBoxSerialPassThru7_stateChanged(int state)
     ui->pushButtonSerialPassThru7->setEnabled(state == Qt::Checked);
 }
 #endif
+
+void
+SettingsPorts::on_pushButtonSerialPassThru1_clicked()
+{
+    DeviceConfig::ConfigureDevice(&serial_passthrough_device, 1, qobject_cast<Settings *>(Settings::settings));
+}
+
+void
+SettingsPorts::on_pushButtonSerialPassThru2_clicked()
+{
+    DeviceConfig::ConfigureDevice(&serial_passthrough_device, 2, qobject_cast<Settings *>(Settings::settings));
+}
+
+void
+SettingsPorts::on_pushButtonSerialPassThru3_clicked()
+{
+    DeviceConfig::ConfigureDevice(&serial_passthrough_device, 3, qobject_cast<Settings *>(Settings::settings));
+}
+
+void
+SettingsPorts::on_pushButtonSerialPassThru4_clicked()
+{
+    DeviceConfig::ConfigureDevice(&serial_passthrough_device, 4, qobject_cast<Settings *>(Settings::settings));
+}
+
+#if 0
+void
+SettingsPorts::on_pushButtonSerialPassThru5_clicked()
+{
+    DeviceConfig::ConfigureDevice(&serial_passthrough_device, 5, qobject_cast<Settings *>(Settings::settings));
+}
+
+void
+SettingsPorts::on_pushButtonSerialPassThru6_clicked()
+{
+    DeviceConfig::ConfigureDevice(&serial_passthrough_device, 6, qobject_cast<Settings *>(Settings::settings));
+}
+
+void
+SettingsPorts::on_pushButtonSerialPassThru7_clicked()
+{
+    DeviceConfig::ConfigureDevice(&serial_passthrough_device, 7, qobject_cast<Settings *>(Settings::settings));
+}
+#endif
diff --git a/src/qt/qt_settingsports.hpp b/src/qt/qt_settingsports.hpp
index 7d332509c..83560914f 100644
--- a/src/qt/qt_settingsports.hpp
+++ b/src/qt/qt_settingsports.hpp
@@ -16,44 +16,47 @@ public:
 
     void save();
 
+public slots:
+    void onCurrentMachineChanged(int machineId);
+
 private slots:
-#if 0
-    void on_checkBoxSerialPassThru7_stateChanged(int state);
-    void on_checkBoxSerialPassThru6_stateChanged(int state);
-    void on_checkBoxSerialPassThru5_stateChanged(int state);
-#endif
-    void on_checkBoxSerialPassThru4_stateChanged(int state);
-    void on_checkBoxSerialPassThru3_stateChanged(int state);
-    void on_checkBoxSerialPassThru2_stateChanged(int state);
-    void on_checkBoxSerialPassThru1_stateChanged(int state);
-
-#if 0
-    void on_checkBoxSerial7_stateChanged(int state);
-    void on_checkBoxSerial6_stateChanged(int state);
-    void on_checkBoxSerial5_stateChanged(int state);
-#endif
-    void on_checkBoxSerial4_stateChanged(int state);
-    void on_checkBoxSerial3_stateChanged(int state);
-    void on_checkBoxSerial2_stateChanged(int state);
-    void on_checkBoxSerial1_stateChanged(int state);
-
-#if 0
-    void on_pushButtonSerialPassThru7_clicked();
-    void on_pushButtonSerialPassThru6_clicked();
-    void on_pushButtonSerialPassThru5_clicked();
-#endif
-    void on_pushButtonSerialPassThru4_clicked();
-    void on_pushButtonSerialPassThru3_clicked();
-    void on_pushButtonSerialPassThru2_clicked();
-    void on_pushButtonSerialPassThru1_clicked();
-
-    void on_checkBoxParallel4_stateChanged(int state);
-    void on_checkBoxParallel3_stateChanged(int state);
-    void on_checkBoxParallel2_stateChanged(int state);
     void on_checkBoxParallel1_stateChanged(int state);
+    void on_checkBoxParallel2_stateChanged(int state);
+    void on_checkBoxParallel3_stateChanged(int state);
+    void on_checkBoxParallel4_stateChanged(int state);
+
+    void on_checkBoxSerial1_stateChanged(int state);
+    void on_checkBoxSerial2_stateChanged(int state);
+    void on_checkBoxSerial3_stateChanged(int state);
+    void on_checkBoxSerial4_stateChanged(int state);
+#if 0
+    void on_checkBoxSerial5_stateChanged(int state);
+    void on_checkBoxSerial6_stateChanged(int state);
+    void on_checkBoxSerial7_stateChanged(int state);
+#endif
+    void on_checkBoxSerialPassThru1_stateChanged(int state);
+    void on_checkBoxSerialPassThru2_stateChanged(int state);
+    void on_checkBoxSerialPassThru3_stateChanged(int state);
+    void on_checkBoxSerialPassThru4_stateChanged(int state);
+#if 0
+    void on_checkBoxSerialPassThru5_stateChanged(int state);
+    void on_checkBoxSerialPassThru6_stateChanged(int state);
+    void on_checkBoxSerialPassThru7_stateChanged(int state);
+#endif
+
+    void on_pushButtonSerialPassThru1_clicked();
+    void on_pushButtonSerialPassThru2_clicked();
+    void on_pushButtonSerialPassThru3_clicked();
+    void on_pushButtonSerialPassThru4_clicked();
+#if 0
+    void on_pushButtonSerialPassThru5_clicked();
+    void on_pushButtonSerialPassThru6_clicked();
+    void on_pushButtonSerialPassThru7_clicked();
+#endif
 
 private:
     Ui::SettingsPorts *ui;
+    int                machineId = 0;
 };
 
 #endif // QT_SETTINGSPORTS_HPP

From 676087bcaecd2986f48fdc919b430481023c06b9 Mon Sep 17 00:00:00 2001
From: Alexander Babikov <lemondrops358@gmail.com>
Date: Fri, 7 Mar 2025 00:11:45 +0500
Subject: [PATCH 57/58] Refactor into a function

---
 src/qt/qt_mainwindow.cpp | 22 +++-------------------
 src/qt/qt_util.cpp       | 22 ++++++++++++++++++++++
 src/qt/qt_util.hpp       |  3 +++
 3 files changed, 28 insertions(+), 19 deletions(-)

diff --git a/src/qt/qt_mainwindow.cpp b/src/qt/qt_mainwindow.cpp
index 036f56375..1cc6c33c9 100644
--- a/src/qt/qt_mainwindow.cpp
+++ b/src/qt/qt_mainwindow.cpp
@@ -122,19 +122,6 @@ extern int qt_nvr_save(void);
 #include "x11_util.h"
 #endif
 
-#ifdef Q_OS_WINDOWS
-#include <dwmapi.h>
-#ifndef DWMWA_WINDOW_CORNER_PREFERENCE
-#define DWMWA_WINDOW_CORNER_PREFERENCE 33
-#endif
-#ifndef DWMWCP_DEFAULT
-#define DWMWCP_DEFAULT 0
-#endif
-#ifndef DWMWCP_DONOTROUND
-#define DWMWCP_DONOTROUND 1
-#endif
-#endif
-
 #ifdef Q_OS_MACOS
 #    include "cocoa_keyboard.hpp"
 // The namespace is required to avoid clashing typedefs; we only use this
@@ -195,8 +182,7 @@ MainWindow::MainWindow(QWidget *parent)
     ui->stackedWidget->setMouseTracking(true);
     statusBar()->setVisible(!hide_status_bar);
 #ifdef Q_OS_WINDOWS
-    auto cornerPreference = (hide_status_bar ? DWMWCP_DONOTROUND : DWMWCP_DEFAULT);
-    DwmSetWindowAttribute((HWND) this->winId(), DWMWA_WINDOW_CORNER_PREFERENCE, (LPCVOID) &cornerPreference, sizeof(cornerPreference));
+    util::setWin11RoundedCorners(this->winId(), (hide_status_bar ? false : true));
 #endif
     statusBar()->setStyleSheet("QStatusBar::item {border: None; } QStatusBar QLabel { margin-right: 2px; margin-bottom: 1px; }");
     this->centralWidget()->setStyleSheet("background-color: black;");
@@ -817,8 +803,7 @@ MainWindow::initRendererMonitorSlot(int monitor_index)
             secondaryRenderer->setFixedSize(fixed_size_x, fixed_size_y);
         secondaryRenderer->setWindowIcon(this->windowIcon());
 #ifdef Q_OS_WINDOWS
-        auto cornerPreference = DWMWCP_DONOTROUND;
-        DwmSetWindowAttribute((HWND) secondaryRenderer->winId(), DWMWA_WINDOW_CORNER_PREFERENCE, (LPCVOID) &cornerPreference, sizeof(cornerPreference));
+        util::setWin11RoundedCorners(secondaryRenderer->winId(), false);
 #endif
         if (show_second_monitors) {
             secondaryRenderer->show();
@@ -1855,8 +1840,7 @@ MainWindow::on_actionHide_status_bar_triggered()
     ui->actionHide_status_bar->setChecked(hide_status_bar);
     statusBar()->setVisible(!hide_status_bar);
 #ifdef Q_OS_WINDOWS
-    auto cornerPreference = (hide_status_bar ? DWMWCP_DONOTROUND : DWMWCP_DEFAULT);
-    DwmSetWindowAttribute((HWND) main_window->winId(), DWMWA_WINDOW_CORNER_PREFERENCE, (LPCVOID) &cornerPreference, sizeof(cornerPreference));
+    util::setWin11RoundedCorners(main_window->winId(), (hide_status_bar ? false : true));
 #endif
     if (vid_resize >= 2) {
         setFixedSize(fixed_size_x, fixed_size_y + menuBar()->height() + (hide_status_bar ? 0 : statusBar()->height()) + (hide_tool_bar ? 0 : ui->toolBar->height()));
diff --git a/src/qt/qt_util.cpp b/src/qt/qt_util.cpp
index 5c9059272..0c78c2f5e 100644
--- a/src/qt/qt_util.cpp
+++ b/src/qt/qt_util.cpp
@@ -26,6 +26,19 @@
 #include <QUuid>
 #include "qt_util.hpp"
 
+#ifdef Q_OS_WINDOWS
+#    include <dwmapi.h>
+#    ifndef DWMWA_WINDOW_CORNER_PREFERENCE
+#        define DWMWA_WINDOW_CORNER_PREFERENCE 33
+#    endif
+#    ifndef DWMWCP_DEFAULT
+#        define DWMWCP_DEFAULT 0
+#    endif
+#    ifndef DWMWCP_DONOTROUND
+#        define DWMWCP_DONOTROUND 1
+#    endif
+#endif
+
 extern "C" {
 #include <86box/86box.h>
 #include <86box/config.h>
@@ -48,6 +61,15 @@ screenOfWidget(QWidget *widget)
 #endif
 }
 
+#ifdef Q_OS_WINDOWS
+void
+setWin11RoundedCorners(WId hwnd, bool enable)
+{
+    auto cornerPreference = (enable ? DWMWCP_DEFAULT : DWMWCP_DONOTROUND);
+    DwmSetWindowAttribute((HWND) hwnd, DWMWA_WINDOW_CORNER_PREFERENCE, (LPCVOID) &cornerPreference, sizeof(cornerPreference));
+}
+#endif
+
 QString
 DlgFilter(std::initializer_list<QString> extensions, bool last)
 {
diff --git a/src/qt/qt_util.hpp b/src/qt/qt_util.hpp
index 6a0bdc30b..9432610b6 100644
--- a/src/qt/qt_util.hpp
+++ b/src/qt/qt_util.hpp
@@ -13,6 +13,9 @@ static constexpr auto UUID_MIN_LENGTH = 36;
 QString DlgFilter(std::initializer_list<QString> extensions, bool last = false);
 /* Returns screen the widget is on */
 QScreen *screenOfWidget(QWidget *widget);
+#ifdef Q_OS_WINDOWS
+void setWin11RoundedCorners(WId hwnd, bool enable);
+#endif
 QString currentUuid();
 void storeCurrentUuid();
 bool compareUuid();

From 05a005b6f0a681b965ab60ae7638c2c2408fd8e7 Mon Sep 17 00:00:00 2001
From: Alexander Babikov <lemondrops358@gmail.com>
Date: Fri, 7 Mar 2025 05:21:16 +0500
Subject: [PATCH 58/58] Initialize the sound icon menu pointer

Fixes a crash when clicking the icon
---
 src/qt/qt_machinestatus.cpp | 1 +
 1 file changed, 1 insertion(+)

diff --git a/src/qt/qt_machinestatus.cpp b/src/qt/qt_machinestatus.cpp
index 3c3f4019f..a00cc032b 100644
--- a/src/qt/qt_machinestatus.cpp
+++ b/src/qt/qt_machinestatus.cpp
@@ -259,6 +259,7 @@ MachineStatus::MachineStatus(QObject *parent)
 {
     d = std::make_unique<MachineStatus::States>(this);
     muteUnmuteAction = nullptr;
+    soundMenu = nullptr;
     connect(refreshTimer, &QTimer::timeout, this, &MachineStatus::refreshIcons);
     refreshTimer->start(75);
 }