Further optimisation

2 files changed, 55 insertions, 40 deletions

diff --git a/README.md b/README.md
index e2e6dc6..c9792a4 100644
--- a/README.md
+++ b/README.md
@@ -8,7 +8,7 @@ Running the game:
 
 - MS/DOS or compatible
 - A DOS Protected Mode Interface, e.g. [CWSDPMI](http://sandmann.dotster.com/cwsdpmi/)
-- 386DX with VGA or better
+- 386DX 33MHz (or better) with VGA
 - At least 4MB of RAM
 - Sound Blaster (or no sound)
 
diff --git a/src/vga.c b/src/vga.c
index a82fca9..1dc77db 100644
--- a/src/vga.c
+++ b/src/vga.c
@@ -122,58 +122,73 @@ void blit(const uint8_t *sprite, const Rect *dst)
 
 void blit_c(const uint8_t *sprite, const Rect *dst, uint8_t c)
 {
-    uint8_t b;
     uint8_t *dtarget = target + dst->x + dst->y * 320;
-    const uint16_t inc = 320 - dst->w;
-
-    for (uint16_t j = dst->h; j > 0; j--)
-    {
-        for (uint16_t i = dst->w; i > 0; i--)
-        {
-            b = *sprite++;
 
-            /* transparent */
-            if (b == TRANSPARENT)
-            {
-                dtarget++;
-                continue;
-            }
+    asm volatile (
+        "movl $320, %%ebp\n\t"
+        "subl %%ecx, %%ebp\n\t"
+        "blit_c_w:\n\t"
+        "movl %%ecx, %%eax\n\t"
+        "blit_c_w_loop:\n\t"
+        "movb (%%esi), %%ah\n\t"
+        "cmpb $%p5, %%ah\n\t"
+        "je blitrc_tr\n\t"
+        "orb %%ah, %%ah\n\t"
+        "je blit_c_store\n\t"
+        "movb %%dl, %%ah\n\t"
+        "blit_c_store:\n\t"
+        "movb %%ah, (%%edi)\n\t"
+        "blit_c_tr:\n\t"
+        "incl %%esi\n\t"
+        "incl %%edi\n\t"
+        "decb %%al\n\t"
+        "jne blit_c_w_loop\n\t"
+        "addl %%ebp, %%edi\n\t"
+        "decl %%ebx\n\t"
+        "jne blit_c_w\n\t"
+        : /*no output */
+        : "S" (sprite), "D" (dtarget),
+        "b" ((uint32_t)dst->h), "d" ((uint32_t)c),
+        "c" ((uint32_t)dst->w), "i" (TRANSPARENT)
+        : "eax", "ebp"
+    );
 
-            *dtarget++ = b ? c : b;
-        }
-        dtarget += inc;
-    }
 }
 
 void blitrc(const uint8_t *sprite, const Rect *src, const Rect *dst)
 {
-    uint8_t b;
     uint8_t *dtarget = target + dst->x + dst->y * 320;
-    const uint16_t inc1 = src->w - dst->w;
-    const uint16_t inc2 = 320 - dst->w;
 
     sprite += src->x + src->y * src->w;
 
-    for (uint16_t j = dst->h; j > 0; j--)
-    {
-        for (uint16_t i = dst->w; i > 0; i--)
-        {
-            b = *sprite++;
-
-            /* transparent */
-            if (b == TRANSPARENT)
-            {
-                dtarget++;
-                continue;
-            }
-
-            *dtarget++ = b;
-        }
-        sprite += inc1;
-        dtarget += inc2;
-    }
+    asm volatile (
+        "movl $320, %%ebp\n\t"
+        "subl %%ecx, %%ebp\n\t"
+        "blitrc_w:\n\t"
+        "movl %%ecx, %%eax\n\t"
+        "blitrc_w_loop:\n\t"
+        "movb (%%esi), %%ah\n\t"
+        "cmpb $%p5, %%ah\n\t"
+        "je blitrc_tr\n\t"
+        "movb %%ah, (%%edi)\n\t"
+        "blitrc_tr:\n\t"
+        "incl %%esi\n\t"
+        "incl %%edi\n\t"
+        "decb %%al\n\t"
+        "jne blitrc_w_loop\n\t"
+        "addl %%edx, %%esi\n\t"
+        "addl %%ebp, %%edi\n\t"
+        "decl %%ebx\n\t"
+        "jne blitrc_w\n\t"
+        : /*no output */
+        : "S" (sprite), "D" (dtarget),
+        "b" ((uint32_t)dst->h), "d" ((uint32_t)(src->w - dst->w)),
+        "c" ((uint32_t)dst->w), "i" (TRANSPARENT)
+        : "eax", "ebp"
+    );
 }
 
+
 void blit_erase(uint8_t c)
 {
     memset(target, c, 320 * 200);