Assembler optimised blit copy hardcoded to 16x16

1 files changed, 25 insertions, 0 deletions

diff --git a/src/vga.c b/src/vga.c
index 70a9a67..c78fcec 100644
--- a/src/vga.c
+++ b/src/vga.c
@@ -196,3 +196,28 @@ void blit_copy(const Rect *dst)
         dtarget += 320;
     }
 }
+
+void blit_copy16x16(const Rect *dst)
+{
+    uint8_t *src = buffer + dst->x + dst->y * 320;
+    uint8_t *dtarget = screen + dst->x + dst->y * 320;
+
+    asm volatile (
+        "movl $16, %%eax\n\t"
+        "movl $0x130, %%ebx\n\t"
+        "cld\n\t"
+        "blit_copy16x16_loop:\n\t"
+        "movsl\n\t"
+        "movsl\n\t"
+        "movsl\n\t"
+        "movsl\n\t"
+        "addl %%ebx, %%esi\n\t"
+        "addl %%ebx, %%edi\n\t"
+        "decl %%eax\n\t"
+        "orl %%eax, %%eax\n\t"
+        "jne blit_copy16x16_loop\n\t"
+        : /* no output */
+        : "S" (src), "D" (dtarget)
+        : "ecx", "ebx", "eax"
+    );
+}