[greatlrd] 15856: Fix align 4 problem with color fill and hline.

Fix align 4 problem with color fill and hline. 
Now OARCH=pentium2 working with 32bpp
Modified: trunk/reactos/subsys/win32k/dib/dib32bpp.c
  _____

Modified: trunk/reactos/subsys/win32k/dib/dib32bpp.c
--- trunk/reactos/subsys/win32k/dib/dib32bpp.c	2005-06-10 22:54:11 UTC
(rev 15855)
+++ trunk/reactos/subsys/win32k/dib/dib32bpp.c	2005-06-11 09:53:14 UTC
(rev 15856)
@@ -48,10 +48,29 @@

LONG cx  = (x2 - x1) ;  
    PBYTE byteaddr = SurfObj->pvScan0 + y * SurfObj->lDelta;  
    PDWORD addr = (PDWORD)byteaddr + x1;
-   memset4(addr, c, cx);  
+
+    __asm__(
+"  cld\n"
+"  mov  %0, %%eax\n"
+"  test $0x03, %%edi\n" /* Align to fullword boundary */
+"  jnz   .L1\n"
+"  mov  %1,%%ecx\n"     /* Setup count of fullwords to fill */
+"  rep stosl\n"         /* The actual fill */
+"  jz   .L2\n"
+".L1:\n"
+"  mov %%eax, %%ecx\n"
+"  stosw\n"
+"  sub  $0x04,%1\n"
+"  mov  %1,%%ecx\n"     /* Setup count of fullwords to fill */
+"  rep stosl\n"         /* The actual fill */
+"  shr $0x08,%%eax\n"
+"  stosw\n"
+".L2:\n"
+  : /* no output */
+  : "r"(c), "r"(cx), "D"(addr)
+  : "%eax", "%ecx");
+

-  
-  
 }
 #else
 VOID
@@ -693,9 +712,28 @@
   pos =  (PULONG) (DestSurface->pvScan0 + DestRect->top * delta +
(DestRect->left<<2));

for (DestY = DestRect->top; DestY< DestRect->bottom; DestY++)
-  {			 				
-    memset4( ((PDWORD)(pos)),  color, width);
-    pos =(PULONG)((ULONG_PTR)pos + delta);	 
+  {			
+    __asm__(
+            "  cld\n"
+            "  mov  %0, %%eax\n"
+            "  test $0x03, %%edi\n" /* Align to fullword boundary */
+            "  jnz   .FL1\n"
+            "  mov  %1,%%ecx\n"     /* Setup count of fullwords to fill
*/
+            "  rep stosl\n"         /* The actual fill */
+            "  jz   .FL2\n"
+            ".FL1:\n"
+            "  mov %%eax, %%ecx\n"
+            "  stosw\n"
+            "  sub  $0x04,%1\n"
+            "  mov  %1,%%ecx\n"     /* Setup count of fullwords to fill
*/
+            "  rep stosl\n"         /* The actual fill */
+            "  shr $0x08,%%eax\n"
+            "  stosw\n"
+            ".FL2:\n"            
+            : /* no output */
+            : "r"(color), "r"(width), "D"(pos)
+            : "%eax", "%ecx");    
+            pos =(PULONG)((ULONG_PTR)pos + delta);	 
   }

#else