[xiph-commits] r11503 - branches/theora-playtime/lib/x86_32_vs

[illiminable at svn.xiph.org]

Author: illiminable
Date: 2006-06-03 05:33:36 -0700 (Sat, 03 Jun 2006)
New Revision: 11503

Modified:
   branches/theora-playtime/lib/x86_32_vs/dsp_sse2.c
Log:
* Reorder reads in sub8x8_128__sse2

Modified: branches/theora-playtime/lib/x86_32_vs/dsp_sse2.c
===================================================================
--- branches/theora-playtime/lib/x86_32_vs/dsp_sse2.c	2006-06-03 12:18:22 UTC (rev 11502)
+++ branches/theora-playtime/lib/x86_32_vs/dsp_sse2.c	2006-06-03 12:33:36 UTC (rev 11503)
@@ -236,15 +236,13 @@
              8x16bits to fill all 128 bits of the register        
         */
         movq        xmm1, QWORD PTR [esi]
+        movq        xmm2, QWORD PTR [esi + eax]
+        movq        xmm3, QWORD PTR [esi + eax * 2]
+        movq        xmm4, QWORD PTR [esi + ebx]
+
         punpcklbw   xmm1, xmm0
-
-        movq        xmm2, QWORD PTR [esi + eax]
         punpcklbw   xmm2, xmm0
-
-        movq        xmm3, QWORD PTR [esi + eax * 2]
         punpcklbw   xmm3, xmm0
-
-        movq        xmm4, QWORD PTR [esi + ebx]
         punpcklbw   xmm4, xmm0
 
         /* Subtract 128 16bitwise and write*/
@@ -258,21 +256,24 @@
         movdqa      [edi + 48], xmm4
 
 
+
+
         /* Advance the source and dest pointer for the next 4 iterations */
         lea         esi, [esi + eax * 4]
         add         edi, 64
 
+
+
+
         /* Repeat of above for second round */
         movq        xmm1, QWORD PTR [esi]
+        movq        xmm2, QWORD PTR [esi + eax]
+        movq        xmm3, QWORD PTR [esi + eax * 2]
+        movq        xmm4, QWORD PTR [esi + ebx]
+
         punpcklbw   xmm1, xmm0
-
-        movq        xmm2, QWORD PTR [esi + eax]
         punpcklbw   xmm2, xmm0
-
-        movq        xmm3, QWORD PTR [esi + eax * 2]
         punpcklbw   xmm3, xmm0
-
-        movq        xmm4, QWORD PTR [esi + ebx]
         punpcklbw   xmm4, xmm0
 
         /* Subtract 128 16bitwise and write*/
@@ -286,6 +287,7 @@
         movdqa      [edi + 48], xmm4
 
     };
+
  
 #endif
 }