Bug 498770 - Enable optimized Theora code in Windows builds - r=kinetik rs=roc

media/libtheora/README_MOZILLA

@@ -3,6 +3,7 @@ source distribution using the update.sh script. The changes made were
 those applied by update.sh, the addition/update of Makefile.in files
 for the Mozilla build system and the patch in bug below.
 
+bug498770.patch - Enable optimized theora code in windows build
 Bug 455357 - WinCE LibTheora Pre-defined Macro usage in local variable
   455357_wince_local_variable_macro_clash_patch
   This patch is needed for building WinCE / WinMobile because the 

media/libtheora/bug498770.patch

@@ -0,0 +1,97 @@
+Index: lib/dec/x86_vc/mmxfrag.c
+===================================================================
+--- lib/dec/x86_vc/mmxfrag.c	(revision 16142)
++++ lib/dec/x86_vc/mmxfrag.c	(working copy)
+@@ -27,12 +27,14 @@
+ 
+ void oc_frag_recon_intra_mmx(unsigned char *_dst,int _dst_ystride,
+  const ogg_int16_t *_residue){
++  int _save_ebx;
+   /* ---------------------------------------------------------------------
+   This function does the inter reconstruction step with 8 iterations
+   unrolled. The iteration for each instruction is noted by the #id in the
+   comments (in case you want to reconstruct it)
+   --------------------------------------------------------------------- */
+   _asm{
++    mov       [_save_ebx], ebx
+     mov       edi, [_residue]     /* load residue ptr     */
+     mov       eax, 0x00800080     /* generate constant    */
+     mov       ebx, [_dst_ystride] /* load dst-stride      */
+@@ -93,6 +95,7 @@
+     packuswb  mm3, mm4            /* #8 pack to byte      */
+     movq      [edx + ecx*2], mm1  /* #7 write row         */
+     movq      [edx + eax], mm3    /* #8 write row         */
++    mov       ebx, [_save_ebx]
+   }
+ }
+ 
+@@ -100,6 +103,7 @@
+ 
+ void oc_frag_recon_inter_mmx (unsigned char *_dst, int _dst_ystride,
+  const unsigned char *_src, int _src_ystride, const ogg_int16_t *_residue){
++  int _save_ebx;
+   /* ---------------------------------------------------------------------
+   This function does the inter reconstruction step with two iterations
+   running in parallel to hide some load-latencies and break the dependency
+@@ -107,6 +111,7 @@
+   comments (in case you want to reconstruct it)
+   --------------------------------------------------------------------- */
+   _asm{
++    mov       [_save_ebx], ebx
+     pxor      mm0, mm0          /* generate constant 0 */
+     mov       esi, [_src]
+     mov       edi, [_residue]
+@@ -143,6 +148,7 @@
+     movq      [edx + ebx], mm7  /* #2 write row          */
+     lea       edx, [edx+ebx*2]  /* dst += stride * 2     */
+     jne       nextchunk
++    mov       ebx, [_save_ebx]
+   }
+ }
+ 
+@@ -150,6 +156,7 @@
+ void oc_frag_recon_inter2_mmx(unsigned char *_dst,  int _dst_ystride,
+  const unsigned char *_src1,  int _src1_ystride, const unsigned char *_src2,
+  int _src2_ystride,const ogg_int16_t *_residue){
++  int _save_ebx;
+   /* ---------------------------------------------------------------------
+   This function does the inter2 reconstruction step.The building of the
+   average is done with a bit-twiddeling trick to avoid excessive register
+@@ -166,6 +173,7 @@
+   using the pavgb instruction let me know and I'll do the 3dnow codepath.
+   --------------------------------------------------------------------- */
+  _asm{
++   mov        [_save_ebx], ebx
+    mov        eax, 0xfefefefe
+    mov        esi, [_src1]
+    mov        edi, [_src2]
+@@ -204,6 +212,7 @@
+    packuswb   mm2,  mm3           /* pack and saturate   */
+    movq       [edx], mm2          /* write row           */
+    jne        nextrow
++   mov        ebx, [_save_ebx]
+  }
+ }
+ 
+Index: lib/dec/x86_vc/mmxloopfilter.c
+===================================================================
+--- lib/dec/x86_vc/mmxloopfilter.c	(revision 16142)
++++ lib/dec/x86_vc/mmxloopfilter.c	(working copy)
+@@ -38,7 +38,7 @@
+   _asm {
+     mov       eax,  [_pix]
+     mov       edx,  [_ystride]
+-    mov       ebx,  [_ll]
++    mov       ecx,  [_ll]
+ 
+     /* _pix -= ystride */
+     sub       eax,   edx
+@@ -104,7 +104,7 @@
+     /*Free up mm5.*/
+     packuswb  mm4, mm5
+     /*mm0=L L L L*/
+-    movq      mm0, [ebx]
++    movq      mm0, [ecx]
+     /*if(R_i<-2L||R_i>2L)R_i=0:*/
+     movq      mm5, mm2
+     pxor      mm6, mm6

media/libtheora/lib/Makefile.in

@@ -48,12 +48,10 @@ FORCE_STATIC_LIB= 1
 DEFINES += -DTHEORA_DISABLE_ENCODE
 
 ifeq ($(findstring 86,$(OS_TEST)), 86)
-ifneq ($(OS_ARCH),WINNT)
 ifneq ($(OS_ARCH),SunOS)
 DEFINES += -DOC_X86ASM -DUSE_ASM
 endif
 endif
-endif
 
 VPATH		:= $(srcdir) $(srcdir)/dec
 
@@ -75,6 +73,17 @@ CSRCS		= \
 		$(NULL)
 
 ifeq ($(findstring 86,$(OS_TEST)), 86)
+ifeq ($(OS_ARCH),WINNT)
+VPATH		+= $(srcdir)/dec/x86_vc
+
+CSRCS		+= \
+		mmxfrag.c \
+		mmxloopfilter.c \
+		x86state.c \
+		mmxstate.c \
+		mmxidct.c \
+		$(NULL)
+else
 VPATH		+= $(srcdir)/dec/x86
 
 CSRCS		+= \
@@ -84,6 +93,7 @@ CSRCS		+= \
 		mmxidct.c \
 		$(NULL)
 endif
+endif
 
 include $(topsrcdir)/config/rules.mk
 

media/libtheora/lib/dec/x86_vc/mmxfrag.c

@@ -27,12 +27,14 @@
 
 void oc_frag_recon_intra_mmx(unsigned char *_dst,int _dst_ystride,
  const ogg_int16_t *_residue){
+  int _save_ebx;
   /* ---------------------------------------------------------------------
   This function does the inter reconstruction step with 8 iterations
   unrolled. The iteration for each instruction is noted by the #id in the
   comments (in case you want to reconstruct it)
   --------------------------------------------------------------------- */
   _asm{
+    mov       [_save_ebx], ebx
     mov       edi, [_residue]     /* load residue ptr     */
     mov       eax, 0x00800080     /* generate constant    */
     mov       ebx, [_dst_ystride] /* load dst-stride      */
@@ -93,6 +95,7 @@ void oc_frag_recon_intra_mmx(unsigned char *_dst,int _dst_ystride,
     packuswb  mm3, mm4            /* #8 pack to byte      */
     movq      [edx + ecx*2], mm1  /* #7 write row         */
     movq      [edx + eax], mm3    /* #8 write row         */
+    mov       ebx, [_save_ebx]
   }
 }
 
@@ -100,6 +103,7 @@ void oc_frag_recon_intra_mmx(unsigned char *_dst,int _dst_ystride,
 
 void oc_frag_recon_inter_mmx (unsigned char *_dst, int _dst_ystride,
  const unsigned char *_src, int _src_ystride, const ogg_int16_t *_residue){
+  int _save_ebx;
   /* ---------------------------------------------------------------------
   This function does the inter reconstruction step with two iterations
   running in parallel to hide some load-latencies and break the dependency
@@ -107,6 +111,7 @@ void oc_frag_recon_inter_mmx (unsigned char *_dst, int _dst_ystride,
   comments (in case you want to reconstruct it)
   --------------------------------------------------------------------- */
   _asm{
+    mov       [_save_ebx], ebx
     pxor      mm0, mm0          /* generate constant 0 */
     mov       esi, [_src]
     mov       edi, [_residue]
@@ -143,6 +148,7 @@ nextchunk:
     movq      [edx + ebx], mm7  /* #2 write row          */
     lea       edx, [edx+ebx*2]  /* dst += stride * 2     */
     jne       nextchunk
+    mov       ebx, [_save_ebx]
   }
 }
 
@@ -150,6 +156,7 @@ nextchunk:
 void oc_frag_recon_inter2_mmx(unsigned char *_dst,  int _dst_ystride,
  const unsigned char *_src1,  int _src1_ystride, const unsigned char *_src2,
  int _src2_ystride,const ogg_int16_t *_residue){
+  int _save_ebx;
   /* ---------------------------------------------------------------------
   This function does the inter2 reconstruction step.The building of the
   average is done with a bit-twiddeling trick to avoid excessive register
@@ -166,6 +173,7 @@ void oc_frag_recon_inter2_mmx(unsigned char *_dst,  int _dst_ystride,
   using the pavgb instruction let me know and I'll do the 3dnow codepath.
   --------------------------------------------------------------------- */
  _asm{
+   mov        [_save_ebx], ebx
    mov        eax, 0xfefefefe
    mov        esi, [_src1]
    mov        edi, [_src2]
@@ -204,6 +212,7 @@ nextrow:
    packuswb   mm2,  mm3           /* pack and saturate   */
    movq       [edx], mm2          /* write row           */
    jne        nextrow
+   mov        ebx, [_save_ebx]
  }
 }
 

media/libtheora/lib/dec/x86_vc/mmxloopfilter.c

@@ -38,7 +38,7 @@ static void loop_filter_v(unsigned char *_pix,int _ystride,
   _asm {
     mov       eax,  [_pix]
     mov       edx,  [_ystride]
-    mov       ebx,  [_ll]
+    mov       ecx,  [_ll]
 
     /* _pix -= ystride */
     sub       eax,   edx
@@ -104,7 +104,7 @@ static void loop_filter_v(unsigned char *_pix,int _ystride,
     /*Free up mm5.*/
     packuswb  mm4, mm5
     /*mm0=L L L L*/
-    movq      mm0, [ebx]
+    movq      mm0, [ecx]
     /*if(R_i<-2L||R_i>2L)R_i=0:*/
     movq      mm5, mm2
     pxor      mm6, mm6

media/libtheora/update.sh

@@ -53,3 +53,4 @@ cp $1/include/theora/theora.h ./include/theora/theora.h
 cp $1/include/theora/theoradec.h ./include/theora/theoradec.h
 cp $1/include/theora/codec.h ./include/theora/codec.h
 patch -p3 <455357_wince_local_variable_macro_clash_patch
+patch -p0 <bug498770.patch