Fix crash and off-by-one error in MMX YUVA->ARGB conversion.

media/base/simd/convert_yuva_to_argb_mmx.inc

Index: media/base/simd/convert_yuva_to_argb_mmx.inc
diff --git a/media/base/simd/convert_yuva_to_argb_mmx.inc b/media/base/simd/convert_yuva_to_argb_mmx.inc
index 5faa6a54236858c03b05c486694baa11d4ba9394..d4933836ca89bec26134e015048ae8151bd3531b 100644
--- a/media/base/simd/convert_yuva_to_argb_mmx.inc
+++ b/media/base/simd/convert_yuva_to_argb_mmx.inc
@@ -9,9 +9,10 @@
 
 mangle(SYMBOL):
   %assign   stack_offset 0
-  PROLOGUE  7, 7, 3, Y, U, V, A, ARGB, WIDTH, TABLE, TEMP
+  PROLOGUE  7, 7, 3, Y, U, V, A, ARGB, WIDTH, TABLE
   PUSH      WIDTHq
   DEFINE_ARGS Y, U, V, A, ARGB, TABLE, TEMP
+  mov       TABLEq, TEMPq
   jmp       .convertend
 
 .convertloop:
@@ -39,11 +40,25 @@ mangle(SYMBOL):
   psraw     mm2, 6
   packuswb  mm1, mm2
 
-  ; Unpack and multiply by alpha value, then repack high bytes of words.
+  ; Unpack
   movq      mm0, mm1
   pxor      mm2, mm2
   punpcklbw mm0, mm2
   punpckhbw mm1, mm2
+
+  ; Add one to our alpha values, this is a somewhat unfortunate hack; while
+  ; the pack/unpack above handle saturating any negative numbers to 0, they also
+  ; truncate the alpha value to 255. The math ahead wants to produce the same
+  ; ARGB alpha value as the source pixel in YUVA, but this depends on the alpha
+  ; value in |mm0| and |mm1| being 256, (let A be the source image alpha,
+  ; 256 * A >> 8 == A, whereas 255 * A >> 8 is off by one except at 0).
+  mov       TEMPq, 0x00010000
+  movd      mm2, TEMPd
+  psllq     mm2, 32
+  paddsw    mm0, mm2
+  paddsw    mm1, mm2
+
+  ; Multiply by alpha value, then repack high bytes of words.
   movzx     TEMPd, BYTE [Aq]
   movq      mm2, [TABLEq + 6144 + 8 * TEMPq]
   pmullw    mm0, mm2
@@ -79,6 +94,13 @@ mangle(SYMBOL):
   ; Multiply ARGB by alpha value.
   pxor      mm0, mm0
   punpcklbw mm1, mm0
+
+  ; See above note about this hack.
+  mov       TEMPq, 0x00010000
+  movd      mm0, TEMPd
+  psllq     mm0, 32
+  paddsw    mm1, mm0
+
   movzx     TEMPd, BYTE [Aq]
   movq      mm0, [TABLEq + 6144 + 8 * TEMPq]
   pmullw    mm1, mm0