Optimize SkAlphaMulQ_SSE2

src/opts/SkColor_opts_SSE2.h

Index: src/opts/SkColor_opts_SSE2.h
diff --git a/src/opts/SkColor_opts_SSE2.h b/src/opts/SkColor_opts_SSE2.h
index 7e61d526b3bf4427462b59e407187d32fb52bcd6..a5d6772b60671562ab148ef216c22c9d60cc30f5 100644
--- a/src/opts/SkColor_opts_SSE2.h
+++ b/src/opts/SkColor_opts_SSE2.h
@@ -52,11 +52,9 @@ static inline __m128i SkAlphaMulQ_SSE2(const __m128i& c, const __m128i& scale) {
 
     // uint32_t ag = ((c >> 8) & mask) * scale
     __m128i ag = _mm_srli_epi16(c, 8);
-    ag = _mm_and_si128(ag, mask);
     ag = _mm_mullo_epi16(ag, s);
 
     // (rb & mask) | (ag & ~mask)
-    rb = _mm_and_si128(mask, rb);

[mtklein, 2014/11/14 14:22:58]

Hmm.  It'd be nice to explain / assert this for readability.  Perhaps,
  static inline void assert_eq(__m128i a, __m128i b) {
    // Compare each byte pair for equality, then check all comparison bits are
true.
    SkASSERT(0xffff == _mm_movemask_epi8(_mm_cmpeq_epi8(a, b)));
  }

Then assert these masks are unnecessary:
  assert_eq(ag, _mm_and_sli128(mask, ag));  // _mm_srli_epi16(c,8) just did this
for us.
  ...
  assert_eq(rb, _mm_and_sli128(mask, rb));   // rb = _mm_srli_epi16(rb, 8) did
this for us.

I'm assuming that the arguments to assert_eq will not even be evaluated in
Release mode where SkASSERT() is a no-op.  If that's not the case, we could of
course write assert_eq as a macro:
  #define ASSERT_EQ(a,b) SkASSERT(0xffff ==
_mm_movemask_epi8(_mm_cmpeq_epi8((a), (b))))

[qiankun, 2014/11/14 15:35:21]

Done.

     ag = _mm_andnot_si128(mask, ag);
     return _mm_or_si128(rb, ag);
 }