trim some fat from SSE2 fixed point alpha code

src/opts/Sk4px_SSE2.h

Index: src/opts/Sk4px_SSE2.h
diff --git a/src/opts/Sk4px_SSE2.h b/src/opts/Sk4px_SSE2.h
index 5e97abf308b3af30c3d6f3301c7101f0030a72c5..9c3eb1210c105f9a9f6ccbf808c4818f59c5e7fa 100644
--- a/src/opts/Sk4px_SSE2.h
+++ b/src/opts/Sk4px_SSE2.h
@@ -63,18 +63,17 @@ inline Sk4px Sk4px::Wide::addNarrowHi(const Sk16h& other) const {
 #else
     inline Sk4px Sk4px::alphas() const {
         static_assert(SK_A32_SHIFT == 24, "Intel's always little-endian.");
-        __m128i as = _mm_srli_epi32(this->fVec, 24);   // ___3 ___2 ___1 ___0
-        as = _mm_or_si128(as, _mm_slli_si128(as, 1));  // __33 __22 __11 __00
-        as = _mm_or_si128(as, _mm_slli_si128(as, 2));  // 3333 2222 1111 0000
+        // We exploit that A >= rgb for any premul pixel.

[msarett, 2015/11/17 19:59:57]

Clever :)

+        __m128i as = fVec;                             // 3xxx 2xxx 1xxx 0xxx
+        as = _mm_max_epu8(as, _mm_srli_epi32(as,  8)); // 33xx 22xx 11xx 00xx
+        as = _mm_max_epu8(as, _mm_srli_epi32(as, 16)); // 3333 2222 1111 0000
         return Sk16b(as);
     }
 
     inline Sk4px Sk4px::Load4Alphas(const SkAlpha a[4]) {
         __m128i as = _mm_cvtsi32_si128(*(const uint32_t*)a);  // ____ ____ ____ 3210
-        as = _mm_unpacklo_epi8 (as, _mm_setzero_si128());     // ____ ____ _3_2 _1_0
-        as = _mm_unpacklo_epi16(as, _mm_setzero_si128());     // ___3 ___2 ___1 ___0
-        as = _mm_or_si128(as, _mm_slli_si128(as, 1));         // __33 __22 __11 __00
-        as = _mm_or_si128(as, _mm_slli_si128(as, 2));         // 3333 2222 1111 0000
+        as = _mm_unpacklo_epi8 (as, as);                      // ____ ____ 3322 1100
+        as = _mm_unpacklo_epi16(as, as);                      // 3333 2222 1111 0000
         return Sk16b(as);
     }
 #endif