trim some fat from SSE2 fixed point alpha code

src/opts/SkPx_sse.h

 /*
  * Copyright 2015 Google Inc.
  *
  * Use of this source code is governed by a BSD-style license that can be
  * found in the LICENSE file.
  */
 
 #ifndef SkPx_sse_DEFINED
 #define SkPx_sse_DEFINED
 
@@ skipping 43 matching lines @@
     struct Alpha {
         __m128i fVec;
         Alpha(__m128i vec) : fVec(vec) {}
 
         static Alpha Dup(uint8_t a) { return _mm_set1_epi8(a); }
         static Alpha Load(const uint8_t* a) {
             __m128i as = _mm_cvtsi32_si128(*(const uint32_t*)a);    // ____ ____ ____ 3210
         #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
             return _mm_shuffle_epi8(as, _mm_set_epi8(3,3,3,3, 2,2,2,2, 1,1,1,1, 0,0,0,0));
         #else
-            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
-            return _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 as;
         #endif
         }
         static Alpha Load(const uint8_t* a, int n) {
             SkASSERT(n > 0 && n < 4);
             uint8_t a4[] = { 0,0,0,0 };
             switch (n) {
                 case 3: a4[2] = a[2];  // fall through
                 case 2: a4[1] = a[1];  // fall through
                 case 1: a4[0] = a[0];
             }
@@ skipping 23 matching lines @@
         SkPx addNarrowHi(const SkPx& o) const {
             Wide sum = (*this + o.widenLo()).shr<8>();
             return _mm_packus_epi16(sum.fLo, sum.fHi);
         }
     };
 
     Alpha alpha() const {
     #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
         return _mm_shuffle_epi8(fVec, _mm_set_epi8(15,15,15,15, 11,11,11,11, 7,7,7,7, 3,3,3,3));
     #else
-        __m128i as = _mm_srli_epi32(fVec, 24);           // ___3 ___2 ___1 ___0
-        as   = _mm_or_si128(as, _mm_slli_si128(as, 1));  // __33 __22 __11 __00
-        return _mm_or_si128(as, _mm_slli_si128(as, 2));  // 3333 2222 1111 0000
+        // We exploit that A >= rgb for any premul pixel.
+        __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 as;
     #endif
     }
 
     Wide widenLo() const {
         return Wide(_mm_unpacklo_epi8(fVec, _mm_setzero_si128()),
                     _mm_unpackhi_epi8(fVec, _mm_setzero_si128()));
     }
     Wide widenHi() const {
         return Wide(_mm_unpacklo_epi8(_mm_setzero_si128(), fVec),
                     _mm_unpackhi_epi8(_mm_setzero_si128(), fVec));
@@ skipping 21 matching lines @@
     SkPx addAlpha(const Alpha& a) const {
         return _mm_add_epi8(fVec, _mm_and_si128(a.fVec, _mm_set1_epi32(0xFF000000)));
     }
 };
 
 }  // namespace sse
 
 typedef sse::SkPx SkPx;
 
 #endif//SkPx_sse_DEFINED