Sk4px blit mask.

src/core/Sk4px.h

Index: src/core/Sk4px.h
diff --git a/src/core/Sk4px.h b/src/core/Sk4px.h
index 24a21c66c1c40ff02a1fada344cf5dd4a713ed01..ffde1af504f682504034839b5db142eecf76fa39 100644
--- a/src/core/Sk4px.h
+++ b/src/core/Sk4px.h
@@ -165,6 +165,34 @@ public:
         }
     }
 
+    // As above, but with dst4' = fn(dst4, alpha4).
+    template <typename Fn, typename Dst>
+    static void MapDstAlpha(int n, Dst* dst, const SkAlpha* a, const Fn& fn) {
+        while (n > 0) {
+            if (n >= 8) {
+                Sk4px dst0 = fn(Load4(dst+0), Load4Alphas(a+0)),
+                      dst4 = fn(Load4(dst+4), Load4Alphas(a+4));
+                dst0.store4(dst+0);
+                dst4.store4(dst+4);
+                dst += 8; a += 8; n -= 8;
+                continue;  // Keep our stride at 8 pixels as long as possible.
+            }
+            SkASSERT(n <= 7);
+            if (n >= 4) {
+                fn(Load4(dst), Load4Alphas(a)).store4(dst);
+                dst += 4; a += 4; n -= 4;
+            }
+            if (n >= 2) {
+                fn(Load2(dst), Load2Alphas(a)).store2(dst);
+                dst += 2; a += 2; n -= 2;
+            }
+            if (n >= 1) {
+                fn(Load1(dst), DupAlpha(*a)).store1(dst);
+            }
+            break;
+        }
+    }
+
     // As above, but with dst4' = fn(dst4, src4, alpha4).
     template <typename Fn, typename Dst>
     static void MapDstSrcAlpha(int n, Dst* dst, const SkPMColor* src, const SkAlpha* a,