SkTreatAsSprite should take AA into account

src/core/SkMatrix.cpp

Index: src/core/SkMatrix.cpp
diff --git a/src/core/SkMatrix.cpp b/src/core/SkMatrix.cpp
index dfeb721d45e68394f4bf2db1906c11d708f044e3..c6c79631929f11ed9cd2bc96bcd883eb76d70ab3 100644
--- a/src/core/SkMatrix.cpp
+++ b/src/core/SkMatrix.cpp
@@ -1623,8 +1623,12 @@ void SkMatrix::toString(SkString* str) const {
 
 #include "SkMatrixUtils.h"
 
-bool SkTreatAsSprite(const SkMatrix& mat, int width, int height,
-                     unsigned subpixelBits) {
+bool SkTreatAsSprite(const SkMatrix& mat, const SkISize& size, const SkPaint& paint) {
+    // AA is applied using alpha modulation => 8 bits.
+    static const unsigned kAntiAliasSubpixelBits = 8;

[reed1, 2016/01/07 14:40:50]

we have various AA techniques: paths get 2 subpixel bits per coordinate, rects
are special and get 8. I wonder if we can lower this const to 2 or 4 and see if
things still look pretty. If they do, we might get a perf win.

Speaking of that, we should consider realistic ways to measure the perf win of
treating-as-sprite, just to be sure all of this complexity is actually worth it.

[f(malita), 2016/01/07 15:19:00]

4 seems to work just as well, and it matches the prev filter subpixel value. 
Updated.
Agreed, we should add a bench if we don't already have one.

+
+    const unsigned subpixelBits = paint.isAntiAlias() ? kAntiAliasSubpixelBits : 0;
+
     // quick reject on affine or perspective
     if (mat.getType() & ~(SkMatrix::kScale_Mask | SkMatrix::kTranslate_Mask)) {
         return false;
@@ -1641,7 +1645,7 @@ bool SkTreatAsSprite(const SkMatrix& mat, int width, int height,
     }
 
     SkRect dst;
-    SkIRect isrc = { 0, 0, width, height };
+    SkIRect isrc = SkIRect::MakeSize(size);
 
     {
         SkRect src;