try to speed-up maprect + round2i + contains

src/core/SkMatrix.cpp

Index: src/core/SkMatrix.cpp
diff --git a/src/core/SkMatrix.cpp b/src/core/SkMatrix.cpp
index 71ef24c5e8e13ea9799ca6cf1f996697c96610ab..2f9d1358b907fc61c6c88c4525dc5cf0a593656c 100644
--- a/src/core/SkMatrix.cpp
+++ b/src/core/SkMatrix.cpp
@@ -1119,6 +1119,28 @@ SkRect SkMatrix::mapRectScaleTranslate(SkRect src) const {
     return dst;
 }
 
+SkRect SkMatrix::mapRectScaleTranslate(SkRect src) const {
+    SkASSERT(this->isScaleTranslate());
+    
+    SkScalar sx = fMat[kMScaleX];
+    SkScalar sy = fMat[kMScaleY];
+    SkScalar tx = fMat[kMTransX];
+    SkScalar ty = fMat[kMTransY];
+    Sk4f scale(sx, sy, sx, sy);
+    Sk4f trans(tx, ty, tx, ty);
+    
+    Sk4f ltrb = Sk4f::Load(&src.fLeft) * scale + trans;
+    // need to sort so we're not inverted
+    Sk4f rblt(ltrb[2], ltrb[3], ltrb[0], ltrb[1]);
+    Sk4f min = Sk4f::Min(ltrb, rblt);
+    Sk4f max = Sk4f::Max(ltrb, rblt);
+    // We can extract either pair [0,1] or [2,3] from min and max and be correct, but on
+    // ARM this sequence generates the fastest (a single instruction).
+    SkRect dst;
+    Sk4f(min[0], min[1], max[0], max[1]).store(&dst.fLeft);
+    return dst;
+}
+
 bool SkMatrix::mapRect(SkRect* dst, const SkRect& src) const {
     SkASSERT(dst);