speed up maprect for scale+trans case

src/core/SkMatrix.cpp

Index: src/core/SkMatrix.cpp
diff --git a/src/core/SkMatrix.cpp b/src/core/SkMatrix.cpp
index 5711bd9aef2f169fba19af723be848dbc0cef584..bad91742ae57b0edefef5854be7cefbb8e8b1c15 100644
--- a/src/core/SkMatrix.cpp
+++ b/src/core/SkMatrix.cpp
@@ -1097,12 +1097,39 @@ void SkMatrix::mapVectors(SkPoint dst[], const SkPoint src[], int count) const {
     }
 }
 
+void SkMatrix::mapRectScaleTranslate(SkRect* dst, const SkRect& src) const {
+    SkASSERT(dst);
+    SkASSERT(this->isScaleTranslate());
+    
+#if 1
+    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);
+#else
+    Sk4f a = Sk4f::Load(&fMat[0]);
+    Sk4f b = Sk4f::Load(&fMat[4]);
+    Sk4f scale(a[0], b[0], a[0], b[0]);
+    Sk4f trans(a[2], b[1], a[2], b[1]);
+#endif
+
+    Sk4f result = Sk4f::Load(&src.fLeft) * scale + trans;
+    // need to sort so we're not inverted
+    Sk4f flipped(result[2], result[3], result[0], result[1]);
+    Sk4f min = Sk4f::Min(result, flipped);
+    Sk4f max = Sk4f::Max(result, flipped);
+    // We can extract either pair [0,1] or [2,3] from min and max and be correct, but on

[mtklein, 2016/06/30 13:24:56]

Neat.

+    // ARM this sequence generates the fastest (a single instruction).
+    Sk4f(min[2], min[3], max[0], max[1]).store(&dst->fLeft);
+}
+
 bool SkMatrix::mapRect(SkRect* dst, const SkRect& src) const {
     SkASSERT(dst);
 
-    if (this->rectStaysRect()) {
-        this->mapPoints((SkPoint*)dst, (const SkPoint*)&src, 2);
-        dst->sort();
+    if (this->isScaleTranslate()) {
+        this->mapRectScaleTranslate(dst, src);
         return true;
     } else {
         SkPoint quad[4];