speed up maprect for scale+trans case

src/core/SkMatrix.cpp

 /*
  * Copyright 2006 The Android Open Source Project
  *
  * Use of this source code is governed by a BSD-style license that can be
  * found in the LICENSE file.
  */
 
 #include "SkFloatBits.h"
 #include "SkMatrix.h"
 #include "SkNx.h"
@@ skipping 1079 matching lines @@
         }
     } else {
         SkMatrix tmp = *this;
 
         tmp.fMat[kMTransX] = tmp.fMat[kMTransY] = 0;
         tmp.clearTypeMask(kTranslate_Mask);
         tmp.mapPoints(dst, src, count);
     }
 }
 
+void SkMatrix::mapRectScaleTranslate(SkRect* dst, const SkRect& src) const {
+    SkASSERT(dst);
+    SkASSERT(this->isScaleTranslate());
+    
+#if 1

[mtklein, 2016/06/30 13:15:39]

You're saying this enabled #if case is faster than the #else?

If so, good that the simple one works best, and good to have tried the #else
experiment.  I don't think there's anything else to try here for now.

[reed1, 2016/06/30 13:20:51]

Done.

+    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;

[mtklein, 2016/06/30 13:15:39]

I'd personally rename result to ltrb and flipped to rblt... makes it just a
little easier to see what's going on.

[reed1, 2016/06/30 13:20:51]

Done.

+    // 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);
+    Sk4f(min[0], min[1], 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];
 
         src.toQuad(quad);
         this->mapPoints(quad, quad, 4);
         dst->set(quad, 4);
         return false;
     }
 }
@@ skipping 741 matching lines @@
     const SkScalar m10 = -m01;
     const SkScalar m11 = m00;
     const SkScalar m12 = fTy;
 
     quad[0].set(m02, m12);
     quad[1].set(m00 * width + m02, m10 * width + m12);
     quad[2].set(m00 * width + m01 * height + m02, m10 * width + m11 * height + m12);
     quad[3].set(m01 * height + m02, m11 * height + m12);
 #endif
 }