check bounds of each cubic segment against clip

src/core/SkEdgeClipper.cpp

Index: src/core/SkEdgeClipper.cpp
diff --git a/src/core/SkEdgeClipper.cpp b/src/core/SkEdgeClipper.cpp
index c6a4fb2971312217fa2eb4eb988e9b5594f671b7..55f9192aea198f77eb61232274c28cf746363aa4 100644
--- a/src/core/SkEdgeClipper.cpp
+++ b/src/core/SkEdgeClipper.cpp
@@ -358,14 +358,19 @@ void SkEdgeClipper::clipMonoCubic(const SkPoint src[4], const SkRect& clip) {
     }
 }
 
+static bool quick_reject_in_y(const SkPoint pts[4], const SkRect& clip) {

[mtklein, 2015/12/17 14:58:15]

Just for fun, let's compare speed with

  SkScalar min = SkTMin(SkTMin(pts[0].fY, pts[1].fY), SkTMin(pts[2].fY,
pts[3].fY));
  SkScalar max = SkTMax(SkTMax(pts[0].fY, pts[1].fY), SkTMax(pts[2].fY,
pts[3].fY));
  return min > clip.bottom() || max < clip.top();

?

If it's roughly the same speed on your desktop, I'd go with that.  Hopping back
and forth between scalar and vector code isn't cheap on ARM, nor is allTrue().

+    Sk4s ys(pts[0].fY, pts[1].fY, pts[2].fY, pts[3].fY);
+    Sk4s t(clip.top());
+    Sk4s b(clip.bottom());
+
+    return (ys < t).allTrue() || (ys > b).allTrue();
+}
+
 bool SkEdgeClipper::clipCubic(const SkPoint srcPts[4], const SkRect& clip) {
     fCurrPoint = fPoints;
     fCurrVerb = fVerbs;
 
-    SkRect  bounds;
-    bounds.set(srcPts, 4);
-
-    if (!quick_reject(bounds, clip)) {
+    if (!quick_reject_in_y(srcPts, clip)) {
         SkPoint monoY[10];
         int countY = SkChopCubicAtYExtrema(srcPts, monoY);
         for (int y = 0; y <= countY; y++) {