work around leftshift for negative values

src/core/SkEdge.cpp

Index: src/core/SkEdge.cpp
diff --git a/src/core/SkEdge.cpp b/src/core/SkEdge.cpp
index c64896f2e0345125ba35d8b5728fb92c612773f9..94b7870b034b66814b89ebc718c74e4861641653 100644
--- a/src/core/SkEdge.cpp
+++ b/src/core/SkEdge.cpp
@@ -326,8 +326,9 @@ static inline int SkFDot6UpShift(SkFDot6 x, int upShift) {
 */
 static SkFDot6 cubic_delta_from_line(SkFDot6 a, SkFDot6 b, SkFDot6 c, SkFDot6 d)
 {
-    SkFDot6 oneThird = ((a << 3) - ((b << 4) - b) + 6*c + d) * 19 >> 9;
-    SkFDot6 twoThird = (a + 6*b - ((c << 4) - c) + (d << 3)) * 19 >> 9;
+    // since our parameters may be negative, we don't use << to avoid ASAN warnings
+    SkFDot6 oneThird = (a*8 - b*15 + 6*c + d) * 19 >> 9;
+    SkFDot6 twoThird = (a + 6*b - c*15 + d*8) * 19 >> 9;

[caryclark, 2016/03/08 14:20:23]

I'm curious how this works. If the bug is that the highest number bit is shifted
into the sign bit by the shift of a, b, c, or d, doesn't this imply that the
number may be so large that it can't be safely multiplied by 19?

[reed1, 2016/03/08 14:26:44]

I don't think that is the bug. I think the bug from ASAN is that one of the
values was negative at all (as that is "undefined"), not that we actually
overflowed.

 
     return SkMax32(SkAbs32(oneThird), SkAbs32(twoThird));
 }