Restrict query bounds for reduce clip to dev bounds

src/gpu/batches/GrTessellatingPathRenderer.cpp

Index: src/gpu/batches/GrTessellatingPathRenderer.cpp
diff --git a/src/gpu/batches/GrTessellatingPathRenderer.cpp b/src/gpu/batches/GrTessellatingPathRenderer.cpp
index 901d38369cfe072496a306b063544a8846a98b39..a3a888302a15ee775f2ba88208f77b0678ff0b06 100644
--- a/src/gpu/batches/GrTessellatingPathRenderer.cpp
+++ b/src/gpu/batches/GrTessellatingPathRenderer.cpp
@@ -1592,9 +1592,17 @@ private:
       , fColor(color)
       , fPath(path)
       , fStroke(stroke)
-      , fViewMatrix(viewMatrix)
-      , fClipBounds(clipBounds) {
-        fBounds = path.getBounds();
+      , fViewMatrix(viewMatrix) {
+        const SkRect& pathBounds = path.getBounds();
+        fClipBounds = clipBounds;
+        // Because the clip bounds are used to add a contour for inverse fills, they must also
+        // include the path bounds.

[robertphillips, 2015/11/23 17:22:55]

// This will increase the size of the resulting geometry but will propagate more
reliable information to the ClipStack (which will need to apply an explicit clip
to the inverse filled geometry).

?

I wonder if adding an extra bit of bloat to the clipBounds (for the inverse fill
case) will make the tesselator happier.

[bsalomon, 2015/11/23 17:59:34]

It's not fundamentally for clipping. The current code is just incorrect even
without changes to the clipping code.
I don't see any artifacts in any of our tests. I can think of cases where it
would simplify things to add pad and cases where it would make it worse (e.g.
the contains a rect where one edge of the rect is along the path bounds.

+        fClipBounds.join(pathBounds);
+        if (path.isInverseFillType()) {
+            fBounds = fClipBounds;
+        } else {
+            fBounds = path.getBounds();
+        }
         if (!stroke.isFillStyle()) {
             SkScalar radius = SkScalarHalf(stroke.getWidth());
             if (stroke.getJoin() == SkPaint::kMiter_Join) {