fix the visual difference of raster vs gpu -- inverse_paths

src/core/SkCanvas.cpp

Index: src/core/SkCanvas.cpp
diff --git a/src/core/SkCanvas.cpp b/src/core/SkCanvas.cpp
index ed8274dc5d3b4a9bf09d6241541620b1a11f9a4f..eded8d80b3af16382782db4e77de46313a08656c 100644
--- a/src/core/SkCanvas.cpp
+++ b/src/core/SkCanvas.cpp
@@ -1737,7 +1737,9 @@ void SkCanvas::drawPath(const SkPath& path, const SkPaint& paint) {
             return;
         }
     }
-    if (path.isEmpty()) {
+
+    SkRect r = path.getBounds();
+    if (SkScalarNearlyZero(r.width()) && SkScalarNearlyZero(r.height())) {

[bsalomon, 2014/02/14 13:56:40]

It's not obvious to me that this is the right check.  Where does the sw
rasterizer check the bounds? Is that check redundant with this change? Also, why
nearly zero and not simply path.getBounds().isEmpty()?

[yunchao, 2014/02/17 09:54:32]

Hi Brian, SW rasterizer check the bounds in some cases, the code scattered in
different files for FillPath, AntiFillPath. But Hairline Path and Anti-Hairline
Path didn't check the bounds for stroked path. HW rasterizer didn't check the
bounds at all.

path.getBounds().isEmpty() basically equals to r.width() == 0 || r.height() ==
0, not r.width() == 0 && r.height() == 0. So it is not suitable for this path
who has only a line. This kind path is not empty. 

Personally, I think check the bounds in SkCanvas::drawPath() is appropriate.
checking path bounds is a revision of path.isEmpty(). It just cover all
practical empty paths -- although these paths have verbs, they are not treated
as empty paths by SkPath::isEmpty(). 

If you prefer to check the bounds in different files for different situation, I
will upload a corresponding patch.

         if (path.isInverseFillType()) {
             this->internalDrawPaint(paint);
         }