add rounding-using-doubles methods on SkScalar and SkRect

include/core/SkScalar.h

Index: include/core/SkScalar.h
diff --git a/include/core/SkScalar.h b/include/core/SkScalar.h
index b9256badb4c02b6524ab075e39310e52fb118ebf..b37cf5c998fd4505f6dfd03b43ac39841827e91a 100644
--- a/include/core/SkScalar.h
+++ b/include/core/SkScalar.h
@@ -83,6 +83,26 @@ static inline bool SkScalarIsFinite(float x) {
 #define SkScalarRoundToInt(x)       sk_float_round2int(x)
 #define SkScalarTruncToInt(x)       static_cast<int>(x)
 
+/**
+ *  Variant of SkScalarRoundToInt, that performs the rounding step (adding 0.5) explicitly using
+ *  double, to avoid possibly losing the low bit(s) of the answer before calling floor().
+ *
+ *  This routine will likely be slower than SkScalarRoundToInt(), and should only be used when the
+ *  extra precision is known to be valuable.
+ *
+ *  In particular, this catches the following case:
+ *      SkScalar x = 0.49999997;
+ *      int ix = SkScalarRoundToInt(x);
+ *      SkASSERT(0 == ix);    // <--- fails
+ *      ix = SkDScalarRoundToInt(x);
+ *      SkASSERT(0 == ix);    // <--- succeeds
+ */
+static inline int SkDScalarRoundToInt(SkScalar x) {
+    double xx = x;
+    xx += 0.5;
+    return (int)floor(xx);

[bungeman-skia, 2014/05/05 15:39:31]

So this is round to nearest, round half up. C++11 has nearbyint with
FE_TONEAREST, but that is round to nearest, round away from zero. At least
that's what the spec seems to say. libc appears to document that it will instead
round to even. In other words, C++11 and C99 have a round to nearest, but the
tie breaking seems to still be lacking.

[reed1, 2014/05/05 15:56:37]

Agreed, that's why unfortunately we have to write our own.

+}
+
 /** Returns the absolute value of the specified SkScalar
 */
 #define SkScalarAbs(x)          sk_float_abs(x)