Add SkDivMod with a special case for ARM.

include/core/SkMath.h

Index: include/core/SkMath.h
diff --git a/include/core/SkMath.h b/include/core/SkMath.h
index 078c8fccdc2206975c07ed830222fe82f7e710af..ba3d36e85f0353df70f89e25593f1e5cd4623887 100644
--- a/include/core/SkMath.h
+++ b/include/core/SkMath.h
@@ -173,4 +173,27 @@ static inline U8CPU SkMulDiv255Round(U16CPU a, U16CPU b) {
     return (prod + (prod >> 8)) >> 8;
 }
 
+/**
+ * Stores x/y and x%y into div and mod respectively.
+ */
+
+template <typename T>
+inline void SkDivMod(T x, T y, T* div, T* mod) {

[reed1, 2013/09/20 19:02:39]

1. SkTDivMode
2. add dox
3. rename x to numer? y to denom?

[mtklein, 2013/09/20 19:29:35]

Done.

+#ifdef SK_CPU_ARM
+    // If we wrote this as in the else branch, GCC won't fuse the two into one
+    // divmod call, but rather a div call followed by a divmod.  Silly!  This
+    // version is just as fast as calling __aeabi_[u]idivmod manually, but with
+    // prettier code.
+    //
+    // This benches as around 2x faster than the code in the else branch.
+    const T d = x/y;
+    *div = d;
+    *mod = x-d*y;
+#else
+    // On x86 this will just be a single idiv.
+    *div = x/y;
+    *mod = x%y;
+#endif  // SK_CPU_ARM
+}
+
 #endif