Both shader and gradient_shader write matrices to the flatten stream. However, they were
just calling write(&matrix, sizeof(SkMatrix)) and the matrix can contain lazily-computed
function ptrs as part of its internal cache. Thus two matrices that are logically the
same may write different bytes.
This is a problem because picture relies on flattening objects and then using the
flatten stream as a key into its cache. This matrix-write bug effectively kills the
effectiveness of the cache for shaders.
The fix is to write proper read/write functions for matrix (and region btw). These
call through to the existing low-level flatten routines (which just write into a
memory ptr).
git-svn-id: http://skia.googlecode.com/svn/trunk@1290 2bbb7eff-a529-9590-31e7-b0007b416f81
f2eb5ab780