Remove all easily-removable uses of setTexture()

[TomH, 2012-07-20 13:58:45]

After this, all we have to do is figure out how color table is messing with
multiple textures & then we should be able to get rid of
GrDrawState::fTextures[].

http://codereview.appspot.com/6425055/diff/1/include/gpu/GrPaint.h
File include/gpu/GrPaint.h (right):

http://codereview.appspot.com/6425055/diff/1/include/gpu/GrPaint.h#newcode56
include/gpu/GrPaint.h:56: //}
Can't enable these until we're storing all the textures on custom stages - which
will be very soon, I believe, but not in this CL.

[bsalomon, 2012-07-20 14:29:01]

Mostly LGTM. But I think a helper setter on GrDrawState would be nice.

http://codereview.appspot.com/6425055/diff/1/src/gpu/SkGpuDevice.cpp
File src/gpu/SkGpuDevice.cpp (right):

http://codereview.appspot.com/6425055/diff/1/src/gpu/SkGpuDevice.cpp#newcode431
src/gpu/SkGpuDevice.cpp:431: // particularly since table color filter requires
multiple textures
I don't get the color table filter part of the comment.

[TomH, 2012-07-20 14:53:50]

Added helper. Half the places we're building things are on GrPaint instead of
GrDrawState, but this is at least somewhat more concise.

[bsalomon, 2012-07-20 15:41:32]

On 2012/07/20 14:53:50, TomH wrote:
> Added helper. Half the places we're building things are on GrPaint instead of
> GrDrawState, but this is at least somewhat more concise.

LGTM. It seems like we could make GrPaint::setTexture/setMask() do the same
thing as GrDrawState::createTextureEffect() once we kill that last use of the
existing setTexture.

[TomH, 2012-07-20 16:27:29]

Closed with r4694.

[TomH, 2012-07-20 19:16:03]

Bug not exposed on Linux tests by missing gms fixed in r4700 /
http://codereview.appspot.com/6419061/.

[TomH, 2012-07-23 14:20:02]

It looks like this caused a pretty severe performance regression of the
picture_playback benchmarks - worst on Linux, mildest on Windows - and will need
further modification or rollback. When I'm back in the office tomorrow I'll see
if we're now hitting memory management and need to test Brian's technique.

[TomH, 2012-08-01 20:14:15]

For the record: using GrMemoryPool recovers most of the regression, but SkTLS is
slowing that down. For now we've accepted the performance loss and filed a bug
to watch it.