Shrink GL ProgramDesc from 88B to 24B

Skia issue 284 (http://code.google.com/p/skia/issues/detail?id=284)
GrGLProgram::ProgramDesc is passed as input to GrBinHashKey::keyData(), which yesterday was identified as the most expensive Skia function in profiles of the FishIETank demo.
Algorithm changes yielded a 6x speedup; this patch reduces the size of the data being hashed 3x for additional speedup.
Bools and small enums are stored in bytes, one bool is merged into a preexisting flag byte.
Many format tweaks (removal of trailing whitespace).

[TomH, 2011-06-02 15:09:19]

Skia issue 284 (http://code.google.com/p/skia/issues/detail?id=284)
GrGLProgram::ProgramDesc is passed as input to GrBinHashKey::keyData(), which
yesterday was identified as the most expensive Skia function in profiles of the
FishIETank demo.
Algorithm changes yielded a 6x speedup; this patch reduces the size of the data
being hashed 3x for additional speedup.
Bools and small enums are stored in bytes, one bool is merged into a preexisting
flag byte.

[junov1, 2011-06-02 15:22:55]

There are still unused bits in here. For example, some of the enum fields could
be bit-shifted and stored in the same uint8_t, but that might be overkill. I
think packing any tighter would risk making the code harder to maintain.  If
hash key computation no longer appears on the profiling radar screen, then this
is the right solution IMHO.
LGTM.

[reed1, 2011-06-02 15:23:48]

we'll want a 2nd CL soon to make it clearer that enabled is overlayed with
fOptFlags, either by having setters/getters for all flags, or by forcing
call-sites to specify all the flags at once.

LGTM

http://codereview.appspot.com/4561058/diff/4001/gpu/src/GrGLProgram.h
File gpu/src/GrGLProgram.h (right):

http://codereview.appspot.com/4561058/diff/4001/gpu/src/GrGLProgram.h#newcode127
gpu/src/GrGLProgram.h:127: inline bool isEnabled () const {
no space before (

[TomH, 2011-06-02 15:41:55]

On 2011/06/02 15:22:55, junov1 wrote:
> There are still unused bits in here. For example, some of the enum fields
could
> be bit-shifted and stored in the same uint8_t, but that might be overkill. I
> think packing any tighter would risk making the code harder to maintain.  If
> hash key computation no longer appears on the profiling radar screen, then
this
> is the right solution IMHO.
> LGTM.

Mike & I spent a few minutes talking about that before I uploaded the patch, and
that was our conclusion.

Before changes, GrBinHashKey::keyData() was over 4% of profiles on both Linux &
Mac. With code changes, it was 0.5%-0.8%, and hopefully  this will put it below
0.3%. There are half a dozen skia functions in the 0.5%-2% range; one looks hard
to optimize, but the rest either have hotspots or unnecessary call sites. We're
tracking those at http://code.google.com/p/skia/issues/detail?id=285.

[TomH, 2011-06-02 15:47:58]

Committed in revision 1480.
Maintainability improvements are skia issue 287.