[Mesa-dev] i965: FS copy propagation across control flow

[Kenneth Graunke]

On 10/30/2012 08:28 PM, Eric Anholt wrote:
> Here's a patch series to clean up the most glaring failures I think we have
> left in FS code generation other than variable-indexed array access.
> Unfortunately, I haven't found a particular testcase to show that it's a
> performance improvement, but I still think it's a good idea since it does
> remove instructions.
>
> I was hoping this series would let me remove the badly-named
> register_coalesce() pass, but it turns out that doing so increases shader-db
> instruction count by a significant fraction of a percent.  A bit of a
> surprise.

Patches 1-3 are:
Reviewed-by: Kenneth Graunke <kenneth at whitecape.org>

(I haven't looked at patch 4 yet, but I will soon.)

This also has another benefit: it cuts compilation time of L4D2's 
largest fragment shader from 10.2 to 4.3 seconds (a 57% reduction!).

We used to do 26 iterations through the brw_fs optimization loop; the 
first two did a bunch of optimizing, but on iterations 3-25 only 
register_coalesce() flagged any progress.  Which also meant 
recalculating live intervals every time.  Absurdly expensive.

With your patch, we do exactly 3 iterations.
1: copy propagation coalesce coalesce2 compute->mrf
2: CSE copy propagation
3: (nothing)

This is vastly more reasonable.

--Ken