Skip to content

intel/fs: Don't munge constant sources to integer multiplication

Ian Romanick requested to merge idr/mesa:review/integer-multiply into master

I started by writing "intel/fs: Promote integer constants after lowering integer multiplication". I had expected it to help more shaders, so then I wrote "intel/fs: Do common optimizations again after lowering integer multiplication". There were some regressions from those two patches (the first, really) in the Intel CI. After some discussion with @mattst88, I wrote "intel/fs: Don't count integer instructions as being possibly coissue".

My gut tells me that we won't want "intel/fs: Promote integer constants after lowering integer multiplication". The first patch makes it so that the integer constants will never get combined, so the ordering should never matter. I have only included it to show the process that I went through.

We also might not want "intel/fs: Do common optimizations again after lowering integer multiplication" since it helps so few shaders on a single, old platform. 🤷

Merge request reports