Skip to content

nir: Some memcpy and pointer optimizations

Faith Ekstrand requested to merge gfxstrand/mesa:review/nir-opt-memcpy into master

With this batch of optimizations, the quite gnarly batch 32 of OpenCL kernels I've been looking at is reduced to spilling no local variables at all. Those kernels have lots of vec3s which LLVM does it's worst to turn into something horrifically unmanageable but this seems to be able to cut through it.

Merge request reports