[PATCH] D130742: [AMDGPU] Extend SILoadStoreOptimizer to s_load instructions

[Carl Ritson via Phabricator via llvm-commits]

critson added a comment.

Compilation testing on graphics shader corpus (~10k pipelines) for Navi10.

~15% reduction in s_load_dword instructions.
~1% reduction in s_waitcnt instructions.
~1% increase in s_mov instructions.
~0.75% overall decrease in instructions.

VGPR usage increase in ~400 pipelines (mostly 1-2 VGPRs).
VGPR usage decrease in ~500 pipelines.
Mean VGPR usage change: -0.02 VGPRs.

SGPR usage increase in ~3000 pipelines.
SGPR usage decrease in ~1000 pipelines.
Mean SGPR usage change: 0.62 SGPRs.

No changes in scratch usage.

Occupancy (executable wavefronts based on VGPR count) increased in 234 pipelines, decreased in 125 pipelines.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D130742/new/

https://reviews.llvm.org/D130742