[PATCH] D106079: [AMDGPU] Divergence-driven compare operations instruction selection

Mon Jul 19 11:02:39 PDT 2021

alex-t added inline comments.

================
Comment at: llvm/test/CodeGen/AMDGPU/addrspacecast.ll:14
+; CI-DAG: s_cmp_lg_u32 [[PTR]], -1
+; CI-DAG: s_cselect_b64 vcc, -1, 0
 ; CI-DAG: v_cndmask_b32_e32 v[[HI:[0-9]+]], 0, [[VAPERTURE]], vcc
----------------
rampitec wrote:
> This also needs a sort of a look ahead.
All such a kind of regressions need to be addressed in the separate patch I believe. Please note here 2 facts:
1. v_cndmask that consumes VCC is uniform and should be selected to S_CSELECT itself but this part of the work is not done yet.
2. To fix all of these patterns I need to add a lookahead in the selection predicate - scan over all CC users and select only if no VALU instructions were found. I was planning this as a separate change.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D106079/new/

https://reviews.llvm.org/D106079