[PATCH] D140546: [AMDGPU] Remove permlane discard vdst_in optimization from isel

[Jay Foad via Phabricator via llvm-commits]

foad added a subscriber: nlopes.
foad added inline comments.


================
Comment at: llvm/test/CodeGen/AMDGPU/llvm.amdgcn.permlane.ll:853
   %tidx = call i32 @llvm.amdgcn.workitem.id.x()
-  %v = call i32 @llvm.amdgcn.permlane16(i32 12345, i32 %tidx, i32 %src1, i32 %src2, i1 1, i1 0)
+  %v = call i32 @llvm.amdgcn.permlane16(i32 undef, i32 %tidx, i32 %src1, i32 %src2, i1 1, i1 0)
   store i32 %v, ptr addrspace(1) %out
----------------
arsenm wrote:
> foad wrote:
> > arsenm wrote:
> > > Should use poison
> > Really? I read that "Most instructions return ‘poison’ when one of their arguments is ‘poison’" and that's not what I want here. Besides I am just mimicking InstCombine which still uses undef for this optimization.
> undef is soft deprecated. Undef should be freeze poison
+ @nlopes So is "freeze poison" the best option here? I guess just "poison" would work, since nothing is actually going to remove the whole intrinsic in a codegen test, but it seems conceptually wrong.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D140546/new/

https://reviews.llvm.org/D140546