[PATCH] D137954: Enable roundeven.
Matt Arsenault via Phabricator via llvm-commits
llvm-commits at lists.llvm.org
Tue Dec 20 09:20:15 PST 2022
arsenm added inline comments.
================
Comment at: llvm/test/CodeGen/AMDGPU/roundeven.ll:479
+; SDAG_GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; SDAG_GFX6-NEXT: v_cvt_f32_f16_e32 v1, v1
+; SDAG_GFX6-NEXT: v_rndne_f32_e32 v0, v0
----------------
Leonc wrote:
> arsenm wrote:
> > bcahoon wrote:
> > > It looks like arguments are converted from f32->f16->f32. Is that correct/efficient?
> > This is the broken ABI the DAG wants to give targets without legal f16. It’s a problem and ends up with different behavior for GlobalISel
> Is there a workaround?
use an i16 argument and bitcast to half in the IR. Should also figure out how to fix the DAG from promoting to float
Repository:
rG LLVM Github Monorepo
CHANGES SINCE LAST ACTION
https://reviews.llvm.org/D137954/new/
https://reviews.llvm.org/D137954
More information about the llvm-commits
mailing list