[PATCH] D101474: [AMDGPU] Make some VOP3 insts commutable

Thu Apr 29 08:34:45 PDT 2021

Joe_Nash added a comment.

In D101474#2725051 <https://reviews.llvm.org/D101474#2725051>, @foad wrote:

>> This patch revises d35d8da7d6ac6c08578ec0569b072292631691e0 <https://reviews.llvm.org/rGd35d8da7d6ac6c08578ec0569b072292631691e0>.
>> It contains the commute opportunities excluding float insts
>
> Maybe mention D99376 <https://reviews.llvm.org/D99376> here rather than just the commit hash?
>
> Why are you excluding float insts now? Is it related to the test failures that caused you to revert D99376 <https://reviews.llvm.org/D99376>, or is it a different problem?

That is what you suggested last time on https://reviews.llvm.org/D99376

foad added inline comments.Mar 30 2021, 2:08 AM
llvm/lib/Target/AMDGPU/VOP3Instructions.td

371
You have to be super careful with fp min/max/med because the NaN handling is not commutative. You could commute them with suitable "nnan" or other IEEE-related flags, but it's probably not worth it. So I would suggest dropping them.

632
Likewise, drop f16 min/max/med.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D101474/new/

https://reviews.llvm.org/D101474