[PATCH] D109658: [X86][FP16] Change the order of the operands in complex FMA intrinsics to allow swap between the mul operands.

Mon Sep 13 07:25:12 PDT 2021

RKSimon added a comment.

> Regarding this case, we already have a different name with GCC, I think it worthwhile to use a different order for the swapping optimization.
> With a bit research on AVX512IFMA, I found:
>
> 1. The use of C, A, B order in GCC is not consistent on its AVX512IFMA builtins. It supposes GCC should change to A, B, C order if considering consistency;
> 2. We aren't consistent on AVX512IFMA builtins with GCC either due to the use of select.

Do we have any builtins with the same name as gcc but different operands/behaviours? Those are the only ones that I'd be worried about.

> By the way, GCC folks told me GCC has ability to specify arbitrary operands that can be commutative. But I found both SDNode and MI only have ability on the first 2 operands, which is insufficient for instruction like CFMA. Do you know if we have other mechanism for commutable operands?

Doesn't X86InstrInfo::findCommutedOpIndices handle these cases?

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D109658/new/

https://reviews.llvm.org/D109658