[PATCH] D142418: AMDGPU: Teach fneg combines that select has source modifiers

Fri Feb 3 03:06:00 PST 2023

arsenm added a comment.

In D142418#4081927 <https://reviews.llvm.org/D142418#4081927>, @Pierre-vh wrote:

> Do we have/want something like this in GISel?

Of course, but the base set of source modifier combines haven't been ported yet

================
Comment at: llvm/test/CodeGen/AMDGPU/fmin_fmax_legacy.amdgcn.ll:1
 ; RUN: llc -march=amdgcn < %s | FileCheck -enable-var-scope -check-prefixes=SI-SAFE,GCN %s
+; RUN: llc -enable-no-nans-fp-math -enable-no-signed-zeros-fp-math -march=amdgcn < %s | FileCheck -enable-var-scope --check-prefixes=GCN %s
----------------
Pierre-vh wrote:
> It would be nice for those tests to be autogenerated to make them easier to update, make review easier and increase our testing coverage.
> Is there a reason why we have some manual tests like these?
> 
> In any case it's not for this diff but I've been wondering about this. If I have some time maybe I can make diffs to autogenerate a few of them because it can be a bit annoying sometimes
Manual test checks are the original way to do it, generating checks is much newer. Blindly generating checks feels a bit dangerous since it's easy to not actually look at the diffs.

Plus generated tests break more frequently, partially because update_llc_test_checks doesn't support matching registers into variables like the x86 one

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D142418/new/

https://reviews.llvm.org/D142418