Bad replacement of `vpmaddwd` + `vpaddd` with `vpdpwssd` - Back in Clang 20.1.0

[TheBlackPlague]

See #62026.

This issue has resurfaced in the latest version of LLVM (Clang).

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: Shaheryar Sohail (TheBlackPlague)

See #62026.

This issue has resurfaced in the latest version of LLVM (Clang).

[phoebewang]

Are you using AMD Zen4 or Zen5? #85375 disabled prior optimization for AMD.

[TheBlackPlague]

I am using Zen 4.

[TheBlackPlague]

@phoebewang Could you clarify how #85375 might be related? I'm a tad bit lost.

[phoebewang]

@phoebewang Could you clarify how #85375 might be related? I'm a tad bit lost.

@TheBlackPlague #85375 chose to generate vpdpwssd for Zen 4.

[TheBlackPlague]

Right, but why was this change approved, given that the original issue showcased that vpdpwssd had dependency issues slowing down the pipeline in a looped case? It is better in straight code but then the compiler should determine if it's in a loop context or not.

I believe that's what's stated here:

• Bad replacement of vpmaddwd + vpaddd with vpdpwssd. #62026 (comment)
• Bad replacement of vpmaddwd + vpaddd with vpdpwssd. #62026 (comment)

Was there another good reason for approving #85375 or was #62026 just not considered at the time?

[ganeshgit]

Was there another good reason for approving #85375 or was #62026 just not considered at the time?

There is a counter case #84182! So, a tuning option was added. We are working on this as stated in #84182 (comment)

[TheBlackPlague]

Was there another good reason for approving #85375 or was #62026 just not considered at the time?

There is a counter case #84182! So, a tuning option was added. We are working on this as stated in #84182 (comment)

I see. Thank you for the clarification.

[TheBlackPlague]

Was there another good reason for approving #85375 or was #62026 just not considered at the time?

There is a counter case #84182! So, a tuning option was added. We are working on this as stated in #84182 (comment)

As a follow-up, I cannot find a method of disabling this tuning option from Clang. Could you clarify how one would disable this from clang, given one is using the -march=native argument?