[X86] 454.calculix miscompare after #143017

[vzakhari]

This may not be reprodusible with the upstream Flang, because I have a customized LLVM pipeline that runs some extra passes. At the same time, I can confirm that #143017 causes the miscompare by reverting it.

The only change in the whole benchmark appears in stiff2mat.f module that contains a single function.

Attaching the assembly for stiff2mat.f:

• With [X86] combineConcatVectorOps - concat per-lane v2f64/v4f64 shuffles into vXf64 vshufpd #143017: bad.s.gz
• Without [X86] combineConcatVectorOps - concat per-lane v2f64/v4f64 shuffles into vXf64 vshufpd #143017: good.s.gz

Here is the LLVM IR and MIR dumps right before and afrer x86-isel:

• With [X86] combineConcatVectorOps - concat per-lane v2f64/v4f64 shuffles into vXf64 vshufpd #143017: isel.bad.log.gz
• Without [X86] combineConcatVectorOps - concat per-lane v2f64/v4f64 shuffles into vXf64 vshufpd #143017: isel.good.log.gz

Please let me know if you need information from more passes.

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: Slava Zakharin (vzakhari)

This may not be reprodusible with the upstream Flang, because I have a customized LLVM pipeline that runs some extra passes. At the same time, I can confirm that #143017 causes the miscompare by reverting it.

The only change in the whole benchmark appears in stiff2mat.f module that contains a single function.

Attaching the assembly for stiff2mat.f:

• With #143017: bad.s.gz
• Without #143017: good.s.gz

Here is the LLVM IR and MIR dumps right before and afrer x86-isel:

• With #143017: isel.bad.log.gz
• Without #143017: isel.good.log.gz

Please let me know if you need information from more passes.

[RKSimon]

Thanks - I think I know whats going on now - will make a fix tomorrow

[RKSimon]

@vzakhari a6ace28 should address the issue - please can you confirm?

[vzakhari]

Yes, it fixes the issue. Thank you, @RKSimon!