[RISCV] Unprofitable vectorization in llvm-test-suite since a49030e493b76bc2b294d36709c7ed84aef557f4

[lukel97]

Between a49030e and e8b5fbd we're seeing some significant regressions on llvm-test-suite, up to 2x slower on MultiSource/Benchmarks/MiBench/security-rijndael/security-rijndael: https://lnt.lukelau.me/db_default/v4/nts/988

From a brief look at the profiles it looks like we're now performing a lot of unprofitable SLP vectorization that involves a lot of vslideups and downs of immediates. Is this related to #153213? #152787 also landed in the meantime but there's no floats in the generated code.

cc) @alexey-bataev

[alexey-bataev]

Between a49030e and e8b5fbd we're seeing some significant regressions on llvm-test-suite, up to 2x slower on MultiSource/Benchmarks/MiBench/security-rijndael/security-rijndael: https://lnt.lukelau.me/db_default/v4/nts/988

From a brief look at the profiles it looks like we're now performing a lot of unprofitable SLP vectorization that involves a lot of vslideups and downs of immediates. Is this related to #153213? #152787 also landed in the meantime but there's no floats in the generated code.

cc) @alexey-bataev

Need a reproducer (LLVM IR) for investigation, cannot say anything without a repro

[llvmbot]

@llvm/issue-subscribers-backend-risc-v

Author: Luke Lau (lukel97)

Between a49030e and e8b5fbd we're seeing some significant regressions on llvm-test-suite, up to 2x slower on MultiSource/Benchmarks/MiBench/security-rijndael/security-rijndael: https://lnt.lukelau.me/db_default/v4/nts/988

From a brief look at the profiles it looks like we're now performing a lot of unprofitable SLP vectorization that involves a lot of vslideups and downs of immediates. Is this related to #153213? #152787 also landed in the meantime but there's no floats in the generated code.

cc) @alexey-bataev

[lukel97]

I've bisected this back to 055bfc0 #155415. I guess SLP is now able to handle these new GEPs? I've attached the reproducer below.

decrypt.ll.txt

[preames]

An observation on the optimized IR from Luke's reproducer - we have a whole bunch of shuffle pairs on either side of gathers which could be combined and eliminated. I don't know if we know how to reshuffle through a gather, but this case points out why doing so might be profitable. In addition to the SLP regression here, maybe a missed VectorCombine?

[hiraditya]

In addition to the SLP regression here, maybe a missed VectorCombine?

Should we create a separate issue for this?