[llvm] [PatternMatch] Use `m_SpecificCmp` matchers. NFC. (PR #100878)

[Nikita Popov via llvm-commits]

nikic wrote:

> ```
> Top 5 improvements:
>   faiss/IndexIVFFastScan.cpp.ll 7502943608 7223027166 -3.73%
>   gromacs/pull.cpp.ll 3581024503 3451962626 -3.60%
>   spike/regnames.ll 496396386 481168580 -3.07%
>   lightgbm/gradient_discretizer.cpp.ll 1364403988 1324787616 -2.90%
>   meshlab/cube_style_precomputation.cpp.ll 29532646099 28707512827 -2.79%
> Top 5 regressions:
>   openjdk/ad_x86_gen.ll 18251379251 20870783067 +14.35%
>   faiss/IndexIVFAdditiveQuantizerFastScan.cpp.ll 825075373 857500006 +3.93%
>   brotli/encoder_dict.c.ll 543355291 560137438 +3.09%
>   gromacs/constr.cpp.ll 2099218140 2163924992 +3.08%
>   lightgbm/dataset_loader.cpp.ll 7317833652 7536019648 +2.98%
> Overall: -0.02132359%
> ```

I don't see how this change can cause any significant compile-time changes (unless it's due to the bug in foldSelectWithBinaryOp). Can you please check what kind of results you get if you run the compile-time test against a no-op change to see what the noise level looks like?

https://github.com/llvm/llvm-project/pull/100878