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

[Nikita Popov via llvm-commits]

nikic wrote:

Okay, so it looks like the top 5 improvements/regressions are very noisy. I did wonder why I wasn't able to reproduce all regressions on previous PRs.

To mostly avoid this problem I collect per-benchmark stddev data and only consider differences starting at 3-4 sigma significant.

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