[llvm] Deduplication of cyclic PHI nodes (PR #86662)

[Marek Sedláček via llvm-commits]

mark-sed wrote:

> It looks like there are some very large regressions in there, e.g. libclamav_nsis_LZMADecode.c by 8% or libclamav_nsis_bzlib.c by 4%.

@nikic I have halved the phi cutoff depth, looking at the new benchmarks, it looks better to me. Some tests are even performing better (which might be because of speeding up some other optimizations). So hopefully you'll be okay with these new results.

https://github.com/llvm/llvm-project/pull/86662