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

[Nikita Popov via llvm-commits]

nikic wrote:

> > Thanks, so the actual results are these: https://llvm-compile-time-tracker.com/compare.php?from=df5252165524fbcabbd7287a8d923b5139247114&to=167f6301e6477af8a0c0db1ecd5bd3e655949f00&stat=instructions:u
> 
> @nikic I can see that the worst score is about +0.7 % and one has improvement by -0.9 %. I would personally think that these are okay results regarding compile time impact, but I have not really worked with these benchmarks and can't quantify them well enough. What is your experienced opinion on this?

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%.

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