[llvm] [SelectOpt] Optimise big select groups in the latch of a non-inner loop to branches (PR #119728)

[David Green via llvm-commits]

https://github.com/davemgreen approved this pull request.

> You can look at the test. SelectGroup consists of 6 instructions at the beginning of the latch, and the latch has 12 instructions in total. It doesn't look like a small latch to me.

1/2 of the instructions sounds like quite a high proportion. I was thinking about larger latches (hundreds of instructions) that are essentially just like any other blocks and not particularly special for being latches.

The tail duplication I was thinking of isn't triggering in the test file though, if that is representative https://godbolt.org/z/jT68zsevE (even with enough options to prevent the ifcvting). It might be nice if it did. I guess that there are paths through the program that can lead to csel stalls where they are avoided with branches, and this might be one of them even though it is not part of an inner loop. In that way latches are special and we have certainly seen similar issues in other (inner) loops before. Trying to calculate all the possible paths sounds too expensive.

In practice I think this sounds OK to me (even if like you say it is a bit narrow) - in that it will likely not do much harm anywhere and should lead to perf improvements in the cases you have seen, and I like the limit of at least 3 instructions in the select group. Big groups will be generally more profitable to convert.

I think this sounds OK to me.

https://github.com/llvm/llvm-project/pull/119728