[llvm] [SLP]Represent externally used values as original scalars, if profitable. (PR #100904)

[via llvm-commits]

dgg5503 wrote:

Hi @alexey-bataev ,

> Hi, I rather doubt it is a good idea to push so many commits to fix a single regression. Better to ask release maintainers if it worth doing for the 19x release or better to wait for 20

That is understandable, thank you for your timely response & suggestion. I will ping the current release manager here which, according to https://llvm.org/docs/HowToReleaseLLVM.html#official-testing, should be Tobias Hieta for LLVM 19.

Hi @tru ,

Per Alexey's response above, may you please provide your opinion on whether or not it would be worth cherry-picking in the commits I've listed to LLVM 19 to resolve a performance regression with particular SPEC CPU tests? We at Sony Interactive Entertainment understand there are quite a few users of the SPEC benchmark who may also notice this regression in their internal testing, however, from our point-of-view, we are by no means pushing for this, we just think it would be overall beneficial.

Please let me know what you think.

Thank you,
Douglas Gliner

https://github.com/llvm/llvm-project/pull/100904