[llvm] [CodeGen] Add an option to skip extTSP BB placement for huge functions. (PR #99310)

[Krzysztof Pszeniczny via llvm-commits]

amharc wrote:

> Is it possible by any chance to play with the data leading to excessive runtimes? We could try to make the implementation faster instead of disabling it

I'll try to come up with something, but unfortunately huge IR is often nontrivial to strip of confidential information.

Usually such huge functions, even if they have a surprisingly large profile coverage, don't account for a huge amount of cores from the whole datacenter perspective. So it's actually likely that spending time on investigating how to apply ext-TSP for such abnormally large functions with a surprising number of hot chains might not be worth the savings.

That said, transformations that are significantly super-linear (e.g. quadratic) will almost certainly run into issues when running on huge functions. Adding a cut-off flag to avoid significantly increasing the compilation time seems like a reasonable action to me: this PR doesn't modify the default behaviour of the compiler (note: the default value of this flag effectively means "no limit"), it just allows imposing a cut-off.

https://github.com/llvm/llvm-project/pull/99310