[llvm] RFC: Argpromotion of externally visible functions (PR #87731)

[Nikita Popov via llvm-commits]

nikic wrote:

I like the general idea, but I'm also somewhat concerned about this being too aggressive in promoting functions that have external callers. As I understand it, the cases where you would want this optimization to apply are cases where a function *could* be internalized (and would be internalized in full LTO), but isn't internalized due to use in multiple ThinLTO modules.

I guess if the performance numbers say this is fine in practice, then we can just do it and see what falls out, though I'd definitely want to see numbers for Rust in that case (which, I think, is even heavier on argpromo than C++).

I have a vague memory of someone doing work on argument promotion integrated in the thin link before -- possibly it was @pcwalton?

https://github.com/llvm/llvm-project/pull/87731