[llvm] [FunctionAttrs] deduce attr `cold` on functions if all CG paths call a `cold` function (PR #101298)

[Nikita Popov via llvm-commits]

nikic wrote:

> Is this potentially acceptable?

Generally yes, second-order regressions are fine. Of course, it depends on the details. How does the presence of cold attributes affect things -- is this "just" because of substantially different inlining decisions, or something else?

https://github.com/llvm/llvm-project/pull/101298