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

[via llvm-commits]

goldsteinn wrote:

> I'm not sure that smaller but more functions necessarily means more compile time, larger functions tend to exasperate superlinear compile times.
> 
> Can you post a llvm-compile-time-tracker link (even though it won't necessarily show the second order effects since you tend to need FDO profiles for mark functions as cold)?

The original impact (of this exact patch):
https://llvm-compile-time-tracker.com/compare.php?from=ffcbfa46b9926bcb42dd7b1f778ada113672595c&to=85efc82366a64fbe5287b7bc9d20853b06d147f6&stat=instructions:u

A test showing the impact as second order:
https://llvm-compile-time-tracker.com/compare.php?from=eb0d0125306e201fbff49c1b00a059c95757f595&to=bb96b08e7c97548bc1242f98b18bf07ba42be1d8&stat=instructions:u

The test basically replaces the setting of `cold` to setting of `nosanitize_bounds` in the first commit but results in the same code executing in `FunctionAttrs`. As you can see in the second commit, when all I do is change the set attr to `cold` the compile time explodes.

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