[clang-tools-extra] [clang] [llvm] Dont alter cold function alignment unless using Os (PR #72387)

[Paul T Robinson via cfe-commits]

pogo59 wrote:

I guess I'm a little confused still. Are you claiming that `cold` has a specified ABI effect? That if some pass decides a function is cold, that decision (correctly) affects its ABI? I'm not seeing any documentation to that effect, either on the clang attribute or the IR attribute. My admittedly aging copy of the psABI doesn't mention function alignment at all, either.

We're not inventing an ABI rule that `cold` functions are special and must have a specific alignment; we just want `cold` functions to have the same alignment as non-`cold` functions. (And if you specify `-Os` then that alignment changes to 1, for cold and non-cold functions alike.)

The only thing that makes `cold` functions special _in this implementation_ is that, as an implementation convenience, clang's `cold` attribute also sets `optsize` which has an effect that we are finding inconvenient. We just want `cold` not to have that particular effect. We played around with several different ways to make that happen, and this seemed the least intrusive.

I suppose another way to go at this would be for `-Os` or clang's `minsize` to add both `optsize` and align(1) to functions, and have LLVM's implementation of `optsize` stop implying align(1). This would make everything more orthogonal and disentangled, yes? And then clang's `cold` could set `optsize` without affecting alignment; if non-PS targets still wanted `cold` to mean align(1), we can set that as well. No new attributes, and that bit of code down in LLVM to compute alignment gets a wee bit cleaner. WDYT?

https://github.com/llvm/llvm-project/pull/72387