[llvm] [AMDGPU] Make maximum hard clause size a subtarget feature (PR #81287)

[Jay Foad via llvm-commits]

jayfoad wrote:

> I was also unsure if, for example, we'd end up with future hardware that allows for a clause length of 64 (instead of the old 63).

Sure, that could happen, but there is no reason that those limits have to be encoded as subtarget features. They can just be hard-coded into the compiler, e.g. in GCNSubtarget::maxHardClauseLength. Even limiting the clause length to 32 does not _have_ to be a subtarget feature, but we seem to have a convention that workarounds for hardware bugs do get subtarget features.

https://github.com/llvm/llvm-project/pull/81287