[clang] [llvm] [AMDGPU][clang][CodeGen][opt] Add late-resolved feature identifying predicates (PR #134016)

[Aaron Ballman via llvm-commits]

AaronBallman wrote:

> > IMO, we are better served by a warning diagnostic if we detect these are misused. `ParseCXXCondition` (or the C equivalent, but since you are returning bool it seems you're not concerned about C?) might be a good place to set a variable to enable the warning.
> 
> This is a very good suggestion, thank you very much for it - it might well be where we end up. My worry is that ignoring warning and diagnostics is rather common. 

You could default the warning to an error to make it more visible to the user. Not certain if that's a bad idea or not though.

> If I may be so bold as to inquire: would you and @AaronBallman be slightly less horrified if the return type variance would be replaced with returning an odd type that only knows how to `bool`ify itself in conditions? More explicitly, if instead of `void __builtin_amdgcn_processor_is(const char*)` what we see is `__amdgpu_predicate_t __builtin_amdgcn_processor_is(const char*)`, would that be somewhat less bad? There is precedent for special-ish builtins returning special-ish target types (please consider `__amdgpu_buffer_rsrc_t` for `__builtin_amdgcn_make_buffer_rsrc` or `svcount_t`)

That would be just as weird, IMO. Having something that's contextually converted to bool but only in some contexts, is going to be confusing in practice.

https://github.com/llvm/llvm-project/pull/134016