[clang] [HIP] diagnose -mwavefrontsize64 for gfx10+ (PR #140185)

[via cfe-commits]

b-sumner wrote:

> > > > I don't think we should be introducing an additional option that behaves identically to another option just differing by a warning. This also isn't really a problem with the language, but the library support. The actual codegen should work just fine. Is there some other existing mechanism to just silence the warning?
> > > 
> > > 
> > > We could emit a warning for -mwavefrontsize64 for HIP on gfx10+, which can be disabled by -Wno-unsupported-wave64. Then we do not need introduce -mforce-unsafe-wavefrontsize64.
> > > @b-sumner are you OK with this?
> > 
> > 
> > I am not. I think we need to emit an error for -mwavefrontsize64 for HIP on gfx10+. If we do this as if -Werror=unsafe-wavefrontsize64 is the default, then users can use -Wno-error=unsafe-wavefrontsize64 I suppose. This seems like a weird use of -Wno-error though.
> 
> We can make the warning an error by default, and users can use -Wno-error=unsafe-wavefrontsize64 to disable it. There is precedence of this kind of use case. Is that OK?

That sounds good to me!

https://github.com/llvm/llvm-project/pull/140185