[llvm] [LLVM][TableGen] Check overloaded intrinsic mangling suffix conflicts (PR #110324)

[Matt Arsenault via llvm-commits]

arsenm wrote:

> > I did actually run into this oddity when removing some of the amdgpu atomic intrinsics
> 
> Could you please elaborate on this? I don't really understand what kind of issue you can hit when _removing_ intrinsics.

In 9d364286f3b63e99ed3838f179aa2223f930f1ab, I removed the int_amdgcn_flat_atomic_fadd_v2bf16 and int_amdgcn_global_atomic_fadd_v2bf16 intrinsics. These should never have been added. We already had int_amdgcn_global_atomic_fadd and int_amdgcn_flat_atomic_fadd overloaded on the type, so we had a collision between the overloaded int_amdgcn_global_atomic_fadd_v2bf16 and int_amdgcn_global_atomic_fadd with v2bf16. 

I don't remember the details, but something confusing was going on. I'm not sure if the int_amdgcn_global_atomic_fadd_v2bf16 case was ever really parsed. You could create the intrinsic with the suffixed ID, but then it would be recognized as the mangled version, or something like that.

https://github.com/llvm/llvm-project/pull/110324