[llvm] [X86][AMX-AVX512] Remove patterns for immediate variants (PR #173545)

[Phoebe Wang via llvm-commits]

phoebewang wrote:

> > > Ideally, they should have different patterns with different intrinsics.
> > 
> > 
> > The difference between immediate and register variants is confusing. I believe it's not only us make such mistake, so I think it's better to not provide them. User who are clear with the differences can use inline assemble instead.
> 
> No objections to dropping this for now, ideally we'd provide some peephole to allow correct use of imm variants in the future, but better to be safe initially than digging ourselves into a hole. Are there C/C++ intrinsics for both?

AFAIK, we don't have precedent that provide 2 intrinsics for both reg and imm variants, mainly because there's no such exception before. GCC has provided 2 intrinsics, but I think we can weaken it by not providing the imm intrinsics definition in Intrinsic Guide.

https://github.com/llvm/llvm-project/pull/173545