[llvm] [AMDGPU] Prevent folding of the negative i32 literals as i64 (PR #70274)

[Matt Arsenault via llvm-commits]

arsenm wrote:

> In a long term I think we need to replace INT64 operands with SINT64 and UINT64. That said there are no MVT::s64 and MVT::u64 to use in an instruction profile.

You would have to bundle that somehow in the profile. 

https://github.com/llvm/llvm-project/pull/70274