[clang] [llvm] [TargetVerifier][AMDGPU] Add TargetVerifier. (PR #123609)

[Eli Friedman via llvm-commits]

efriedma-quic wrote:

> This would be run during the backend after we've already selected the target, so it can be target-dependent.

All IR is "target-dependent" in the sense that it depends on the datalayout.  But it's target-independent in the sense that the validity rules don't depend on the triple.  That separation of concerns is what allows people to write IR optimizations without worrying about specific targets: the rules are the same everywhere.  You break that when you forbid the use of `i24`, or restrict what control flow is allowed after calls to specific intrinsics.

https://github.com/llvm/llvm-project/pull/123609