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

[Jon Chesterfield via llvm-commits]

JonChesterfield wrote:

Just to say I *love* this premise. More checks in the verifier is a great way to narrow down which pass has created broken code.

The alloca-must-be-const looks slightly optimistic to me - is that an invariant clang sema enforces for c++ input? I can see how we'd struggle to lower it, just not sure whether it's illegal in all IR or only in some of it. Target verifier should probably be invariants that hold between every pass.

Main thing is it would be really nice for this to be a hook that the usual verifier calls into as opposed to a separate pass, as verify is already in the debugging flow for lots of people. Means these checks can be easily inserted between every IR pass etc without altering infrastructure elsewhere.

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