[llvm] [AMDGPU] Allow rematerialization of instructions with virtual register uses (PR #124327)

[Jeffrey Byrnes via llvm-commits]

jrbyrnes wrote:

In the latest I disallowed rematerializing an instruction if it has a dependency on a planned remat.

The main challenge with such remats is doing the actual rematerialization in correct order. Since we walk over the region to calculate RP after each remat, we may end up temporariliy producing illegal code.

After https://github.com/llvm/llvm-project/pull/118722 , we can remat in batches before doing RP calculations, so this will not be a problem.

https://github.com/llvm/llvm-project/pull/124327