[clang] [llvm] [AMDGPU] Change CF intrinsics lowering to reconverge on predecessors (PR #108596)

[via llvm-commits]

alex-t wrote:

> Thank you for going into a bit more detail. It does seem like distinguishing between SGPR reloads and VGPR reloads would help.
> 
> Hmm, what about a case where we run out of VGPRs to spill SGPRs into? What do we actually do in that case today?

Not sure if I understand your concern. We allocate SGPRs in a separate pass before VGPRs. So, we have the whole amount of VGPRs available on a given target to spill SGPRs. I hardly can imagine a kernel that requires 256x32 SGPRs.

https://github.com/llvm/llvm-project/pull/108596