[llvm] [AMDGPU] Generalize scratch SGPR CFI emission (PR #217066)

[Emma Pilkington via llvm-commits]

epilk wrote:

> > It's possible to spill to vcc_hi in wave32, however that isn't representable in dwarf since 64-bit vcc is only usable in wave64.
> 
> This isn't true? It's a usable 64-bit SGPR tuple like any other

I mean usable in dwarf, see: https://github.com/ROCm/ROCdbgapi/blob/ccb7f486ebcc284e7cd3cb29eba71b39ff2232e7/src/architecture.cpp#L2304 . That being said, we should probably just come up with a dwarf design that can actually represent vcc_hi on wave32. (@lancesix: WDYT?)

https://github.com/llvm/llvm-project/pull/217066