[llvm] [AMDGPU] Multi dword spilling for unaligned tuples (PR #183701)

[via llvm-commits]

================

----------------
easyonaadit wrote:

> Isn't only the first piece a special case?

Not exactly. 
Suppose its `SPILL $v1_v2_v3`, after spilling `v1`, `EltSize` needs to be recomputed for `v2_v3`.
Similarly, for `SPILL $v1_v2_v3_v4_v5_v6` translates to:
1. spill `v1`.
2. recompute `EltSize` for `v2_v3_v4_v5` tuple.
3. recompute `EltSize` for `v6` (LastChunk).

`LastChunk` for unaligned tuples is "somewhat" analogous to `NumRemSubRegs` for aligned tuples.


https://github.com/llvm/llvm-project/pull/183701