[llvm] [RISCV] Recursively split concat_vector into smaller LMULs (PR #83035)

[Philip Reames via llvm-commits]

https://github.com/preames approved this pull request.

LGTM

Just because it doesn't seem to be reflected otherwise in the review clearly... I talked with Luke about this one offline on Monday and pushed strongly for a restriction on the minimum lmul we'd split.  The concern was to avoid increasing register pressure in an unbounded manner.  The exact threshold value isn't particular important.  For the corresponding build_vector lowering change, we also chose mf2 with the intention to (someday) explore that tradeoff point more carefully.  

I'm a bit hesitant about this being a combine instead of a lowering.  In particular, I'm concerned about the split form inhibiting other combines - such as e.g. the load combine.  However, I'm choosing to approve for now, we can revisit if this turns out to be an actual problem.  Moving this into lowering shouldn't be too hard.  

https://github.com/llvm/llvm-project/pull/83035