[llvm] [RISCV] Use Zilsd Pseudos in ISel (PR #169580)

[Sam Elliott via llvm-commits]

lenary wrote:

> Do we need to prevent isel from setting the immediate larger than 2043 so we can add 4 to it if it gets split?

I don't know.

I have punted on this so far - including for mergebaseoffset - but that might mean there are correctness issues? I haven't verified whether there actually are or not though.

I will do some correctness runs for the embench-iot benchmarks I looked at to check we're not introducing miscompiles but I don't think we are?

https://github.com/llvm/llvm-project/pull/169580