[llvm] [RISCV] Restrict when we fold an ADD_LO into a load/store address. (PR #93129)

[Craig Topper via llvm-commits]

topperc wrote:

> I think I'm missing something here.  This doesn't seem like a profitable transform, and I'm not understanding why it is required for the pseudo.  I get that we'd have two values live through regalloc (the lui and the lui/addi pseudo), but is the impact really that bad?  That seems analogous to what we have now?

We end up with 2 LUIs after the pseudo gets expanded. They won't get CSEd since we are post regalloc at that point. I'll post a version of the patch without this and we can decide.

https://github.com/llvm/llvm-project/pull/93129