[llvm] [SelectionDAG] Fold (icmp eq/ne (shift X, C), 0) -> (icmp eq/ne X, 0) (PR #88801)

[Björn Pettersson via llvm-commits]

================
@@ -586,6 +586,7 @@ define i1 @add_ultcmp_i32_i16(i32 %x) nounwind {
 ; RV64I-NEXT:    lui a1, 8
 ; RV64I-NEXT:    add a0, a0, a1
 ; RV64I-NEXT:    srliw a0, a0, 16
+; RV64I-NEXT:    slli a0, a0, 16
----------------
bjope wrote:

I don't think shouldFoldConstantShiftPairToMask has anything to do with it here, since I can't see any shift pair being created in the first place.

We could add a new hook to avoid the simplification done here for RISCV. But I really think that this is something that RISCV should handle at selection (i.e. allowing the simplified DAG, but doing a better selection).
No idea if the regression here is acceptable until someone that knows more about RISCV would deal with that. I don't feel comfortable doing that (and I do not really have time to work on that as it isn't important for my downstream target).


https://github.com/llvm/llvm-project/pull/88801