[llvm] [mlir] Add usub_cond and usub_sat operations to atomicrmw (PR #105568)

[Matt Arsenault via llvm-commits]

================
@@ -41,6 +41,8 @@ bool isIdempotentRMW(AtomicRMWInst& RMWI) {
     case AtomicRMWInst::Sub:
     case AtomicRMWInst::Or:
     case AtomicRMWInst::Xor:
+    case AtomicRMWInst::USubCond:
+    case AtomicRMWInst::USubSat:
----------------
arsenm wrote:

The atomicrmw -> atomicrmw idempotent case handled here is fine, it's the atomicrmw -> load atomic case that's problematic in the lowering.

Also in a follow up, the AMDGPU lowering should have the idempotent case to add 0 since that works over PCIe

https://github.com/llvm/llvm-project/pull/105568