[llvm] [AMDGPU] Re-enable atomic optimization of uniform fadd/fsub with result (PR #97604)

[Matt Arsenault via llvm-commits]

================
@@ -937,18 +930,25 @@ void AMDGPUAtomicOptimizerImpl::optimizeAtomic(Instruction &I,
         break;
       case AtomicRMWInst::FAdd:
       case AtomicRMWInst::FSub: {
-        // FIXME: This path is currently disabled in visitAtomicRMWInst because
-        // of problems calculating the first active lane of the result (where
-        // Mbcnt is 0):
-        // - If V is infinity or NaN we will return NaN instead of BroadcastI.
-        // - If BroadcastI is -0.0 and V is positive we will return +0.0 instead
-        //   of -0.0.
         LaneOffset = B.CreateFMul(V, Mbcnt);
         break;
       }
       }
     }
-    Value *const Result = buildNonAtomicBinOp(B, Op, BroadcastI, LaneOffset);
+    Value *Result = buildNonAtomicBinOp(B, Op, BroadcastI, LaneOffset);
+    if (isAtomicFloatingPointTy) {
----------------
arsenm wrote:

Can you skip this if the function has the no-signed-zeros attribute? 

https://github.com/llvm/llvm-project/pull/97604