[llvm] [AMDGPU] Optimize out s_barrier_signal/_wait (PR #116993)

[Carl Ritson via llvm-commits]

================
@@ -9605,7 +9605,9 @@ SDValue SITargetLowering::LowerINTRINSIC_VOID(SDValue Op,
     unsigned Opc = Done->isZero() ? AMDGPU::EXP : AMDGPU::EXP_DONE;
     return SDValue(DAG.getMachineNode(Opc, DL, Op->getVTList(), Ops), 0);
   }
-  case Intrinsic::amdgcn_s_barrier: {
+  case Intrinsic::amdgcn_s_barrier:
+  case Intrinsic::amdgcn_s_barrier_signal:
----------------
perlfu wrote:

This is lowering both the signal and the wait to `WAVE_BARRIER`, but I wonder if really the signal should just be removed entirely?

https://github.com/llvm/llvm-project/pull/116993