Do not generate wave_barrier for s_barrier_signal

Author: piotrAMD

llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp

@@ -1832,18 +1832,23 @@ bool AMDGPUInstructionSelector::selectInitWholeWave(MachineInstr &MI) const {
 }
 
 bool AMDGPUInstructionSelector::selectSBarrier(MachineInstr &MI) const {
+  Intrinsic::ID IntrinsicID = cast<GIntrinsic>(MI).getIntrinsicID();
   if (TM.getOptLevel() > CodeGenOptLevel::None) {
     unsigned WGSize = STI.getFlatWorkGroupSizes(MF->getFunction()).second;
     if (WGSize <= STI.getWavefrontSize()) {
-      MachineBasicBlock *MBB = MI.getParent();
-      const DebugLoc &DL = MI.getDebugLoc();
-      BuildMI(*MBB, &MI, DL, TII.get(AMDGPU::WAVE_BARRIER));
+      // If the workgroup fits in a wave, remove s_barrier_signal and lower
+      // s_barrier/s_barrier_wait to wave_barrier.
+      if (IntrinsicID == Intrinsic::amdgcn_s_barrier ||
+          IntrinsicID == Intrinsic::amdgcn_s_barrier_wait) {
+        MachineBasicBlock *MBB = MI.getParent();
+        const DebugLoc &DL = MI.getDebugLoc();
+        BuildMI(*MBB, &MI, DL, TII.get(AMDGPU::WAVE_BARRIER));
+      }
       MI.eraseFromParent();
       return true;
     }
   }
 
-  Intrinsic::ID IntrinsicID = cast<GIntrinsic>(MI).getIntrinsicID();
   if (STI.hasSplitBarriers() && IntrinsicID == Intrinsic::amdgcn_s_barrier) {
     // On GFX12 lower s_barrier into s_barrier_signal_imm and s_barrier_wait
     MachineBasicBlock *MBB = MI.getParent();

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

@@ -9611,10 +9611,16 @@ SDValue SITargetLowering::LowerINTRINSIC_VOID(SDValue Op,
     const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
     if (getTargetMachine().getOptLevel() > CodeGenOptLevel::None) {
       unsigned WGSize = ST.getFlatWorkGroupSizes(MF.getFunction()).second;
-      if (WGSize <= ST.getWavefrontSize())
-        return SDValue(DAG.getMachineNode(AMDGPU::WAVE_BARRIER, DL, MVT::Other,
-                                          Op.getOperand(0)),
-                       0);
+      if (WGSize <= ST.getWavefrontSize()) {
+        // If the workgroup fits in a wave, remove s_barrier_signal and lower
+        // s_barrier/s_barrier_wait to wave_barrier.
+        if (IntrinsicID == Intrinsic::amdgcn_s_barrier_signal)
+          return Op.getOperand(0);
+        else
+          return SDValue(DAG.getMachineNode(AMDGPU::WAVE_BARRIER, DL,
+                                            MVT::Other, Op.getOperand(0)),
+                         0);
+      }
     }
 
     if (ST.hasSplitBarriers() && IntrinsicID == Intrinsic::amdgcn_s_barrier) {

llvm/test/CodeGen/AMDGPU/barrier-elimination-gfx12.ll

@@ -17,7 +17,6 @@ define amdgpu_kernel void @signal_flat_wgs_attr_32_128() #1 {
 
 ; CHECK-LABEL: {{^}}signal_flat_wgs_attr_16_32:
 ; CHECK: :
-; CHECK-NEXT: ; wave barrier
 ; CHECK-NEXT: s_endpgm
 define amdgpu_kernel void @signal_flat_wgs_attr_16_32() #2 {
   tail call void @llvm.amdgcn.s.barrier.signal(i32 -1)