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

Extend the optimization that converts s_barrier to wave_barrier (nop)
when the number of work items is not larger than wave size.
    
This handles the "split barrier" form of s_barrier where the barrier
is represented by separate intrinsics (s_barrier_signal/s_barrier_wait).
Note: the version where s_barrier is used in gfx12 (and later split)
has the optimization already, but some front-ends may prefer to use
split intrinsics and this is being addressed by the patch.

Author: piotrAMD

llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp

@@ -1878,19 +1878,25 @@ 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;
     }
   }
 
-  // On GFX12 lower s_barrier into s_barrier_signal_imm and s_barrier_wait
-  if (STI.hasSplitBarriers()) {
+  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();
     const DebugLoc &DL = MI.getDebugLoc();
     BuildMI(*MBB, &MI, DL, TII.get(AMDGPU::S_BARRIER_SIGNAL_IMM))
@@ -2207,6 +2213,8 @@ bool AMDGPUInstructionSelector::selectG_INTRINSIC_W_SIDE_EFFECTS(
   case Intrinsic::amdgcn_init_whole_wave:
     return selectInitWholeWave(I);
   case Intrinsic::amdgcn_s_barrier:
+  case Intrinsic::amdgcn_s_barrier_signal:
+  case Intrinsic::amdgcn_s_barrier_wait:
     return selectSBarrier(I);
   case Intrinsic::amdgcn_raw_buffer_load_lds:
   case Intrinsic::amdgcn_raw_ptr_buffer_load_lds:

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

@@ -9614,18 +9614,26 @@ 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:
+  case Intrinsic::amdgcn_s_barrier_wait: {
     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);
+      }
     }
 
-    // On GFX12 lower s_barrier into s_barrier_signal_imm and s_barrier_wait
-    if (ST.hasSplitBarriers()) {
+    if (ST.hasSplitBarriers() && IntrinsicID == Intrinsic::amdgcn_s_barrier) {
+      // On GFX12 lower s_barrier into s_barrier_signal_imm and s_barrier_wait
       SDValue K =
           DAG.getSignedTargetConstant(AMDGPU::Barrier::WORKGROUP, DL, MVT::i32);
       SDValue BarSignal =

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

@@ -0,0 +1,64 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx1200 < %s | FileCheck %s
+; RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx1200 < %s | FileCheck %s
+
+define amdgpu_kernel void @signal_unknown_wgs() {
+; CHECK-LABEL: signal_unknown_wgs:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_barrier_signal -1
+; CHECK-NEXT:    s_endpgm
+  tail call void @llvm.amdgcn.s.barrier.signal(i32 -1)
+  ret void
+}
+
+define amdgpu_kernel void @signal_flat_wgs_attr_32_128() #1 {
+; CHECK-LABEL: signal_flat_wgs_attr_32_128:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_barrier_signal -1
+; CHECK-NEXT:    s_endpgm
+  tail call void @llvm.amdgcn.s.barrier.signal(i32 -1)
+  ret void
+}
+
+define amdgpu_kernel void @signal_flat_wgs_attr_16_32() #2 {
+; CHECK-LABEL: signal_flat_wgs_attr_16_32:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_endpgm
+  tail call void @llvm.amdgcn.s.barrier.signal(i32 -1)
+  ret void
+}
+
+
+define amdgpu_kernel void @wait_unknown_wgs() {
+; CHECK-LABEL: wait_unknown_wgs:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_barrier_wait -1
+; CHECK-NEXT:    s_endpgm
+  tail call void @llvm.amdgcn.s.barrier.wait(i16 -1)
+  ret void
+}
+
+define amdgpu_kernel void @wait_flat_wgs_attr_32_128() #1 {
+; CHECK-LABEL: wait_flat_wgs_attr_32_128:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_barrier_wait -1
+; CHECK-NEXT:    s_endpgm
+  tail call void @llvm.amdgcn.s.barrier.wait(i16 -1)
+  ret void
+}
+
+define amdgpu_kernel void @wait_flat_wgs_attr_16_32() #2 {
+; CHECK-LABEL: wait_flat_wgs_attr_16_32:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    ; wave barrier
+; CHECK-NEXT:    s_endpgm
+  tail call void @llvm.amdgcn.s.barrier.wait(i16 -1)
+  ret void
+}
+
+declare void @llvm.amdgcn.s.barrier.signal(i32 immarg) #0
+declare void @llvm.amdgcn.s.barrier.wait(i16 immarg) #0
+
+attributes #0 = { convergent nounwind }
+attributes #1 = { nounwind "amdgpu-flat-work-group-size"="32,128" }
+attributes #2 = { nounwind "amdgpu-flat-work-group-size"="16,32" }