[AMDGPU][GlobalISel] Eliminate barrier if workgroup size is not greater than wavefront size

[foad]

Summary

If a workgroup size is known to be not greater than wavefront size
the s_barrier instruction is not needed since all threads are guaranteed
to come to the same point at the same time.

This is the same optimization that was implemented for SelectionDAG in
D31731.

[arsenm]

This could be a pattern predicate but I guess it's not important

llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp

1426

There's no real reason to check the opt level but I guess it does this already

llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.h

private:
   bool selectGroupStaticSize(MachineInstr &I) const;
   bool selectReturnAddress(MachineInstr &I) const;
   bool selectG_INTRINSIC(MachineInstr &I) const;
 
   bool selectEndCfIntrinsic(MachineInstr &MI) const;
   bool selectDSOrderedIntrinsic(MachineInstr &MI, Intrinsic::ID IID) const;
   bool selectDSGWSIntrinsic(MachineInstr &MI, Intrinsic::ID IID) const;
   bool selectDSAppendConsume(MachineInstr &MI, bool IsAppend) const;
+  bool selectSBarrier(MachineInstr &MI) const;
 
   bool selectImageIntrinsic(MachineInstr &MI,
                             const AMDGPU::ImageDimIntrinsicInfo *Intr) const;
   bool selectG_INTRINSIC_W_SIDE_EFFECTS(MachineInstr &I) const;
   int getS_CMPOpcode(CmpInst::Predicate P, unsigned Size) const;
   bool selectG_ICMP(MachineInstr &I) const;
   bool hasVgprParts(ArrayRef<GEPInfo> AddrInfo) const;
   void getAddrModeInfo(const MachineInstr &Load, const MachineRegisterInfo &MRI,

llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp

bool AMDGPUInstructionSelector::selectDSAppendConsume(MachineInstr &MI,
   auto MIB = BuildMI(*MBB, &MI, DL, TII.get(Opc), MI.getOperand(0).getReg())
     .addImm(Offset)
     .addImm(IsGDS ? -1 : 0)
     .cloneMemRefs(MI);
   MI.eraseFromParent();
   return constrainSelectedInstRegOperands(*MIB, TII, TRI, RBI);
 }
 
+bool AMDGPUInstructionSelector::selectSBarrier(MachineInstr &MI) const {
+  if (TM.getOptLevel() > CodeGenOpt::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));
+      MI.eraseFromParent();
+      return true;
+    }
+  }
+  return selectImpl(MI, *CoverageInfo);
+}
+
 static bool parseTexFail(uint64_t TexFailCtrl, bool &TFE, bool &LWE,
                          bool &IsTexFail) {
   if (TexFailCtrl)
     IsTexFail = true;
 
   TFE = (TexFailCtrl & 0x1) ? 1 : 0;
   TexFailCtrl &= ~(uint64_t)0x1;
   LWE = (TexFailCtrl & 0x2) ? 1 : 0;
bool AMDGPUInstructionSelector::selectG_INTRINSIC_W_SIDE_EFFECTS(
   case Intrinsic::amdgcn_ds_gws_sema_br:
   case Intrinsic::amdgcn_ds_gws_sema_p:
   case Intrinsic::amdgcn_ds_gws_sema_release_all:
     return selectDSGWSIntrinsic(I, IntrinsicID);
   case Intrinsic::amdgcn_ds_append:
     return selectDSAppendConsume(I, true);
   case Intrinsic::amdgcn_ds_consume:
     return selectDSAppendConsume(I, false);
+  case Intrinsic::amdgcn_s_barrier:
+    return selectSBarrier(I);
   default: {
     return selectImpl(I, *CoverageInfo);
   }
   }
 }
 
 bool AMDGPUInstructionSelector::selectG_SELECT(MachineInstr &I) const {
   if (selectImpl(I, *CoverageInfo))

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

 ; RUN: llc -march=amdgcn < %s | FileCheck %s
+; RUN: llc -march=amdgcn < %s -global-isel | FileCheck %s
 
 ; CHECK-LABEL: {{^}}unknown_wgs:
 ; CHECK: s_barrier
 define amdgpu_kernel void @unknown_wgs() {
   tail call void @llvm.amdgcn.s.barrier() #0
   ret void
 }