1 files changed, 124 insertions, 0 deletions

diff --git a/llvm/lib/Target/AMDGPU/SIOptimizeExecMaskingPreRA.cpp b/llvm/lib/Target/AMDGPU/SIOptimizeExecMaskingPreRA.cpp
index e53d876f232..a61e6d72eb9 100644
--- a/llvm/lib/Target/AMDGPU/SIOptimizeExecMaskingPreRA.cpp
+++ b/llvm/lib/Target/AMDGPU/SIOptimizeExecMaskingPreRA.cpp
@@ -103,6 +103,122 @@ static MachineInstr* getOrExecSource(const MachineInstr &MI,
   return SaveExecInst;
 }
 
+// Optimize sequence
+//    %sel = V_CNDMASK_B32_e64 0, 1, %cc
+//    %cmp = V_CMP_NE_U32 1, %1
+//    $vcc = S_AND_B64 $exec, %cmp
+//    S_CBRANCH_VCC[N]Z
+// =>
+//    $vcc = S_ANDN2_B64 $exec, %cc
+//    S_CBRANCH_VCC[N]Z
+//
+// It is the negation pattern inserted by DAGCombiner::visitBRCOND() in the
+// rebuildSetCC(). We start with S_CBRANCH to avoid exhaustive search, but
+// only 3 first instructions are really needed. S_AND_B64 with exec is a
+// required part of the pattern since V_CNDMASK_B32 writes zeroes for inactive
+// lanes.
+//
+// Returns %cc register on success.
+static unsigned optimizeVcndVcmpPair(MachineBasicBlock &MBB,
+                                     const GCNSubtarget &ST,
+                                     MachineRegisterInfo &MRI,
+                                     LiveIntervals *LIS) {
+  const SIRegisterInfo *TRI = ST.getRegisterInfo();
+  const SIInstrInfo *TII = ST.getInstrInfo();
+  const unsigned AndOpc = AMDGPU::S_AND_B64;
+  const unsigned Andn2Opc = AMDGPU::S_ANDN2_B64;
+  const unsigned CondReg = AMDGPU::VCC;
+  const unsigned ExecReg = AMDGPU::EXEC;
+
+  auto I = llvm::find_if(MBB.terminators(), [](const MachineInstr &MI) {
+                           unsigned Opc = MI.getOpcode();
+                           return Opc == AMDGPU::S_CBRANCH_VCCZ ||
+                                  Opc == AMDGPU::S_CBRANCH_VCCNZ; });
+  if (I == MBB.terminators().end())
+    return AMDGPU::NoRegister;
+
+  auto *And = TRI->findReachingDef(CondReg, AMDGPU::NoSubRegister,
+                                   *I, MRI, LIS);
+  if (!And || And->getOpcode() != AndOpc ||
+      !And->getOperand(1).isReg() || !And->getOperand(2).isReg())
+    return AMDGPU::NoRegister;
+
+  MachineOperand *AndCC = &And->getOperand(1);
+  unsigned CmpReg = AndCC->getReg();
+  unsigned CmpSubReg = AndCC->getSubReg();
+  if (CmpReg == ExecReg) {
+    AndCC = &And->getOperand(2);
+    CmpReg = AndCC->getReg();
+    CmpSubReg = AndCC->getSubReg();
+  } else if (And->getOperand(2).getReg() != ExecReg) {
+    return AMDGPU::NoRegister;
+  }
+
+  auto *Cmp = TRI->findReachingDef(CmpReg, CmpSubReg, *And, MRI, LIS);
+  if (!Cmp || !(Cmp->getOpcode() == AMDGPU::V_CMP_NE_U32_e32 ||
+                Cmp->getOpcode() == AMDGPU::V_CMP_NE_U32_e64) ||
+      Cmp->getParent() != And->getParent())
+    return AMDGPU::NoRegister;
+
+  MachineOperand *Op1 = TII->getNamedOperand(*Cmp, AMDGPU::OpName::src0);
+  MachineOperand *Op2 = TII->getNamedOperand(*Cmp, AMDGPU::OpName::src1);
+  if (Op1->isImm() && Op2->isReg())
+    std::swap(Op1, Op2);
+  if (!Op1->isReg() || !Op2->isImm() || Op2->getImm() != 1)
+    return AMDGPU::NoRegister;
+
+  unsigned SelReg = Op1->getReg();
+  auto *Sel = TRI->findReachingDef(SelReg, Op1->getSubReg(), *Cmp, MRI, LIS);
+  if (!Sel || Sel->getOpcode() != AMDGPU::V_CNDMASK_B32_e64)
+    return AMDGPU::NoRegister;
+
+  Op1 = TII->getNamedOperand(*Sel, AMDGPU::OpName::src0);
+  Op2 = TII->getNamedOperand(*Sel, AMDGPU::OpName::src1);
+  MachineOperand *CC = TII->getNamedOperand(*Sel, AMDGPU::OpName::src2);
+  if (!Op1->isImm() || !Op2->isImm() || !CC->isReg() ||
+      Op1->getImm() != 0 || Op2->getImm() != 1)
+    return AMDGPU::NoRegister;
+
+  LLVM_DEBUG(dbgs() << "Folding sequence:\n\t" << *Sel << '\t'
+                    << *Cmp << '\t' << *And);
+
+  unsigned CCReg = CC->getReg();
+  LIS->RemoveMachineInstrFromMaps(*And);
+  MachineInstr *Andn2 = BuildMI(MBB, *And, And->getDebugLoc(),
+                                TII->get(Andn2Opc), And->getOperand(0).getReg())
+                            .addReg(ExecReg)
+                            .addReg(CCReg, CC->getSubReg());
+  And->eraseFromParent();
+  LIS->InsertMachineInstrInMaps(*Andn2);
+
+  LLVM_DEBUG(dbgs() << "=>\n\t" << *Andn2 << '\n');
+
+  // Try to remove compare. Cmp value should not used in between of cmp
+  // and s_and_b64 if VCC or just unused if any other register.
+  if ((TargetRegisterInfo::isVirtualRegister(CmpReg) &&
+       MRI.use_nodbg_empty(CmpReg)) ||
+      (CmpReg == CondReg &&
+       std::none_of(std::next(Cmp->getIterator()), Andn2->getIterator(),
+                    [TRI, CondReg](const MachineInstr &MI) {
+                      return MI.readsRegister(CondReg, TRI); }))) {
+    LLVM_DEBUG(dbgs() << "Erasing: " << *Cmp << '\n');
+
+    LIS->RemoveMachineInstrFromMaps(*Cmp);
+    Cmp->eraseFromParent();
+
+    // Try to remove v_cndmask_b32.
+    if (TargetRegisterInfo::isVirtualRegister(SelReg) &&
+        MRI.use_nodbg_empty(SelReg)) {
+      LLVM_DEBUG(dbgs() << "Erasing: " << *Sel << '\n');
+
+      LIS->RemoveMachineInstrFromMaps(*Sel);
+      Sel->eraseFromParent();
+    }
+  }
+
+  return CCReg;
+}
+
 bool SIOptimizeExecMaskingPreRA::runOnMachineFunction(MachineFunction &MF) {
   if (skipFunction(MF.getFunction()))
     return false;
@@ -117,6 +233,14 @@ bool SIOptimizeExecMaskingPreRA::runOnMachineFunction(MachineFunction &MF) {
 
   for (MachineBasicBlock &MBB : MF) {
 
+    if (unsigned Reg = optimizeVcndVcmpPair(MBB, ST, MRI, LIS)) {
+      RecalcRegs.insert(Reg);
+      RecalcRegs.insert(AMDGPU::VCC_LO);
+      RecalcRegs.insert(AMDGPU::VCC_HI);
+      RecalcRegs.insert(AMDGPU::SCC);
+      Changed = true;
+    }
+
     // Try to remove unneeded instructions before s_endpgm.
     if (MBB.succ_empty()) {
       if (MBB.empty())