[llvm] [AMDGPU] Switch V_CNDMASK operands to shrink it into VOP2 (PR #135162)

[Jay Foad via llvm-commits]

================
@@ -831,6 +836,109 @@ bool SIShrinkInstructions::tryReplaceDeadSDST(MachineInstr &MI) const {
   return true;
 }
 
+unsigned SIShrinkInstructions::getInverseCompareOpcode(MachineInstr &MI) const {
+  switch (MI.getOpcode()) {
+  // unsigned 32
+  case AMDGPU::V_CMP_EQ_U32_e64:
+    return AMDGPU::V_CMP_NE_U32_e64;
+  case AMDGPU::V_CMP_NE_U32_e64:
+    return AMDGPU::V_CMP_EQ_U32_e64;
+  case AMDGPU::V_CMP_GE_U32_e64:
+    return AMDGPU::V_CMP_LT_U32_e64;
+  case AMDGPU::V_CMP_LE_U32_e64:
+    return AMDGPU::V_CMP_GT_U32_e64;
+  case AMDGPU::V_CMP_GT_U32_e64:
+    return AMDGPU::V_CMP_LE_U32_e64;
+  case AMDGPU::V_CMP_LT_U32_e64:
+    return AMDGPU::V_CMP_GE_U32_e64;
+    // float 32
+  case AMDGPU::V_CMP_EQ_F32_e64:
+    return AMDGPU::V_CMP_NEQ_F32_e64;
+  case AMDGPU::V_CMP_NEQ_F32_e64:
+    return AMDGPU::V_CMP_EQ_F32_e64;
+  case AMDGPU::V_CMP_GE_F32_e64:
+    return AMDGPU::V_CMP_LT_F32_e64;
+  case AMDGPU::V_CMP_LE_F32_e64:
+    return AMDGPU::V_CMP_GT_F32_e64;
+  case AMDGPU::V_CMP_GT_F32_e64:
+    return AMDGPU::V_CMP_LE_F32_e64;
+  case AMDGPU::V_CMP_LT_F32_e64:
+    return AMDGPU::V_CMP_GE_F32_e64;
+  default:
+    return 0;
+  }
+}
+
+bool SIShrinkInstructions::shouldSwitchOperands(MachineRegisterInfo &MRI,
+                                                MachineInstr &MI,
+                                                const SIInstrInfo &TII) const {
+  auto allUses = MRI.use_nodbg_operands(MI.getOperand(5).getReg());
+  unsigned Count = 0;
+
+  for (auto &Use : allUses) {
+    if (Use.getParent()->getOpcode() != AMDGPU::V_CNDMASK_B32_e64)
+      return false;
+    MachineOperand *Src0 =
+        TII.getNamedOperand(*Use.getParent(), AMDGPU::OpName::src0);
+    MachineOperand *Src1 =
+        TII.getNamedOperand(*Use.getParent(), AMDGPU::OpName::src1);
+
+    auto Src0Imm = Src0->isImm();
+    auto Src1Imm = Src1->isImm();
+
+    if (!Src1Imm && Src0Imm)
+      return false;
+    if (Src1Imm && !Src0Imm)
+      Count++;
+  }
+  return (Count >= 1);
+}
+
+// OldVCC and NewVCC are used to remember VCC after inverting comparison
+bool SIShrinkInstructions::trySwitchOperands(MachineInstr &MI, Register *OldVCC,
+                                             Register *NewVCC) const {
+  const DebugLoc &DL = MI.getDebugLoc();
+  auto Reg = MI.getOperand(5).getReg();
+  if (!Reg.isVirtual())
+    return false;
+
+  if (*OldVCC != Reg) {
+    MachineInstr *DefMI = MRI->getVRegDef(Reg);
+    if (DefMI) {
+      unsigned Opcode = getInverseCompareOpcode(*DefMI);
+      if (Opcode &&
+          SIShrinkInstructions::shouldSwitchOperands(*MRI, MI, *TII)) {
+        auto cmpDL = DefMI->getDebugLoc();
----------------
jayfoad wrote:

`CmpDL`

https://github.com/llvm/llvm-project/pull/135162