[llvm] [AMDGPU] Merge two V_CNDMASK instructions into V_DUAL_CNDMASK (PR #135007)

[Jay Foad via llvm-commits]

================
@@ -1459,13 +1482,73 @@ bool SIFoldOperandsImpl::tryConstantFoldOp(MachineInstr *MI) const {
   return false;
 }
 
+bool SIFoldOperandsImpl::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 = getImmOrMaterializedImm(*Src0);
+    auto src1Imm = getImmOrMaterializedImm(*Src1);
+
+    if (!src1Imm && src0Imm)
+      return false;
+    if (src1Imm && !src0Imm)
+      count++;
+  }
+  return (count >= 2);
----------------
jayfoad wrote:

It is probably still worth doing this if count == 1, to reduce code size.

https://github.com/llvm/llvm-project/pull/135007