[llvm-branch-commits] [llvm] MachineUniformityAnalysis: Improve isConstantOrUndefValuePhi (PR #112866)

[Petar Avramovic via llvm-branch-commits]

================
@@ -54,9 +54,28 @@ const MachineBasicBlock *MachineSSAContext::getDefBlock(Register value) const {
   return F->getRegInfo().getVRegDef(value)->getParent();
 }
 
+static bool isUndef(const MachineInstr &MI) {
+  return MI.getOpcode() == TargetOpcode::G_IMPLICIT_DEF ||
+         MI.getOpcode() == TargetOpcode::IMPLICIT_DEF;
+}
+
+/// MachineInstr equivalent of PHINode::hasConstantOrUndefValue()
 template <>
-bool MachineSSAContext::isConstantOrUndefValuePhi(const MachineInstr &Phi) {
-  return Phi.isConstantValuePHI();
+bool MachineSSAContext::isConstantOrUndefValuePhi(const MachineInstr &MI) {
+  if (!MI.isPHI())
+    return false;
+  const MachineRegisterInfo &MRI = MI.getMF()->getRegInfo();
+  Register This = MI.getOperand(0).getReg();
+  Register ConstantValue;
+  for (unsigned i = 1, e = MI.getNumOperands(); i < e; i += 2) {
+    Register Incoming = MI.getOperand(i).getReg();
+    if (Incoming != This && !isUndef(*MRI.getVRegDef(Incoming))) {
----------------
petar-avramovic wrote:

Should we try to do something about PHIs then? Machine uniformity analysis in later stages is not used really. It gives different results then LLVM-IR and G-MIR versions, mostly because sgpr register class without LLT stops propagation of divergence/uniformity

https://github.com/llvm/llvm-project/pull/112866