[llvm-branch-commits] [llvm] AMDGPU/GlobalISel: Temporal divergence lowering (non i1) (PR #124298)

[Sameer Sahasrabuddhe via llvm-branch-commits]

================
@@ -188,6 +190,37 @@ void DivergenceLoweringHelper::constrainAsLaneMask(Incoming &In) {
   In.Reg = Copy.getReg(0);
 }
 
+void replaceUsesOfRegInInstWith(Register Reg, MachineInstr *Inst,
+                                Register NewReg) {
+  for (MachineOperand &Op : Inst->operands()) {
+    if (Op.isReg() && Op.getReg() == Reg)
+      Op.setReg(NewReg);
+  }
+}
+
+bool DivergenceLoweringHelper::lowerTempDivergence() {
+  AMDGPU::IntrinsicLaneMaskAnalyzer ILMA(*MF);
+
+  for (auto [Inst, UseInst, _] : MUI->getUsesOutsideCycleWithDivergentExit()) {
+    Register Reg = Inst->getOperand(0).getReg();
+    if (MRI->getType(Reg) == LLT::scalar(1) || MUI->isDivergent(Reg) ||
+        ILMA.isS32S64LaneMask(Reg))
+      continue;
+
+    MachineInstr *MI = const_cast<MachineInstr *>(Inst);
----------------
ssahasra wrote:

I lean on the other side. If you look at LoopInfoBase or LoopBase, their functions take const pointers as arguments but return non-const pointers when asked. Sure, an analysis should treat its inputs as const, but when it returns something to the client, that client owns it anyway, so forcing that to be const is just an inconvenience. I would rather have the analysis do the const_cast before returning a list of pointers to something I already own.

This seems to be the first time that uniformity analysis is returning something. Until now, the public interface has simply been a bunch of predicates like "isUniform" that take a const pointer as arguments.

https://github.com/llvm/llvm-project/pull/124298