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

[Nicolai Hähnle via llvm-branch-commits]

================
@@ -188,6 +190,35 @@ 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::lowerTemporalDivergence() {
+  AMDGPU::IntrinsicLaneMaskAnalyzer ILMA(*MF);
+
+  for (auto [Inst, UseInst, _] : MUI->getTemporalDivergenceList()) {
+    Register Reg = Inst->getOperand(0).getReg();
+    if (MRI->getType(Reg) == LLT::scalar(1) || MUI->isDivergent(Reg) ||
+        ILMA.isS32S64LaneMask(Reg))
+      continue;
+
+    MachineBasicBlock *MBB = Inst->getParent();
+    B.setInsertPt(*MBB, MBB->SkipPHIsAndLabels(std::next(Inst->getIterator())));
+
+    Register VgprReg = MRI->createGenericVirtualRegister(MRI->getType(Reg));
+    B.buildInstr(AMDGPU::COPY, {VgprReg}, {Reg})
+        .addUse(ExecReg, RegState::Implicit);
+
+    replaceUsesOfRegInInstWith(Reg, UseInst, VgprReg);
+  }
+  return false;
+}
----------------
nhaehnle wrote:

I do have one high-level comment about this.

Every `Inst` may potentially appear with many `UseInst`s in the temporal divergence list. The current code will create multiple new registers and multiple `COPY` instructions, which seems wasteful even if downstream passes can often clean it up.

I would suggest capturing the created register in a `DenseMap<Instruction *, Register>` for re-use.

Also, how about inserting the `COPY` at the end of `Inst->getParent()`? That way, the live range of the VGPR is reduced.

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