[llvm] [LV][VPlan] Add initial support for CSA vectorization (PR #106560)

[Elvis Wang via llvm-commits]

================
@@ -2346,6 +2444,222 @@ void VPScalarCastRecipe ::print(raw_ostream &O, const Twine &Indent,
 }
 #endif
 
+#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
+void VPCSAHeaderPHIRecipe::print(raw_ostream &O, const Twine &Indent,
+                                 VPSlotTracker &SlotTracker) const {
+  O << Indent << "EMIT ";
+  printAsOperand(O, SlotTracker);
+  O << " = csa-data-phi ";
+  printOperands(O, SlotTracker);
+}
+#endif
+
+void VPCSAHeaderPHIRecipe::execute(VPTransformState &State) {
+  // PrevBB is this BB
+  IRBuilder<>::InsertPointGuard Guard(State.Builder);
+  State.Builder.SetInsertPoint(State.CFG.PrevBB->getFirstNonPHI());
+
+  Value *InitData = State.get(getVPInitData(), 0);
+  PHINode *DataPhi =
+      State.Builder.CreatePHI(InitData->getType(), 2, "csa.data.phi");
+  BasicBlock *PreheaderBB = State.CFG.getPreheaderBBFor(this);
+  DataPhi->addIncoming(InitData, PreheaderBB);
+  // Note: We didn't add Incoming for the new data since VPCSADataUpdateRecipe
+  // may not have been executed. We let VPCSADataUpdateRecipe::execute add the
+  // incoming operand to DataPhi.
+
+  State.set(this, DataPhi);
+}
+
+InstructionCost VPCSAHeaderPHIRecipe::computeCost(ElementCount VF,
+                                                  VPCostContext &Ctx) const {
+  if (VF.isScalar())
+    return 0;
+
+  InstructionCost C = 0;
+  auto *VTy = VectorType::get(getUnderlyingValue()->getType(), VF);
----------------
ElvisWang123 wrote:

Using `inferScalarType(this)` instead of `getUnderlyingValue()->getType()` would be better.
And I think you also need to implement `inferScalarType()` in `VPlanAnalysis` for these CSA recipes.

https://github.com/llvm/llvm-project/pull/106560