[llvm] [VPlan] Explicitly unroll replicate-regions without live-outs by VF. (PR #170212)

[via llvm-commits]

================
@@ -543,6 +543,42 @@ void VPlanTransforms::unrollByUF(VPlan &Plan, unsigned UF) {
   VPlanTransforms::removeDeadRecipes(Plan);
 }
 
+/// Add a lane offset to the start index of \p Steps.
+static void addLaneToStartIndex(VPScalarIVStepsRecipe *Steps, unsigned Lane,
+                                VPlan &Plan, VPRecipeBase *InsertPt) {
+  assert(Lane > 0 && "Zero lane adds no offset to start index");
+  VPTypeAnalysis TypeInfo(Plan);
+  Type *BaseIVTy = TypeInfo.inferScalarType(Steps->getOperand(0));
+
+  VPValue *OldStartIndex = Steps->getStartIndex();
+  VPValue *LaneOffset;
+  unsigned AddOpcode;
+  // TODO: Retrieve the flags from Steps unconditionally.
+  VPIRFlags Flags;
+  if (BaseIVTy->isFloatingPointTy()) {
+    int SignedLane = static_cast<int>(Lane);
+    if (!OldStartIndex && Steps->getInductionOpcode() == Instruction::FSub)
+      SignedLane = -SignedLane;
+    LaneOffset = Plan.getOrAddLiveIn(ConstantFP::get(BaseIVTy, SignedLane));
+    AddOpcode = Steps->getInductionOpcode();
+    Flags = VPIRFlags(FastMathFlags());
+  } else {
+    unsigned BaseIVBits = BaseIVTy->getScalarSizeInBits();
+    LaneOffset = Plan.getConstantInt(
+        APInt(BaseIVBits, Lane, /*isSigned*/ false, /*implicitTrunc*/ true));
+    AddOpcode = Instruction::Add;
+    Flags = VPIRFlags(VPIRFlags::WrapFlagsTy(false, false));
+  }
+
+  VPValue *NewStartIndex = LaneOffset;
+  if (OldStartIndex) {
+    VPBuilder Builder(InsertPt);
----------------
ayalz wrote:

Ahh, understand now why Steps was inserted before calling - so it could be used here as the insertion point. It would good if `addLaneToStartIndex()` folded the constant Lane with the constant start index (Part * FixedVF), w/o having to introduce any additional recipe.

https://github.com/llvm/llvm-project/pull/170212