[VPlan] Replace UnrollPart for VPScalarIVSteps with start index op (NFC)

Replace the unroll part operand for VPScalarIVStepsRecipe with the start
index. This simplifies #170053
and is also a first step to break down the recipe into its components.

Author: fhahn

llvm/lib/Transforms/Vectorize/LoopVectorizationPlanner.h

@@ -322,6 +322,17 @@ class VPBuilder {
     return createScalarCast(CastOp, Op, ResultTy, DL);
   }
 
+  VPValue *createScalarSExtOrTrunc(VPValue *Op, Type *ResultTy, Type *SrcTy,
+                                   DebugLoc DL) {
+    if (ResultTy == SrcTy)
+      return Op;
+    Instruction::CastOps CastOp =
+        ResultTy->getScalarSizeInBits() < SrcTy->getScalarSizeInBits()
+            ? Instruction::Trunc
+            : Instruction::SExt;
+    return createScalarCast(CastOp, Op, ResultTy, DL);
+  }
+
   VPWidenCastRecipe *createWidenCast(Instruction::CastOps Opcode, VPValue *Op,
                                      Type *ResultTy) {
     VPIRFlags Flags;

llvm/lib/Transforms/Vectorize/VPlan.h

@@ -3783,9 +3783,9 @@ class VPDerivedIVRecipe : public VPSingleDefRecipe {
 };
 
 /// A recipe for handling phi nodes of integer and floating-point inductions,
-/// producing their scalar values.
-class LLVM_ABI_FOR_TEST VPScalarIVStepsRecipe : public VPRecipeWithIRFlags,
-                                                public VPUnrollPartAccessor<3> {
+/// producing their scalar values. Before unrolling the recipe has 3 operands:
+/// IV, step and VF. Unrolling adds an extra operand StartIndex.
+class LLVM_ABI_FOR_TEST VPScalarIVStepsRecipe : public VPRecipeWithIRFlags {
   Instruction::BinaryOps InductionOpcode;
 
 public:
@@ -3815,10 +3815,6 @@ class LLVM_ABI_FOR_TEST VPScalarIVStepsRecipe : public VPRecipeWithIRFlags,
         getDebugLoc());
   }
 
-  /// Return true if this VPScalarIVStepsRecipe corresponds to part 0. Note that
-  /// this is only accurate after the VPlan has been unrolled.
-  bool isPart0() const { return getUnrollPart(*this) == 0; }
-
   VP_CLASSOF_IMPL(VPDef::VPScalarIVStepsSC)
 
   /// Generate the scalarized versions of the phi node as needed by their users.

llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp

@@ -2380,8 +2380,6 @@ void VPScalarIVStepsRecipe::execute(VPTransformState &State) {
   // iteration.
   bool FirstLaneOnly = vputils::onlyFirstLaneUsed(this);
   // Compute the scalar steps and save the results in State.
-  Type *IntStepTy =
-      IntegerType::get(BaseIVTy->getContext(), BaseIVTy->getScalarSizeInBits());
 
   unsigned StartLane = 0;
   unsigned EndLane = FirstLaneOnly ? 1 : State.VF.getKnownMinValue();
@@ -2390,20 +2388,10 @@ void VPScalarIVStepsRecipe::execute(VPTransformState &State) {
     EndLane = StartLane + 1;
   }
   Value *StartIdx0;
-  if (getUnrollPart(*this) == 0)
-    StartIdx0 = ConstantInt::get(IntStepTy, 0);
-  else {
-    StartIdx0 = State.get(getOperand(2), true);
-    if (getUnrollPart(*this) != 1) {
-      StartIdx0 =
-          Builder.CreateMul(StartIdx0, ConstantInt::get(StartIdx0->getType(),
-                                                        getUnrollPart(*this)));
-    }
-    StartIdx0 = Builder.CreateSExtOrTrunc(StartIdx0, IntStepTy);
-  }
-
-  if (BaseIVTy->isFloatingPointTy())
-    StartIdx0 = Builder.CreateSIToFP(StartIdx0, BaseIVTy);
+  if (getNumOperands() == 3) {
+    StartIdx0 = getSignedIntOrFpConstant(BaseIVTy, 0);
+  } else
+    StartIdx0 = State.get(getOperand(3), true);
 
   for (unsigned Lane = StartLane; Lane < EndLane; ++Lane) {
     Value *StartIdx = Builder.CreateBinOp(

llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp

@@ -1459,7 +1459,9 @@ static void simplifyRecipe(VPSingleDefRecipe *Def, VPTypeAnalysis &TypeInfo) {
   // VPScalarIVSteps for part 0 can be replaced by their start value, if only
   // the first lane is demanded.
   if (auto *Steps = dyn_cast<VPScalarIVStepsRecipe>(Def)) {
-    if (Steps->isPart0() && vputils::onlyFirstLaneUsed(Steps)) {
+    if ((Steps->getNumOperands() == 3 ||
+         match(Steps->getOperand(3), m_ZeroInt())) &&
+        vputils::onlyFirstLaneUsed(Steps)) {
       Steps->replaceAllUsesWith(Steps->getOperand(0));
       return;
     }

llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp

@@ -53,6 +53,9 @@ class UnrollState {
   /// Unroll replicate region \p VPR by cloning the region UF - 1 times.
   void unrollReplicateRegionByUF(VPRegionBlock *VPR);
 
+  /// Add a start index operand to \p Steps for \p Part.
+  void addStartIndexForScalarSteps(VPScalarIVStepsRecipe *Steps, unsigned Part);
+
   /// Unroll recipe \p R by cloning it UF - 1 times, unless it is uniform across
   /// all parts.
   void unrollRecipeByUF(VPRecipeBase &R);
@@ -123,6 +126,32 @@ class UnrollState {
 };
 } // namespace
 
+void UnrollState::addStartIndexForScalarSteps(VPScalarIVStepsRecipe *Steps, unsigned Part) {
+  if (Part == 0) {
+    Steps->addOperand(getConstantInt(Part));
+    return;
+  }
+
+  VPBuilder Builder(Steps);
+  Type *BaseIVTy = TypeInfo.inferScalarType(Steps->getOperand(0));
+  Type *IntStepTy =
+      IntegerType::get(BaseIVTy->getContext(), BaseIVTy->getScalarSizeInBits());
+  VPValue *StartIdx0 = Steps->getOperand(2);
+  StartIdx0 = Builder.createOverflowingOp(
+      Instruction::Mul,
+      {StartIdx0,
+       Plan.getConstantInt(TypeInfo.inferScalarType(StartIdx0), Part)});
+  StartIdx0 = Builder.createScalarSExtOrTrunc(
+      StartIdx0, IntStepTy, TypeInfo.inferScalarType(StartIdx0),
+      DebugLoc::getUnknown());
+
+  if (BaseIVTy->isFloatingPointTy())
+    StartIdx0 = Builder.createScalarCast(Instruction::SIToFP, StartIdx0,
+                                         BaseIVTy, DebugLoc::getUnknown());
+
+  Steps->addOperand(StartIdx0);
+}
+
 void UnrollState::unrollReplicateRegionByUF(VPRegionBlock *VPR) {
   VPBlockBase *InsertPt = VPR->getSingleSuccessor();
   for (unsigned Part = 1; Part != UF; ++Part) {
@@ -136,9 +165,8 @@ void UnrollState::unrollReplicateRegionByUF(VPRegionBlock *VPR) {
              VPBlockUtils::blocksOnly<VPBasicBlock>(Part0))) {
       for (const auto &[PartIR, Part0R] : zip(*PartIVPBB, *Part0VPBB)) {
         remapOperands(&PartIR, Part);
-        if (auto *ScalarIVSteps = dyn_cast<VPScalarIVStepsRecipe>(&PartIR)) {
-          ScalarIVSteps->addOperand(getConstantInt(Part));
-        }
+        if (auto *Steps = dyn_cast<VPScalarIVStepsRecipe>(&PartIR))
+          addStartIndexForScalarSteps(Steps, Part);
 
         addRecipeForPart(&Part0R, &PartIR, Part);
       }
@@ -311,10 +339,13 @@ void UnrollState::unrollRecipeByUF(VPRecipeBase &R) {
     }
     remapOperands(Copy, Part);
 
+    if (auto *ScalarIVSteps = dyn_cast<VPScalarIVStepsRecipe>(Copy))
+      addStartIndexForScalarSteps(ScalarIVSteps, Part);
+
     // Add operand indicating the part to generate code for, to recipes still
     // requiring it.
-    if (isa<VPScalarIVStepsRecipe, VPWidenCanonicalIVRecipe,
-            VPVectorPointerRecipe, VPVectorEndPointerRecipe>(Copy) ||
+    if (isa<VPWidenCanonicalIVRecipe, VPVectorPointerRecipe,
+            VPVectorEndPointerRecipe>(Copy) ||
         match(Copy,
               m_VPInstruction<VPInstruction::CanonicalIVIncrementForPart>()))
       Copy->addOperand(getConstantInt(Part));