Skip to content

[VPlan] Move getCanonicalIV to VPRegionBlock (NFC). #163020

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Open
wants to merge 3 commits into
base: main
Choose a base branch
from
Open

[VPlan] Move getCanonicalIV to VPRegionBlock (NFC). #163020

wants to merge 3 commits into from
+78 −58

Conversation

fhahn
Copy link
Contributor

@fhahn fhahn commented Oct 11, 2025

The canonical IV is tied to region blocks; move getCanonicalIV there and update all users.

@fhahn fhahn requested review from aniragil, ayalz and rengolin October 11, 2025 21:17
@llvmbot
Copy link
Member

llvmbot commented Oct 11, 2025

@llvm/pr-subscribers-llvm-transforms

Author: Florian Hahn (fhahn)

Changes

The canonical IV is tied to region blocks; move getCanonicalIV there and update all users.

Full diff: https://github.com/llvm/llvm-project/pull/163020.diff

7 Files Affected:

  • (modified) llvm/lib/Transforms/Vectorize/LoopVectorize.cpp (+12-11)
  • (modified) llvm/lib/Transforms/Vectorize/VPlan.h (+10-10)
  • (modified) llvm/lib/Transforms/Vectorize/VPlanConstruction.cpp (+9-6)
  • (modified) llvm/lib/Transforms/Vectorize/VPlanPredicator.cpp (+2-1)
  • (modified) llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp (+31-23)
  • (modified) llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp (+2-1)
  • (modified) llvm/lib/Transforms/Vectorize/VPlanUtils.cpp (+6-3)
diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index e62d57e6920b7..7a041fda07321 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -8242,12 +8242,12 @@ VPlanPtr LoopVectorizationPlanner::tryToBuildVPlanWithVPRecipes(
   // count is >= increment and a multiple of the increment.
   bool HasNUW = !IVUpdateMayOverflow || Style == TailFoldingStyle::None;
   if (!HasNUW) {
-    auto *IVInc = Plan->getVectorLoopRegion()
-                      ->getExitingBasicBlock()
-                      ->getTerminator()
-                      ->getOperand(0);
-    assert(match(IVInc, m_VPInstruction<Instruction::Add>(
-                            m_Specific(Plan->getCanonicalIV()), m_VPValue())) &&
+    VPRegionBlock *LoopRegion = Plan->getVectorLoopRegion();
+    auto *IVInc =
+        LoopRegion->getExitingBasicBlock()->getTerminator()->getOperand(0);
+    assert(match(IVInc,
+                 m_VPInstruction<Instruction::Add>(
+                     m_Specific(LoopRegion->getCanonicalIV()), m_VPValue())) &&
            "Did not find the canonical IV increment");
     cast<VPRecipeWithIRFlags>(IVInc)->dropPoisonGeneratingFlags();
   }
@@ -8377,8 +8377,8 @@ VPlanPtr LoopVectorizationPlanner::tryToBuildVPlanWithVPRecipes(
   for (VPValue *Old : Old2New.keys())
     Old->getDefiningRecipe()->eraseFromParent();
 
-  assert(isa<VPRegionBlock>(Plan->getVectorLoopRegion()) &&
-         !Plan->getVectorLoopRegion()->getEntryBasicBlock()->empty() &&
+  assert(isa<VPRegionBlock>(LoopRegion) &&
+         !LoopRegion->getEntryBasicBlock()->empty() &&
          "entry block must be set to a VPRegionBlock having a non-empty entry "
          "VPBasicBlock");
 
@@ -9320,8 +9320,9 @@ static void preparePlanForMainVectorLoop(VPlan &MainPlan, VPlan &EpiPlan) {
   if (ResumePhiIter == MainScalarPH->phis().end()) {
     VPBuilder ScalarPHBuilder(MainScalarPH, MainScalarPH->begin());
     ResumePhi = ScalarPHBuilder.createScalarPhi(
-        {VectorTC, MainPlan.getCanonicalIV()->getStartValue()}, {},
-        "vec.epilog.resume.val");
+        {VectorTC,
+         MainPlan.getVectorLoopRegion()->getCanonicalIV()->getStartValue()},
+        {}, "vec.epilog.resume.val");
   } else {
     ResumePhi = cast<VPPhi>(&*ResumePhiIter);
     if (MainScalarPH->begin() == MainScalarPH->end())
@@ -9350,7 +9351,7 @@ static SmallVector<Instruction *> preparePlanForEpilogueVectorLoop(
 
   // Ensure that the start values for all header phi recipes are updated before
   // vectorizing the epilogue loop.
-  VPCanonicalIVPHIRecipe *IV = Plan.getCanonicalIV();
+  VPCanonicalIVPHIRecipe *IV = VectorLoop->getCanonicalIV();
   // When vectorizing the epilogue loop, the canonical induction start
   // value needs to be changed from zero to the value after the main
   // vector loop. Find the resume value created during execution of the main
diff --git a/llvm/lib/Transforms/Vectorize/VPlan.h b/llvm/lib/Transforms/Vectorize/VPlan.h
index fb696bea671af..8db626fe9bf58 100644
--- a/llvm/lib/Transforms/Vectorize/VPlan.h
+++ b/llvm/lib/Transforms/Vectorize/VPlan.h
@@ -4066,6 +4066,16 @@ class LLVM_ABI_FOR_TEST VPRegionBlock : public VPBlockBase {
   /// Remove the current region from its VPlan, connecting its predecessor to
   /// its entry, and its exiting block to its successor.
   void dissolveToCFGLoop();
+
+  /// Returns the canonical induction recipe of the region.
+  VPCanonicalIVPHIRecipe *getCanonicalIV() {
+    VPBasicBlock *EntryVPBB = getEntryBasicBlock();
+    if (EntryVPBB->empty()) {
+      // VPlan native path.
+      EntryVPBB = cast<VPBasicBlock>(EntryVPBB->getSingleSuccessor());
+    }
+    return cast<VPCanonicalIVPHIRecipe>(&*EntryVPBB->begin());
+  }
 };
 
 /// VPlan models a candidate for vectorization, encoding various decisions take
@@ -4377,16 +4387,6 @@ class VPlan {
   LLVM_DUMP_METHOD void dump() const;
 #endif
 
-  /// Returns the canonical induction recipe of the vector loop.
-  VPCanonicalIVPHIRecipe *getCanonicalIV() {
-    VPBasicBlock *EntryVPBB = getVectorLoopRegion()->getEntryBasicBlock();
-    if (EntryVPBB->empty()) {
-      // VPlan native path.
-      EntryVPBB = cast<VPBasicBlock>(EntryVPBB->getSingleSuccessor());
-    }
-    return cast<VPCanonicalIVPHIRecipe>(&*EntryVPBB->begin());
-  }
-
   VPValue *getSCEVExpansion(const SCEV *S) const {
     return SCEVToExpansion.lookup(S);
   }
diff --git a/llvm/lib/Transforms/Vectorize/VPlanConstruction.cpp b/llvm/lib/Transforms/Vectorize/VPlanConstruction.cpp
index 81deba2932ef8..55ec19281c876 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanConstruction.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanConstruction.cpp
@@ -661,9 +661,11 @@ void VPlanTransforms::attachCheckBlock(VPlan &Plan, Value *Cond,
   }
 
   VPIRMetadata VPBranchWeights;
-  auto *Term = VPBuilder(CheckBlockVPBB)
-                   .createNaryOp(VPInstruction::BranchOnCond, {CondVPV},
-                                 Plan.getCanonicalIV()->getDebugLoc());
+  auto *Term =
+      VPBuilder(CheckBlockVPBB)
+          .createNaryOp(
+              VPInstruction::BranchOnCond, {CondVPV},
+              Plan.getVectorLoopRegion()->getCanonicalIV()->getDebugLoc());
   if (AddBranchWeights) {
     MDBuilder MDB(Plan.getContext());
     MDNode *BranchWeights =
@@ -925,8 +927,8 @@ bool VPlanTransforms::handleMaxMinNumReductions(VPlan &Plan) {
     if (auto *DerivedIV = dyn_cast<VPDerivedIVRecipe>(VecV)) {
       if (DerivedIV->getNumUsers() == 1 &&
           DerivedIV->getOperand(1) == &Plan.getVectorTripCount()) {
-        auto *NewSel = Builder.createSelect(AnyNaN, Plan.getCanonicalIV(),
-                                            &Plan.getVectorTripCount());
+        auto *NewSel = Builder.createSelect(
+            AnyNaN, LoopRegion->getCanonicalIV(), &Plan.getVectorTripCount());
         DerivedIV->moveAfter(&*Builder.getInsertPoint());
         DerivedIV->setOperand(1, NewSel);
         continue;
@@ -939,7 +941,8 @@ bool VPlanTransforms::handleMaxMinNumReductions(VPlan &Plan) {
                            "FMaxNum/FMinNum reduction.\n");
       return false;
     }
-    auto *NewSel = Builder.createSelect(AnyNaN, Plan.getCanonicalIV(), VecV);
+    auto *NewSel =
+        Builder.createSelect(AnyNaN, LoopRegion->getCanonicalIV(), VecV);
     ResumeR->setOperand(0, NewSel);
   }
 
diff --git a/llvm/lib/Transforms/Vectorize/VPlanPredicator.cpp b/llvm/lib/Transforms/Vectorize/VPlanPredicator.cpp
index 0c27d535b680e..1d0472df1762e 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanPredicator.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanPredicator.cpp
@@ -168,7 +168,8 @@ void VPPredicator::createHeaderMask(VPBasicBlock *HeaderVPBB, bool FoldTail) {
   // non-phi instructions.
 
   auto &Plan = *HeaderVPBB->getPlan();
-  auto *IV = new VPWidenCanonicalIVRecipe(Plan.getCanonicalIV());
+  auto *IV = new VPWidenCanonicalIVRecipe(
+      Plan.getVectorLoopRegion()->getCanonicalIV());
   Builder.setInsertPoint(HeaderVPBB, HeaderVPBB->getFirstNonPhi());
   Builder.insert(IV);
 
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
index c8a2d84a535d3..ef1492320e1c3 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
@@ -501,7 +501,8 @@ static void removeRedundantInductionCasts(VPlan &Plan) {
 /// Try to replace VPWidenCanonicalIVRecipes with a widened canonical IV
 /// recipe, if it exists.
 static void removeRedundantCanonicalIVs(VPlan &Plan) {
-  VPCanonicalIVPHIRecipe *CanonicalIV = Plan.getCanonicalIV();
+  VPRegionBlock *LoopRegion = Plan.getVectorLoopRegion();
+  VPCanonicalIVPHIRecipe *CanonicalIV = LoopRegion->getCanonicalIV();
   VPWidenCanonicalIVRecipe *WidenNewIV = nullptr;
   for (VPUser *U : CanonicalIV->users()) {
     WidenNewIV = dyn_cast<VPWidenCanonicalIVRecipe>(U);
@@ -512,7 +513,7 @@ static void removeRedundantCanonicalIVs(VPlan &Plan) {
   if (!WidenNewIV)
     return;
 
-  VPBasicBlock *HeaderVPBB = Plan.getVectorLoopRegion()->getEntryBasicBlock();
+  VPBasicBlock *HeaderVPBB = LoopRegion->getEntryBasicBlock();
   for (VPRecipeBase &Phi : HeaderVPBB->phis()) {
     auto *WidenOriginalIV = dyn_cast<VPWidenIntOrFpInductionRecipe>(&Phi);
 
@@ -582,8 +583,9 @@ createScalarIVSteps(VPlan &Plan, InductionDescriptor::InductionKind Kind,
                     FPMathOperator *FPBinOp, Instruction *TruncI,
                     VPValue *StartV, VPValue *Step, DebugLoc DL,
                     VPBuilder &Builder) {
-  VPBasicBlock *HeaderVPBB = Plan.getVectorLoopRegion()->getEntryBasicBlock();
-  VPCanonicalIVPHIRecipe *CanonicalIV = Plan.getCanonicalIV();
+  VPRegionBlock *LoopRegion = Plan.getVectorLoopRegion();
+  VPBasicBlock *HeaderVPBB = LoopRegion->getEntryBasicBlock();
+  VPCanonicalIVPHIRecipe *CanonicalIV = LoopRegion->getCanonicalIV();
   VPSingleDefRecipe *BaseIV = Builder.createDerivedIV(
       Kind, FPBinOp, StartV, CanonicalIV, Step, "offset.idx");
 
@@ -800,8 +802,9 @@ static VPValue *optimizeEarlyExitInductionUser(VPlan &Plan,
     return nullptr;
 
   // Calculate the final index.
-  VPValue *EndValue = Plan.getCanonicalIV();
-  auto CanonicalIVType = Plan.getCanonicalIV()->getScalarType();
+  VPRegionBlock *LoopRegion = Plan.getVectorLoopRegion();
+  VPValue *EndValue = LoopRegion->getCanonicalIV();
+  auto CanonicalIVType = LoopRegion->getCanonicalIV()->getScalarType();
   VPBuilder B(cast<VPBasicBlock>(PredVPBB));
 
   DebugLoc DL = cast<VPInstruction>(Op)->getDebugLoc();
@@ -1520,7 +1523,7 @@ static bool isConditionTrueViaVFAndUF(VPValue *Cond, VPlan &Plan,
       return isConditionTrueViaVFAndUF(C, Plan, BestVF, BestUF, SE);
     });
 
-  auto *CanIV = Plan.getCanonicalIV();
+  auto *CanIV = Plan.getVectorLoopRegion()->getCanonicalIV();
   if (!match(Cond, m_SpecificICmp(CmpInst::ICMP_EQ,
                                   m_Specific(CanIV->getBackedgeValue()),
                                   m_Specific(&Plan.getVectorTripCount()))))
@@ -2304,7 +2307,7 @@ static VPActiveLaneMaskPHIRecipe *addVPLaneMaskPhiAndUpdateExitBranch(
     VPlan &Plan, bool DataAndControlFlowWithoutRuntimeCheck) {
   VPRegionBlock *TopRegion = Plan.getVectorLoopRegion();
   VPBasicBlock *EB = TopRegion->getExitingBasicBlock();
-  auto *CanonicalIVPHI = Plan.getCanonicalIV();
+  auto *CanonicalIVPHI = TopRegion->getCanonicalIV();
   VPValue *StartV = CanonicalIVPHI->getStartValue();
 
   auto *CanonicalIVIncrement =
@@ -2343,7 +2346,7 @@ static VPActiveLaneMaskPHIRecipe *addVPLaneMaskPhiAndUpdateExitBranch(
 
   // Create the active lane mask instruction in the VPlan preheader.
   VPValue *ALMMultiplier = Plan.getOrAddLiveIn(
-      ConstantInt::get(Plan.getCanonicalIV()->getScalarType(), 1));
+      ConstantInt::get(TopRegion->getCanonicalIV()->getScalarType(), 1));
   auto *EntryALM = Builder.createNaryOp(VPInstruction::ActiveLaneMask,
                                         {EntryIncrement, TC, ALMMultiplier}, DL,
                                         "active.lane.mask.entry");
@@ -2379,13 +2382,15 @@ static VPActiveLaneMaskPHIRecipe *addVPLaneMaskPhiAndUpdateExitBranch(
 /// TODO: Introduce explicit recipe for header-mask instead of searching
 /// for the header-mask pattern manually.
 static VPSingleDefRecipe *findHeaderMask(VPlan &Plan) {
+  VPRegionBlock *LoopRegion = Plan.getVectorLoopRegion();
   SmallVector<VPValue *> WideCanonicalIVs;
-  auto *FoundWidenCanonicalIVUser = find_if(Plan.getCanonicalIV()->users(),
-                                            IsaPred<VPWidenCanonicalIVRecipe>);
-  assert(count_if(Plan.getCanonicalIV()->users(),
+  auto *FoundWidenCanonicalIVUser = find_if(
+      LoopRegion->getCanonicalIV()->users(), IsaPred<VPWidenCanonicalIVRecipe>);
+  assert(count_if(LoopRegion->getCanonicalIV()->users(),
                   IsaPred<VPWidenCanonicalIVRecipe>) <= 1 &&
          "Must have at most one VPWideCanonicalIVRecipe");
-  if (FoundWidenCanonicalIVUser != Plan.getCanonicalIV()->users().end()) {
+  if (FoundWidenCanonicalIVUser !=
+      LoopRegion->getCanonicalIV()->users().end()) {
     auto *WideCanonicalIV =
         cast<VPWidenCanonicalIVRecipe>(*FoundWidenCanonicalIVUser);
     WideCanonicalIVs.push_back(WideCanonicalIV);
@@ -2393,7 +2398,7 @@ static VPSingleDefRecipe *findHeaderMask(VPlan &Plan) {
 
   // Also include VPWidenIntOrFpInductionRecipes that represent a widened
   // version of the canonical induction.
-  VPBasicBlock *HeaderVPBB = Plan.getVectorLoopRegion()->getEntryBasicBlock();
+  VPBasicBlock *HeaderVPBB = LoopRegion->getEntryBasicBlock();
   for (VPRecipeBase &Phi : HeaderVPBB->phis()) {
     auto *WidenOriginalIV = dyn_cast<VPWidenIntOrFpInductionRecipe>(&Phi);
     if (WidenOriginalIV && WidenOriginalIV->isCanonical())
@@ -2426,8 +2431,9 @@ void VPlanTransforms::addActiveLaneMask(
          "DataAndControlFlowWithoutRuntimeCheck implies "
          "UseActiveLaneMaskForControlFlow");
 
-  auto *FoundWidenCanonicalIVUser = find_if(Plan.getCanonicalIV()->users(),
-                                            IsaPred<VPWidenCanonicalIVRecipe>);
+  VPRegionBlock *LoopRegion = Plan.getVectorLoopRegion();
+  auto *FoundWidenCanonicalIVUser = find_if(
+      LoopRegion->getCanonicalIV()->users(), IsaPred<VPWidenCanonicalIVRecipe>);
   assert(FoundWidenCanonicalIVUser &&
          "Must have widened canonical IV when tail folding!");
   VPSingleDefRecipe *HeaderMask = findHeaderMask(Plan);
@@ -2440,7 +2446,7 @@ void VPlanTransforms::addActiveLaneMask(
   } else {
     VPBuilder B = VPBuilder::getToInsertAfter(WideCanonicalIV);
     VPValue *ALMMultiplier = Plan.getOrAddLiveIn(
-        ConstantInt::get(Plan.getCanonicalIV()->getScalarType(), 1));
+        ConstantInt::get(LoopRegion->getCanonicalIV()->getScalarType(), 1));
     LaneMask =
         B.createNaryOp(VPInstruction::ActiveLaneMask,
                        {WideCanonicalIV, Plan.getTripCount(), ALMMultiplier},
@@ -2550,9 +2556,10 @@ static void transformRecipestoEVLRecipes(VPlan &Plan, VPValue &EVL) {
   });
 
   assert(all_of(Plan.getVFxUF().users(),
-                [&Plan](VPUser *U) {
-                  return match(U, m_c_Add(m_Specific(Plan.getCanonicalIV()),
-                                          m_Specific(&Plan.getVFxUF()))) ||
+                [&LoopRegion, &Plan](VPUser *U) {
+                  return match(U,
+                               m_c_Add(m_Specific(LoopRegion->getCanonicalIV()),
+                                       m_Specific(&Plan.getVFxUF()))) ||
                          isa<VPWidenPointerInductionRecipe>(U);
                 }) &&
          "Only users of VFxUF should be VPWidenPointerInductionRecipe and the "
@@ -2707,9 +2714,10 @@ void VPlanTransforms::addExplicitVectorLength(
     VPlan &Plan, const std::optional<unsigned> &MaxSafeElements) {
   if (Plan.hasScalarVFOnly())
     return;
-  VPBasicBlock *Header = Plan.getVectorLoopRegion()->getEntryBasicBlock();
+  VPRegionBlock *LoopRegion = Plan.getVectorLoopRegion();
+  VPBasicBlock *Header = LoopRegion->getEntryBasicBlock();
 
-  auto *CanonicalIVPHI = Plan.getCanonicalIV();
+  auto *CanonicalIVPHI = LoopRegion->getCanonicalIV();
   auto *CanIVTy = CanonicalIVPHI->getScalarType();
   VPValue *StartV = CanonicalIVPHI->getStartValue();
 
@@ -4153,7 +4161,7 @@ void VPlanTransforms::narrowInterleaveGroups(VPlan &Plan, ElementCount VF,
 
   // Adjust induction to reflect that the transformed plan only processes one
   // original iteration.
-  auto *CanIV = Plan.getCanonicalIV();
+  auto *CanIV = VectorLoop->getCanonicalIV();
   auto *Inc = cast<VPInstruction>(CanIV->getBackedgeValue());
   VPBuilder PHBuilder(Plan.getVectorPreheader());
 
diff --git a/llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp b/llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp
index 5e7f19faebb56..14682d851e765 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp
@@ -69,7 +69,8 @@ class UnrollState {
                                 VPBasicBlock::iterator InsertPtForPhi);
 
   VPValue *getConstantVPV(unsigned Part) {
-    Type *CanIVIntTy = Plan.getCanonicalIV()->getScalarType();
+    Type *CanIVIntTy =
+        Plan.getVectorLoopRegion()->getCanonicalIV()->getScalarType();
     return Plan.getOrAddLiveIn(ConstantInt::get(CanIVIntTy, Part));
   }
 
diff --git a/llvm/lib/Transforms/Vectorize/VPlanUtils.cpp b/llvm/lib/Transforms/Vectorize/VPlanUtils.cpp
index 66748c534f10b..553731731a1d8 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanUtils.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanUtils.cpp
@@ -67,8 +67,10 @@ bool vputils::isHeaderMask(const VPValue *V, VPlan &Plan) {
 
   if (match(V, m_ActiveLaneMask(m_VPValue(A), m_VPValue(B), m_One())))
     return B == Plan.getTripCount() &&
-           (match(A, m_ScalarIVSteps(m_Specific(Plan.getCanonicalIV()), m_One(),
-                                     m_Specific(&Plan.getVF()))) ||
+           (match(A,
+                  m_ScalarIVSteps(
+                      m_Specific(Plan.getVectorLoopRegion()->getCanonicalIV()),
+                      m_One(), m_Specific(&Plan.getVF()))) ||
             IsWideCanonicalIV(A));
 
   return match(V, m_ICmp(m_VPValue(A), m_VPValue(B))) && IsWideCanonicalIV(A) &&
@@ -102,7 +104,8 @@ bool vputils::isUniformAcrossVFsAndUFs(VPValue *V) {
     return all_of(R->operands(), isUniformAcrossVFsAndUFs);
   }
 
-  auto *CanonicalIV = R->getParent()->getPlan()->getCanonicalIV();
+  auto *CanonicalIV =
+      R->getParent()->getPlan()->getVectorLoopRegion()->getCanonicalIV();
   // Canonical IV chain is uniform.
   if (V == CanonicalIV || V == CanonicalIV->getBackedgeValue())
     return true;

@llvmbot
Copy link
Member

llvmbot commented Oct 11, 2025

@llvm/pr-subscribers-vectorizers

Author: Florian Hahn (fhahn)

Changes

The canonical IV is tied to region blocks; move getCanonicalIV there and update all users.

Full diff: https://github.com/llvm/llvm-project/pull/163020.diff

7 Files Affected:

  • (modified) llvm/lib/Transforms/Vectorize/LoopVectorize.cpp (+12-11)
  • (modified) llvm/lib/Transforms/Vectorize/VPlan.h (+10-10)
  • (modified) llvm/lib/Transforms/Vectorize/VPlanConstruction.cpp (+9-6)
  • (modified) llvm/lib/Transforms/Vectorize/VPlanPredicator.cpp (+2-1)
  • (modified) llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp (+31-23)
  • (modified) llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp (+2-1)
  • (modified) llvm/lib/Transforms/Vectorize/VPlanUtils.cpp (+6-3)
diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index e62d57e6920b7..7a041fda07321 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -8242,12 +8242,12 @@ VPlanPtr LoopVectorizationPlanner::tryToBuildVPlanWithVPRecipes(
   // count is >= increment and a multiple of the increment.
   bool HasNUW = !IVUpdateMayOverflow || Style == TailFoldingStyle::None;
   if (!HasNUW) {
-    auto *IVInc = Plan->getVectorLoopRegion()
-                      ->getExitingBasicBlock()
-                      ->getTerminator()
-                      ->getOperand(0);
-    assert(match(IVInc, m_VPInstruction<Instruction::Add>(
-                            m_Specific(Plan->getCanonicalIV()), m_VPValue())) &&
+    VPRegionBlock *LoopRegion = Plan->getVectorLoopRegion();
+    auto *IVInc =
+        LoopRegion->getExitingBasicBlock()->getTerminator()->getOperand(0);
+    assert(match(IVInc,
+                 m_VPInstruction<Instruction::Add>(
+                     m_Specific(LoopRegion->getCanonicalIV()), m_VPValue())) &&
            "Did not find the canonical IV increment");
     cast<VPRecipeWithIRFlags>(IVInc)->dropPoisonGeneratingFlags();
   }
@@ -8377,8 +8377,8 @@ VPlanPtr LoopVectorizationPlanner::tryToBuildVPlanWithVPRecipes(
   for (VPValue *Old : Old2New.keys())
     Old->getDefiningRecipe()->eraseFromParent();
 
-  assert(isa<VPRegionBlock>(Plan->getVectorLoopRegion()) &&
-         !Plan->getVectorLoopRegion()->getEntryBasicBlock()->empty() &&
+  assert(isa<VPRegionBlock>(LoopRegion) &&
+         !LoopRegion->getEntryBasicBlock()->empty() &&
          "entry block must be set to a VPRegionBlock having a non-empty entry "
          "VPBasicBlock");
 
@@ -9320,8 +9320,9 @@ static void preparePlanForMainVectorLoop(VPlan &MainPlan, VPlan &EpiPlan) {
   if (ResumePhiIter == MainScalarPH->phis().end()) {
     VPBuilder ScalarPHBuilder(MainScalarPH, MainScalarPH->begin());
     ResumePhi = ScalarPHBuilder.createScalarPhi(
-        {VectorTC, MainPlan.getCanonicalIV()->getStartValue()}, {},
-        "vec.epilog.resume.val");
+        {VectorTC,
+         MainPlan.getVectorLoopRegion()->getCanonicalIV()->getStartValue()},
+        {}, "vec.epilog.resume.val");
   } else {
     ResumePhi = cast<VPPhi>(&*ResumePhiIter);
     if (MainScalarPH->begin() == MainScalarPH->end())
@@ -9350,7 +9351,7 @@ static SmallVector<Instruction *> preparePlanForEpilogueVectorLoop(
 
   // Ensure that the start values for all header phi recipes are updated before
   // vectorizing the epilogue loop.
-  VPCanonicalIVPHIRecipe *IV = Plan.getCanonicalIV();
+  VPCanonicalIVPHIRecipe *IV = VectorLoop->getCanonicalIV();
   // When vectorizing the epilogue loop, the canonical induction start
   // value needs to be changed from zero to the value after the main
   // vector loop. Find the resume value created during execution of the main
diff --git a/llvm/lib/Transforms/Vectorize/VPlan.h b/llvm/lib/Transforms/Vectorize/VPlan.h
index fb696bea671af..8db626fe9bf58 100644
--- a/llvm/lib/Transforms/Vectorize/VPlan.h
+++ b/llvm/lib/Transforms/Vectorize/VPlan.h
@@ -4066,6 +4066,16 @@ class LLVM_ABI_FOR_TEST VPRegionBlock : public VPBlockBase {
   /// Remove the current region from its VPlan, connecting its predecessor to
   /// its entry, and its exiting block to its successor.
   void dissolveToCFGLoop();
+
+  /// Returns the canonical induction recipe of the region.
+  VPCanonicalIVPHIRecipe *getCanonicalIV() {
+    VPBasicBlock *EntryVPBB = getEntryBasicBlock();
+    if (EntryVPBB->empty()) {
+      // VPlan native path.
+      EntryVPBB = cast<VPBasicBlock>(EntryVPBB->getSingleSuccessor());
+    }
+    return cast<VPCanonicalIVPHIRecipe>(&*EntryVPBB->begin());
+  }
 };
 
 /// VPlan models a candidate for vectorization, encoding various decisions take
@@ -4377,16 +4387,6 @@ class VPlan {
   LLVM_DUMP_METHOD void dump() const;
 #endif
 
-  /// Returns the canonical induction recipe of the vector loop.
-  VPCanonicalIVPHIRecipe *getCanonicalIV() {
-    VPBasicBlock *EntryVPBB = getVectorLoopRegion()->getEntryBasicBlock();
-    if (EntryVPBB->empty()) {
-      // VPlan native path.
-      EntryVPBB = cast<VPBasicBlock>(EntryVPBB->getSingleSuccessor());
-    }
-    return cast<VPCanonicalIVPHIRecipe>(&*EntryVPBB->begin());
-  }
-
   VPValue *getSCEVExpansion(const SCEV *S) const {
     return SCEVToExpansion.lookup(S);
   }
diff --git a/llvm/lib/Transforms/Vectorize/VPlanConstruction.cpp b/llvm/lib/Transforms/Vectorize/VPlanConstruction.cpp
index 81deba2932ef8..55ec19281c876 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanConstruction.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanConstruction.cpp
@@ -661,9 +661,11 @@ void VPlanTransforms::attachCheckBlock(VPlan &Plan, Value *Cond,
   }
 
   VPIRMetadata VPBranchWeights;
-  auto *Term = VPBuilder(CheckBlockVPBB)
-                   .createNaryOp(VPInstruction::BranchOnCond, {CondVPV},
-                                 Plan.getCanonicalIV()->getDebugLoc());
+  auto *Term =
+      VPBuilder(CheckBlockVPBB)
+          .createNaryOp(
+              VPInstruction::BranchOnCond, {CondVPV},
+              Plan.getVectorLoopRegion()->getCanonicalIV()->getDebugLoc());
   if (AddBranchWeights) {
     MDBuilder MDB(Plan.getContext());
     MDNode *BranchWeights =
@@ -925,8 +927,8 @@ bool VPlanTransforms::handleMaxMinNumReductions(VPlan &Plan) {
     if (auto *DerivedIV = dyn_cast<VPDerivedIVRecipe>(VecV)) {
       if (DerivedIV->getNumUsers() == 1 &&
           DerivedIV->getOperand(1) == &Plan.getVectorTripCount()) {
-        auto *NewSel = Builder.createSelect(AnyNaN, Plan.getCanonicalIV(),
-                                            &Plan.getVectorTripCount());
+        auto *NewSel = Builder.createSelect(
+            AnyNaN, LoopRegion->getCanonicalIV(), &Plan.getVectorTripCount());
         DerivedIV->moveAfter(&*Builder.getInsertPoint());
         DerivedIV->setOperand(1, NewSel);
         continue;
@@ -939,7 +941,8 @@ bool VPlanTransforms::handleMaxMinNumReductions(VPlan &Plan) {
                            "FMaxNum/FMinNum reduction.\n");
       return false;
     }
-    auto *NewSel = Builder.createSelect(AnyNaN, Plan.getCanonicalIV(), VecV);
+    auto *NewSel =
+        Builder.createSelect(AnyNaN, LoopRegion->getCanonicalIV(), VecV);
     ResumeR->setOperand(0, NewSel);
   }
 
diff --git a/llvm/lib/Transforms/Vectorize/VPlanPredicator.cpp b/llvm/lib/Transforms/Vectorize/VPlanPredicator.cpp
index 0c27d535b680e..1d0472df1762e 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanPredicator.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanPredicator.cpp
@@ -168,7 +168,8 @@ void VPPredicator::createHeaderMask(VPBasicBlock *HeaderVPBB, bool FoldTail) {
   // non-phi instructions.
 
   auto &Plan = *HeaderVPBB->getPlan();
-  auto *IV = new VPWidenCanonicalIVRecipe(Plan.getCanonicalIV());
+  auto *IV = new VPWidenCanonicalIVRecipe(
+      Plan.getVectorLoopRegion()->getCanonicalIV());
   Builder.setInsertPoint(HeaderVPBB, HeaderVPBB->getFirstNonPhi());
   Builder.insert(IV);
 
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
index c8a2d84a535d3..ef1492320e1c3 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
@@ -501,7 +501,8 @@ static void removeRedundantInductionCasts(VPlan &Plan) {
 /// Try to replace VPWidenCanonicalIVRecipes with a widened canonical IV
 /// recipe, if it exists.
 static void removeRedundantCanonicalIVs(VPlan &Plan) {
-  VPCanonicalIVPHIRecipe *CanonicalIV = Plan.getCanonicalIV();
+  VPRegionBlock *LoopRegion = Plan.getVectorLoopRegion();
+  VPCanonicalIVPHIRecipe *CanonicalIV = LoopRegion->getCanonicalIV();
   VPWidenCanonicalIVRecipe *WidenNewIV = nullptr;
   for (VPUser *U : CanonicalIV->users()) {
     WidenNewIV = dyn_cast<VPWidenCanonicalIVRecipe>(U);
@@ -512,7 +513,7 @@ static void removeRedundantCanonicalIVs(VPlan &Plan) {
   if (!WidenNewIV)
     return;
 
-  VPBasicBlock *HeaderVPBB = Plan.getVectorLoopRegion()->getEntryBasicBlock();
+  VPBasicBlock *HeaderVPBB = LoopRegion->getEntryBasicBlock();
   for (VPRecipeBase &Phi : HeaderVPBB->phis()) {
     auto *WidenOriginalIV = dyn_cast<VPWidenIntOrFpInductionRecipe>(&Phi);
 
@@ -582,8 +583,9 @@ createScalarIVSteps(VPlan &Plan, InductionDescriptor::InductionKind Kind,
                     FPMathOperator *FPBinOp, Instruction *TruncI,
                     VPValue *StartV, VPValue *Step, DebugLoc DL,
                     VPBuilder &Builder) {
-  VPBasicBlock *HeaderVPBB = Plan.getVectorLoopRegion()->getEntryBasicBlock();
-  VPCanonicalIVPHIRecipe *CanonicalIV = Plan.getCanonicalIV();
+  VPRegionBlock *LoopRegion = Plan.getVectorLoopRegion();
+  VPBasicBlock *HeaderVPBB = LoopRegion->getEntryBasicBlock();
+  VPCanonicalIVPHIRecipe *CanonicalIV = LoopRegion->getCanonicalIV();
   VPSingleDefRecipe *BaseIV = Builder.createDerivedIV(
       Kind, FPBinOp, StartV, CanonicalIV, Step, "offset.idx");
 
@@ -800,8 +802,9 @@ static VPValue *optimizeEarlyExitInductionUser(VPlan &Plan,
     return nullptr;
 
   // Calculate the final index.
-  VPValue *EndValue = Plan.getCanonicalIV();
-  auto CanonicalIVType = Plan.getCanonicalIV()->getScalarType();
+  VPRegionBlock *LoopRegion = Plan.getVectorLoopRegion();
+  VPValue *EndValue = LoopRegion->getCanonicalIV();
+  auto CanonicalIVType = LoopRegion->getCanonicalIV()->getScalarType();
   VPBuilder B(cast<VPBasicBlock>(PredVPBB));
 
   DebugLoc DL = cast<VPInstruction>(Op)->getDebugLoc();
@@ -1520,7 +1523,7 @@ static bool isConditionTrueViaVFAndUF(VPValue *Cond, VPlan &Plan,
       return isConditionTrueViaVFAndUF(C, Plan, BestVF, BestUF, SE);
     });
 
-  auto *CanIV = Plan.getCanonicalIV();
+  auto *CanIV = Plan.getVectorLoopRegion()->getCanonicalIV();
   if (!match(Cond, m_SpecificICmp(CmpInst::ICMP_EQ,
                                   m_Specific(CanIV->getBackedgeValue()),
                                   m_Specific(&Plan.getVectorTripCount()))))
@@ -2304,7 +2307,7 @@ static VPActiveLaneMaskPHIRecipe *addVPLaneMaskPhiAndUpdateExitBranch(
     VPlan &Plan, bool DataAndControlFlowWithoutRuntimeCheck) {
   VPRegionBlock *TopRegion = Plan.getVectorLoopRegion();
   VPBasicBlock *EB = TopRegion->getExitingBasicBlock();
-  auto *CanonicalIVPHI = Plan.getCanonicalIV();
+  auto *CanonicalIVPHI = TopRegion->getCanonicalIV();
   VPValue *StartV = CanonicalIVPHI->getStartValue();
 
   auto *CanonicalIVIncrement =
@@ -2343,7 +2346,7 @@ static VPActiveLaneMaskPHIRecipe *addVPLaneMaskPhiAndUpdateExitBranch(
 
   // Create the active lane mask instruction in the VPlan preheader.
   VPValue *ALMMultiplier = Plan.getOrAddLiveIn(
-      ConstantInt::get(Plan.getCanonicalIV()->getScalarType(), 1));
+      ConstantInt::get(TopRegion->getCanonicalIV()->getScalarType(), 1));
   auto *EntryALM = Builder.createNaryOp(VPInstruction::ActiveLaneMask,
                                         {EntryIncrement, TC, ALMMultiplier}, DL,
                                         "active.lane.mask.entry");
@@ -2379,13 +2382,15 @@ static VPActiveLaneMaskPHIRecipe *addVPLaneMaskPhiAndUpdateExitBranch(
 /// TODO: Introduce explicit recipe for header-mask instead of searching
 /// for the header-mask pattern manually.
 static VPSingleDefRecipe *findHeaderMask(VPlan &Plan) {
+  VPRegionBlock *LoopRegion = Plan.getVectorLoopRegion();
   SmallVector<VPValue *> WideCanonicalIVs;
-  auto *FoundWidenCanonicalIVUser = find_if(Plan.getCanonicalIV()->users(),
-                                            IsaPred<VPWidenCanonicalIVRecipe>);
-  assert(count_if(Plan.getCanonicalIV()->users(),
+  auto *FoundWidenCanonicalIVUser = find_if(
+      LoopRegion->getCanonicalIV()->users(), IsaPred<VPWidenCanonicalIVRecipe>);
+  assert(count_if(LoopRegion->getCanonicalIV()->users(),
                   IsaPred<VPWidenCanonicalIVRecipe>) <= 1 &&
          "Must have at most one VPWideCanonicalIVRecipe");
-  if (FoundWidenCanonicalIVUser != Plan.getCanonicalIV()->users().end()) {
+  if (FoundWidenCanonicalIVUser !=
+      LoopRegion->getCanonicalIV()->users().end()) {
     auto *WideCanonicalIV =
         cast<VPWidenCanonicalIVRecipe>(*FoundWidenCanonicalIVUser);
     WideCanonicalIVs.push_back(WideCanonicalIV);
@@ -2393,7 +2398,7 @@ static VPSingleDefRecipe *findHeaderMask(VPlan &Plan) {
 
   // Also include VPWidenIntOrFpInductionRecipes that represent a widened
   // version of the canonical induction.
-  VPBasicBlock *HeaderVPBB = Plan.getVectorLoopRegion()->getEntryBasicBlock();
+  VPBasicBlock *HeaderVPBB = LoopRegion->getEntryBasicBlock();
   for (VPRecipeBase &Phi : HeaderVPBB->phis()) {
     auto *WidenOriginalIV = dyn_cast<VPWidenIntOrFpInductionRecipe>(&Phi);
     if (WidenOriginalIV && WidenOriginalIV->isCanonical())
@@ -2426,8 +2431,9 @@ void VPlanTransforms::addActiveLaneMask(
          "DataAndControlFlowWithoutRuntimeCheck implies "
          "UseActiveLaneMaskForControlFlow");
 
-  auto *FoundWidenCanonicalIVUser = find_if(Plan.getCanonicalIV()->users(),
-                                            IsaPred<VPWidenCanonicalIVRecipe>);
+  VPRegionBlock *LoopRegion = Plan.getVectorLoopRegion();
+  auto *FoundWidenCanonicalIVUser = find_if(
+      LoopRegion->getCanonicalIV()->users(), IsaPred<VPWidenCanonicalIVRecipe>);
   assert(FoundWidenCanonicalIVUser &&
          "Must have widened canonical IV when tail folding!");
   VPSingleDefRecipe *HeaderMask = findHeaderMask(Plan);
@@ -2440,7 +2446,7 @@ void VPlanTransforms::addActiveLaneMask(
   } else {
     VPBuilder B = VPBuilder::getToInsertAfter(WideCanonicalIV);
     VPValue *ALMMultiplier = Plan.getOrAddLiveIn(
-        ConstantInt::get(Plan.getCanonicalIV()->getScalarType(), 1));
+        ConstantInt::get(LoopRegion->getCanonicalIV()->getScalarType(), 1));
     LaneMask =
         B.createNaryOp(VPInstruction::ActiveLaneMask,
                        {WideCanonicalIV, Plan.getTripCount(), ALMMultiplier},
@@ -2550,9 +2556,10 @@ static void transformRecipestoEVLRecipes(VPlan &Plan, VPValue &EVL) {
   });
 
   assert(all_of(Plan.getVFxUF().users(),
-                [&Plan](VPUser *U) {
-                  return match(U, m_c_Add(m_Specific(Plan.getCanonicalIV()),
-                                          m_Specific(&Plan.getVFxUF()))) ||
+                [&LoopRegion, &Plan](VPUser *U) {
+                  return match(U,
+                               m_c_Add(m_Specific(LoopRegion->getCanonicalIV()),
+                                       m_Specific(&Plan.getVFxUF()))) ||
                          isa<VPWidenPointerInductionRecipe>(U);
                 }) &&
          "Only users of VFxUF should be VPWidenPointerInductionRecipe and the "
@@ -2707,9 +2714,10 @@ void VPlanTransforms::addExplicitVectorLength(
     VPlan &Plan, const std::optional<unsigned> &MaxSafeElements) {
   if (Plan.hasScalarVFOnly())
     return;
-  VPBasicBlock *Header = Plan.getVectorLoopRegion()->getEntryBasicBlock();
+  VPRegionBlock *LoopRegion = Plan.getVectorLoopRegion();
+  VPBasicBlock *Header = LoopRegion->getEntryBasicBlock();
 
-  auto *CanonicalIVPHI = Plan.getCanonicalIV();
+  auto *CanonicalIVPHI = LoopRegion->getCanonicalIV();
   auto *CanIVTy = CanonicalIVPHI->getScalarType();
   VPValue *StartV = CanonicalIVPHI->getStartValue();
 
@@ -4153,7 +4161,7 @@ void VPlanTransforms::narrowInterleaveGroups(VPlan &Plan, ElementCount VF,
 
   // Adjust induction to reflect that the transformed plan only processes one
   // original iteration.
-  auto *CanIV = Plan.getCanonicalIV();
+  auto *CanIV = VectorLoop->getCanonicalIV();
   auto *Inc = cast<VPInstruction>(CanIV->getBackedgeValue());
   VPBuilder PHBuilder(Plan.getVectorPreheader());
 
diff --git a/llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp b/llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp
index 5e7f19faebb56..14682d851e765 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp
@@ -69,7 +69,8 @@ class UnrollState {
                                 VPBasicBlock::iterator InsertPtForPhi);
 
   VPValue *getConstantVPV(unsigned Part) {
-    Type *CanIVIntTy = Plan.getCanonicalIV()->getScalarType();
+    Type *CanIVIntTy =
+        Plan.getVectorLoopRegion()->getCanonicalIV()->getScalarType();
     return Plan.getOrAddLiveIn(ConstantInt::get(CanIVIntTy, Part));
   }
 
diff --git a/llvm/lib/Transforms/Vectorize/VPlanUtils.cpp b/llvm/lib/Transforms/Vectorize/VPlanUtils.cpp
index 66748c534f10b..553731731a1d8 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanUtils.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanUtils.cpp
@@ -67,8 +67,10 @@ bool vputils::isHeaderMask(const VPValue *V, VPlan &Plan) {
 
   if (match(V, m_ActiveLaneMask(m_VPValue(A), m_VPValue(B), m_One())))
     return B == Plan.getTripCount() &&
-           (match(A, m_ScalarIVSteps(m_Specific(Plan.getCanonicalIV()), m_One(),
-                                     m_Specific(&Plan.getVF()))) ||
+           (match(A,
+                  m_ScalarIVSteps(
+                      m_Specific(Plan.getVectorLoopRegion()->getCanonicalIV()),
+                      m_One(), m_Specific(&Plan.getVF()))) ||
             IsWideCanonicalIV(A));
 
   return match(V, m_ICmp(m_VPValue(A), m_VPValue(B))) && IsWideCanonicalIV(A) &&
@@ -102,7 +104,8 @@ bool vputils::isUniformAcrossVFsAndUFs(VPValue *V) {
     return all_of(R->operands(), isUniformAcrossVFsAndUFs);
   }
 
-  auto *CanonicalIV = R->getParent()->getPlan()->getCanonicalIV();
+  auto *CanonicalIV =
+      R->getParent()->getPlan()->getVectorLoopRegion()->getCanonicalIV();
   // Canonical IV chain is uniform.
   if (V == CanonicalIV || V == CanonicalIV->getBackedgeValue())
     return true;

@fhahn fhahn mentioned this pull request Oct 11, 2025
Open
@fhahn fhahn force-pushed the vplan-getcanonicaliv-region branch from 56853f5 to f8b7eff Compare October 12, 2025 07:35
@fhahn
[VPlan] Move getCanonicalIV to VPRegionBlock (NFC).
96263b3 
The canonical IV is tied to region blocks; move getCanonicalIV there and
update all users.
ayalz
ayalz approved these changes Oct 12, 2025
View reviewed changes
Copy link
Collaborator

@ayalz ayalz left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This LGTM, raises various suggestions.

llvm/lib/Transforms/Vectorize/VPlan.h Outdated
VPCanonicalIVPHIRecipe *getCanonicalIV() {
VPBasicBlock *EntryVPBB = getEntryBasicBlock();
if (EntryVPBB->empty()) {
// VPlan native path.
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

TODO to resolve this discrepancy?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Added thanks

llvm/lib/Transforms/Vectorize/VPlan.h

/// Remove the current region from its VPlan, connecting its predecessor to
/// its entry, and its exiting block to its successor.
void dissolveToCFGLoop();
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Should dissolveToCFGLoop() also call getCanonicalIV(), possibly asserting that it resides in the-header/its-front/the-region?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Currently getCanonicalIV assumes that the region must have a canonical IV, which may not be the case at the moment (e.g. the inner loop region when vectorizing the outer loop does not have a canonical IV at the moment).

llvm/lib/Transforms/Vectorize/VPlan.h
/// Remove the current region from its VPlan, connecting its predecessor to
/// its entry, and its exiting block to its successor.
void dissolveToCFGLoop();

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Should VPWidenIntOrFpInductionRecipe::isCanonical() also call getCanonicalIV() instead of checking first recipe of header directly?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Updated, thanks

llvm/lib/Transforms/Vectorize/VPlan.h
/// its entry, and its exiting block to its successor.
void dissolveToCFGLoop();

/// Returns the canonical induction recipe of the region.
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Should VPlanVerifier::verify() (continue to) avoid calling getCanonicalIV() as it would assert rather than return false in case of failure?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yep.

llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp Outdated
Comment on lines 806 to 807
VPValue *EndValue = LoopRegion->getCanonicalIV();
auto CanonicalIVType = LoopRegion->getCanonicalIV()->getScalarType();
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Independent. resolving the repeated invocations of getCanonicalIV():

Suggested change
VPValue *EndValue = LoopRegion->getCanonicalIV();
auto CanonicalIVType = LoopRegion->getCanonicalIV()->getScalarType();
auto *CanonicalIV = LoopRegion->getCanonicalIV();
auto CanonicalIVType = CanonicalIV->getScalarType();
VPValue *EndValue = CanonicalIV;

or better define EndValue below using
VPValue *EndValue = B.createNaryOp(Instruction::Add, {CanonicalIV, FirstActiveLane}, DL);

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done thanks

llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp

/// Try to replace VPWidenCanonicalIVRecipes with a widened canonical IV
/// recipe, if it exists.
static void removeRedundantCanonicalIVs(VPlan &Plan) {
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Independent: shorten name or extend implementation to match it

Suggested change
static void removeRedundantCanonicalIVs(VPlan &Plan) {
static void removeRedundantCanonicalIV(VPlan &Plan) {

llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
VPBasicBlock *HeaderVPBB = LoopRegion->getEntryBasicBlock();
for (VPRecipeBase &Phi : HeaderVPBB->phis()) {
auto *WidenOriginalIV = dyn_cast<VPWidenIntOrFpInductionRecipe>(&Phi);

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Independent: if "all users of WidenNewIV demand the first lane only" below - then why is it widened? Holds to recipes generating vectors in general when only their first lane is needed.

llvm/lib/Transforms/Vectorize/VPlanUtils.cpp Outdated

auto *CanonicalIV = R->getParent()->getPlan()->getCanonicalIV();
auto *CanonicalIV =
R->getParent()->getPlan()->getVectorLoopRegion()->getCanonicalIV();
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
R->getParent()->getPlan()->getVectorLoopRegion()->getCanonicalIV();
R->getParent()->getEnclosingLoopRegion()->getCanonicalIV();

?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

done thanks

llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp
Comment on lines +72 to +73
Type *CanIVIntTy =
Plan.getVectorLoopRegion()->getCanonicalIV()->getScalarType();
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Independent: better have TypeInfo supply CanIVIntTy?

llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp
void unrollWidenInductionByUF(VPWidenInductionRecipe *IV,
VPBasicBlock::iterator InsertPtForPhi);

VPValue *getConstantVPV(unsigned Part) {
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Independent:

Suggested change
VPValue *getConstantVPV(unsigned Part) {
VPValue *getConstantIntVPV(unsigned Part) {

@fhahn fhahn force-pushed the vplan-getcanonicaliv-region branch from f8b7eff to 96263b3 Compare October 12, 2025 09:41
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@ayalz ayalz ayalz approved these changes

@rengolin rengolin Awaiting requested review from rengolin

@aniragil aniragil Awaiting requested review from aniragil

Assignees

No one assigned

Labels

llvm:transforms vectorizers

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

3 participants

@fhahn @llvmbot @ayalz