[llvm] [VPlan] Add hidden `-vplan-print-after-all` option (PR #175839)
Andrei Elovikov via llvm-commits
llvm-commits at lists.llvm.org
Tue Jan 27 14:46:31 PST 2026
https://github.com/eas updated https://github.com/llvm/llvm-project/pull/175839
>From a50cbf408691601b062a4b82c53310d15ea5d6de Mon Sep 17 00:00:00 2001
From: Andrei Elovikov <andrei.elovikov at sifive.com>
Date: Tue, 27 Jan 2026 08:35:51 -0800
Subject: [PATCH 1/5] Limit to existing runPass lines only, use macro to get
the name
---
.../Transforms/Vectorize/LoopVectorize.cpp | 68 ++++++++---------
llvm/lib/Transforms/Vectorize/VPlan.cpp | 8 ++
.../Vectorize/VPlanConstruction.cpp | 5 ++
.../Transforms/Vectorize/VPlanTransforms.cpp | 34 ++++-----
.../Transforms/Vectorize/VPlanTransforms.h | 49 +++++++-----
.../LoopVectorize/vplan-print-after-all.ll | 76 +++++++++++++++++++
.../Vectorize/VPlanUncountableExitTest.cpp | 4 +-
7 files changed, 172 insertions(+), 72 deletions(-)
create mode 100644 llvm/test/Transforms/LoopVectorize/vplan-print-after-all.ll
diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index 3cf5e5fb3f818..9016c4aa984e5 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -366,6 +366,10 @@ cl::opt<bool>
cl::Hidden,
cl::desc("Verfiy VPlans after VPlan transforms."));
+cl::opt<bool> llvm::PrintAfterEachVPlanPass(
+ "vplan-print-after-all", cl::init(false), cl::Hidden,
+ cl::desc("Print after each VPlanTransforms::runPass."));
+
// This flag enables the stress testing of the VPlan H-CFG construction in the
// VPlan-native vectorization path. It must be used in conjuction with
// -enable-vplan-native-path. -vplan-verify-hcfg can also be used to enable the
@@ -7380,17 +7384,16 @@ DenseMap<const SCEV *, Value *> LoopVectorizationPlanner::executePlan(
++LoopsEarlyExitVectorized;
// TODO: Move to VPlan transform stage once the transition to the VPlan-based
// cost model is complete for better cost estimates.
- VPlanTransforms::runPass(VPlanTransforms::unrollByUF, BestVPlan, BestUF);
- VPlanTransforms::runPass(VPlanTransforms::materializePacksAndUnpacks,
- BestVPlan);
- VPlanTransforms::runPass(VPlanTransforms::materializeBroadcasts, BestVPlan);
- VPlanTransforms::runPass(VPlanTransforms::replicateByVF, BestVPlan, BestVF);
+ RUN_VPLAN_PASS(VPlanTransforms::unrollByUF, BestVPlan, BestUF);
+ RUN_VPLAN_PASS(VPlanTransforms::materializePacksAndUnpacks, BestVPlan);
+ RUN_VPLAN_PASS(VPlanTransforms::materializeBroadcasts, BestVPlan);
+ RUN_VPLAN_PASS(VPlanTransforms::replicateByVF, BestVPlan, BestVF);
bool HasBranchWeights =
hasBranchWeightMD(*OrigLoop->getLoopLatch()->getTerminator());
if (HasBranchWeights) {
std::optional<unsigned> VScale = CM.getVScaleForTuning();
- VPlanTransforms::runPass(VPlanTransforms::addBranchWeightToMiddleTerminator,
- BestVPlan, BestVF, VScale);
+ RUN_VPLAN_PASS(VPlanTransforms::addBranchWeightToMiddleTerminator,
+ BestVPlan, BestVF, VScale);
}
// Checks are the same for all VPlans, added to BestVPlan only for
@@ -8349,14 +8352,14 @@ void LoopVectorizationPlanner::buildVPlansWithVPRecipes(ElementCount MinVF,
// Now optimize the initial VPlan.
VPlanTransforms::hoistPredicatedLoads(*Plan, PSE, OrigLoop);
VPlanTransforms::sinkPredicatedStores(*Plan, PSE, OrigLoop);
- VPlanTransforms::runPass(VPlanTransforms::truncateToMinimalBitwidths,
- *Plan, CM.getMinimalBitwidths());
- VPlanTransforms::runPass(VPlanTransforms::optimize, *Plan);
+ RUN_VPLAN_PASS(VPlanTransforms::truncateToMinimalBitwidths, *Plan,
+ CM.getMinimalBitwidths());
+ RUN_VPLAN_PASS(VPlanTransforms::optimize, *Plan);
// TODO: try to put addExplicitVectorLength close to addActiveLaneMask
if (CM.foldTailWithEVL()) {
- VPlanTransforms::runPass(VPlanTransforms::addExplicitVectorLength,
- *Plan, CM.getMaxSafeElements());
- VPlanTransforms::runPass(VPlanTransforms::optimizeEVLMasks, *Plan);
+ RUN_VPLAN_PASS(VPlanTransforms::addExplicitVectorLength, *Plan,
+ CM.getMaxSafeElements());
+ RUN_VPLAN_PASS(VPlanTransforms::optimizeEVLMasks, *Plan);
}
assert(verifyVPlanIsValid(*Plan) && "VPlan is invalid");
VPlans.push_back(std::move(Plan));
@@ -8563,18 +8566,15 @@ VPlanPtr LoopVectorizationPlanner::tryToBuildVPlanWithVPRecipes(
// Apply mandatory transformation to handle reductions with multiple in-loop
// uses if possible, bail out otherwise.
- if (!VPlanTransforms::runPass(VPlanTransforms::handleMultiUseReductions,
- *Plan))
+ if (!RUN_VPLAN_PASS(VPlanTransforms::handleMultiUseReductions, *Plan))
return nullptr;
// Apply mandatory transformation to handle FP maxnum/minnum reduction with
// NaNs if possible, bail out otherwise.
- if (!VPlanTransforms::runPass(VPlanTransforms::handleMaxMinNumReductions,
- *Plan))
+ if (!RUN_VPLAN_PASS(VPlanTransforms::handleMaxMinNumReductions, *Plan))
return nullptr;
// Create whole-vector selects for find-last recurrences.
- if (!VPlanTransforms::runPass(VPlanTransforms::handleFindLastReductions,
- *Plan))
+ if (!RUN_VPLAN_PASS(VPlanTransforms::handleFindLastReductions, *Plan))
return nullptr;
// Transform recipes to abstract recipes if it is legal and beneficial and
@@ -8584,8 +8584,8 @@ VPlanPtr LoopVectorizationPlanner::tryToBuildVPlanWithVPRecipes(
if (!CM.foldTailWithEVL()) {
VPCostContext CostCtx(CM.TTI, *CM.TLI, *Plan, CM, CM.CostKind, CM.PSE,
OrigLoop);
- VPlanTransforms::runPass(VPlanTransforms::convertToAbstractRecipes, *Plan,
- CostCtx, Range);
+ RUN_VPLAN_PASS(VPlanTransforms::convertToAbstractRecipes, *Plan, CostCtx,
+ Range);
}
for (ElementCount VF : Range)
@@ -8595,24 +8595,23 @@ VPlanPtr LoopVectorizationPlanner::tryToBuildVPlanWithVPRecipes(
// Interleave memory: for each Interleave Group we marked earlier as relevant
// for this VPlan, replace the Recipes widening its memory instructions with a
// single VPInterleaveRecipe at its insertion point.
- VPlanTransforms::runPass(VPlanTransforms::createInterleaveGroups, *Plan,
- InterleaveGroups, RecipeBuilder,
- CM.isScalarEpilogueAllowed());
+ RUN_VPLAN_PASS(VPlanTransforms::createInterleaveGroups, *Plan,
+ InterleaveGroups, RecipeBuilder, CM.isScalarEpilogueAllowed());
// Replace VPValues for known constant strides.
- VPlanTransforms::runPass(VPlanTransforms::replaceSymbolicStrides, *Plan, PSE,
- Legal->getLAI()->getSymbolicStrides());
+ RUN_VPLAN_PASS(VPlanTransforms::replaceSymbolicStrides, *Plan, PSE,
+ Legal->getLAI()->getSymbolicStrides());
auto BlockNeedsPredication = [this](BasicBlock *BB) {
return Legal->blockNeedsPredication(BB);
};
- VPlanTransforms::runPass(VPlanTransforms::dropPoisonGeneratingRecipes, *Plan,
- BlockNeedsPredication);
+ RUN_VPLAN_PASS(VPlanTransforms::dropPoisonGeneratingRecipes, *Plan,
+ BlockNeedsPredication);
// Sink users of fixed-order recurrence past the recipe defining the previous
// value and introduce FirstOrderRecurrenceSplice VPInstructions.
- if (!VPlanTransforms::runPass(VPlanTransforms::adjustFixedOrderRecurrences,
- *Plan, Builder))
+ if (!RUN_VPLAN_PASS(VPlanTransforms::adjustFixedOrderRecurrences, *Plan,
+ Builder))
return nullptr;
if (useActiveLaneMask(Style)) {
@@ -8895,7 +8894,7 @@ void LoopVectorizationPlanner::addReductionResultComputation(
for (VPRecipeBase *R : ToDelete)
R->eraseFromParent();
- VPlanTransforms::runPass(VPlanTransforms::clearReductionWrapFlags, *Plan);
+ RUN_VPLAN_PASS(VPlanTransforms::clearReductionWrapFlags, *Plan);
}
void LoopVectorizationPlanner::attachRuntimeChecks(
@@ -9283,7 +9282,7 @@ static void preparePlanForMainVectorLoop(VPlan &MainPlan, VPlan &EpiPlan) {
VPIRInst->eraseFromParent();
ResumePhi->eraseFromParent();
}
- VPlanTransforms::runPass(VPlanTransforms::removeDeadRecipes, MainPlan);
+ RUN_VPLAN_PASS(VPlanTransforms::removeDeadRecipes, MainPlan);
using namespace VPlanPatternMatch;
// When vectorizing the epilogue, FindFirstIV & FindLastIV reductions can
@@ -10120,9 +10119,8 @@ bool LoopVectorizePass::processLoop(Loop *L) {
BestPlan);
// TODO: Move to general VPlan pipeline once epilogue loops are also
// supported.
- VPlanTransforms::runPass(
- VPlanTransforms::materializeConstantVectorTripCount, BestPlan, VF.Width,
- IC, PSE);
+ RUN_VPLAN_PASS(VPlanTransforms::materializeConstantVectorTripCount,
+ BestPlan, VF.Width, IC, PSE);
LVP.addMinimumIterationCheck(BestPlan, VF.Width, IC,
VF.MinProfitableTripCount);
diff --git a/llvm/lib/Transforms/Vectorize/VPlan.cpp b/llvm/lib/Transforms/Vectorize/VPlan.cpp
index d1aa1b35ac632..06014ce8afa59 100644
--- a/llvm/lib/Transforms/Vectorize/VPlan.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlan.cpp
@@ -912,6 +912,9 @@ bool VPlan::isExitBlock(VPBlockBase *VPBB) {
return is_contained(ExitBlocks, VPBB);
}
+/// To make RUN_VPLAN_PASS print final VPlan:
+static void printFinalVPlan(VPlan &) {}
+
/// Generate the code inside the preheader and body of the vectorized loop.
/// Assumes a single pre-header basic-block was created for this. Introduce
/// additional basic-blocks as needed, and fill them all.
@@ -933,7 +936,12 @@ void VPlan::execute(VPTransformState *State) {
LLVM_DEBUG(dbgs() << "Executing best plan with VF=" << State->VF
<< ", UF=" << getUF() << '\n');
setName("Final VPlan");
+ // TODO: -debug-only=loop-vectorize should probably just imply
+ // `-vplan-print-after-all`, but the latter is much more verbose than the
+ // former, so doing that would require updating existing tests and that hasn't
+ // been done (yet). For now, just do both:
LLVM_DEBUG(dump());
+ RUN_VPLAN_PASS(printFinalVPlan, *this);
BasicBlock *ScalarPh = State->CFG.ExitBB;
VPBasicBlock *ScalarPhVPBB = getScalarPreheader();
diff --git a/llvm/lib/Transforms/Vectorize/VPlanConstruction.cpp b/llvm/lib/Transforms/Vectorize/VPlanConstruction.cpp
index 4fc7b2fe1c2de..d40404b6a1e4e 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanConstruction.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanConstruction.cpp
@@ -586,6 +586,9 @@ static void simplifyLiveInsWithSCEV(VPlan &Plan,
}
}
+/// To make RUN_VPLAN_PASS print initial VPlan:
+static void printAfterInitialConstruction(VPlan &) {}
+
std::unique_ptr<VPlan>
VPlanTransforms::buildVPlan0(Loop *TheLoop, LoopInfo &LI, Type *InductionTy,
DebugLoc IVDL, PredicatedScalarEvolution &PSE,
@@ -594,6 +597,8 @@ VPlanTransforms::buildVPlan0(Loop *TheLoop, LoopInfo &LI, Type *InductionTy,
std::unique_ptr<VPlan> VPlan0 = Builder.buildPlainCFG();
addInitialSkeleton(*VPlan0, InductionTy, IVDL, PSE, TheLoop);
simplifyLiveInsWithSCEV(*VPlan0, PSE);
+
+ RUN_VPLAN_PASS(printAfterInitialConstruction, *VPlan0);
return VPlan0;
}
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
index 90e32b9e1ce1d..4f5635d9341b2 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
@@ -2675,23 +2675,23 @@ void VPlanTransforms::removeBranchOnConst(VPlan &Plan) {
}
void VPlanTransforms::optimize(VPlan &Plan) {
- runPass(removeRedundantCanonicalIVs, Plan);
- runPass(removeRedundantInductionCasts, Plan);
-
- runPass(simplifyRecipes, Plan);
- runPass(removeDeadRecipes, Plan);
- runPass(simplifyBlends, Plan);
- runPass(legalizeAndOptimizeInductions, Plan);
- runPass(narrowToSingleScalarRecipes, Plan);
- runPass(removeRedundantExpandSCEVRecipes, Plan);
- runPass(simplifyRecipes, Plan);
- runPass(removeBranchOnConst, Plan);
- runPass(removeDeadRecipes, Plan);
-
- runPass(createAndOptimizeReplicateRegions, Plan);
- runPass(hoistInvariantLoads, Plan);
- runPass(mergeBlocksIntoPredecessors, Plan);
- runPass(licm, Plan);
+ RUN_VPLAN_PASS(removeRedundantCanonicalIVs, Plan);
+ RUN_VPLAN_PASS(removeRedundantInductionCasts, Plan);
+
+ RUN_VPLAN_PASS(simplifyRecipes, Plan);
+ RUN_VPLAN_PASS(removeDeadRecipes, Plan);
+ RUN_VPLAN_PASS(simplifyBlends, Plan);
+ RUN_VPLAN_PASS(legalizeAndOptimizeInductions, Plan);
+ RUN_VPLAN_PASS(narrowToSingleScalarRecipes, Plan);
+ RUN_VPLAN_PASS(removeRedundantExpandSCEVRecipes, Plan);
+ RUN_VPLAN_PASS(simplifyRecipes, Plan);
+ RUN_VPLAN_PASS(removeBranchOnConst, Plan);
+ RUN_VPLAN_PASS(removeDeadRecipes, Plan);
+
+ RUN_VPLAN_PASS(createAndOptimizeReplicateRegions, Plan);
+ RUN_VPLAN_PASS(hoistInvariantLoads, Plan);
+ RUN_VPLAN_PASS(mergeBlocksIntoPredecessors, Plan);
+ RUN_VPLAN_PASS(licm, Plan);
}
// Add a VPActiveLaneMaskPHIRecipe and related recipes to \p Plan and replace
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.h b/llvm/lib/Transforms/Vectorize/VPlanTransforms.h
index e0d09a099647a..15f03fb9d3b19 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.h
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.h
@@ -33,28 +33,41 @@ class VPRecipeBuilder;
struct VFRange;
LLVM_ABI_FOR_TEST extern cl::opt<bool> VerifyEachVPlan;
+LLVM_ABI_FOR_TEST extern cl::opt<bool> PrintAfterEachVPlanPass;
LLVM_ABI_FOR_TEST extern cl::opt<bool> EnableWideActiveLaneMask;
struct VPlanTransforms {
- /// Helper to run a VPlan transform \p Transform on \p VPlan, forwarding extra
- /// arguments to the transform. Returns the boolean returned by the transform.
- template <typename... ArgsTy>
- static bool runPass(bool (*Transform)(VPlan &, ArgsTy...), VPlan &Plan,
- typename std::remove_reference<ArgsTy>::type &...Args) {
- bool Res = Transform(Plan, Args...);
- if (VerifyEachVPlan)
- verifyVPlanIsValid(Plan);
- return Res;
- }
- /// Helper to run a VPlan transform \p Transform on \p VPlan, forwarding extra
- /// arguments to the transform.
- template <typename... ArgsTy>
- static void runPass(void (*Fn)(VPlan &, ArgsTy...), VPlan &Plan,
- typename std::remove_reference<ArgsTy>::type &...Args) {
- Fn(Plan, Args...);
- if (VerifyEachVPlan)
- verifyVPlanIsValid(Plan);
+ /// Helper to run a VPlan pass \p Pass on \p VPlan, forwarding extra arguments
+ /// to the pass. Performs verification/printing after each VPlan pass if
+ /// requested via command line options.
+ template <typename PassTy, typename... ArgsTy>
+ static decltype(auto) runPass(StringRef PassName, PassTy &&Pass, VPlan &Plan,
+ ArgsTy &&...Args) {
+ auto PostTransformActions = [&]() {
+ // Make sure to print before verification, so that output is more useful
+ // in case of failures:
+ if (PrintAfterEachVPlanPass) {
+ dbgs() << "VPlan after " << PassName << '\n';
+ dbgs() << Plan << '\n';
+ }
+ if (VerifyEachVPlan)
+ verifyVPlanIsValid(Plan);
+ };
+
+ using ResTy = decltype(std::forward<PassTy>(Pass)(
+ Plan, std::forward<ArgsTy>(Args)...));
+ if constexpr (std::is_same_v<ResTy, void>) {
+ std::forward<PassTy>(Pass)(Plan, std::forward<ArgsTy>(Args)...);
+ PostTransformActions();
+ } else {
+ decltype(auto) Res =
+ std::forward<PassTy>(Pass)(Plan, std::forward<ArgsTy>(Args)...);
+ PostTransformActions();
+ return Res;
+ }
}
+#define RUN_VPLAN_PASS(PASS, ...) \
+ llvm::VPlanTransforms::runPass(#PASS, PASS, __VA_ARGS__)
/// Create a base VPlan0, serving as the common starting point for all later
/// candidates. It consists of an initial plain CFG loop with loop blocks from
diff --git a/llvm/test/Transforms/LoopVectorize/vplan-print-after-all.ll b/llvm/test/Transforms/LoopVectorize/vplan-print-after-all.ll
new file mode 100644
index 0000000000000..34f576ed31da5
--- /dev/null
+++ b/llvm/test/Transforms/LoopVectorize/vplan-print-after-all.ll
@@ -0,0 +1,76 @@
+; REQUIRES: asserts
+
+; Verify that `-vplan-print-after-all` option works.
+
+; RUN: opt -passes=loop-vectorize -disable-output -vplan-print-after-all -force-vector-width=4 < %s 2>&1 | FileCheck %s --implicit-check-not "VPlan after"
+
+; CHECK: VPlan after printAfterInitialConstruction
+; CHECK: VPlan after VPlanTransforms::clearReductionWrapFlags
+; CHECK: VPlan after VPlanTransforms::handleMultiUseReductions
+; CHECK: VPlan after VPlanTransforms::handleMaxMinNumReductions
+; CHECK: VPlan after VPlanTransforms::handleFindLastReductions
+; CHECK: VPlan after VPlanTransforms::convertToAbstractRecipes
+; CHECK: VPlan after VPlanTransforms::createInterleaveGroups
+; CHECK: VPlan after VPlanTransforms::replaceSymbolicStrides
+; CHECK: VPlan after VPlanTransforms::dropPoisonGeneratingRecipes
+; CHECK: VPlan after VPlanTransforms::adjustFixedOrderRecurrences
+; CHECK: VPlan after VPlanTransforms::truncateToMinimalBitwidths
+; CHECK: VPlan after removeRedundantCanonicalIVs
+; CHECK: VPlan after removeRedundantInductionCasts
+; CHECK: VPlan after simplifyRecipes
+; CHECK: VPlan after removeDeadRecipes
+; CHECK: VPlan after simplifyBlends
+; CHECK: VPlan after legalizeAndOptimizeInductions
+; CHECK: VPlan after narrowToSingleScalarRecipes
+; CHECK: VPlan after removeRedundantExpandSCEVRecipes
+; CHECK: VPlan after simplifyRecipes
+; CHECK: VPlan after removeBranchOnConst
+; CHECK: VPlan after removeDeadRecipes
+; CHECK: VPlan after createAndOptimizeReplicateRegions
+; CHECK: VPlan after hoistInvariantLoads
+; CHECK: VPlan after mergeBlocksIntoPredecessors
+; CHECK: VPlan after licm
+; CHECK: VPlan after VPlanTransforms::optimize
+; CHECK: VPlan after VPlanTransforms::materializeConstantVectorTripCount
+; CHECK: VPlan after VPlanTransforms::unrollByUF
+; CHECK: VPlan after VPlanTransforms::materializePacksAndUnpacks
+; CHECK: VPlan after VPlanTransforms::materializeBroadcasts
+; CHECK: VPlan after VPlanTransforms::replicateByVF
+; CHECK: VPlan after printFinalVPlan
+; Also verify that VPlan is actually printed:
+; CHECK-NEXT: VPlan 'Final VPlan for VF={4},UF={1}' {
+; CHECK-NEXT: Live-in ir<%smax> = original trip-count
+; CHECK-EMPTY:
+; CHECK-NEXT: ir-bb<entry>:
+; CHECK-NEXT: IR %smax = call i64 @llvm.smax.i64(i64 %n, i64 1)
+; CHECK-NEXT: EMIT vp<%min.iters.check> = icmp ult ir<%smax>, ir<4>
+; CHECK-NEXT: EMIT branch-on-cond vp<%min.iters.check>
+; CHECK-NEXT: Successor(s): ir-bb<scalar.ph>, vector.ph
+; CHECK-EMPTY:
+; CHECK-NEXT: vector.ph:
+; CHECK-NEXT: EMIT vp<%n.mod.vf> = urem ir<%smax>, ir<4>
+; CHECK-NEXT: EMIT vp<%n.vec> = sub ir<%smax>, vp<%n.mod.vf>
+; CHECK-NEXT: EMIT vp<%3> = step-vector i64
+; CHECK-NEXT: EMIT vp<%4> = broadcast ir<4>
+; CHECK-NEXT: Successor(s): vector.body
+; CHECK-EMPTY:
+; CHECK-NEXT: vector.body:
+; CHECK-NEXT: EMIT-SCALAR vp<%index> = phi [ ir<0>, vector.ph ], [ vp<%index.next>, vector.body ]
+; CHECK-NEXT: WIDEN-PHI ir<%iv> = phi [ vp<%3>, vector.ph ], [ vp<%vec.ind.next>, vector.body ]
+; CHECK-NEXT: CLONE ir<%gep> = getelementptr ir<%ptr>, vp<%index>
+
+define void @foo(ptr dereferenceable(1024) %ptr, i64 %n) {
+entry:
+ br label %header
+
+header:
+ %iv = phi i64 [ 0, %entry ], [ %iv.next, %header ]
+ %gep = getelementptr i64, ptr %ptr, i64 %iv
+ store i64 %iv, ptr %gep
+ %iv.next = add nsw i64 %iv, 1
+ %exitcond = icmp slt i64 %iv.next, %n
+ br i1 %exitcond, label %header, label %exit
+
+exit:
+ ret void
+}
diff --git a/llvm/unittests/Transforms/Vectorize/VPlanUncountableExitTest.cpp b/llvm/unittests/Transforms/Vectorize/VPlanUncountableExitTest.cpp
index 270537a333b8e..f895a04bff786 100644
--- a/llvm/unittests/Transforms/Vectorize/VPlanUncountableExitTest.cpp
+++ b/llvm/unittests/Transforms/Vectorize/VPlanUncountableExitTest.cpp
@@ -48,7 +48,7 @@ TEST_F(VPUncountableExitTest, FindUncountableExitRecipes) {
BasicBlock *LoopHeader = F->getEntryBlock().getSingleSuccessor();
auto Plan = buildVPlan(LoopHeader, /*HasUncountableExit=*/true);
VPlanTransforms::tryToConvertVPInstructionsToVPRecipes(*Plan, *TLI);
- VPlanTransforms::runPass(VPlanTransforms::optimize, *Plan);
+ RUN_VPLAN_PASS(VPlanTransforms::optimize, *Plan);
SmallVector<VPRecipeBase *> Recipes;
SmallVector<VPRecipeBase *> GEPs;
@@ -84,7 +84,7 @@ TEST_F(VPUncountableExitTest, NoUncountableExit) {
BasicBlock *LoopHeader = F->getEntryBlock().getSingleSuccessor();
auto Plan = buildVPlan(LoopHeader);
VPlanTransforms::tryToConvertVPInstructionsToVPRecipes(*Plan, *TLI);
- VPlanTransforms::runPass(VPlanTransforms::optimize, *Plan);
+ RUN_VPLAN_PASS(VPlanTransforms::optimize, *Plan);
SmallVector<VPRecipeBase *> Recipes;
SmallVector<VPRecipeBase *> GEPs;
>From aea06bd607dd4e6691500a4e70c597eca65f1e6a Mon Sep 17 00:00:00 2001
From: Andrei Elovikov <andrei.elovikov at sifive.com>
Date: Tue, 27 Jan 2026 09:41:52 -0800
Subject: [PATCH 2/5] Initial VPlan won't pass verifier
---
llvm/lib/Transforms/Vectorize/VPlanConstruction.cpp | 2 +-
llvm/lib/Transforms/Vectorize/VPlanTransforms.h | 6 ++++--
llvm/test/Transforms/LoopVectorize/vplan-print-after-all.ll | 2 +-
3 files changed, 6 insertions(+), 4 deletions(-)
diff --git a/llvm/lib/Transforms/Vectorize/VPlanConstruction.cpp b/llvm/lib/Transforms/Vectorize/VPlanConstruction.cpp
index d40404b6a1e4e..379d9d3e15795 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanConstruction.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanConstruction.cpp
@@ -598,7 +598,7 @@ VPlanTransforms::buildVPlan0(Loop *TheLoop, LoopInfo &LI, Type *InductionTy,
addInitialSkeleton(*VPlan0, InductionTy, IVDL, PSE, TheLoop);
simplifyLiveInsWithSCEV(*VPlan0, PSE);
- RUN_VPLAN_PASS(printAfterInitialConstruction, *VPlan0);
+ RUN_VPLAN_PASS_NO_VERIFY(printAfterInitialConstruction, *VPlan0);
return VPlan0;
}
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.h b/llvm/lib/Transforms/Vectorize/VPlanTransforms.h
index 15f03fb9d3b19..11544d95dc3c5 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.h
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.h
@@ -40,7 +40,7 @@ struct VPlanTransforms {
/// Helper to run a VPlan pass \p Pass on \p VPlan, forwarding extra arguments
/// to the pass. Performs verification/printing after each VPlan pass if
/// requested via command line options.
- template <typename PassTy, typename... ArgsTy>
+ template <bool EnableVerify = true, typename PassTy, typename... ArgsTy>
static decltype(auto) runPass(StringRef PassName, PassTy &&Pass, VPlan &Plan,
ArgsTy &&...Args) {
auto PostTransformActions = [&]() {
@@ -50,7 +50,7 @@ struct VPlanTransforms {
dbgs() << "VPlan after " << PassName << '\n';
dbgs() << Plan << '\n';
}
- if (VerifyEachVPlan)
+ if (VerifyEachVPlan && EnableVerify)
verifyVPlanIsValid(Plan);
};
@@ -68,6 +68,8 @@ struct VPlanTransforms {
}
#define RUN_VPLAN_PASS(PASS, ...) \
llvm::VPlanTransforms::runPass(#PASS, PASS, __VA_ARGS__)
+#define RUN_VPLAN_PASS_NO_VERIFY(PASS, ...) \
+ llvm::VPlanTransforms::runPass<false>(#PASS, PASS, __VA_ARGS__)
/// Create a base VPlan0, serving as the common starting point for all later
/// candidates. It consists of an initial plain CFG loop with loop blocks from
diff --git a/llvm/test/Transforms/LoopVectorize/vplan-print-after-all.ll b/llvm/test/Transforms/LoopVectorize/vplan-print-after-all.ll
index 34f576ed31da5..9b59afbf8a6c6 100644
--- a/llvm/test/Transforms/LoopVectorize/vplan-print-after-all.ll
+++ b/llvm/test/Transforms/LoopVectorize/vplan-print-after-all.ll
@@ -2,7 +2,7 @@
; Verify that `-vplan-print-after-all` option works.
-; RUN: opt -passes=loop-vectorize -disable-output -vplan-print-after-all -force-vector-width=4 < %s 2>&1 | FileCheck %s --implicit-check-not "VPlan after"
+; RUN: opt -passes=loop-vectorize -disable-output -vplan-print-after-all -force-vector-width=4 -vplan-verify-each < %s 2>&1 | FileCheck %s --implicit-check-not "VPlan after"
; CHECK: VPlan after printAfterInitialConstruction
; CHECK: VPlan after VPlanTransforms::clearReductionWrapFlags
>From 047bc7096288046cd184104627061c5c8d20b24c Mon Sep 17 00:00:00 2001
From: Andrei Elovikov <andrei.elovikov at sifive.com>
Date: Tue, 27 Jan 2026 14:29:58 -0800
Subject: [PATCH 3/5] Update TODO comment about "-debug"
---
llvm/lib/Transforms/Vectorize/VPlan.cpp | 7 +++----
1 file changed, 3 insertions(+), 4 deletions(-)
diff --git a/llvm/lib/Transforms/Vectorize/VPlan.cpp b/llvm/lib/Transforms/Vectorize/VPlan.cpp
index 06014ce8afa59..3055d5aac74ae 100644
--- a/llvm/lib/Transforms/Vectorize/VPlan.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlan.cpp
@@ -936,10 +936,9 @@ void VPlan::execute(VPTransformState *State) {
LLVM_DEBUG(dbgs() << "Executing best plan with VF=" << State->VF
<< ", UF=" << getUF() << '\n');
setName("Final VPlan");
- // TODO: -debug-only=loop-vectorize should probably just imply
- // `-vplan-print-after-all`, but the latter is much more verbose than the
- // former, so doing that would require updating existing tests and that hasn't
- // been done (yet). For now, just do both:
+ // TODO: RUN_VPLAN_PASS/VPlanTransforms::runPass should automatically dump
+ // VPlans after some specific stages when "-debug" is specified, but that
+ // hasn't been implemented yet. For now, just do both:
LLVM_DEBUG(dump());
RUN_VPLAN_PASS(printFinalVPlan, *this);
>From 2146c90f781de6a428c30486bb9315dbd1ba0aa9 Mon Sep 17 00:00:00 2001
From: Andrei Elovikov <andrei.elovikov at sifive.com>
Date: Tue, 27 Jan 2026 14:31:19 -0800
Subject: [PATCH 4/5] Don't use RUN_VPLAN_PASS in llvm/unittests/**
---
.../Transforms/Vectorize/VPlanUncountableExitTest.cpp | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
diff --git a/llvm/unittests/Transforms/Vectorize/VPlanUncountableExitTest.cpp b/llvm/unittests/Transforms/Vectorize/VPlanUncountableExitTest.cpp
index f895a04bff786..c701382c54e6b 100644
--- a/llvm/unittests/Transforms/Vectorize/VPlanUncountableExitTest.cpp
+++ b/llvm/unittests/Transforms/Vectorize/VPlanUncountableExitTest.cpp
@@ -48,7 +48,7 @@ TEST_F(VPUncountableExitTest, FindUncountableExitRecipes) {
BasicBlock *LoopHeader = F->getEntryBlock().getSingleSuccessor();
auto Plan = buildVPlan(LoopHeader, /*HasUncountableExit=*/true);
VPlanTransforms::tryToConvertVPInstructionsToVPRecipes(*Plan, *TLI);
- RUN_VPLAN_PASS(VPlanTransforms::optimize, *Plan);
+ VPlanTransforms::optimize(*Plan);
SmallVector<VPRecipeBase *> Recipes;
SmallVector<VPRecipeBase *> GEPs;
@@ -84,7 +84,7 @@ TEST_F(VPUncountableExitTest, NoUncountableExit) {
BasicBlock *LoopHeader = F->getEntryBlock().getSingleSuccessor();
auto Plan = buildVPlan(LoopHeader);
VPlanTransforms::tryToConvertVPInstructionsToVPRecipes(*Plan, *TLI);
- RUN_VPLAN_PASS(VPlanTransforms::optimize, *Plan);
+ VPlanTransforms::optimize(*Plan);
SmallVector<VPRecipeBase *> Recipes;
SmallVector<VPRecipeBase *> GEPs;
>From 7cc41355a06e69acef43b1ddb963bff86eeaea6a Mon Sep 17 00:00:00 2001
From: Andrei Elovikov <andrei.elovikov at sifive.com>
Date: Tue, 27 Jan 2026 14:44:37 -0800
Subject: [PATCH 5/5] Update test
---
.../LoopVectorize/vplan-print-after-all.ll | 39 +++++++------------
1 file changed, 15 insertions(+), 24 deletions(-)
diff --git a/llvm/test/Transforms/LoopVectorize/vplan-print-after-all.ll b/llvm/test/Transforms/LoopVectorize/vplan-print-after-all.ll
index 9b59afbf8a6c6..96e15740cd413 100644
--- a/llvm/test/Transforms/LoopVectorize/vplan-print-after-all.ll
+++ b/llvm/test/Transforms/LoopVectorize/vplan-print-after-all.ll
@@ -1,9 +1,9 @@
+; RUN: opt -passes=loop-vectorize -disable-output -vplan-print-after-all -force-vector-width=4 -vplan-verify-each < %s 2>&1 | FileCheck %s --implicit-check-not "VPlan after"
+; RUN: opt -passes=loop-vectorize -disable-output -vplan-print-after-all -force-vector-width=4 -vplan-verify-each < %s 2>&1 | FileCheck %s --check-prefix CHECK-DUMP
; REQUIRES: asserts
; Verify that `-vplan-print-after-all` option works.
-; RUN: opt -passes=loop-vectorize -disable-output -vplan-print-after-all -force-vector-width=4 -vplan-verify-each < %s 2>&1 | FileCheck %s --implicit-check-not "VPlan after"
-
; CHECK: VPlan after printAfterInitialConstruction
; CHECK: VPlan after VPlanTransforms::clearReductionWrapFlags
; CHECK: VPlan after VPlanTransforms::handleMultiUseReductions
@@ -37,29 +37,20 @@
; CHECK: VPlan after VPlanTransforms::materializeBroadcasts
; CHECK: VPlan after VPlanTransforms::replicateByVF
; CHECK: VPlan after printFinalVPlan
-; Also verify that VPlan is actually printed:
-; CHECK-NEXT: VPlan 'Final VPlan for VF={4},UF={1}' {
-; CHECK-NEXT: Live-in ir<%smax> = original trip-count
-; CHECK-EMPTY:
-; CHECK-NEXT: ir-bb<entry>:
-; CHECK-NEXT: IR %smax = call i64 @llvm.smax.i64(i64 %n, i64 1)
-; CHECK-NEXT: EMIT vp<%min.iters.check> = icmp ult ir<%smax>, ir<4>
-; CHECK-NEXT: EMIT branch-on-cond vp<%min.iters.check>
-; CHECK-NEXT: Successor(s): ir-bb<scalar.ph>, vector.ph
-; CHECK-EMPTY:
-; CHECK-NEXT: vector.ph:
-; CHECK-NEXT: EMIT vp<%n.mod.vf> = urem ir<%smax>, ir<4>
-; CHECK-NEXT: EMIT vp<%n.vec> = sub ir<%smax>, vp<%n.mod.vf>
-; CHECK-NEXT: EMIT vp<%3> = step-vector i64
-; CHECK-NEXT: EMIT vp<%4> = broadcast ir<4>
-; CHECK-NEXT: Successor(s): vector.body
-; CHECK-EMPTY:
-; CHECK-NEXT: vector.body:
-; CHECK-NEXT: EMIT-SCALAR vp<%index> = phi [ ir<0>, vector.ph ], [ vp<%index.next>, vector.body ]
-; CHECK-NEXT: WIDEN-PHI ir<%iv> = phi [ vp<%3>, vector.ph ], [ vp<%vec.ind.next>, vector.body ]
-; CHECK-NEXT: CLONE ir<%gep> = getelementptr ir<%ptr>, vp<%index>
-define void @foo(ptr dereferenceable(1024) %ptr, i64 %n) {
+; Also verify that VPlans are actually printed (we aren't interested in the
+; exact dump content, just that it's performed):
+
+; CHECK-DUMP: VPlan after printAfterInitialConstruction
+; CHECK-DUMP-NEXT: VPlan ' for UF>=1' {
+;
+; CHECK-DUMP: VPlan after VPlanTransforms::optimize
+; CHECK-DUMP-NEXT: VPlan 'Initial VPlan for VF={4},UF>=1' {
+;
+; CHECK-DUMP: VPlan after printFinalVPlan
+; CHECK-DUMP-NEXT: VPlan 'Final VPlan for VF={4},UF={1}' {
+
+define void @foo(ptr %ptr, i64 %n) {
entry:
br label %header
More information about the llvm-commits
mailing list