[VPlan] Add ExtractLastLanePerPart, use in narrowToSingleScalar. #163056
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
When narrowing stores of a single-scalar, we currently use ExtractLastElement, which extracts the last element across all parts. This is not correct if the store's address is not uniform across all parts. If it is only uniform-per-part, the last lane per part must be extracted. Add a new ExtractLastLanePerPart opcode to handle this correctly. Most transforms apply to both ExtractLastElement and ExtractLastLanePerPart, with the only difference being their treatment during unrolling. Fixes llvm#162498.
|
@llvm/pr-subscribers-vectorizers @llvm/pr-subscribers-llvm-transforms Author: Florian Hahn (fhahn) ChangesWhen narrowing stores of a single-scalar, we currently use ExtractLastElement, which extracts the last element across all parts. This is not correct if the store's address is not uniform across all parts. If it is only uniform-per-part, the last lane per part must be extracted. Add a new ExtractLastLanePerPart opcode to handle this correctly. Most transforms apply to both ExtractLastElement and ExtractLastLanePerPart, with the only difference being their treatment during unrolling. Fixes #162498. Full diff: https://github.com/llvm/llvm-project/pull/163056.diff 7 Files Affected:
diff --git a/llvm/lib/Transforms/Vectorize/VPlan.h b/llvm/lib/Transforms/Vectorize/VPlan.h
index fb696bea671af..fc2e7f81baa31 100644
--- a/llvm/lib/Transforms/Vectorize/VPlan.h
+++ b/llvm/lib/Transforms/Vectorize/VPlan.h
@@ -1021,6 +1021,8 @@ class LLVM_ABI_FOR_TEST VPInstruction : public VPRecipeWithIRFlags,
// part if scalar. In the latter case, the recipe will be removed during
// unrolling.
ExtractLastElement,
+ // Extracts the last lane for each part from its operand.
+ ExtractLastLanePerPart,
// Extracts the second-to-last lane from its operand or the second-to-last
// part if it is scalar. In the latter case, the recipe will be removed
// during unrolling.
diff --git a/llvm/lib/Transforms/Vectorize/VPlanAnalysis.cpp b/llvm/lib/Transforms/Vectorize/VPlanAnalysis.cpp
index 07bfe7a896d86..f413c63c6d14c 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanAnalysis.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanAnalysis.cpp
@@ -116,6 +116,7 @@ Type *VPTypeAnalysis::inferScalarTypeForRecipe(const VPInstruction *R) {
case VPInstruction::FirstActiveLane:
return Type::getIntNTy(Ctx, 64);
case VPInstruction::ExtractLastElement:
+ case VPInstruction::ExtractLastLanePerPart:
case VPInstruction::ExtractPenultimateElement: {
Type *BaseTy = inferScalarType(R->getOperand(0));
if (auto *VecTy = dyn_cast<VectorType>(BaseTy))
diff --git a/llvm/lib/Transforms/Vectorize/VPlanPatternMatch.h b/llvm/lib/Transforms/Vectorize/VPlanPatternMatch.h
index 555efea1ea840..ef51d46c65c5c 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanPatternMatch.h
+++ b/llvm/lib/Transforms/Vectorize/VPlanPatternMatch.h
@@ -368,6 +368,12 @@ m_ExtractLastElement(const Op0_t &Op0) {
return m_VPInstruction<VPInstruction::ExtractLastElement>(Op0);
}
+template <typename Op0_t>
+inline VPInstruction_match<VPInstruction::ExtractLastLanePerPart, Op0_t>
+m_ExtractLastLanePerPart(const Op0_t &Op0) {
+ return m_VPInstruction<VPInstruction::ExtractLastLanePerPart>(Op0);
+}
+
template <typename Op0_t, typename Op1_t, typename Op2_t>
inline VPInstruction_match<VPInstruction::ActiveLaneMask, Op0_t, Op1_t, Op2_t>
m_ActiveLaneMask(const Op0_t &Op0, const Op1_t &Op1, const Op2_t &Op2) {
diff --git a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
index 8e91677292788..585c84b938a71 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
@@ -511,6 +511,7 @@ unsigned VPInstruction::getNumOperandsForOpcode(unsigned Opcode) {
case VPInstruction::CanonicalIVIncrementForPart:
case VPInstruction::ExplicitVectorLength:
case VPInstruction::ExtractLastElement:
+ case VPInstruction::ExtractLastLanePerPart:
case VPInstruction::ExtractPenultimateElement:
case VPInstruction::FirstActiveLane:
case VPInstruction::Not:
@@ -878,9 +879,11 @@ Value *VPInstruction::generate(VPTransformState &State) {
return ReducedPartRdx;
}
+ case VPInstruction::ExtractLastLanePerPart:
case VPInstruction::ExtractLastElement:
case VPInstruction::ExtractPenultimateElement: {
- unsigned Offset = getOpcode() == VPInstruction::ExtractLastElement ? 1 : 2;
+ unsigned Offset =
+ getOpcode() == VPInstruction::ExtractPenultimateElement ? 2 : 1;
Value *Res;
if (State.VF.isVector()) {
assert(Offset <= State.VF.getKnownMinValue() &&
@@ -1166,6 +1169,7 @@ InstructionCost VPInstruction::computeCost(ElementCount VF,
bool VPInstruction::isVectorToScalar() const {
return getOpcode() == VPInstruction::ExtractLastElement ||
+ getOpcode() == VPInstruction::ExtractLastLanePerPart ||
getOpcode() == VPInstruction::ExtractPenultimateElement ||
getOpcode() == Instruction::ExtractElement ||
getOpcode() == VPInstruction::ExtractLane ||
@@ -1229,6 +1233,7 @@ bool VPInstruction::opcodeMayReadOrWriteFromMemory() const {
case VPInstruction::CanonicalIVIncrementForPart:
case VPInstruction::ExtractLane:
case VPInstruction::ExtractLastElement:
+ case VPInstruction::ExtractLastLanePerPart:
case VPInstruction::ExtractPenultimateElement:
case VPInstruction::ActiveLaneMask:
case VPInstruction::FirstActiveLane:
@@ -1376,6 +1381,9 @@ void VPInstruction::print(raw_ostream &O, const Twine &Indent,
case VPInstruction::ExtractLastElement:
O << "extract-last-element";
break;
+ case VPInstruction::ExtractLastLanePerPart:
+ O << "extract-last-lane-per-part";
+ break;
case VPInstruction::ExtractPenultimateElement:
O << "extract-penultimate-element";
break;
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
index 7563cd719b19c..3bc7f58449e29 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
@@ -1207,7 +1207,8 @@ static void simplifyRecipe(VPRecipeBase &R, VPTypeAnalysis &TypeInfo) {
}
// Look through ExtractLastElement (BuildVector ....).
- if (match(&R, m_ExtractLastElement(m_BuildVector()))) {
+ if (match(&R, m_CombineOr(m_ExtractLastElement(m_BuildVector()),
+ m_ExtractLastLanePerPart(m_BuildVector())))) {
auto *BuildVector = cast<VPInstruction>(R.getOperand(0));
Def->replaceAllUsesWith(
BuildVector->getOperand(BuildVector->getNumOperands() - 1));
@@ -1273,13 +1274,16 @@ static void simplifyRecipe(VPRecipeBase &R, VPTypeAnalysis &TypeInfo) {
return;
}
- if (match(Def, m_ExtractLastElement(m_Broadcast(m_VPValue(A))))) {
+ if (match(Def,
+ m_CombineOr(m_ExtractLastElement(m_Broadcast(m_VPValue(A))),
+ m_ExtractLastLanePerPart(m_Broadcast(m_VPValue(A)))))) {
Def->replaceAllUsesWith(A);
return;
}
- if (match(Def,
- m_VPInstruction<VPInstruction::ExtractLastElement>(m_VPValue(A))) &&
+ if (match(Def, m_CombineOr(m_VPInstruction<VPInstruction::ExtractLastElement>(
+ m_VPValue(A)),
+ m_ExtractLastLanePerPart(m_VPValue(A)))) &&
((isa<VPInstruction>(A) && vputils::isSingleScalar(A)) ||
(isa<VPReplicateRecipe>(A) &&
cast<VPReplicateRecipe>(A)->isSingleScalar())) &&
@@ -1287,6 +1291,12 @@ static void simplifyRecipe(VPRecipeBase &R, VPTypeAnalysis &TypeInfo) {
[Def, A](VPUser *U) { return U->usesScalars(A) || Def == U; })) {
return Def->replaceAllUsesWith(A);
}
+
+ if (Plan->getUF() == 1 &&
+ match(Def, m_ExtractLastLanePerPart(m_VPValue(A)))) {
+ return Def->replaceAllUsesWith(
+ Builder.createNaryOp(VPInstruction::ExtractLastElement, {A}));
+ }
}
void VPlanTransforms::simplifyRecipes(VPlan &Plan) {
@@ -1324,8 +1334,11 @@ static void narrowToSingleScalarRecipes(VPlan &Plan) {
RepOrWidenR->getUnderlyingInstr(), RepOrWidenR->operands(),
true /*IsSingleScalar*/, nullptr /*Mask*/, *RepR /*Metadata*/);
Clone->insertBefore(RepOrWidenR);
- auto *Ext = new VPInstruction(VPInstruction::ExtractLastElement,
- {Clone->getOperand(0)});
+ unsigned ExtractOpc =
+ vputils::isUniformAcrossVFsAndUFs(RepR->getOperand(1))
+ ? VPInstruction::ExtractLastElement
+ : VPInstruction::ExtractLastLanePerPart;
+ auto *Ext = new VPInstruction(ExtractOpc, {Clone->getOperand(0)});
Ext->insertBefore(Clone);
Clone->setOperand(0, Ext);
RepR->eraseFromParent();
@@ -1339,7 +1352,8 @@ static void narrowToSingleScalarRecipes(VPlan &Plan) {
!all_of(RepOrWidenR->users(), [RepOrWidenR](const VPUser *U) {
return U->usesScalars(RepOrWidenR) ||
match(cast<VPRecipeBase>(U),
- m_ExtractLastElement(m_VPValue()));
+ m_CombineOr(m_ExtractLastElement(m_VPValue()),
+ m_ExtractLastLanePerPart(m_VPValue())));
}))
continue;
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/replicating-load-store-costs.ll b/llvm/test/Transforms/LoopVectorize/AArch64/replicating-load-store-costs.ll
index ab9b48fb68f6b..aff2c4cb644eb 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/replicating-load-store-costs.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/replicating-load-store-costs.ll
@@ -153,17 +153,20 @@ define void @uniform_gep_for_replicating_gep(ptr %dst) {
; CHECK-NEXT: [[VEC_IND:%.*]] = phi <2 x i32> [ <i32 0, i32 1>, %[[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], %[[VECTOR_BODY]] ]
; CHECK-NEXT: [[STEP_ADD:%.*]] = add <2 x i32> [[VEC_IND]], splat (i32 2)
; CHECK-NEXT: [[TMP2:%.*]] = add i32 [[INDEX]], 2
-; CHECK-NEXT: [[TMP5:%.*]] = icmp eq <2 x i32> [[STEP_ADD]], zeroinitializer
+; CHECK-NEXT: [[TMP5:%.*]] = icmp eq <2 x i32> [[VEC_IND]], zeroinitializer
+; CHECK-NEXT: [[TMP3:%.*]] = icmp eq <2 x i32> [[STEP_ADD]], zeroinitializer
; CHECK-NEXT: [[TMP8:%.*]] = lshr i32 [[INDEX]], 1
; CHECK-NEXT: [[TMP9:%.*]] = lshr i32 [[TMP2]], 1
; CHECK-NEXT: [[TMP11:%.*]] = zext <2 x i1> [[TMP5]] to <2 x i8>
+; CHECK-NEXT: [[TMP6:%.*]] = zext <2 x i1> [[TMP3]] to <2 x i8>
; CHECK-NEXT: [[TMP14:%.*]] = zext i32 [[TMP8]] to i64
; CHECK-NEXT: [[TMP15:%.*]] = zext i32 [[TMP9]] to i64
; CHECK-NEXT: [[TMP18:%.*]] = getelementptr i64, ptr [[DST]], i64 [[TMP14]]
; CHECK-NEXT: [[TMP19:%.*]] = getelementptr i64, ptr [[DST]], i64 [[TMP15]]
; CHECK-NEXT: [[TMP22:%.*]] = extractelement <2 x i8> [[TMP11]], i32 1
+; CHECK-NEXT: [[TMP12:%.*]] = extractelement <2 x i8> [[TMP6]], i32 1
; CHECK-NEXT: store i8 [[TMP22]], ptr [[TMP18]], align 1
-; CHECK-NEXT: store i8 [[TMP22]], ptr [[TMP19]], align 1
+; CHECK-NEXT: store i8 [[TMP12]], ptr [[TMP19]], align 1
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 4
; CHECK-NEXT: [[VEC_IND_NEXT]] = add <2 x i32> [[STEP_ADD]], splat (i32 2)
; CHECK-NEXT: [[TMP24:%.*]] = icmp eq i32 [[INDEX_NEXT]], 128
diff --git a/llvm/test/Transforms/LoopVectorize/narrow-to-single-scalar.ll b/llvm/test/Transforms/LoopVectorize/narrow-to-single-scalar.ll
index 1533906247739..53dad3a74482c 100644
--- a/llvm/test/Transforms/LoopVectorize/narrow-to-single-scalar.ll
+++ b/llvm/test/Transforms/LoopVectorize/narrow-to-single-scalar.ll
@@ -74,8 +74,7 @@ exit:
ret void
}
-; FIXME: Currently this mis-compiled when interleaving; all stores store the
-; last lane of the last part, instead of the last lane per part.
+; Check each unrolled store stores the last lane of the corresponding part.
; Test case for https://github.com/llvm/llvm-project/issues/162498.
define void @narrow_to_single_scalar_store_address_not_uniform_across_all_parts(ptr %dst) {
; VF4IC1-LABEL: define void @narrow_to_single_scalar_store_address_not_uniform_across_all_parts(
@@ -121,13 +120,15 @@ define void @narrow_to_single_scalar_store_address_not_uniform_across_all_parts(
; VF2IC2-NEXT: br label %[[VECTOR_BODY:.*]]
; VF2IC2: [[VECTOR_BODY]]:
; VF2IC2-NEXT: [[INDEX:%.*]] = phi i32 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; VF2IC2-NEXT: [[TMP7:%.*]] = add i32 [[INDEX]], 0
+; VF2IC2-NEXT: [[TMP8:%.*]] = add i32 [[INDEX]], 1
; VF2IC2-NEXT: [[TMP0:%.*]] = add i32 [[INDEX]], 2
; VF2IC2-NEXT: [[TMP1:%.*]] = add i32 [[INDEX]], 3
-; VF2IC2-NEXT: [[TMP2:%.*]] = lshr i32 [[INDEX]], 1
+; VF2IC2-NEXT: [[TMP2:%.*]] = lshr i32 [[TMP7]], 1
; VF2IC2-NEXT: [[TMP3:%.*]] = lshr i32 [[TMP0]], 1
; VF2IC2-NEXT: [[TMP4:%.*]] = getelementptr i32, ptr [[DST]], i32 [[TMP2]]
; VF2IC2-NEXT: [[TMP5:%.*]] = getelementptr i32, ptr [[DST]], i32 [[TMP3]]
-; VF2IC2-NEXT: store i32 [[TMP1]], ptr [[TMP4]], align 4
+; VF2IC2-NEXT: store i32 [[TMP8]], ptr [[TMP4]], align 4
; VF2IC2-NEXT: store i32 [[TMP1]], ptr [[TMP5]], align 4
; VF2IC2-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 4
; VF2IC2-NEXT: [[TMP6:%.*]] = icmp eq i32 [[INDEX_NEXT]], 100
|
david-arm
approved these changes
Oct 13, 2025
|
|
||
| if (match(Def, | ||
| m_VPInstruction<VPInstruction::ExtractLastElement>(m_VPValue(A))) && | ||
| if (match(Def, m_CombineOr(m_VPInstruction<VPInstruction::ExtractLastElement>( |
There was a problem hiding this comment.
nit: Since you're here, perhaps worth changing m_VPInstruction<VPInstruction::ExtractLastElement> to m_ExtractLastElement as well?
ayalz
reviewed
Oct 13, 2025
Comment on lines
1021
to
+1025
| // part if scalar. In the latter case, the recipe will be removed during | ||
| // unrolling. | ||
| ExtractLastElement, | ||
| // Extracts the last lane for each part from its operand. | ||
| ExtractLastLanePerPart, |
There was a problem hiding this comment.
Given ExtractLastLanePerPart which produces a scalar per part, should ExtractLastPart also be added - which together could then replace ExtractLastElement? Perhaps suffice to call it ExtractLastLane(*) - producing a scalar from (each) vector, orthogonal of how many parts/vectors it is given. I.e., ExtractLastElement could then be replaced by ExtractLastPart if its operand is scalar or else by ExtractLastPart(ExtractLastLane) or ExtractLastLane(ExtractLastPart) if it's vector.
(*) ExtractLane seems to apply similar double extraction, which could also be simplified by combining with some "de-splat" reduction of a uniform-across-UF value that provides a(ny) single part?
There was a problem hiding this comment.
Ah yes, I will check that as follow-up, thanks!
llvm-sync bot
pushed a commit
to arm/arm-toolchain
that referenced
this pull request
Oct 15, 2025
…calar. (#163056) When narrowing stores of a single-scalar, we currently use ExtractLastElement, which extracts the last element across all parts. This is not correct if the store's address is not uniform across all parts. If it is only uniform-per-part, the last lane per part must be extracted. Add a new ExtractLastLanePerPart opcode to handle this correctly. Most transforms apply to both ExtractLastElement and ExtractLastLanePerPart, with the only difference being their treatment during unrolling. Fixes llvm/llvm-project#162498. PR: llvm/llvm-project#163056
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
When narrowing stores of a single-scalar, we currently use ExtractLastElement, which extracts the last element across all parts. This is not correct if the store's address is not uniform across all parts. If it is only uniform-per-part, the last lane per part must be extracted. Add a new ExtractLastLanePerPart opcode to handle this correctly. Most transforms apply to both ExtractLastElement and ExtractLastLanePerPart, with the only difference being their treatment during unrolling.
Fixes #162498.