[LV] Support scalable interleave groups for factors 3,5,6 and 7 #141865
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
|
@llvm/pr-subscribers-llvm-analysis @llvm/pr-subscribers-backend-aarch64 Author: Luke Lau (lukel97) ChangesCurrently the loop vectorizer can only vectorize interleave groups for power-of-2 factors at scalable VFs by recursively interleaving [de]interleave2 intrinsics. However after #124825 and #139893, we now have [de]interleave intrinsics for all factors up to 8, which is enough to support all types of segmented loads and stores on RISC-V. Now that the interleaved access pass has been taught to lower these in #139373 and #141512, this patch teaches the loop vectorizer to emit these intrinsics for factors up to 8, which enables scalable vectorization for non-power-of-2 factors. The existing recursive [de]interleaving has been kept for power-of-2 factors > 8. The idea is that some targets might still want to perform a wide load/store and a series of [de]interleaves, even if there's no interleaved memory operation that interleaved access pass can lower to. As far as I'm aware, no in-tree target will trigger this though because the maximum interleave factor is capped at 4 on AArch64 and 8 on RISC-V. This code could be removed if reviewers prefer, and added back when needed. This patch shouldn't be landed until #141513 has landed, so that on the off-chance that the interleaved access pass doesn't convert an [de]interleave4 intrinsic it can still be lowered. Factors of 3 with scalable VFs are also turned off in AArch64 since there's no lowering for [de]interleave3 just yet either. Patch is 118.46 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/141865.diff 6 Files Affected:
diff --git a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
index 68aec80f07e1d..8cbc30b071d6b 100644
--- a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
+++ b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
@@ -4575,6 +4575,12 @@ InstructionCost AArch64TTIImpl::getInterleavedMemoryOpCost(
if (VecTy->isScalableTy() && !ST->hasSVE())
return InstructionCost::getInvalid();
+ // Currently factors 2 and 4 can be de[interleaved] with scalable vectors.
+ // TODO: Add lowering for vector.[de]interleave3 intrinsics and
+ // support in InterleavedAccessPass for ld3/st3
+ if (VecTy->isScalableTy() && Factor != 2 && Factor != 4)
+ return InstructionCost::getInvalid();
+
// Vectorization for masked interleaved accesses is only enabled for scalable
// VF.
if (!VecTy->isScalableTy() && (UseMaskForCond || UseMaskForGaps))
diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index 8040d375f0dbd..2b440c778ec1a 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -3193,10 +3193,10 @@ bool LoopVectorizationCostModel::interleavedAccessCanBeWidened(
if (hasIrregularType(ScalarTy, DL))
return false;
- // For scalable vectors, the only interleave factor currently supported
- // must be power of 2 since we require the (de)interleave2 intrinsics
- // instead of shufflevectors.
- if (VF.isScalable() && !isPowerOf2_32(InterleaveFactor))
+ // For scalable vectors, the interleave factors must be <= 8 or a power of 2
+ // since we require the (de)interleaveN intrinsics instead of shufflevectors.
+ if (VF.isScalable() &&
+ !(InterleaveFactor <= 8 || isPowerOf2_32(InterleaveFactor)))
return false;
// If the group involves a non-integral pointer, we may not be able to
@@ -9058,10 +9058,11 @@ VPlanPtr LoopVectorizationPlanner::tryToBuildVPlanWithVPRecipes(
bool Result = (VF.isVector() && // Query is illegal for VF == 1
CM.getWideningDecision(IG->getInsertPos(), VF) ==
LoopVectorizationCostModel::CM_Interleave);
- // For scalable vectors, the only interleave factor currently supported
- // must be power of 2 since we require the (de)interleave2 intrinsics
- // instead of shufflevectors.
- assert((!Result || !VF.isScalable() || isPowerOf2_32(IG->getFactor())) &&
+ // For scalable vectors, the interleave factors must be <= 8 or a power of
+ // 2 since we require the (de)interleaveN intrinsics instead of
+ // shufflevectors.
+ assert((!Result || !VF.isScalable() ||
+ (IG->getFactor() <= 8 || isPowerOf2_32(IG->getFactor()))) &&
"Unsupported interleave factor for scalable vectors");
return Result;
};
diff --git a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
index 266f658b1f9c1..1d4c95633a680 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
@@ -3217,6 +3217,62 @@ static Value *createBitOrPointerCast(IRBuilderBase &Builder, Value *V,
return Builder.CreateBitOrPointerCast(CastVal, DstVTy);
}
+static Intrinsic::ID getInterleaveIntrinsicID(unsigned Factor) {
+ switch (Factor) {
+ case 2:
+ return Intrinsic::vector_interleave2;
+ break;
+ case 3:
+ return Intrinsic::vector_interleave3;
+ break;
+ case 4:
+ return Intrinsic::vector_interleave4;
+ break;
+ case 5:
+ return Intrinsic::vector_interleave5;
+ break;
+ case 6:
+ return Intrinsic::vector_interleave6;
+ break;
+ case 7:
+ return Intrinsic::vector_interleave7;
+ break;
+ case 8:
+ return Intrinsic::vector_interleave8;
+ break;
+ default:
+ llvm_unreachable("Unexpected factor");
+ }
+}
+
+static Intrinsic::ID getDeinterleaveIntrinsicID(unsigned Factor) {
+ switch (Factor) {
+ case 2:
+ return Intrinsic::vector_deinterleave2;
+ break;
+ case 3:
+ return Intrinsic::vector_deinterleave3;
+ break;
+ case 4:
+ return Intrinsic::vector_deinterleave4;
+ break;
+ case 5:
+ return Intrinsic::vector_deinterleave5;
+ break;
+ case 6:
+ return Intrinsic::vector_deinterleave6;
+ break;
+ case 7:
+ return Intrinsic::vector_deinterleave7;
+ break;
+ case 8:
+ return Intrinsic::vector_deinterleave8;
+ break;
+ default:
+ llvm_unreachable("Unexpected factor");
+ }
+}
+
/// Return a vector containing interleaved elements from multiple
/// smaller input vectors.
static Value *interleaveVectors(IRBuilderBase &Builder, ArrayRef<Value *> Vals,
@@ -3233,6 +3289,14 @@ static Value *interleaveVectors(IRBuilderBase &Builder, ArrayRef<Value *> Vals,
// Scalable vectors cannot use arbitrary shufflevectors (only splats), so
// must use intrinsics to interleave.
if (VecTy->isScalableTy()) {
+ if (Factor <= 8) {
+ VectorType *InterleaveTy = VectorType::get(
+ VecTy->getElementType(),
+ VecTy->getElementCount().multiplyCoefficientBy(Factor));
+ return Builder.CreateIntrinsic(InterleaveTy,
+ getInterleaveIntrinsicID(Factor), Vals,
+ /*FMFSource=*/nullptr, Name);
+ }
assert(isPowerOf2_32(Factor) && "Unsupported interleave factor for "
"scalable vectors, must be power of 2");
SmallVector<Value *> InterleavingValues(Vals);
@@ -3333,7 +3397,7 @@ void VPInterleaveRecipe::execute(VPTransformState &State) {
&InterleaveFactor](Value *MaskForGaps) -> Value * {
if (State.VF.isScalable()) {
assert(!MaskForGaps && "Interleaved groups with gaps are not supported.");
- assert(isPowerOf2_32(InterleaveFactor) &&
+ assert((InterleaveFactor <= 8 || isPowerOf2_32(InterleaveFactor)) &&
"Unsupported deinterleave factor for scalable vectors");
auto *ResBlockInMask = State.get(BlockInMask);
SmallVector<Value *> Ops(InterleaveFactor, ResBlockInMask);
@@ -3377,34 +3441,45 @@ void VPInterleaveRecipe::execute(VPTransformState &State) {
ArrayRef<VPValue *> VPDefs = definedValues();
const DataLayout &DL = State.CFG.PrevBB->getDataLayout();
if (VecTy->isScalableTy()) {
- assert(isPowerOf2_32(InterleaveFactor) &&
- "Unsupported deinterleave factor for scalable vectors");
-
// Scalable vectors cannot use arbitrary shufflevectors (only splats),
// so must use intrinsics to deinterleave.
SmallVector<Value *> DeinterleavedValues(InterleaveFactor);
- DeinterleavedValues[0] = NewLoad;
- // For the case of InterleaveFactor > 2, we will have to do recursive
- // deinterleaving, because the current available deinterleave intrinsic
- // supports only Factor of 2, otherwise it will bailout after first
- // iteration.
- // When deinterleaving, the number of values will double until we
- // have "InterleaveFactor".
- for (unsigned NumVectors = 1; NumVectors < InterleaveFactor;
- NumVectors *= 2) {
- // Deinterleave the elements within the vector
- SmallVector<Value *> TempDeinterleavedValues(NumVectors);
- for (unsigned I = 0; I < NumVectors; ++I) {
- auto *DiTy = DeinterleavedValues[I]->getType();
- TempDeinterleavedValues[I] = State.Builder.CreateIntrinsic(
- Intrinsic::vector_deinterleave2, DiTy, DeinterleavedValues[I],
- /*FMFSource=*/nullptr, "strided.vec");
+
+ if (InterleaveFactor <= 8) {
+ Value *Deinterleave = State.Builder.CreateIntrinsic(
+ getDeinterleaveIntrinsicID(InterleaveFactor), NewLoad->getType(),
+ NewLoad,
+ /*FMFSource=*/nullptr, "strided.vec");
+ for (unsigned I = 0; I < InterleaveFactor; I++)
+ DeinterleavedValues[I] =
+ State.Builder.CreateExtractValue(Deinterleave, I);
+ } else {
+ assert(isPowerOf2_32(InterleaveFactor) &&
+ "Unsupported deinterleave factor for scalable vectors");
+ DeinterleavedValues[0] = NewLoad;
+ // For the case of InterleaveFactor > 2, we will have to do recursive
+ // deinterleaving, because the current available deinterleave intrinsic
+ // supports only Factor of 2, otherwise it will bailout after first
+ // iteration.
+ // When deinterleaving, the number of values will double until we
+ // have "InterleaveFactor".
+ for (unsigned NumVectors = 1; NumVectors < InterleaveFactor;
+ NumVectors *= 2) {
+ // Deinterleave the elements within the vector
+ SmallVector<Value *> TempDeinterleavedValues(NumVectors);
+ for (unsigned I = 0; I < NumVectors; ++I) {
+ auto *DiTy = DeinterleavedValues[I]->getType();
+ TempDeinterleavedValues[I] = State.Builder.CreateIntrinsic(
+ Intrinsic::vector_deinterleave2, DiTy, DeinterleavedValues[I],
+ /*FMFSource=*/nullptr, "strided.vec");
+ }
+ // Extract the deinterleaved values:
+ for (unsigned I = 0; I < 2; ++I)
+ for (unsigned J = 0; J < NumVectors; ++J)
+ DeinterleavedValues[NumVectors * I + J] =
+ State.Builder.CreateExtractValue(TempDeinterleavedValues[J],
+ I);
}
- // Extract the deinterleaved values:
- for (unsigned I = 0; I < 2; ++I)
- for (unsigned J = 0; J < NumVectors; ++J)
- DeinterleavedValues[NumVectors * I + J] =
- State.Builder.CreateExtractValue(TempDeinterleavedValues[J], I);
}
#ifndef NDEBUG
diff --git a/llvm/tes
8000
t/Transforms/LoopVectorize/AArch64/sve-interleaved-accesses.ll b/llvm/test/Transforms/LoopVectorize/AArch64/sve-interleaved-accesses.ll
index 6861644fc9969..10939de3e9fe0 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/sve-interleaved-accesses.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/sve-interleaved-accesses.ll
@@ -1479,34 +1479,24 @@ define void @interleave_deinterleave(ptr writeonly noalias %dst, ptr readonly %a
; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
; CHECK-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_XYZT:%.*]], ptr [[A:%.*]], i64 [[INDEX]]
; CHECK-NEXT: [[WIDE_VEC:%.*]] = load <vscale x 16 x i32>, ptr [[TMP6]], align 4
-; CHECK-NEXT: [[STRIDED_VEC:%.*]] = call { <vscale x 8 x i32>, <vscale x 8 x i32> } @llvm.vector.deinterleave2.nxv16i32(<vscale x 16 x i32> [[WIDE_VEC]])
-; CHECK-NEXT: [[TMP7:%.*]] = extractvalue { <vscale x 8 x i32>, <vscale x 8 x i32> } [[STRIDED_VEC]], 0
-; CHECK-NEXT: [[TMP8:%.*]] = extractvalue { <vscale x 8 x i32>, <vscale x 8 x i32> } [[STRIDED_VEC]], 1
-; CHECK-NEXT: [[STRIDED_VEC6:%.*]] = call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.vector.deinterleave2.nxv8i32(<vscale x 8 x i32> [[TMP7]])
-; CHECK-NEXT: [[STRIDED_VEC7:%.*]] = call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.vector.deinterleave2.nxv8i32(<vscale x 8 x i32> [[TMP8]])
-; CHECK-NEXT: [[TMP9:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC6]], 0
-; CHECK-NEXT: [[TMP10:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC7]], 0
-; CHECK-NEXT: [[TMP11:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC6]], 1
-; CHECK-NEXT: [[TMP12:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC7]], 1
+; CHECK-NEXT: [[STRIDED_VEC:%.*]] = call { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.vector.deinterleave4.nxv16i32(<vscale x 16 x i32> [[WIDE_VEC]])
+; CHECK-NEXT: [[TMP9:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 0
+; CHECK-NEXT: [[TMP10:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 1
+; CHECK-NEXT: [[TMP11:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 2
+; CHECK-NEXT: [[TMP12:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 3
; CHECK-NEXT: [[TMP13:%.*]] = getelementptr inbounds [[STRUCT_XYZT]], ptr [[B:%.*]], i64 [[INDEX]]
; CHECK-NEXT: [[WIDE_VEC8:%.*]] = load <vscale x 16 x i32>, ptr [[TMP13]], align 4
-; CHECK-NEXT: [[STRIDED_VEC9:%.*]] = call { <vscale x 8 x i32>, <vscale x 8 x i32> } @llvm.vector.deinterleave2.nxv16i32(<vscale x 16 x i32> [[WIDE_VEC8]])
-; CHECK-NEXT: [[TMP14:%.*]] = extractvalue { <vscale x 8 x i32>, <vscale x 8 x i32> } [[STRIDED_VEC9]], 0
-; CHECK-NEXT: [[TMP15:%.*]] = extractvalue { <vscale x 8 x i32>, <vscale x 8 x i32> } [[STRIDED_VEC9]], 1
-; CHECK-NEXT: [[STRIDED_VEC10:%.*]] = call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.vector.deinterleave2.nxv8i32(<vscale x 8 x i32> [[TMP14]])
-; CHECK-NEXT: [[STRIDED_VEC11:%.*]] = call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.vector.deinterleave2.nxv8i32(<vscale x 8 x i32> [[TMP15]])
-; CHECK-NEXT: [[TMP16:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC10]], 0
-; CHECK-NEXT: [[TMP17:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC11]], 0
-; CHECK-NEXT: [[TMP18:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC10]], 1
-; CHECK-NEXT: [[TMP19:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC11]], 1
+; CHECK-NEXT: [[STRIDED_VEC2:%.*]] = call { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.vector.deinterleave4.nxv16i32(<vscale x 16 x i32> [[WIDE_VEC8]])
+; CHECK-NEXT: [[TMP16:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC2]], 0
+; CHECK-NEXT: [[TMP17:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC2]], 1
+; CHECK-NEXT: [[TMP18:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC2]], 2
+; CHECK-NEXT: [[TMP19:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC2]], 3
; CHECK-NEXT: [[TMP20:%.*]] = add nsw <vscale x 4 x i32> [[TMP16]], [[TMP9]]
; CHECK-NEXT: [[TMP21:%.*]] = getelementptr inbounds [[STRUCT_XYZT]], ptr [[DST:%.*]], i64 [[INDEX]]
; CHECK-NEXT: [[TMP22:%.*]] = sub nsw <vscale x 4 x i32> [[TMP10]], [[TMP17]]
; CHECK-NEXT: [[TMP23:%.*]] = shl <vscale x 4 x i32> [[TMP11]], [[TMP18]]
; CHECK-NEXT: [[TMP24:%.*]] = ashr <vscale x 4 x i32> [[TMP12]], [[TMP19]]
-; CHECK-NEXT: [[INTERLEAVED_VEC:%.*]] = call <vscale x 8 x i32> @llvm.vector.interleave2.nxv8i32(<vscale x 4 x i32> [[TMP20]], <vscale x 4 x i32> [[TMP23]])
-; CHECK-NEXT: [[INTERLEAVED_VEC12:%.*]] = call <vscale x 8 x i32> @llvm.vector.interleave2.nxv8i32(<vscale x 4 x i32> [[TMP22]], <vscale x 4 x i32> [[TMP24]])
-; CHECK-NEXT: [[INTERLEAVED_VEC13:%.*]] = call <vscale x 16 x i32> @llvm.vector.interleave2.nxv16i32(<vscale x 8 x i32> [[INTERLEAVED_VEC]], <vscale x 8 x i32> [[INTERLEAVED_VEC12]])
+; CHECK-NEXT: [[INTERLEAVED_VEC13:%.*]] = call <vscale x 16 x i32> @llvm.vector.interleave4.nxv16i32(<vscale x 4 x i32> [[TMP20]], <vscale x 4 x i32> [[TMP22]], <vscale x 4 x i32> [[TMP23]], <vscale x 4 x i32> [[TMP24]])
; CHECK-NEXT: store <vscale x 16 x i32> [[INTERLEAVED_VEC13]], ptr [[TMP21]], align 4
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP5]]
; CHECK-NEXT: [[TMP25:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
@@ -1595,15 +1585,11 @@ define void @interleave_deinterleave_reverse(ptr noalias nocapture readonly %A,
; CHECK-NEXT: [[TMP9:%.*]] = sext i32 [[TMP8]] to i64
; CHECK-NEXT: [[TMP10:%.*]] = getelementptr inbounds i32, ptr [[TMP5]], i64 [[TMP9]]
; CHECK-NEXT: [[WIDE_VEC:%.*]] = load <vscale x 16 x i32>, ptr [[TMP10]], align 4
-; CHECK-NEXT: [[STRIDED_VEC:%.*]] = call { <vscale x 8 x i32>, <vscale x 8 x i32> } @llvm.vector.deinterleave2.nxv16i32(<vscale x 16 x i32> [[WIDE_VEC]])
-; CHECK-NEXT: [[TMP11:%.*]] = extractvalue { <vscale x 8 x i32>, <vscale x 8 x i32> } [[STRIDED_VEC]], 0
-; CHECK-NEXT: [[TMP12:%.*]] = extractvalue { <vscale x 8 x i32>, <vscale x 8 x i32> } [[STRIDED_VEC]], 1
-; CHECK-NEXT: [[STRIDED_VEC1:%.*]] = call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.vector.deinterleave2.nxv8i32(<vscale x 8 x i32> [[TMP11]])
-; CHECK-NEXT: [[STRIDED_VEC2:%.*]] = call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.vector.deinterleave2.nxv8i32(<vscale x 8 x i32> [[TMP12]])
-; CHECK-NEXT: [[TMP13:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC1]], 0
-; CHECK-NEXT: [[TMP14:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC2]], 0
-; CHECK-NEXT: [[TMP15:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC1]], 1
-; CHECK-NEXT: [[TMP16:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC2]], 1
+; CHECK-NEXT: [[STRIDED_VEC:%.*]] = call { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.vector.deinterleave4.nxv16i32(<vscale x 16 x i32> [[WIDE_VEC]])
+; CHECK-NEXT: [[TMP13:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 0
+; CHECK-NEXT: [[TMP14:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 1
+; CHECK-NEXT: [[TMP15:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 2
+; CHECK-NEXT: [[TMP16:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 3
; CHECK-NEXT: [[REVERSE:%.*]] = call <vscale x 4 x i32> @llvm.vector.reverse.nxv4i32(<vscale x 4 x i32> [[TMP13]])
; CHECK-NEXT: [[REVERSE3:%.*]] = call <vscale x 4 x i32> @llvm.vector.reverse.nxv4i32(<vscale x 4 x i32> [[TMP14]])
; CHECK-NEXT: [[REVERSE4:%.*]] = call <vscale x 4 x i32> @llvm.vector.reverse.nxv4i32(<vscale x 4 x i32> [[TMP15]])
@@ -1622,9 +1608,7 @@ define void @interleave_deinterleave_reverse(ptr noalias nocapture readonly %A,
; CHECK-NEXT: [[REVERSE7:%.*]] = call <vscale x 4 x i32> @llvm.vector.reverse.nxv4i32(<vscale x 4 x i32> [[TMP18]])
; CHECK-NEXT: [[REVERSE8:%.*]] = call <vscale x 4 x i32> @llvm.vector.reverse.nxv4i32(<vscale x 4 x i32> [[TMP19]])
; CHECK-NEXT: [[REVERSE9:%.*]] = call <vscale x 4 x i32> @llvm.vector.reverse.nxv4i32(<vscale x 4 x i32> [[TMP20]])
-; CHECK-NEXT: [[INTERLEAVED_VEC:%.*]] = call <vscale x 8 x i32> @llvm.vector.interleave2.nxv8i32(<vscale x 4 x i32> [[REVERSE6]], <vscale x 4 x i32> [[REVERSE8]])
-; CHECK-NEXT: [[INTERLEAVED_VEC10:%.*]] = call <vscale x 8 x i32> @llvm.vector.interleave2.nxv8i32(<vscale x 4 x i32> [[REVERSE7]], <vscale x 4 x i32> [[REVERSE9]])
-; CHECK-NEXT: [[INTERLEAVED_VEC11:%.*]] = call <vscale x 16 x i32> @llvm.vector.interleave2.nxv16i32(<vscale x 8 x i32> [[INTERLEAVED_VEC]], <vscale x 8 x i32> [[INTERLEAVED_VEC10]])
+; CHECK-NEXT: [[INTERLEAVED_VEC11:%.*]] = call <vscale x 16 x i32> @llvm.vector.interleave4.nxv16i32(<vscale x 4 x i32> [[REVERSE6]], <vscale x 4 x i32> [[REVERSE7]], <vscale x 4 x i32> [[REVERSE8]], <vscale x 4 x i32> [[REVERSE9]])
; CHECK-NEXT: store <vscale x 16 x i32> [[INTERLEAVED_VEC11]], ptr [[TMP26]], align 4
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP1]]
; CHECK-NEXT: [[VEC_IND_NEXT]] = add <vscale x 4 x i32> [[VEC_IND]], [[DOTSPLAT]]
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/sve-interleaved-masked-accesses.ll b/llvm/test/Transforms/LoopVectorize/AArch64/sve-interleaved-masked-accesses.ll
index 469faf67a71b3..3567aff0ace4e 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/sve-interleaved-masked-accesses.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/sve-interleaved-masked-ac...
[truncated]
|
|
@llvm/pr-subscribers-llvm-transforms Author: Luke Lau (lukel97) ChangesCurrently the loop vectorizer can only vectorize interleave groups for power-of-2 factors at scalable VFs by recursively interleaving [de]interleave2 intrinsics. However after #124825 and #139893, we now have [de]interleave intrinsics for all factors up to 8, which is enough to support all types of segmented loads and stores on RISC-V. Now that the interleaved access pass has been taught to lower these in #139373 and #141512, this patch teaches the loop vectorizer to emit these intrinsics for factors up to 8, which enables scalable vectorization for non-power-of-2 factors. The existing recursive [de]interleaving has been kept for power-of-2 factors > 8. The idea is that some targets might still want to perform a wide load/store and a series of [de]interleaves, even if there's no interleaved memory operation that interleaved access pass can lower to. As far as I'm aware, no in-tree target will trigger this though because the maximum interleave factor is capped at 4 on AArch64 and 8 on RISC-V. This code could be removed if reviewers prefer, and added back when needed. This patch shouldn't be landed until #141513 has landed, so that on the off-chance that the interleaved access pass doesn't convert an [de]interleave4 intrinsic it can still be lowered. Factors of 3 with scalable VFs are also turned off in AArch64 since there's no lowering for [de]interleave3 just yet either. Patch is 118.46 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/141865.diff 6 Files Affected:
diff --git a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
index 68aec80f07e1d..8cbc30b071d6b 100644
--- a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
+++ b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
@@ -4575,6 +4575,12 @@ InstructionCost AArch64TTIImpl::getInterleavedMemoryOpCost(
if (VecTy->isScalableTy() && !ST->hasSVE())
return InstructionCost::getInvalid();
+ // Currently factors 2 and 4 can be de[interleaved] with scalable vectors.
+ // TODO: Add lowering for vector.[de]interleave3 intrinsics and
+ // support in InterleavedAccessPass for ld3/st3
+ if (VecTy->isScalableTy() && Factor != 2 && Factor != 4)
+ return InstructionCost::getInvalid();
+
// Vectorization for masked interleaved accesses is only enabled for scalable
// VF.
if (!VecTy->isScalableTy() && (UseMaskForCond || UseMaskForGaps))
diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index 8040d375f0dbd..2b440c778ec1a 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -3193,10 +3193,10 @@ bool LoopVectorizationCostModel::interleavedAccessCanBeWidened(
if (hasIrregularType(ScalarTy, DL))
return false;
- // For scalable vectors, the only interleave factor currently supported
- // must be power of 2 since we require the (de)interleave2 intrinsics
- // instead of shufflevectors.
- if (VF.isScalable() && !isPowerOf2_32(InterleaveFactor))
+ // For scalable vectors, the interleave factors must be <= 8 or a power of 2
+ // since we require the (de)interleaveN intrinsics instead of shufflevectors.
+ if (VF.isScalable() &&
+ !(InterleaveFactor <= 8 || isPowerOf2_32(InterleaveFactor)))
return false;
// If the group involves a non-integral pointer, we may not be able to
@@ -9058,10 +9058,11 @@ VPlanPtr LoopVectorizationPlanner::tryToBuildVPlanWithVPRecipes(
bool Result = (VF.isVector() && // Query is illegal for VF == 1
CM.getWideningDecision(IG->getInsertPos(), VF) ==
LoopVectorizationCostModel::CM_Interleave);
- // For scalable vectors, the only interleave factor currently supported
- // must be power of 2 since we require the (de)interleave2 intrinsics
- // instead of shufflevectors.
- assert((!Result || !VF.isScalable() || isPowerOf2_32(IG->getFactor())) &&
+ // For scalable vectors, the interleave factors must be <= 8 or a power of
+ // 2 since we require the (de)interleaveN intrinsics instead of
+ // shufflevectors.
+ assert((!Result || !VF.isScalable() ||
+ (IG->getFactor() <= 8 || isPowerOf2_32(IG->getFactor()))) &&
"Unsupported interleave factor for scalable vectors");
return Result;
};
diff --git a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
index 266f658b1f9c1..1d4c95633a680 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
@@ -3217,6 +3217,62 @@ static Value *createBitOrPointerCast(IRBuilderBase &Builder, Value *V,
return Builder.CreateBitOrPointerCast(CastVal, DstVTy);
}
+static Intrinsic::ID getInterleaveIntrinsicID(unsigned Factor) {
+ switch (Factor) {
+ case 2:
+ return Intrinsic::vector_interleave2;
+ break;
+ case 3:
+ return Intrinsic::vector_interleave3;
+ break;
+ case 4:
+ return Intrinsic::vector_interleave4;
+ break;
+ case 5:
+ return Intrinsic::vector_interleave5;
+ break;
+ case 6:
+ return Intrinsic::vector_interleave6;
+ break;
+ case 7:
+ return Intrinsic::vector_interleave7;
+ break;
+ case 8:
+ return Intrinsic::vector_interleave8;
+ break;
+ default:
+ llvm_unreachable("Unexpected factor");
+ }
+}
+
+static Intrinsic::ID getDeinterleaveIntrinsicID(unsigned Factor) {
+ switch (Factor) {
+ case 2:
+ return Intrinsic::vector_deinterleave2;
+ break;
+ case 3:
+ return Intrinsic::vector_deinterleave3;
+ break;
+ case 4:
+ return Intrinsic::vector_deinterleave4;
+ break;
+ case 5:
+ return Intrinsic::vector_deinterleave5;
+ break;
+ case 6:
+ return Intrinsic::vector_deinterleave6;
+ break;
+ case 7:
+ return Intrinsic::vector_deinterleave7;
+ break;
+ case 8:
+ return Intrinsic::vector_deinterleave8;
+ break;
+ default:
+ llvm_unreachable("Unexpected factor");
+ }
+}
+
/// Return a vector containing interleaved elements from multiple
/// smaller input vectors.
static Value *interleaveVectors(IRBuilderBase &Builder, ArrayRef<Value *> Vals,
@@ -3233,6 +3289,14 @@ static Value *interleaveVectors(IRBuilderBase &Builder, ArrayRef<Value *> Vals,
// Scalable vectors cannot use arbitrary shufflevectors (only splats), so
// must use intrinsics to interleave.
if (VecTy->isScalableTy()) {
+ if (Factor <= 8) {
+ VectorType *InterleaveTy = VectorType::get(
+ VecTy->getElementType(),
+ VecTy->getElementCount().multiplyCoefficientBy(Factor));
+ return Builder.CreateIntrinsic(InterleaveTy,
+ getInterleaveIntrinsicID(Factor), Vals,
+ /*FMFSource=*/nullptr, Name);
+ }
assert(isPowerOf2_32(Factor) && "Unsupported interleave factor for "
"scalable vectors, must be power of 2");
SmallVector<Value *> InterleavingValues(Vals);
@@ -3333,7 +3397,7 @@ void VPInterleaveRecipe::execute(VPTransformState &State) {
&InterleaveFactor](Value *MaskForGaps) -> Value * {
if (State.VF.isScalable()) {
assert(!MaskForGaps && "Interleaved groups with gaps are not supported.");
- assert(isPowerOf2_32(InterleaveFactor) &&
+ assert((InterleaveFactor <= 8 || isPowerOf2_32(InterleaveFactor)) &&
"Unsupported deinterleave factor for scalable vectors");
auto *ResBlockInMask = State.get(BlockInMask);
SmallVector<Value *> Ops(InterleaveFactor, ResBlockInMask);
@@ -3377,34 +3441,45 @@ void VPInterleaveRecipe::execute(VPTransformState &State) {
ArrayRef<VPValue *> VPDefs = definedValues();
const DataLayout &DL = State.CFG.PrevBB->getDataLayout();
if (VecTy->isScalableTy()) {
- assert(isPowerOf2_32(InterleaveFactor) &&
- "Unsupported deinterleave factor for scalable vectors");
-
// Scalable vectors cannot use arbitrary shufflevectors (only splats),
// so must use intrinsics to deinterleave.
SmallVector<Value *> DeinterleavedValues(InterleaveFactor);
- DeinterleavedValues[0] = NewLoad;
- // For the case of InterleaveFactor > 2, we will have to do recursive
- // deinterleaving, because the current available deinterleave intrinsic
- // supports only Factor of 2, otherwise it will bailout after first
- // iteration.
- // When deinterleaving, the number of values will double until we
- // have "InterleaveFactor".
- for (unsigned NumVectors = 1; NumVectors < InterleaveFactor;
- NumVectors *= 2) {
- // Deinterleave the elements within the vector
- SmallVector<Value *> TempDeinterleavedValues(NumVectors);
- for (unsigned I = 0; I < NumVectors; ++I) {
- auto *DiTy = DeinterleavedValues[I]->getType();
- TempDeinterleavedValues[I] = State.Builder.CreateIntrinsic(
- Intrinsic::vector_deinterleave2, DiTy, DeinterleavedValues[I],
- /*FMFSource=*/nullptr, "strided.vec");
+
+ if (InterleaveFactor <= 8) {
+ Value *Deinterleave = State.Builder.CreateIntrinsic(
+ getDeinterleaveIntrinsicID(InterleaveFactor), NewLoad->getType(),
+ NewLoad,
+ /*FMFSource=*/nullptr, "strided.vec");
+ for (unsigned I = 0; I < InterleaveFactor; I++)
+ DeinterleavedValues[I] =
+ State.Builder.CreateExtractValue(Deinterleave, I);
+ } else {
+ assert(isPowerOf2_32(InterleaveFactor) &&
+ "Unsupported deinterleave factor for scalable vectors");
+ DeinterleavedValues[0] = NewLoad;
+ // For the case of InterleaveFactor > 2, we will have to do recursive
+ // deinterleaving, because the current available deinterleave intrinsic
+ // supports only Factor of 2, otherwise it will bailout after first
+ // iteration.
+ // When deinterleaving, the number of values will double until we
+ // have "InterleaveFactor".
+ for (unsigned NumVectors = 1; NumVectors < InterleaveFactor;
+ NumVectors *= 2) {
+ // Deinterleave the elements within the vector
+ SmallVector<Value *> TempDeinterleavedValues(NumVectors);
+ for (unsigned I = 0; I < NumVectors; ++I) {
+ auto *DiTy = DeinterleavedValues[I]->getType();
+ TempDeinterleavedValues[I] = State.Builder.CreateIntrinsic(
+ Intrinsic::vector_deinterleave2, DiTy, DeinterleavedValues[I],
+ /*FMFSource=*/nullptr, "strided.vec");
+ }
+ // Extract the deinterleaved values:
+ for (unsigned I = 0; I < 2; ++I)
+ for (unsigned J = 0; J < NumVectors; ++J)
+ DeinterleavedValues[NumVectors *
8000
I + J] =
+ State.Builder.CreateExtractValue(TempDeinterleavedValues[J],
+ I);
}
- // Extract the deinterleaved values:
- for (unsigned I = 0; I < 2; ++I)
- for (unsigned J = 0; J < NumVectors; ++J)
- DeinterleavedValues[NumVectors * I + J] =
- State.Builder.CreateExtractValue(TempDeinterleavedValues[J], I);
}
#ifndef NDEBUG
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/sve-interleaved-accesses.ll b/llvm/test/Transforms/LoopVectorize/AArch64/sve-interleaved-accesses.ll
index 6861644fc9969..10939de3e9fe0 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/sve-interleaved-accesses.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/sve-interleaved-accesses.ll
@@ -1479,34 +1479,24 @@ define void @interleave_deinterleave(ptr writeonly noalias %dst, ptr readonly %a
; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
; CHECK-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_XYZT:%.*]], ptr [[A:%.*]], i64 [[INDEX]]
; CHECK-NEXT: [[WIDE_VEC:%.*]] = load <vscale x 16 x i32>, ptr [[TMP6]], align 4
-; CHECK-NEXT: [[STRIDED_VEC:%.*]] = call { <vscale x 8 x i32>, <vscale x 8 x i32> } @llvm.vector.deinterleave2.nxv16i32(<vscale x 16 x i32> [[WIDE_VEC]])
-; CHECK-NEXT: [[TMP7:%.*]] = extractvalue { <vscale x 8 x i32>, <vscale x 8 x i32> } [[STRIDED_VEC]], 0
-; CHECK-NEXT: [[TMP8:%.*]] = extractvalue { <vscale x 8 x i32>, <vscale x 8 x i32> } [[STRIDED_VEC]], 1
-; CHECK-NEXT: [[STRIDED_VEC6:%.*]] = call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.vector.deinterleave2.nxv8i32(<vscale x 8 x i32> [[TMP7]])
-; CHECK-NEXT: [[STRIDED_VEC7:%.*]] = call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.vector.deinterleave2.nxv8i32(<vscale x 8 x i32> [[TMP8]])
-; CHECK-NEXT: [[TMP9:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC6]], 0
-; CHECK-NEXT: [[TMP10:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC7]], 0
-; CHECK-NEXT: [[TMP11:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC6]], 1
-; CHECK-NEXT: [[TMP12:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC7]], 1
+; CHECK-NEXT: [[STRIDED_VEC:%.*]] = call { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.vector.deinterleave4.nxv16i32(<vscale x 16 x i32> [[WIDE_VEC]])
+; CHECK-NEXT: [[TMP9:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 0
+; CHECK-NEXT: [[TMP10:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 1
+; CHECK-NEXT: [[TMP11:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 2
+; CHECK-NEXT: [[TMP12:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 3
; CHECK-NEXT: [[TMP13:%.*]] = getelementptr inbounds [[STRUCT_XYZT]], ptr [[B:%.*]], i64 [[INDEX]]
; CHECK-NEXT: [[WIDE_VEC8:%.*]] = load <vscale x 16 x i32>, ptr [[TMP13]], align 4
-; CHECK-NEXT: [[STRIDED_VEC9:%.*]] = call { <vscale x 8 x i32>, <vscale x 8 x i32> } @llvm.vector.deinterleave2.nxv16i32(<vscale x 16 x i32> [[WIDE_VEC8]])
-; CHECK-NEXT: [[TMP14:%.*]] = extractvalue { <vscale x 8 x i32>, <vscale x 8 x i32> } [[STRIDED_VEC9]], 0
-; CHECK-NEXT: [[TMP15:%.*]] = extractvalue { <vscale x 8 x i32>, <vscale x 8 x i32> } [[STRIDED_VEC9]], 1
-; CHECK-NEXT: [[STRIDED_VEC10:%.*]] = call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.vector.deinterleave2.nxv8i32(<vscale x 8 x i32> [[TMP14]])
-; CHECK-NEXT: [[STRIDED_VEC11:%.*]] = call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.vector.deinterleave2.nxv8i32(<vscale x 8 x i32> [[TMP15]])
-; CHECK-NEXT: [[TMP16:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC10]], 0
-; CHECK-NEXT: [[TMP17:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC11]], 0
-; CHECK-NEXT: [[TMP18:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC10]], 1
-; CHECK-NEXT: [[TMP19:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC11]], 1
+; CHECK-NEXT: [[STRIDED_VEC2:%.*]] = call { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.vector.deinterleave4.nxv16i32(<vscale x 16 x i32> [[WIDE_VEC8]])
+; CHECK-NEXT: [[TMP16:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC2]], 0
+; CHECK-NEXT: [[TMP17:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC2]], 1
+; CHECK-NEXT: [[TMP18:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC2]], 2
+; CHECK-NEXT: [[TMP19:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC2]], 3
; CHECK-NEXT: [[TMP20:%.*]] = add nsw <vscale x 4 x i32> [[TMP16]], [[TMP9]]
; CHECK-NEXT: [[TMP21:%.*]] = getelementptr inbounds [[STRUCT_XYZT]], ptr [[DST:%.*]], i64 [[INDEX]]
; CHECK-NEXT: [[TMP22:%.*]] = sub nsw <vscale x 4 x i32> [[TMP10]], [[TMP17]]
; CHECK-NEXT: [[TMP23:%.*]] = shl <vscale x 4 x i32> [[TMP11]], [[TMP18]]
; CHECK-NEXT: [[TMP24:%.*]] = ashr <vscale x 4 x i32> [[TMP12]], [[TMP19]]
-; CHECK-NEXT: [[INTERLEAVED_VEC:%.*]] = call <vscale x 8 x i32> @llvm.vector.interleave2.nxv8i32(<vscale x 4 x i32> [[TMP20]], <vscale x 4 x i32> [[TMP23]])
-; CHECK-NEXT: [[INTERLEAVED_VEC12:%.*]] = call <vscale x 8 x i32> @llvm.vector.interleave2.nxv8i32(<vscale x 4 x i32> [[TMP22]], <vscale x 4 x i32> [[TMP24]])
-; CHECK-NEXT: [[INTERLEAVED_VEC13:%.*]] = call <vscale x 16 x i32> @llvm.vector.interleave2.nxv16i32(<vscale x 8 x i32> [[INTERLEAVED_VEC]], <vscale x 8 x i32> [[INTERLEAVED_VEC12]])
+; CHECK-NEXT: [[INTERLEAVED_VEC13:%.*]] = call <vscale x 16 x i32> @llvm.vector.interleave4.nxv16i32(<vscale x 4 x i32> [[TMP20]], <vscale x 4 x i32> [[TMP22]], <vscale x 4 x i32> [[TMP23]], <vscale x 4 x i32> [[TMP24]])
; CHECK-NEXT: store <vscale x 16 x i32> [[INTERLEAVED_VEC13]], ptr [[TMP21]], align 4
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP5]]
; CHECK-NEXT: [[TMP25:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
@@ -1595,15 +1585,11 @@ define void @interleave_deinterleave_reverse(ptr noalias nocapture readonly %A,
; CHECK-NEXT: [[TMP9:%.*]] = sext i32 [[TMP8]] to i64
; CHECK-NEXT: [[TMP10:%.*]] = getelementptr inbounds i32, ptr [[TMP5]], i64 [[TMP9]]
; CHECK-NEXT: [[WIDE_VEC:%.*]] = load <vscale x 16 x i32>, ptr [[TMP10]], align 4
-; CHECK-NEXT: [[STRIDED_VEC:%.*]] = call { <vscale x 8 x i32>, <vscale x 8 x i32> } @llvm.vector.deinterleave2.nxv16i32(<vscale x 16 x i32> [[WIDE_VEC]])
-; CHECK-NEXT: [[TMP11:%.*]] = extractvalue { <vscale x 8 x i32>, <vscale x 8 x i32> } [[STRIDED_VEC]], 0
-; CHECK-NEXT: [[TMP12:%.*]] = extractvalue { <vscale x 8 x i32>, <vscale x 8 x i32> } [[STRIDED_VEC]], 1
-; CHECK-NEXT: [[STRIDED_VEC1:%.*]] = call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.vector.deinterleave2.nxv8i32(<vscale x 8 x i32> [[TMP11]])
-; CHECK-NEXT: [[STRIDED_VEC2:%.*]] = call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.vector.deinterleave2.nxv8i32(<vscale x 8 x i32> [[TMP12]])
-; CHECK-NEXT: [[TMP13:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC1]], 0
-; CHECK-NEXT: [[TMP14:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC2]], 0
-; CHECK-NEXT: [[TMP15:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC1]], 1
-; CHECK-NEXT: [[TMP16:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC2]], 1
+; CHECK-NEXT: [[STRIDED_VEC:%.*]] = call { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.vector.deinterleave4.nxv16i32(<vscale x 16 x i32> [[WIDE_VEC]])
+; CHECK-NEXT: [[TMP13:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 0
+; CHECK-NEXT: [[TMP14:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 1
+; CHECK-NEXT: [[TMP15:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 2
+; CHECK-NEXT: [[TMP16:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 3
; CHECK-NEXT: [[REVERSE:%.*]] = call <vscale x 4 x i32> @llvm.vector.reverse.nxv4i32(<vscale x 4 x i32> [[TMP13]])
; CHECK-NEXT: [[REVERSE3:%.*]] = call <vscale x 4 x i32> @llvm.vector.reverse.nxv4i32(<vscale x 4 x i32> [[TMP14]])
; CHECK-NEXT: [[REVERSE4:%.*]] = call <vscale x 4 x i32> @llvm.vector.reverse.nxv4i32(<vscale x 4 x i32> [[TMP15]])
@@ -1622,9 +1608,7 @@ define void @interleave_deinterleave_reverse(ptr noalias nocapture readonly %A,
; CHECK-NEXT: [[REVERSE7:%.*]] = call <vscale x 4 x i32> @llvm.vector.reverse.nxv4i32(<vscale x 4 x i32> [[TMP18]])
; CHECK-NEXT: [[REVERSE8:%.*]] = call <vscale x 4 x i32> @llvm.vector.reverse.nxv4i32(<vscale x 4 x i32> [[TMP19]])
; CHECK-NEXT: [[REVERSE9:%.*]] = call <vscale x 4 x i32> @llvm.vector.reverse.nxv4i32(<vscale x 4 x i32> [[TMP20]])
-; CHECK-NEXT: [[INTERLEAVED_VEC:%.*]] = call <vscale x 8 x i32> @llvm.vector.interleave2.nxv8i32(<vscale x 4 x i32> [[REVERSE6]], <vscale x 4 x i32> [[REVERSE8]])
-; CHECK-NEXT: [[INTERLEAVED_VEC10:%.*]] = call <vscale x 8 x i32> @llvm.vector.interleave2.nxv8i32(<vscale x 4 x i32> [[REVERSE7]], <vscale x 4 x i32> [[REVERSE9]])
-; CHECK-NEXT: [[INTERLEAVED_VEC11:%.*]] = call <vscale x 16 x i32> @llvm.vector.interleave2.nxv16i32(<vscale x 8 x i32> [[INTERLEAVED_VEC]], <vscale x 8 x i32> [[INTERLEAVED_VEC10]])
+; CHECK-NEXT: [[INTERLEAVED_VEC11:%.*]] = call <vscale x 16 x i32> @llvm.vector.interleave4.nxv16i32(<vscale x 4 x i32> [[REVERSE6]], <vscale x 4 x i32> [[REVERSE7]], <vscale x 4 x i32> [[REVERSE8]], <vscale x 4 x i32> [[REVERSE9]])
; CHECK-NEXT: store <vscale x 16 x i32> [[INTERLEAVED_VEC11]], ptr [[TMP26]], align 4
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP1]]
; CHECK-NEXT: [[VEC_IND_NEXT]] = add <vscale x 4 x i32> [[VEC_IND]], [[DOTSPLAT]]
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/sve-interleaved-masked-accesses.ll b/llvm/test/Transforms/LoopVectorize/AArch64/sve-interleaved-masked-accesses.ll
index 469faf67a71b3..3567aff0ace4e 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/sve-interleaved-masked-accesses.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/sve-interleaved-masked-ac...
[truncated]
|
mshockwave
reviewed
May 29, 2025
| @@ -3217,6 +3217,62 @@ static Value *createBitOrPointerCast(IRBuilderBase &Builder, Value *V, | |||
| return Builder.CreateBitOrPointerCast(CastVal, DstVTy); | |||
| } | |||
|
|
|||
| static Intrinsic::ID getInterleaveIntrinsicID(unsigned Factor) { | |||
There was a problem hiding this comment.
not this patch, but perhaps we can factor out utility functions like this and getIntrinsicFactor from InterleavedAccessPass into a public header file in the future
There was a problem hiding this comment.
Yeah, I could imagine SLP might need a function like this eventually?
There was a problem hiding this comment.
Maybe put this in VectorUtils?
fhahn
reviewed
May 30, 2025
| for (unsigned NumVectors = 1; NumVectors < InterleaveFactor; | ||
| NumVectors *= 2) { | ||
|
|
||
| if (InterleaveFactor <= 8) { |
There was a problem hiding this comment.
As the complexity grows here it would probably be good to lower scalable VPInterleaveRecipes to a set of VPWidenIntrinsic recipes before execute?
There was a problem hiding this comment.
Yeah, I think this would allow us to deduplicate some of the reversing logic too.
It's probably also worthwhile raising the question if we should just take out the recursive interleaving logic for now, and only support factors <= 8.
I think this is now dead code for in-tree targets, since the only targets that have scalable vectorization are AArch64 and RISC-V, which limit the factor to 4 and 8 respectively in TTI.
There was a problem hiding this comment.
I don't think there's any point supporting deinterleave factors > 8 until there is hardware that supports it. The isPowerOf2_32 calls can all be stripped, I think. 9E88
There was a problem hiding this comment.
To clarify, even though there's no hardware with instructions for [de]interleaving memory ops @ factor > 8 (and thus no intrinsics for the interleaved access pass to lower to), we can still generate a wide load/store and perform a series of [de]interleave permutations.
But I somewhat agree that if we're not exercising that code path in any target it should be stripped. Can I defer removing it to a separate PR so we can gain consensus there?
There was a problem hiding this comment.
Sounds good to me. You might want to put in a TODO comment in this PR, so that we don't forget?
|
FYI we might start looking into allowing interleaving up to a factor of 8 on AArch64, so what is currently dead code might not be so for very long after this is merged. It's not confirmed though. |
Up to 8 will still be covered without the need for recursive interleaving, do you plan on handling interleave factors > 8? I've added a test to exercise a scalable VF with factor 16 in 203d204 anyway. It turns out that even if the TTI hooks allow a higher factor in |
artagnon
reviewed
Jun 3, 2025
Comment on lines
357
to
360
| ; CHECK-NEXT: br i1 false, label [[SCALAR_PH:%.*]], label [[ENTRY:%.*]] | ||
| ; CHECK-NEXT: [[TMP0:%.*]] = call i64 @llvm.vscale.i64() | ||
| ; CHECK-NEXT: [[TMP1:%.*]] = mul i64 [[TMP0]], 4 | ||
| ; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 1024, [[TMP1]] | ||
| ; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[ENTRY:%.*]] |
There was a problem hiding this comment.
I'm confused about why this is jumping to entry? Is it an outdated UTC pattern?
There was a problem hiding this comment.
I think this is just an outdated FileCheck label name, it doesn't affect the correctness.
There was a problem hiding this comment.
Would be good to re-gen these tests?
| for (unsigned NumVectors = 1; NumVectors < InterleaveFactor; | ||
| NumVectors *= 2) { | ||
|
|
||
| if (InterleaveFactor <= 8) { |
There was a problem hiding this comment.
I don't think there's any point supporting deinterleave factors > 8 until there is hardware that supports it. The isPowerOf2_32 calls can all be stripped, I think.
| @@ -3217,6 +3217,62 @@ static Value *createBitOrPointerCast(IRBuilderBase &Builder, Value *V, | |||
| return Builder.CreateBitOrPointerCast(CastVal, DstVTy); | |||
| } | |||
|
|
|||
| static Intrinsic::ID getInterleaveIntrinsicID(unsigned Factor) { | |||
There was a problem hiding this comment.
Maybe put this in VectorUtils?
Comment on lines
+4582
to
+4583
| if (VecTy->isScalableTy() && !isPowerOf2_32(Factor)) | ||
| return InstructionCost::getInvalid(); |
There was a problem hiding this comment.
I'm confused about why factors 5 and 7 seem to be working fine?
There was a problem hiding this comment.
I guess you meant why factors or 5 and 7 used to work fine? I think that's because vectorizer has never asked for interleaving non-power-of-two factors before (until now, ofc)
There was a problem hiding this comment.
I think the TODO comment above confused me: why is factor 3 a TODO, even though the test seems to be fine?
There was a problem hiding this comment.
I don't think there is a factor 3 test for AArch64 in this patch.
There was a problem hiding this comment.
The loop vectorizer never called getInterleavedMemoryOpCost with a non-power-of-2 factor w/ scalable VF before, it bailed earlier in interleavedAccessCanBeWidened, so this is to preserve the existing behaviour. I.e. it currently only scalably vectorizes factors 2 and 4
artagnon
reviewed
Jun 3, 2025
| @@ -4575,6 +4575,13 @@ InstructionCost AArch64TTIImpl::getInterleavedMemoryOpCost( | |||
| if (VecTy->isScalableTy() && !ST->hasSVE()) | |||
| return InstructionCost::getInvalid(); | |||
|
|
|||
| // Scalable VFs will emit vector.de[interleave] intrinsics, and currently we | |||
There was a problem hiding this comment.
s/de[interleave]/[de]interleave/?
…ze/interleave-factors-3567
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Currently the loop vectorizer can only vectorize interleave groups for power-of-2 factors at scalable VFs by recursively interleaving [de]interleave2 intrinsics.
However after #124825 and #139893, we now have [de]interleave intrinsics for all factors up to 8, which is enough to support all types of segmented loads and stores on RISC-V.
Now that the interleaved access pass has been taught to lower these in #139373 and #141512, this patch teaches the loop vectorizer to emit these intrinsics for factors up to 8, which enables scalable vectorization for non-power-of-2 factors.
The existing recursive [ 8000 de]interleaving has been kept for power-of-2 factors > 8. The idea is that some targets might still want to perform a wide load/store and a series of [de]interleaves, even if there's no interleaved memory operation that interleaved access pass can lower to.
As far as I'm aware, no in-tree target will trigger this though because the maximum interleave factor is capped at 4 on AArch64 and 8 on RISC-V. This code could be removed if reviewers prefer, and added back when needed.
This patch shouldn't be landed until #141513 has landed, so that on the off-chance that the interleaved access pass doesn't convert an [de]interleave4 intrinsic it can still be lowered.
Factors of 3 with scalable VFs are also turned off in AArch64 since there's no lowering for [de]interleave3 just yet either.