[llvm] [SLP]Initial support for non-power-of-2 (but still whole register) number of elements in operands. (PR #107273)

Alexey Bataev via llvm-commits llvm-commits at lists.llvm.org
Mon Sep 9 10:22:45 PDT 2024


https://github.com/alexey-bataev updated https://github.com/llvm/llvm-project/pull/107273

>From f927502a0586c490d08b0491781deab6de9fd548 Mon Sep 17 00:00:00 2001
From: Alexey Bataev <a.bataev at outlook.com>
Date: Wed, 4 Sep 2024 17:27:55 +0000
Subject: [PATCH] =?UTF-8?q?[=F0=9D=98=80=F0=9D=97=BD=F0=9D=97=BF]=20initia?=
 =?UTF-8?q?l=20version?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Created using spr 1.3.5
---
 llvm/include/llvm/CodeGen/BasicTTIImpl.h      |  14 ++-
 .../Transforms/Vectorize/SLPVectorizer.cpp    | 109 ++++++++++++------
 .../reduction-whole-regs-loads.ll             |  28 +++--
 3 files changed, 104 insertions(+), 47 deletions(-)

diff --git a/llvm/include/llvm/CodeGen/BasicTTIImpl.h b/llvm/include/llvm/CodeGen/BasicTTIImpl.h
index 50dc7d5c54c54a..67ded1e21c4835 100644
--- a/llvm/include/llvm/CodeGen/BasicTTIImpl.h
+++ b/llvm/include/llvm/CodeGen/BasicTTIImpl.h
@@ -2531,7 +2531,19 @@ class BasicTTIImplBase : public TargetTransformInfoImplCRTPBase<T> {
 
   unsigned getNumberOfParts(Type *Tp) {
     std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Tp);
-    return LT.first.isValid() ? *LT.first.getValue() : 0;
+    if (!LT.first.isValid())
+      return 0;
+    // Try to find actual number of parts for non-power-of-2 elements as
+    // ceil(num-of-elements/num-of-subtype-elements).
+    if (auto *FTp = dyn_cast<FixedVectorType>(Tp);
+        Tp && LT.second.isFixedLengthVector() &&
+        !has_single_bit(FTp->getNumElements())) {
+      if (auto *SubTp = dyn_cast_if_present<FixedVectorType>(
+              EVT(LT.second).getTypeForEVT(Tp->getContext()));
+          SubTp && SubTp->getElementType() == FTp->getElementType())
+        return divideCeil(FTp->getNumElements(), SubTp->getNumElements());
+    }
+    return *LT.first.getValue();
   }
 
   InstructionCost getAddressComputationCost(Type *Ty, ScalarEvolution *,
diff --git a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
index 60476398e5ca75..3c647b36e98490 100644
--- a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
+++ b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
@@ -260,6 +260,20 @@ static FixedVectorType *getWidenedType(Type *ScalarTy, unsigned VF) {
                               VF * getNumElements(ScalarTy));
 }
 
+/// Returns the number of elements of the given type \p Ty, not less than \p Sz,
+/// which forms type, which splits by \p TTI into whole vector types during
+/// legalization.
+static unsigned getFullVectorNumberOfElements(const TargetTransformInfo &TTI,
+                                              Type *Ty, unsigned Sz) {
+  if (!isValidElementType(Ty))
+    return bit_ceil(Sz);
+  // Find the number of elements, which forms full vectors.
+  const unsigned NumParts = TTI.getNumberOfParts(getWidenedType(Ty, Sz));
+  if (NumParts == 0 || NumParts >= Sz)
+    return bit_ceil(Sz);
+  return bit_ceil(divideCeil(Sz, NumParts)) * NumParts;
+}
+
 static void transformScalarShuffleIndiciesToVector(unsigned VecTyNumElements,
                                                    SmallVectorImpl<int> &Mask) {
   // The ShuffleBuilder implementation use shufflevector to splat an "element".
@@ -394,7 +408,7 @@ static bool isVectorLikeInstWithConstOps(Value *V) {
 /// total number of elements \p Size and number of registers (parts) \p
 /// NumParts.
 static unsigned getPartNumElems(unsigned Size, unsigned NumParts) {
-  return PowerOf2Ceil(divideCeil(Size, NumParts));
+  return std::min<unsigned>(Size, bit_ceil(divideCeil(Size, NumParts)));
 }
 
 /// Returns correct remaining number of elements, considering total amount \p
@@ -1223,6 +1237,22 @@ static bool doesNotNeedToSchedule(ArrayRef<Value *> VL) {
          (all_of(VL, isUsedOutsideBlock) || all_of(VL, areAllOperandsNonInsts));
 }
 
+/// Returns true if widened type of \p Ty elements with size \p Sz represents
+/// full vector type, i.e. adding extra element results in extra parts upon type
+/// legalization.
+static bool hasFullVectorsOnly(const TargetTransformInfo &TTI, Type *Ty,
+                               unsigned Sz) {
+  if (Sz <= 1)
+    return false;
+  if (!isValidElementType(Ty) && !isa<FixedVectorType>(Ty))
+    return false;
+  if (has_single_bit(Sz))
+    return true;
+  const unsigned NumParts = TTI.getNumberOfParts(getWidenedType(Ty, Sz));
+  return NumParts > 0 && NumParts < Sz && has_single_bit(Sz / NumParts) &&
+         Sz % NumParts == 0;
+}
+
 namespace slpvectorizer {
 
 /// Bottom Up SLP Vectorizer.
@@ -2466,7 +2496,9 @@ class BoUpSLP {
         }
         // TODO: Check if we can remove a check for non-power-2 number of
         // scalars after full support of non-power-2 vectorization.
-        return UniqueValues.size() != 2 && has_single_bit(UniqueValues.size());
+        return UniqueValues.size() != 2 &&
+               hasFullVectorsOnly(*R.TTI, (*UniqueValues.begin())->getType(),
+                                  UniqueValues.size());
       };
 
       // If the initial strategy fails for any of the operand indexes, then we
@@ -3275,8 +3307,9 @@ class BoUpSLP {
                           SmallVectorImpl<Value *> *AltScalars = nullptr) const;
 
     /// Return true if this is a non-power-of-2 node.
-    bool isNonPowOf2Vec() const {
-      bool IsNonPowerOf2 = !has_single_bit(Scalars.size());
+    bool isNonPowOf2Vec(const TargetTransformInfo &TTI) const {
+      bool IsNonPowerOf2 = !hasFullVectorsOnly(
+          TTI, getValueType(Scalars.front()), Scalars.size());
       assert((!IsNonPowerOf2 || ReuseShuffleIndices.empty()) &&
              "Reshuffling not supported with non-power-of-2 vectors yet.");
       return IsNonPowerOf2;
@@ -3454,7 +3487,7 @@ class BoUpSLP {
 
     if (UserTreeIdx.UserTE) {
       Last->UserTreeIndices.push_back(UserTreeIdx);
-      assert((!Last->isNonPowOf2Vec() || Last->ReorderIndices.empty()) &&
+      assert((!Last->isNonPowOf2Vec(*TTI) || Last->ReorderIndices.empty()) &&
              "Reordering isn't implemented for non-power-of-2 nodes yet");
     }
     return Last;
@@ -4732,7 +4765,7 @@ BoUpSLP::LoadsState BoUpSLP::canVectorizeLoads(
   // Check the order of pointer operands or that all pointers are the same.
   bool IsSorted = sortPtrAccesses(PointerOps, ScalarTy, *DL, *SE, Order);
   // FIXME: Reordering isn't implemented for non-power-of-2 nodes yet.
-  if (!Order.empty() && !has_single_bit(VL.size())) {
+  if (!Order.empty() && !hasFullVectorsOnly(*TTI, ScalarTy, Sz)) {
     assert(VectorizeNonPowerOf2 && "non-power-of-2 number of loads only "
                                    "supported with VectorizeNonPowerOf2");
     return LoadsState::Gather;
@@ -4786,12 +4819,13 @@ BoUpSLP::LoadsState BoUpSLP::canVectorizeLoads(
                  });
         });
     const unsigned AbsoluteDiff = std::abs(*Diff);
-    if (IsPossibleStrided && (IsAnyPointerUsedOutGraph ||
-                              ((Sz > MinProfitableStridedLoads ||
-                                (AbsoluteDiff <= MaxProfitableLoadStride * Sz &&
-                                 has_single_bit(AbsoluteDiff))) &&
-                               AbsoluteDiff > Sz) ||
-                              *Diff == -(static_cast<int>(Sz) - 1))) {
+    if (IsPossibleStrided &&
+        (IsAnyPointerUsedOutGraph ||
+         ((Sz > MinProfitableStridedLoads ||
+           (AbsoluteDiff <= MaxProfitableLoadStride * Sz &&
+            hasFullVectorsOnly(*TTI, ScalarTy, AbsoluteDiff))) &&
+          AbsoluteDiff > Sz) ||
+         *Diff == -(static_cast<int>(Sz) - 1))) {
       int Stride = *Diff / static_cast<int>(Sz - 1);
       if (*Diff == Stride * static_cast<int>(Sz - 1)) {
         Align Alignment =
@@ -5196,7 +5230,7 @@ static bool areTwoInsertFromSameBuildVector(
 std::optional<BoUpSLP::OrdersType>
 BoUpSLP::getReorderingData(const TreeEntry &TE, bool TopToBottom) {
   // FIXME: Vectorizing is not supported yet for non-power-of-2 ops.
-  if (TE.isNonPowOf2Vec())
+  if (TE.isNonPowOf2Vec(*TTI))
     return std::nullopt;
 
   // No need to reorder if need to shuffle reuses, still need to shuffle the
@@ -5580,7 +5614,7 @@ void BoUpSLP::reorderTopToBottom() {
 
   // Reorder the graph nodes according to their vectorization factor.
   for (unsigned VF = VectorizableTree.front()->getVectorFactor(); VF > 1;
-       VF /= 2) {
+       VF -= 2) {
     auto It = VFToOrderedEntries.find(VF);
     if (It == VFToOrderedEntries.end())
       continue;
@@ -5753,7 +5787,7 @@ bool BoUpSLP::canReorderOperands(
     ArrayRef<TreeEntry *> ReorderableGathers,
     SmallVectorImpl<TreeEntry *> &GatherOps) {
   // FIXME: Reordering isn't implemented for non-power-of-2 nodes yet.
-  if (UserTE->isNonPowOf2Vec())
+  if (UserTE->isNonPowOf2Vec(*TTI))
     return false;
 
   for (unsigned I = 0, E = UserTE->getNumOperands(); I < E; ++I) {
@@ -5928,7 +5962,7 @@ void BoUpSLP::reorderBottomToTop(bool IgnoreReorder) {
         auto Res = OrdersUses.insert(std::make_pair(OrdersType(), 0));
         const auto AllowsReordering = [&](const TreeEntry *TE) {
           // FIXME: Reordering isn't implemented for non-power-of-2 nodes yet.
-          if (TE->isNonPowOf2Vec())
+          if (TE->isNonPowOf2Vec(*TTI))
             return false;
           if (!TE->ReorderIndices.empty() || !TE->ReuseShuffleIndices.empty() ||
               (TE->State == TreeEntry::Vectorize && TE->isAltShuffle()) ||
@@ -6581,7 +6615,7 @@ BoUpSLP::TreeEntry::EntryState BoUpSLP::getScalarsVectorizationState(
   case Instruction::ExtractElement: {
     bool Reuse = canReuseExtract(VL, VL0, CurrentOrder);
     // FIXME: Vectorizing is not supported yet for non-power-of-2 ops.
-    if (!has_single_bit(VL.size()))
+    if (!hasFullVectorsOnly(*TTI, VL0->getType(), VL.size()))
       return TreeEntry::NeedToGather;
     if (Reuse || !CurrentOrder.empty())
       return TreeEntry::Vectorize;
@@ -6987,24 +7021,25 @@ void BoUpSLP::buildTree_rec(ArrayRef<Value *> VL, unsigned Depth,
         UniqueValues.emplace_back(V);
     }
     size_t NumUniqueScalarValues = UniqueValues.size();
-    if (NumUniqueScalarValues == VL.size()) {
+    bool IsFullVectors =
+        hasFullVectorsOnly(*TTI, UniqueValues.front()->getType(),
+                           NumUniqueScalarValues);
+    if (NumUniqueScalarValues == VL.size() &&
+        (VectorizeNonPowerOf2 || IsFullVectors)) {
       ReuseShuffleIndices.clear();
     } else {
       // FIXME: Reshuffing scalars is not supported yet for non-power-of-2 ops.
-      if (UserTreeIdx.UserTE && UserTreeIdx.UserTE->isNonPowOf2Vec()) {
+      if (UserTreeIdx.UserTE && UserTreeIdx.UserTE->isNonPowOf2Vec(*TTI)) {
         LLVM_DEBUG(dbgs() << "SLP: Reshuffling scalars not yet supported "
                              "for nodes with padding.\n");
         newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx);
         return false;
       }
       LLVM_DEBUG(dbgs() << "SLP: Shuffle for reused scalars.\n");
-      if (NumUniqueScalarValues <= 1 ||
-          (UniquePositions.size() == 1 && all_of(UniqueValues,
-                                                 [](Value *V) {
-                                                   return isa<UndefValue>(V) ||
-                                                          !isConstant(V);
-                                                 })) ||
-          !llvm::has_single_bit<uint32_t>(NumUniqueScalarValues)) {
+      if (NumUniqueScalarValues <= 1 || !IsFullVectors ||
+          (UniquePositions.size() == 1 && all_of(UniqueValues, [](Value *V) {
+             return isa<UndefValue>(V) || !isConstant(V);
+           }))) {
         if (DoNotFail && UniquePositions.size() > 1 &&
             NumUniqueScalarValues > 1 && S.MainOp->isSafeToRemove() &&
             all_of(UniqueValues, [=](Value *V) {
@@ -7012,7 +7047,9 @@ void BoUpSLP::buildTree_rec(ArrayRef<Value *> VL, unsigned Depth,
                      areAllUsersVectorized(cast<Instruction>(V),
                                            UserIgnoreList);
             })) {
-          unsigned PWSz = PowerOf2Ceil(UniqueValues.size());
+          // Find the number of elements, which forms full vectors.
+          unsigned PWSz = getFullVectorNumberOfElements(
+              *TTI, UniqueValues.front()->getType(), UniqueValues.size());
           if (PWSz == VL.size()) {
             ReuseShuffleIndices.clear();
           } else {
@@ -9107,9 +9144,6 @@ class BoUpSLP::ShuffleCostEstimator : public BaseShuffleAnalysis {
       return nullptr;
     Value *VecBase = nullptr;
     ArrayRef<Value *> VL = E->Scalars;
-    // If the resulting type is scalarized, do not adjust the cost.
-    if (NumParts == VL.size())
-      return nullptr;
     // Check if it can be considered reused if same extractelements were
     // vectorized already.
     bool PrevNodeFound = any_of(
@@ -9762,7 +9796,7 @@ BoUpSLP::getEntryCost(const TreeEntry *E, ArrayRef<Value *> VectorizedVals,
       InsertMask[Idx] = I + 1;
     }
     unsigned VecScalarsSz = PowerOf2Ceil(NumElts);
-    if (NumOfParts > 0)
+    if (NumOfParts > 0 && NumOfParts < NumElts)
       VecScalarsSz = PowerOf2Ceil((NumElts + NumOfParts - 1) / NumOfParts);
     unsigned VecSz = (1 + OffsetEnd / VecScalarsSz - OffsetBeg / VecScalarsSz) *
                      VecScalarsSz;
@@ -10991,7 +11025,9 @@ InstructionCost BoUpSLP::getTreeCost(ArrayRef<Value *> VectorizedVals) {
           // Keep original scalar if number of externally used instructions in
           // the same entry is not power of 2. It may help to do some extra
           // vectorization for now.
-          KeepScalar = ScalarUsesCount <= 1 || !has_single_bit(ScalarUsesCount);
+          KeepScalar =
+              ScalarUsesCount <= 1 ||
+              !hasFullVectorsOnly(*TTI, EU.Scalar->getType(), ScalarUsesCount);
         }
         if (KeepScalar) {
           ExternalUsesAsOriginalScalar.insert(EU.Scalar);
@@ -11688,13 +11724,14 @@ BoUpSLP::isGatherShuffledEntry(
   if (TE == VectorizableTree.front().get())
     return {};
   // FIXME: Gathering for non-power-of-2 nodes not implemented yet.
-  if (TE->isNonPowOf2Vec())
+  if (TE->isNonPowOf2Vec(*TTI))
     return {};
   Mask.assign(VL.size(), PoisonMaskElem);
   assert(TE->UserTreeIndices.size() == 1 &&
          "Expected only single user of the gather node.");
-  assert(VL.size() % NumParts == 0 &&
-         "Number of scalars must be divisible by NumParts.");
+  // Number of scalars must be divisible by NumParts.
+  if (VL.size() % NumParts != 0)
+    return {};
   unsigned SliceSize = getPartNumElems(VL.size(), NumParts);
   SmallVector<std::optional<TTI::ShuffleKind>> Res;
   for (unsigned Part : seq<unsigned>(NumParts)) {
@@ -17005,7 +17042,7 @@ bool SLPVectorizerPass::tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R,
     for (unsigned I = NextInst; I < MaxInst; ++I) {
       unsigned ActualVF = std::min(MaxInst - I, VF);
 
-      if (!has_single_bit(ActualVF))
+      if (!hasFullVectorsOnly(*TTI, ScalarTy, ActualVF))
         continue;
 
       if (MaxVFOnly && ActualVF < MaxVF)
diff --git a/llvm/test/Transforms/SLPVectorizer/reduction-whole-regs-loads.ll b/llvm/test/Transforms/SLPVectorizer/reduction-whole-regs-loads.ll
index 281b5f99540eab..4074b8654362e0 100644
--- a/llvm/test/Transforms/SLPVectorizer/reduction-whole-regs-loads.ll
+++ b/llvm/test/Transforms/SLPVectorizer/reduction-whole-regs-loads.ll
@@ -1,21 +1,29 @@
 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
-; RUN: opt < %s -passes=slp-vectorizer -S -mtriple=riscv64-unknown-linux -mattr=+v -slp-threshold=-100 | FileCheck %s
+; RUN: opt < %s -passes=slp-vectorizer -S -mtriple=riscv64-unknown-linux -mattr=+v -slp-threshold=-100 | FileCheck %s --check-prefix=RISCV
 ; RUN: opt < %s -passes=slp-vectorizer -S -mtriple=x86_64-unknown-linux -slp-threshold=-100 | FileCheck %s
 ; RUN: opt < %s -passes=slp-vectorizer -S -mtriple=aarch64-unknown-linux -slp-threshold=-100 | FileCheck %s
 ; REQUIRES: aarch64-registered-target, x86-registered-target, riscv-registered-target
 
 define i64 @test(ptr %p) {
+; RISCV-LABEL: @test(
+; RISCV-NEXT:  entry:
+; RISCV-NEXT:    [[ARRAYIDX_4:%.*]] = getelementptr inbounds i64, ptr [[P:%.*]], i64 4
+; RISCV-NEXT:    [[TMP0:%.*]] = load <4 x i64>, ptr [[P]], align 4
+; RISCV-NEXT:    [[TMP1:%.*]] = load <2 x i64>, ptr [[ARRAYIDX_4]], align 4
+; RISCV-NEXT:    [[TMP2:%.*]] = shufflevector <4 x i64> [[TMP0]], <4 x i64> poison, <8 x i32> <i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 0, i32 0>
+; RISCV-NEXT:    [[TMP3:%.*]] = call <8 x i64> @llvm.vector.insert.v8i64.v4i64(<8 x i64> [[TMP2]], <4 x i64> [[TMP0]], i64 0)
+; RISCV-NEXT:    [[TMP4:%.*]] = call <8 x i64> @llvm.vector.insert.v8i64.v2i64(<8 x i64> [[TMP3]], <2 x i64> [[TMP1]], i64 4)
+; RISCV-NEXT:    [[TMP5:%.*]] = mul <8 x i64> [[TMP4]], <i64 42, i64 42, i64 42, i64 42, i64 42, i64 42, i64 42, i64 42>
+; RISCV-NEXT:    [[TMP6:%.*]] = call i64 @llvm.vector.reduce.add.v8i64(<8 x i64> [[TMP5]])
+; RISCV-NEXT:    ret i64 [[TMP6]]
+;
 ; CHECK-LABEL: @test(
 ; CHECK-NEXT:  entry:
-; CHECK-NEXT:    [[ARRAYIDX_4:%.*]] = getelementptr inbounds i64, ptr [[P:%.*]], i64 4
-; CHECK-NEXT:    [[TMP0:%.*]] = load <4 x i64>, ptr [[P]], align 4
-; CHECK-NEXT:    [[TMP1:%.*]] = load <2 x i64>, ptr [[ARRAYIDX_4]], align 4
-; CHECK-NEXT:    [[TMP2:%.*]] = shufflevector <4 x i64> [[TMP0]], <4 x i64> poison, <8 x i32> <i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 0, i32 0>
-; CHECK-NEXT:    [[TMP3:%.*]] = call <8 x i64> @llvm.vector.insert.v8i64.v4i64(<8 x i64> [[TMP2]], <4 x i64> [[TMP0]], i64 0)
-; CHECK-NEXT:    [[TMP4:%.*]] = call <8 x i64> @llvm.vector.insert.v8i64.v2i64(<8 x i64> [[TMP3]], <2 x i64> [[TMP1]], i64 4)
-; CHECK-NEXT:    [[TMP5:%.*]] = mul <8 x i64> [[TMP4]], <i64 42, i64 42, i64 42, i64 42, i64 42, i64 42, i64 42, i64 42>
-; CHECK-NEXT:    [[TMP6:%.*]] = call i64 @llvm.vector.reduce.add.v8i64(<8 x i64> [[TMP5]])
-; CHECK-NEXT:    ret i64 [[TMP6]]
+; CHECK-NEXT:    [[TMP0:%.*]] = load <6 x i64>, ptr [[P:%.*]], align 4
+; CHECK-NEXT:    [[TMP1:%.*]] = shufflevector <6 x i64> [[TMP0]], <6 x i64> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 0, i32 0>
+; CHECK-NEXT:    [[TMP2:%.*]] = mul <8 x i64> [[TMP1]], <i64 42, i64 42, i64 42, i64 42, i64 42, i64 42, i64 42, i64 42>
+; CHECK-NEXT:    [[TMP3:%.*]] = call i64 @llvm.vector.reduce.add.v8i64(<8 x i64> [[TMP2]])
+; CHECK-NEXT:    ret i64 [[TMP3]]
 ;
 entry:
   %arrayidx.1 = getelementptr inbounds i64, ptr %p, i64 1



More information about the llvm-commits mailing list