[llvm] [AArch64] Add MATCH loops to LoopIdiomVectorizePass (PR #101976)

via llvm-commits llvm-commits at lists.llvm.org
Mon Aug 5 06:28:26 PDT 2024


llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-llvm-transforms

Author: Ricardo Jesus (rj-jesus)

<details>
<summary>Changes</summary>

This patch adds a new loop to LoopIdiomVectorizePass, enabling it to
recognise and vectorise loops such as:
```cpp
template<class InputIt, class ForwardIt>
InputIt find_first_of(InputIt first, InputIt last,
                      ForwardIt s_first, ForwardIt s_last)
{
  for (; first != last; ++first)
    for (ForwardIt it = s_first; it != s_last; ++it)
      if (*first == *it)
        return first;
  return last;
}
```

These loops match the C++ standard library function std::find_first_of.

The loops are vectorised using `@<!-- -->experimental.vector.match` in #<!-- -->101974.

---

Patch is 41.43 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/101976.diff


12 Files Affected:

- (modified) llvm/docs/LangRef.rst (+45) 
- (modified) llvm/include/llvm/Analysis/TargetTransformInfo.h (+9) 
- (modified) llvm/include/llvm/Analysis/TargetTransformInfoImpl.h (+2) 
- (modified) llvm/include/llvm/IR/Intrinsics.td (+10) 
- (modified) llvm/lib/Analysis/TargetTransformInfo.cpp (+5) 
- (modified) llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp (+9) 
- (modified) llvm/lib/Target/AArch64/AArch64ISelLowering.cpp (+46) 
- (modified) llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp (+12) 
- (modified) llvm/lib/Target/AArch64/AArch64TargetTransformInfo.h (+2) 
- (modified) llvm/lib/Transforms/Vectorize/LoopIdiomVectorize.cpp (+441-1) 
- (added) llvm/test/CodeGen/AArch64/find-first-byte.ll (+120) 
- (added) llvm/test/CodeGen/AArch64/intrinsic-vector-match-sve2.ll (+57) 


``````````diff
diff --git a/llvm/docs/LangRef.rst b/llvm/docs/LangRef.rst
index b17e3c828ed3d..dd9851d1af078 100644
--- a/llvm/docs/LangRef.rst
+++ b/llvm/docs/LangRef.rst
@@ -19637,6 +19637,51 @@ are undefined.
     }
 
 
+'``llvm.experimental.vector.match.*``' Intrinsic
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+Syntax:
+"""""""
+
+This is an overloaded intrinsic. Support for specific vector types is target
+dependent.
+
+::
+
+    declare <<n> x i1> @llvm.experimental.vector.match(<<n> x <ty>> %op1, <<n> x <ty>> %op2, <<n> x i1> %mask, i32 <segsize>)
+    declare <vscale x <n> x i1> @llvm.experimental.vector.match(<vscale x <n> x <ty>> %op1, <vscale x <n> x <ty>> %op2, <vscale x <n> x i1> %mask, i32 <segsize>)
+
+Overview:
+"""""""""
+
+Find elements of the first argument matching any elements of the second.
+
+Arguments:
+""""""""""
+
+The first argument is the search vector, the second argument is the vector of
+elements we are searching for (i.e. for which we consider a match successful),
+and the third argument is a mask that controls which elements of the first
+argument are active. The fourth argument is an immediate that sets the segment
+size for the search window.
+
+Semantics:
+""""""""""
+
+The '``llvm.experimental.vector.match``' intrinsic compares each element in the
+first argument against potentially several elements of the second, placing
+``1`` in the corresponding element of the output vector if any comparison is
+successful, and ``0`` otherwise. Inactive elements in the mask are set to ``0``
+in the output. The segment size controls the number of elements of the second
+argument that are compared against.
+
+For example, for vectors with 16 elements, if ``segsize = 16`` then each
+element of the first argument is compared against all 16 elements of the second
+argument; but if ``segsize = 4``, then each of the first four elements of the
+first argument is compared against the first four elements of the second
+argument, each of the second four elements of the first argument is compared
+against the second four elements of the second argument, and so forth.
+
 Matrix Intrinsics
 -----------------
 
diff --git a/llvm/include/llvm/Analysis/TargetTransformInfo.h b/llvm/include/llvm/Analysis/TargetTransformInfo.h
index 38e8b9da21397..786c13a177ccf 100644
--- a/llvm/include/llvm/Analysis/TargetTransformInfo.h
+++ b/llvm/include/llvm/Analysis/TargetTransformInfo.h
@@ -1746,6 +1746,10 @@ class TargetTransformInfo {
   bool hasActiveVectorLength(unsigned Opcode, Type *DataType,
                              Align Alignment) const;
 
+  /// \returns Returns true if the target supports vector match operations for
+  /// the vector type `VT` using a segment size of `SegSize`.
+  bool hasVectorMatch(VectorType *VT, unsigned SegSize) const;
+
   struct VPLegalization {
     enum VPTransform {
       // keep the predicating parameter
@@ -2184,6 +2188,7 @@ class TargetTransformInfo::Concept {
   virtual bool supportsScalableVectors() const = 0;
   virtual bool hasActiveVectorLength(unsigned Opcode, Type *DataType,
                                      Align Alignment) const = 0;
+  virtual bool hasVectorMatch(VectorType *VT, unsigned SegSize) const = 0;
   virtual VPLegalization
   getVPLegalizationStrategy(const VPIntrinsic &PI) const = 0;
   virtual bool hasArmWideBranch(bool Thumb) const = 0;
@@ -2952,6 +2957,10 @@ class TargetTransformInfo::Model final : public TargetTransformInfo::Concept {
     return Impl.hasActiveVectorLength(Opcode, DataType, Alignment);
   }
 
+  bool hasVectorMatch(VectorType *VT, unsigned SegSize) const override {
+    return Impl.hasVectorMatch(VT, SegSize);
+  }
+
   VPLegalization
   getVPLegalizationStrategy(const VPIntrinsic &PI) const override {
     return Impl.getVPLegalizationStrategy(PI);
diff --git a/llvm/include/llvm/Analysis/TargetTransformInfoImpl.h b/llvm/include/llvm/Analysis/TargetTransformInfoImpl.h
index d208a710bb27f..36621861ab8c8 100644
--- a/llvm/include/llvm/Analysis/TargetTransformInfoImpl.h
+++ b/llvm/include/llvm/Analysis/TargetTransformInfoImpl.h
@@ -958,6 +958,8 @@ class TargetTransformInfoImplBase {
     return false;
   }
 
+  bool hasVectorMatch(VectorType *VT, unsigned SegSize) const { return false; }
+
   TargetTransformInfo::VPLegalization
   getVPLegalizationStrategy(const VPIntrinsic &PI) const {
     return TargetTransformInfo::VPLegalization(
diff --git a/llvm/include/llvm/IR/Intrinsics.td b/llvm/include/llvm/IR/Intrinsics.td
index b4e758136b39f..f6d77aa596f60 100644
--- a/llvm/include/llvm/IR/Intrinsics.td
+++ b/llvm/include/llvm/IR/Intrinsics.td
@@ -1892,6 +1892,16 @@ def int_experimental_vector_histogram_add : DefaultAttrsIntrinsic<[],
                                LLVMScalarOrSameVectorWidth<0, llvm_i1_ty>], // Mask
                              [ IntrArgMemOnly ]>;
 
+// Experimental match
+def int_experimental_vector_match : DefaultAttrsIntrinsic<
+                             [ LLVMScalarOrSameVectorWidth<0, llvm_i1_ty> ],
+                             [ llvm_anyvector_ty,
+                               LLVMMatchType<0>,
+                               LLVMScalarOrSameVectorWidth<0, llvm_i1_ty>,  // Mask
+                               llvm_i32_ty ],  // Segment size
+                             [ IntrNoMem, IntrNoSync, IntrWillReturn,
+                               ImmArg<ArgIndex<3>> ]>;
+
 // Operators
 let IntrProperties = [IntrNoMem, IntrNoSync, IntrWillReturn] in {
   // Integer arithmetic
diff --git a/llvm/lib/Analysis/TargetTransformInfo.cpp b/llvm/lib/Analysis/TargetTransformInfo.cpp
index dcde78925bfa9..d8314af0537fe 100644
--- a/llvm/lib/Analysis/TargetTransformInfo.cpp
+++ b/llvm/lib/Analysis/TargetTransformInfo.cpp
@@ -1352,6 +1352,11 @@ bool TargetTransformInfo::hasActiveVectorLength(unsigned Opcode, Type *DataType,
   return TTIImpl->hasActiveVectorLength(Opcode, DataType, Alignment);
 }
 
+bool TargetTransformInfo::hasVectorMatch(VectorType *VT,
+                                         unsigned SegSize) const {
+  return TTIImpl->hasVectorMatch(VT, SegSize);
+}
+
 TargetTransformInfo::Concept::~Concept() = default;
 
 TargetIRAnalysis::TargetIRAnalysis() : TTICallback(&getDefaultTTI) {}
diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
index 9d617c7acd13c..9cb7d65975b9f 100644
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
@@ -8096,6 +8096,15 @@ void SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I,
              DAG.getNode(ISD::EXTRACT_SUBVECTOR, sdl, ResultVT, Vec, Index));
     return;
   }
+  case Intrinsic::experimental_vector_match: {
+    auto *VT = dyn_cast<VectorType>(I.getOperand(0)->getType());
+    auto SegmentSize = cast<ConstantInt>(I.getOperand(3))->getLimitedValue();
+    const auto &TTI =
+        TLI.getTargetMachine().getTargetTransformInfo(*I.getFunction());
+    assert(VT && TTI.hasVectorMatch(VT, SegmentSize) && "Unsupported type!");
+    visitTargetIntrinsic(I, Intrinsic);
+    return;
+  }
   case Intrinsic::vector_reverse:
     visitVectorReverse(I);
     return;
diff --git a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
index 7704321a0fc3a..050807142fc0a 100644
--- a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
+++ b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
@@ -6106,6 +6106,51 @@ SDValue AArch64TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
         DAG.getNode(AArch64ISD::CTTZ_ELTS, dl, MVT::i64, CttzOp);
     return DAG.getZExtOrTrunc(NewCttzElts, dl, Op.getValueType());
   }
+  case Intrinsic::experimental_vector_match: {
+    SDValue ID =
+        DAG.getTargetConstant(Intrinsic::aarch64_sve_match, dl, MVT::i64);
+
+    auto Op1 = Op.getOperand(1);
+    auto Op2 = Op.getOperand(2);
+    auto Mask = Op.getOperand(3);
+    auto SegmentSize =
+        cast<ConstantSDNode>(Op.getOperand(4))->getLimitedValue();
+
+    EVT VT = Op.getValueType();
+    auto MinNumElts = VT.getVectorMinNumElements();
+
+    assert(Op1.getValueType() == Op2.getValueType() && "Type mismatch.");
+    assert(Op1.getValueSizeInBits().getKnownMinValue() == 128 &&
+           "Custom lower only works on 128-bit segments.");
+    assert((Op1.getValueType().getVectorElementType() == MVT::i8  ||
+            Op1.getValueType().getVectorElementType() == MVT::i16) &&
+           "Custom lower only supports 8-bit or 16-bit characters.");
+    assert(SegmentSize == MinNumElts && "Custom lower needs segment size to "
+                                        "match minimum number of elements.");
+
+    if (VT.isScalableVector())
+      return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT, ID, Mask, Op1, Op2);
+
+    // We can use the SVE2 match instruction to lower this intrinsic by
+    // converting the operands to scalable vectors, doing a match, and then
+    // extracting a fixed-width subvector from the scalable vector.
+
+    EVT OpVT = Op1.getValueType();
+    EVT OpContainerVT = getContainerForFixedLengthVector(DAG, OpVT);
+    EVT MatchVT = OpContainerVT.changeElementType(MVT::i1);
+
+    auto ScalableOp1 = convertToScalableVector(DAG, OpContainerVT, Op1);
+    auto ScalableOp2 = convertToScalableVector(DAG, OpContainerVT, Op2);
+    auto ScalableMask = DAG.getNode(ISD::SIGN_EXTEND, dl, OpVT, Mask);
+    ScalableMask = convertFixedMaskToScalableVector(ScalableMask, DAG);
+
+    SDValue Match = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MatchVT, ID,
+                                ScalableMask, ScalableOp1, ScalableOp2);
+
+    return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, OpVT,
+                       DAG.getNode(ISD::SIGN_EXTEND, dl, OpContainerVT, Match),
+                       DAG.getVectorIdxConstant(0, dl));
+  }
   }
 }
 
@@ -26544,6 +26589,7 @@ void AArch64TargetLowering::ReplaceNodeResults(
       Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, VT, V));
       return;
     }
+    case Intrinsic::experimental_vector_match:
     case Intrinsic::get_active_lane_mask: {
       if (!VT.isFixedLengthVector() || VT.getVectorElementType() != MVT::i1)
         return;
diff --git a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
index b8f19fa87e2ab..806dc856c5862 100644
--- a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
+++ b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
@@ -3835,6 +3835,18 @@ bool AArch64TTIImpl::isLegalToVectorizeReduction(
   }
 }
 
+bool AArch64TTIImpl::hasVectorMatch(VectorType *VT, unsigned SegSize) const {
+  // Check that the target has SVE2 (and SVE is available), that `VT' is a
+  // legal type for MATCH, and that the segment size is 128-bit.
+  if (ST->hasSVE2() && ST->isSVEAvailable() &&
+      VT->getPrimitiveSizeInBits().getKnownMinValue() == 128 &&
+      VT->getElementCount().getKnownMinValue() == SegSize &&
+      (VT->getElementCount().getKnownMinValue() ==  8 ||
+       VT->getElementCount().getKnownMinValue() == 16))
+    return true;
+  return false;
+}
+
 InstructionCost
 AArch64TTIImpl::getMinMaxReductionCost(Intrinsic::ID IID, VectorType *Ty,
                                        FastMathFlags FMF,
diff --git a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.h b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.h
index a9189fd53f40b..6ad21a9e0a77a 100644
--- a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.h
+++ b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.h
@@ -391,6 +391,8 @@ class AArch64TTIImpl : public BasicTTIImplBase<AArch64TTIImpl> {
     return ST->hasSVE();
   }
 
+  bool hasVectorMatch(VectorType *VT, unsigned SegSize) const;
+
   InstructionCost getArithmeticReductionCost(unsigned Opcode, VectorType *Ty,
                                              std::optional<FastMathFlags> FMF,
                                              TTI::TargetCostKind CostKind);
diff --git a/llvm/lib/Transforms/Vectorize/LoopIdiomVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopIdiomVectorize.cpp
index cb31e2a2ecaec..a9683f08c5ab9 100644
--- a/llvm/lib/Transforms/Vectorize/LoopIdiomVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopIdiomVectorize.cpp
@@ -79,6 +79,12 @@ static cl::opt<unsigned>
               cl::desc("The vectorization factor for byte-compare patterns."),
               cl::init(16));
 
+static cl::opt<bool>
+    DisableFindFirstByte("disable-loop-idiom-vectorize-find-first-byte",
+                         cl::Hidden, cl::init(false),
+                         cl::desc("Proceed with Loop Idiom Vectorize Pass, but "
+                                  "do not convert find-first-byte loop(s)."));
+
 static cl::opt<bool>
     VerifyLoops("loop-idiom-vectorize-verify", cl::Hidden, cl::init(false),
                 cl::desc("Verify loops generated Loop Idiom Vectorize Pass."));
@@ -136,6 +142,21 @@ class LoopIdiomVectorize {
                             PHINode *IndPhi, Value *MaxLen, Instruction *Index,
                             Value *Start, bool IncIdx, BasicBlock *FoundBB,
                             BasicBlock *EndBB);
+
+  bool recognizeFindFirstByte();
+
+  Value *expandFindFirstByte(IRBuilder<> &Builder, DomTreeUpdater &DTU,
+                             unsigned VF, unsigned CharWidth,
+                             BasicBlock *ExitSucc, BasicBlock *ExitFail,
+                             GetElementPtrInst *GEPA, GetElementPtrInst *GEPB,
+                             Value *StartA, Value *EndA,
+                             Value *StartB, Value *EndB);
+
+  void transformFindFirstByte(PHINode *IndPhi, unsigned VF, unsigned CharWidth,
+                              BasicBlock *ExitSucc, BasicBlock *ExitFail,
+                              GetElementPtrInst *GEPA, GetElementPtrInst *GEPB,
+                              Value *StartA, Value *EndA,
+                              Value *StartB, Value *EndB);
   /// @}
 };
 } // anonymous namespace
@@ -190,7 +211,13 @@ bool LoopIdiomVectorize::run(Loop *L) {
   LLVM_DEBUG(dbgs() << DEBUG_TYPE " Scanning: F[" << F.getName() << "] Loop %"
                     << CurLoop->getHeader()->getName() << "\n");
 
-  return recognizeByteCompare();
+  if (recognizeByteCompare())
+    return true;
+
+  if (recognizeFindFirstByte())
+    return true;
+
+  return false;
 }
 
 bool LoopIdiomVectorize::recognizeByteCompare() {
@@ -941,3 +968,416 @@ void LoopIdiomVectorize::transformByteCompare(GetElementPtrInst *GEPA,
       report_fatal_error("Loops must remain in LCSSA form!");
   }
 }
+
+bool LoopIdiomVectorize::recognizeFindFirstByte() {
+  // Currently the transformation only works on scalable vector types, although
+  // there is no fundamental reason why it cannot be made to work for fixed
+  // width too.
+  if (!TTI->supportsScalableVectors() || DisableFindFirstByte)
+    return false;
+
+  // Define some constants we need throughout.
+  // TODO: Some of these could be made configurable parameters. For example, we
+  // could allow CharWidth = 16 (and VF = 8).
+  unsigned VF = 16;
+  unsigned CharWidth = 8;
+  BasicBlock *Header = CurLoop->getHeader();
+  LLVMContext &Ctx = Header->getContext();
+  auto *CharTy = Type::getIntNTy(Ctx, CharWidth);
+  auto *CharVTy = ScalableVectorType::get(CharTy, VF);
+
+  // Check if the target supports efficient vector matches for vectors of
+  // bytes.
+  if (!TTI->hasVectorMatch(CharVTy, VF))
+    return false;
+
+  // In LoopIdiomVectorize::run we have already checked that the loop has a
+  // preheader so we can assume it's in a canonical form.
+  if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 4)
+    return false;
+
+  // We expect this loop to have one nested loop.
+  if (CurLoop->getSubLoops().size() != 1)
+    return false;
+
+  auto *InnerLoop = CurLoop->getSubLoops().front();
+  PHINode *IndPhi = dyn_cast<PHINode>(&Header->front());
+
+  if (!IndPhi || IndPhi->getNumIncomingValues() != 2)
+    return false;
+
+  auto LoopBlocks = CurLoop->getBlocks();
+  // We are expecting the following blocks below. For now, we will bail out for
+  // anything deviating from this.
+  //
+  // .preheader:                                       ; preds = %.preheader.preheader, %23
+  //   %14 = phi ptr [ %24, %23 ], [ %3, %.preheader.preheader ]
+  //   %15 = load i8, ptr %14, align 1, !tbaa !14
+  //   br label %19
+  //
+  // 19:                                               ; preds = %16, %.preheader
+  //   %20 = phi ptr [ %7, %.preheader ], [ %17, %16 ]
+  //   %21 = load i8, ptr %20, align 1, !tbaa !14
+  //   %22 = icmp eq i8 %15, %21
+  //   br i1 %22, label %.loopexit.loopexit, label %16
+  //
+  // 16:                                               ; preds = %19
+  //   %17 = getelementptr inbounds i8, ptr %20, i64 1
+  //   %18 = icmp eq ptr %17, %10
+  //   br i1 %18, label %23, label %19, !llvm.loop !15
+  //
+  // 23:                                               ; preds = %16
+  //   %24 = getelementptr inbounds i8, ptr %14, i64 1
+  //   %25 = icmp eq ptr %24, %6
+  //   br i1 %25, label %.loopexit.loopexit5, label %.preheader, !llvm.loop !17
+  //
+  if (LoopBlocks[0]->sizeWithoutDebug() > 3 ||
+      LoopBlocks[1]->sizeWithoutDebug() > 4 ||
+      LoopBlocks[2]->sizeWithoutDebug() > 3 ||
+      LoopBlocks[3]->sizeWithoutDebug() > 3)
+    return false;
+
+  // If we match the pattern, IndPhi is going to be replaced. We cannot replace
+  // the loop if any other of its instructions are used outside of it.
+  for (BasicBlock *BB : LoopBlocks)
+    for (Instruction &I : *BB)
+      if (&I != IndPhi)
+        for (User *U : I.users())
+          if (!CurLoop->contains(cast<Instruction>(U)))
+            return false;
+
+  // Match the branch instruction for the header. We are expecting an
+  // unconditional branch to the inner loop.
+  BasicBlock *MatchBB;
+  if (!match(Header->getTerminator(), m_UnconditionalBr(MatchBB)) ||
+      !InnerLoop->contains(MatchBB))
+    return false;
+
+  // MatchBB should be the entrypoint into the inner loop containing the
+  // comparison between a search item and a valid/successful match.
+  ICmpInst::Predicate MatchPred;
+  BasicBlock *ExitSucc;
+  BasicBlock *InnerBB;
+  Value *LoadA, *LoadB;
+  if (!match(MatchBB->getTerminator(),
+             m_Br(m_ICmp(MatchPred, m_Value(LoadA), m_Value(LoadB)),
+                  m_BasicBlock(ExitSucc), m_BasicBlock(InnerBB))) ||
+      MatchPred != ICmpInst::Predicate::ICMP_EQ ||
+      !InnerLoop->contains(InnerBB))
+    return false;
+
+  // We expect a single use of IndPhi outside of CurLoop. The outside use
+  // should be a PHINode in ExitSucc coming from MatchBB.
+  // Note: Strictly speaking we are not checking for a *single* use of IndPhi
+  // outside of CurLoop here, but below we check that we only exit CurLoop to
+  // ExitSucc in one place, so by construction this should be true. Besides, in
+  // the event it is not, as long as the use is a PHINode in ExitSucc and comes
+  // from MatchBB, the transformation should still be valid in any case.
+  for (Use &U : IndPhi->uses())
+    if (CurLoop->contains(cast<Instruction>(U.getUser())))
+      continue;
+    else if (auto *PN = dyn_cast<PHINode>(U.getUser());
+             !PN || PN->getParent() != ExitSucc ||
+             PN->getIncomingBlock(U) != MatchBB)
+      return false;
+
+  // Match the loads.
+  Value *A, *B;
+  if (!match(LoadA, m_Load(m_Value(A))) || !match(LoadB, m_Load(m_Value(B))))
+    return false;
+
+  // Make sure they are simple.
+  LoadInst *LoadAI = cast<LoadInst>(LoadA);
+  LoadInst *LoadBI = cast<LoadInst>(LoadB);
+  if (!LoadAI->isSimple() || !LoadBI->isSimple())
+    return false;
+
+  // The values loaded come from two PHIs that can only have two incoming
+  // values.
+  PHINode *PNA = dyn_cast<PHINode>(A);
+  PHINode *PNB = dyn_cast<PHINode>(B);
+  if (!PNA || PNA->getNumIncomingValues() != 2 ||
+      !PNB || PNB->getNumIncomingValues() != 2)
+    return false;
+
+  // One PHI comes from the outer loop, the other one from the inner loop.
+  // CurLoop contains PNA, InnerLoop PNB.
+  if (InnerLoop->contains(PNA))
+    std::swap(PNA, PNB);
+  if (PNA != &Header->front() || PNB != &MatchBB->front())
+    return false;
+
+  // The incoming values of both PHI nodes should be a gep of 1.
+  Value *StartA = PNA->getIncomingValue(0);
+  Value *IndexA = PNA->getIncomingValue(1);
+  if (CurLoop->contains(PNA->getIncomingBlock(0)))
+    std::swap(StartA, IndexA);...
[truncated]

``````````

</details>


https://github.com/llvm/llvm-project/pull/101976


More information about the llvm-commits mailing list