[CostModel][AArch64] Make extractelement, with fmul user, free whenever possible

sushgokh · sushgokh · commit 242576af8fed · 2024-11-06T16:59:30.000+05:30
In case of Neon, if there exists extractelement from lane != 0 such that 1. extractelement does not necessitate a move from vector_reg -> GPR. 2. extractelement result feeds into fmul. 3. Other operand of fmul is a scalar or extractelement from lane 0 or lane equivalent to 0. then the extractelement can be merged with fmul in the backend and it incurs no cost. e.g. define double @foo(<2 x double> %a) { %1 = extractelement <2 x double> %a, i32 0 %2 = extractelement <2 x double> %a, i32 1 %res = fmul double %1, %2 ret double %res } %2 and %res can be merged in the backend to generate: fmul d0, d0, v0.d[1] The change was tested with SPEC FP(C/C++) on Neoverse-v2. Compile time impact: None Performance impact: Observing 1.3-1.7% uplift on lbm benchmark with -flto depending upon the config.
diff --git a/llvm/include/llvm/Analysis/TargetTransformInfo.h b/llvm/include/llvm/Analysis/TargetTransformInfo.h
@@ -22,6 +22,7 @@
 #define LLVM_ANALYSIS_TARGETTRANSFORMINFO_H
 
 #include "llvm/ADT/APInt.h"
+#include "llvm/ADT/ArrayRef.h"
 #include "llvm/IR/FMF.h"
 #include "llvm/IR/InstrTypes.h"
 #include "llvm/IR/PassManager.h"
@@ -1404,6 +1405,20 @@ class TargetTransformInfo {
                                      unsigned Index = -1, Value *Op0 = nullptr,
                                      Value *Op1 = nullptr) const;
 
+  /// \return The expected cost of vector Insert and Extract.
+  /// Use -1 to indicate that there is no information on the index value.
+  /// This is used when the instruction is not available; a typical use
+  /// case is to provision the cost of vectorization/scalarization in
+  /// vectorizer passes.
+  /// \param ScalarUserAndIdx encodes the information about extracts from a
+  /// vector with 'Scalar' being the value being extracted,'User' being the user
+  /// of the extract(nullptr if user is not known before vectorization) and
+  /// 'Idx' being the extract lane.
+  InstructionCost getVectorInstrCost(
+      unsigned Opcode, Type *Val, TTI::TargetCostKind CostKind, unsigned Index,
+      Value *Scalar,
+      ArrayRef<std::tuple<Value *, User *, int>> ScalarUserAndIdx) const;
+
   /// \return The expected cost of vector Insert and Extract.
   /// This is used when instruction is available, and implementation
   /// asserts 'I' is not nullptr.
@@ -2100,6 +2115,16 @@ class TargetTransformInfo::Concept {
                                              TTI::TargetCostKind CostKind,
                                              unsigned Index, Value *Op0,
                                              Value *Op1) = 0;
+
+  /// \param ScalarUserAndIdx encodes the information about extracts from a
+  /// vector with 'Scalar' being the value being extracted,'User' being the user
+  /// of the extract(nullptr if user is not known before vectorization) and
+  /// 'Idx' being the extract lane.
+  virtual InstructionCost getVectorInstrCost(
+      unsigned Opcode, Type *Val, TTI::TargetCostKind CostKind, unsigned Index,
+      Value *Scalar,
+      ArrayRef<std::tuple<Value *, User *, int>> ScalarUserAndIdx) = 0;
+
   virtual InstructionCost getVectorInstrCost(const Instruction &I, Type *Val,
                                              TTI::TargetCostKind CostKind,
                                              unsigned Index) = 0;
@@ -2785,6 +2810,13 @@ class TargetTransformInfo::Model final : public TargetTransformInfo::Concept {
                                      Value *Op1) override {
     return Impl.getVectorInstrCost(Opcode, Val, CostKind, Index, Op0, Op1);
   }
+  InstructionCost getVectorInstrCost(
+      unsigned Opcode, Type *Val, TTI::TargetCostKind CostKind, unsigned Index,
+      Value *Scalar,
+      ArrayRef<std::tuple<Value *, User *, int>> ScalarUserAndIdx) override {
+    return Impl.getVectorInstrCost(Opcode, Val, CostKind, Index, Scalar,
+                                   ScalarUserAndIdx);
+  }
   InstructionCost getVectorInstrCost(const Instruction &I, Type *Val,
                                      TTI::TargetCostKind CostKind,
                                      unsigned Index) override {
diff --git a/llvm/include/llvm/Analysis/TargetTransformInfoImpl.h b/llvm/include/llvm/Analysis/TargetTransformInfoImpl.h
@@ -699,6 +699,17 @@ class TargetTransformInfoImplBase {
     return 1;
   }
 
+  /// \param ScalarUserAndIdx encodes the information about extracts from a
+  /// vector with 'Scalar' being the value being extracted,'User' being the user
+  /// of the extract(nullptr if user is not known before vectorization) and
+  /// 'Idx' being the extract lane.
+  InstructionCost getVectorInstrCost(
+      unsigned Opcode, Type *Val, TTI::TargetCostKind CostKind, unsigned Index,
+      Value *Scalar,
+      ArrayRef<std::tuple<Value *, User *, int>> ScalarUserAndIdx) const {
+    return 1;
+  }
+
   InstructionCost getVectorInstrCost(const Instruction &I, Type *Val,
                                      TTI::TargetCostKind CostKind,
                                      unsigned Index) const {
diff --git a/llvm/include/llvm/CodeGen/BasicTTIImpl.h b/llvm/include/llvm/CodeGen/BasicTTIImpl.h
@@ -17,7 +17,6 @@
 #define LLVM_CODEGEN_BASICTTIIMPL_H
 
 #include "llvm/ADT/APInt.h"
-#include "llvm/ADT/ArrayRef.h"
 #include "llvm/ADT/BitVector.h"
 #include "llvm/ADT/SmallPtrSet.h"
 #include "llvm/ADT/SmallVector.h"
@@ -1282,12 +1281,24 @@ class BasicTTIImplBase : public TargetTransformInfoImplCRTPBase<T> {
     return 1;
   }
 
-  InstructionCost getVectorInstrCost(unsigned Opcode, Type *Val,
-                                     TTI::TargetCostKind CostKind,
-                                     unsigned Index, Value *Op0, Value *Op1) {
+  virtual InstructionCost getVectorInstrCost(unsigned Opcode, Type *Val,
+                                             TTI::TargetCostKind CostKind,
+                                             unsigned Index, Value *Op0,
+                                             Value *Op1) {
     return getRegUsageForType(Val->getScalarType());
   }
 
+  /// \param ScalarUserAndIdx encodes the information about extracts from a
+  /// vector with 'Scalar' being the value being extracted,'User' being the user
+  /// of the extract(nullptr if user is not known before vectorization) and
+  /// 'Idx' being the extract lane.
+  InstructionCost getVectorInstrCost(
+      unsigned Opcode, Type *Val, TTI::TargetCostKind CostKind, unsigned Index,
+      Value *Scalar,
+      ArrayRef<std::tuple<Value *, User *, int>> ScalarUserAndIdx) {
+    return getVectorInstrCost(Opcode, Val, CostKind, Index, nullptr, nullptr);
+  }
+
   InstructionCost getVectorInstrCost(const Instruction &I, Type *Val,
                                      TTI::TargetCostKind CostKind,
                                      unsigned Index) {
diff --git a/llvm/lib/Analysis/TargetTransformInfo.cpp b/llvm/lib/Analysis/TargetTransformInfo.cpp
@@ -1056,6 +1056,19 @@ InstructionCost TargetTransformInfo::getVectorInstrCost(
   return Cost;
 }
 
+InstructionCost TargetTransformInfo::getVectorInstrCost(
+    unsigned Opcode, Type *Val, TTI::TargetCostKind CostKind, unsigned Index,
+    Value *Scalar,
+    ArrayRef<std::tuple<Value *, User *, int>> ScalarUserAndIdx) const {
+  // FIXME: Assert that Opcode is either InsertElement or ExtractElement.
+  // This is mentioned in the interface description and respected by all
+  // callers, but never asserted upon.
+  InstructionCost Cost = TTIImpl->getVectorInstrCost(
+      Opcode, Val, CostKind, Index, Scalar, ScalarUserAndIdx);
+  assert(Cost >= 0 && "TTI should not produce negative costs!");
+  return Cost;
+}
+
 InstructionCost
 TargetTransformInfo::getVectorInstrCost(const Instruction &I, Type *Val,
                                         TTI::TargetCostKind CostKind,
diff --git a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
@@ -11,6 +11,7 @@
 #include "AArch64PerfectShuffle.h"
 #include "MCTargetDesc/AArch64AddressingModes.h"
 #include "Utils/AArch64SMEAttributes.h"
+#include "llvm/ADT/DenseMap.h"
 #include "llvm/Analysis/IVDescriptors.h"
 #include "llvm/Analysis/LoopInfo.h"
 #include "llvm/Analysis/TargetTransformInfo.h"
@@ -3177,10 +3178,10 @@ InstructionCost AArch64TTIImpl::getCFInstrCost(unsigned Opcode,
   return 0;
 }
 
-InstructionCost AArch64TTIImpl::getVectorInstrCostHelper(const Instruction *I,
-                                                         Type *Val,
-                                                         unsigned Index,
-                                                         bool HasRealUse) {
+InstructionCost AArch64TTIImpl::getVectorInstrCostHelper(
+    Type *Val, unsigned Index, bool HasRealUse, const Instruction *I,
+    std::optional<unsigned> Opcode, Value *Scalar,
+    ArrayRef<std::tuple<Value *, User *, int>> ScalarUserAndIdx) {
   assert(Val->isVectorTy() && "This must be a vector type");
 
   if (Index != -1U) {
@@ -3226,6 +3227,128 @@ InstructionCost AArch64TTIImpl::getVectorInstrCostHelper(const Instruction *I,
     // compile-time considerations.
   }
 
+  // In case of Neon, if there exists extractelement from lane != 0 such that
+  // 1. extractelement does not necessitate a move from vector_reg -> GPR.
+  // 2. extractelement result feeds into fmul.
+  // 3. Other operand of fmul is an extractelement from lane 0 or lane
+  // equivalent to 0.
+  // then the extractelement can be merged with fmul in the backend and it
+  // incurs no cost.
+  // e.g.
+  // define double @foo(<2 x double> %a) {
+  //   %1 = extractelement <2 x double> %a, i32 0
+  //   %2 = extractelement <2 x double> %a, i32 1
+  //   %res = fmul double %1, %2
+  //   ret double %res
+  // }
+  // %2 and %res can be merged in the backend to generate fmul d0, d0, v1.d[1]
+  auto ExtractCanFuseWithFmul = [&]() {
+    // We bail out if the extract is from lane 0.
+    if (Index == 0)
+      return false;
+
+    // Check if the scalar element type of the vector operand of ExtractElement
+    // instruction is one of the allowed types.
+    auto IsAllowedScalarTy = [&](const Type *T) {
+      return T->isFloatTy() || T->isDoubleTy() ||
+             (T->isHalfTy() && ST->hasFullFP16());
+    };
+
+    // Check if the extractelement user is scalar fmul.
+    auto IsUserFMulScalarTy = [](const Value *EEUser) {
+      // Check if the user is scalar fmul.
+      const auto *BO = dyn_cast_if_present<BinaryOperator>(EEUser);
+      return BO && BO->getOpcode() == BinaryOperator::FMul &&
+             !BO->getType()->isVectorTy();
+    };
+
+    // Check if the type constraints on input vector type and result scalar type
+    // of extractelement instruction are satisfied.
+    auto TypeConstraintsOnEESatisfied =
+        [&IsAllowedScalarTy](const Type *VectorTy, const Type *ScalarTy) {
+          return isa<FixedVectorType>(VectorTy) && IsAllowedScalarTy(ScalarTy);
+        };
+
+    // Check if the extract index is from lane 0 or lane equivalent to 0 for a
+    // certain scalar type and a certain vector register width.
+    auto IsExtractLaneEquivalentToZero = [&](const unsigned &Idx,
+                                             const unsigned &EltSz) {
+      auto RegWidth =
+          getRegisterBitWidth(TargetTransformInfo::RGK_FixedWidthVector)
+              .getFixedValue();
+      return (Idx == 0 || (Idx * EltSz) % RegWidth == 0);
+    };
+
+    if (Opcode.has_value()) {
+      if (!TypeConstraintsOnEESatisfied(Val, Val->getScalarType()))
+        return false;
+
+      DenseMap<User *, unsigned> UserToExtractIdx;
+      for (auto *U : Scalar->users()) {
+        if (!IsUserFMulScalarTy(U))
+          return false;
+        // Recording entry for the user is important. Index value is not
+        // important.
+        UserToExtractIdx[U];
+      }
+      for (auto &[S, U, L] : ScalarUserAndIdx) {
+        for (auto *U : S->users()) {
+          if (UserToExtractIdx.find(U) != UserToExtractIdx.end()) {
+            auto *FMul = cast<BinaryOperator>(U);
+            auto *Op0 = FMul->getOperand(0);
+            auto *Op1 = FMul->getOperand(1);
+            if ((Op0 == S && Op1 == S) || (Op0 != S) || (Op1 != S)) {
+              UserToExtractIdx[U] = L;
+              break;
+            }
+          }
+        }
+      }
+      for (auto &[U, L] : UserToExtractIdx) {
+        if (!IsExtractLaneEquivalentToZero(Index, Val->getScalarSizeInBits()) &&
+            !IsExtractLaneEquivalentToZero(L, Val->getScalarSizeInBits()))
+          return false;
+      }
+    } else {
+      const auto *EE = cast<ExtractElementInst>(I);
+
+      const auto *IdxOp = dyn_cast<ConstantInt>(EE->getIndexOperand());
+      if (!IdxOp)
+        return false;
+
+      if (!TypeConstraintsOnEESatisfied(EE->getVectorOperand()->getType(),
+                                        EE->getType()))
+        return false;
+
+      return !EE->users().empty() && all_of(EE->users(), [&](const User *U) {
+        if (!IsUserFMulScalarTy(U))
+          return false;
+
+        // Check if the other operand of extractelement is also extractelement
+        // from lane equivalent to 0.
+        const auto *BO = cast<BinaryOperator>(U);
+        const auto *OtherEE = dyn_cast<ExtractElementInst>(
+            BO->getOperand(0) == EE ? BO->getOperand(1) : BO->getOperand(0));
+        if (OtherEE) {
+          const auto *IdxOp = dyn_cast<ConstantInt>(OtherEE->getIndexOperand());
+          if (!IdxOp)
+            return false;
+          return IsExtractLaneEquivalentToZero(
+              cast<ConstantInt>(OtherEE->getIndexOperand())
+                  ->getValue()
+                  .getZExtValue(),
+              OtherEE->getType()->getScalarSizeInBits());
+        }
+        return true;
+      });
+    }
+    return true;
+  };
+
+  unsigned InstOpcode = I ? I->getOpcode() : Opcode.value();
+  if (InstOpcode == Instruction::ExtractElement && ExtractCanFuseWithFmul())
+    return 0;
+
   // All other insert/extracts cost this much.
   return ST->getVectorInsertExtractBaseCost();
 }
@@ -3236,14 +3359,22 @@ InstructionCost AArch64TTIImpl::getVectorInstrCost(unsigned Opcode, Type *Val,
                                                    Value *Op1) {
   bool HasRealUse =
       Opcode == Instruction::InsertElement && Op0 && !isa<UndefValue>(Op0);
-  return getVectorInstrCostHelper(nullptr, Val, Index, HasRealUse);
+  return getVectorInstrCostHelper(Val, Index, HasRealUse);
+}
+
+InstructionCost AArch64TTIImpl::getVectorInstrCost(
+    unsigned Opcode, Type *Val, TTI::TargetCostKind CostKind, unsigned Index,
+    Value *Scalar,
+    ArrayRef<std::tuple<Value *, User *, int>> ScalarUserAndIdx) {
+  return getVectorInstrCostHelper(Val, Index, false, nullptr, Opcode, Scalar,
+                                  ScalarUserAndIdx);
 }
 
 InstructionCost AArch64TTIImpl::getVectorInstrCost(const Instruction &I,
                                                    Type *Val,
                                                    TTI::TargetCostKind CostKind,
                                                    unsigned Index) {
-  return getVectorInstrCostHelper(&I, Val, Index, true /* HasRealUse */);
+  return getVectorInstrCostHelper(Val, Index, true /* HasRealUse */, &I);
 }
 
 InstructionCost AArch64TTIImpl::getScalarizationOverhead(
diff --git a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.h b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.h
@@ -19,7 +19,6 @@
 #include "AArch64.h"
 #include "AArch64Subtarget.h"
 #include "AArch64TargetMachine.h"
-#include "llvm/ADT/ArrayRef.h"
 #include "llvm/Analysis/TargetTransformInfo.h"
 #include "llvm/CodeGen/BasicTTIImpl.h"
 #include "llvm/IR/Function.h"
@@ -66,8 +65,15 @@ class AArch64TTIImpl : public BasicTTIImplBase<AArch64TTIImpl> {
   // 'Val' and 'Index' are forwarded from 'getVectorInstrCost'; 'HasRealUse'
   // indicates whether the vector instruction is available in the input IR or
   // just imaginary in vectorizer passes.
-  InstructionCost getVectorInstrCostHelper(const Instruction *I, Type *Val,
-                                           unsigned Index, bool HasRealUse);
+  /// \param ScalarUserAndIdx encodes the information about extracts from a
+  /// vector with 'Scalar' being the value being extracted,'User' being the user
+  /// of the extract(nullptr if user is not known before vectorization) and
+  /// 'Idx' being the extract lane.
+  InstructionCost getVectorInstrCostHelper(
+      Type *Val, unsigned Index, bool HasRealUse,
+      const Instruction *I = nullptr,
+      std::optional<unsigned> Opcode = std::nullopt, Value *Scalar = nullptr,
+      ArrayRef<std::tuple<Value *, User *, int>> ScalarUserAndIdx = {});
 
 public:
   explicit AArch64TTIImpl(const AArch64TargetMachine *TM, const Function &F)
@@ -185,6 +191,16 @@ class AArch64TTIImpl : public BasicTTIImplBase<AArch64TTIImpl> {
   InstructionCost getVectorInstrCost(unsigned Opcode, Type *Val,
                                      TTI::TargetCostKind CostKind,
                                      unsigned Index, Value *Op0, Value *Op1);
+
+  /// \param ScalarUserAndIdx encodes the information about extracts from a
+  /// vector with 'Scalar' being the value being extracted,'User' being the user
+  /// of the extract(nullptr if user is not known before vectorization) and
+  /// 'Idx' being the extract lane.
+  InstructionCost getVectorInstrCost(
+      unsigned Opcode, Type *Val, TTI::TargetCostKind CostKind, unsigned Index,
+      Value *Scalar,
+      ArrayRef<std::tuple<Value *, User *, int>> ScalarUserAndIdx);
+
   InstructionCost getVectorInstrCost(const Instruction &I, Type *Val,
                                      TTI::TargetCostKind CostKind,
                                      unsigned Index);
diff --git a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
@@ -11977,6 +11977,13 @@ InstructionCost BoUpSLP::getTreeCost(ArrayRef<Value *> VectorizedVals) {
   std::optional<DenseMap<Value *, unsigned>> ValueToExtUses;
   DenseMap<const TreeEntry *, DenseSet<Value *>> ExtractsCount;
   SmallPtrSet<Value *, 4> ScalarOpsFromCasts;
+  // Keep track {Scalar, Index, User} tuple.
+  // On AArch64, this helps in fusing a mov instruction, associated with
+  // extractelement, with fmul in the backend so that extractelement is free.
+  SmallVector<std::tuple<Value *, User *, int>, 4> ScalarUserAndIdx;
+  for (ExternalUser &EU : ExternalUses) {
+    ScalarUserAndIdx.emplace_back(EU.Scalar, EU.User, EU.Lane);
+  }
   for (ExternalUser &EU : ExternalUses) {
     // Uses by ephemeral values are free (because the ephemeral value will be
     // removed prior to code generation, and so the extraction will be
@@ -12089,8 +12096,9 @@ InstructionCost BoUpSLP::getTreeCost(ArrayRef<Value *> VectorizedVals) {
       ExtraCost = TTI->getExtractWithExtendCost(Extend, EU.Scalar->getType(),
                                                 VecTy, EU.Lane);
     } else {
-      ExtraCost = TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy,
-                                          CostKind, EU.Lane);
+      ExtraCost =
+          TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, CostKind,
+                                  EU.Lane, EU.Scalar, ScalarUserAndIdx);
     }
     // Leave the scalar instructions as is if they are cheaper than extracts.
     if (Entry->Idx != 0 || Entry->getOpcode() == Instruction::GetElementPtr ||
diff --git a/llvm/test/Analysis/CostModel/AArch64/extract_float.ll b/llvm/test/Analysis/CostModel/AArch64/extract_float.ll
diff --git a/llvm/test/Transforms/SLPVectorizer/consecutive-access.ll b/llvm/test/Transforms/SLPVectorizer/consecutive-access.ll
