Google Git
Sign in
llvm / llvm-project / 9e3e1aad3161f4ce5301c3a59c7313ad83240a6d / . / llvm / lib / Transforms / InstCombine / InstructionCombining.cpp
blob: 1f81624f79e782cdb38effc25fb1c855eb6e1674 [file] [log] [blame]
//===- InstructionCombining.cpp - Combine multiple instructions -----------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// InstructionCombining - Combine instructions to form fewer, simple
// instructions. This pass does not modify the CFG. This pass is where
// algebraic simplification happens.
//
// This pass combines things like:
// %Y = add i32 %X, 1
// %Z = add i32 %Y, 1
// into:
// %Z = add i32 %X, 2
//
// This is a simple worklist driven algorithm.
//
// This pass guarantees that the following canonicalizations are performed on
// the program:
// 1. If a binary operator has a constant operand, it is moved to the RHS
// 2. Bitwise operators with constant operands are always grouped so that
// shifts are performed first, then or's, then and's, then xor's.
// 3. Compare instructions are converted from <,>,<=,>= to ==,!= if possible
// 4. All cmp instructions on boolean values are replaced with logical ops
// 5. add X, X is represented as (X*2) => (X << 1)
// 6. Multiplies with a power-of-two constant argument are transformed into
// shifts.
// ... etc.
//
//===----------------------------------------------------------------------===//
#include "InstCombineInternal.h"
#include "llvm-c/Initialization.h"
#include "llvm-c/Transforms/InstCombine.h"
#include "llvm/ADT/APInt.h"
#include "llvm/ADT/ArrayRef.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/None.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/TinyPtrVector.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Analysis/AssumptionCache.h"
#include "llvm/Analysis/BasicAliasAnalysis.h"
#include "llvm/Analysis/BlockFrequencyInfo.h"
#include "llvm/Analysis/CFG.h"
#include "llvm/Analysis/ConstantFolding.h"
#include "llvm/Analysis/EHPersonalities.h"
#include "llvm/Analysis/GlobalsModRef.h"
#include "llvm/Analysis/InstructionSimplify.h"
#include "llvm/Analysis/LazyBlockFrequencyInfo.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Analysis/MemoryBuiltins.h"
#include "llvm/Analysis/OptimizationRemarkEmitter.h"
#include "llvm/Analysis/ProfileSummaryInfo.h"
#include "llvm/Analysis/TargetFolder.h"
#include "llvm/Analysis/TargetLibraryInfo.h"
#include "llvm/Analysis/TargetTransformInfo.h"
#include "llvm/Analysis/ValueTracking.h"
#include "llvm/Analysis/VectorUtils.h"
#include "llvm/IR/BasicBlock.h"
#include "llvm/IR/CFG.h"
#include "llvm/IR/Constant.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/DIBuilder.h"
#include "llvm/IR/DataLayout.h"
#include "llvm/IR/DerivedTypes.h"
#include "llvm/IR/Dominators.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/GetElementPtrTypeIterator.h"
#include "llvm/IR/IRBuilder.h"
#include "llvm/IR/InstrTypes.h"
#include "llvm/IR/Instruction.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/Intrinsics.h"
#include "llvm/IR/LegacyPassManager.h"
#include "llvm/IR/Metadata.h"
#include "llvm/IR/Operator.h"
#include "llvm/IR/PassManager.h"
#include "llvm/IR/PatternMatch.h"
#include "llvm/IR/Type.h"
#include "llvm/IR/Use.h"
#include "llvm/IR/User.h"
#include "llvm/IR/Value.h"
#include "llvm/IR/ValueHandle.h"
#include "llvm/InitializePasses.h"
#include "llvm/Pass.h"
#include "llvm/Support/CBindingWrapping.h"
#include "llvm/Support/Casting.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/DebugCounter.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/KnownBits.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Transforms/InstCombine/InstCombine.h"
#include "llvm/Transforms/Utils/Local.h"
#include <algorithm>
#include <cassert>
#include <cstdint>
#include <memory>
#include <string>
#include <utility>
#define DEBUG_TYPE "instcombine"
#include "llvm/Transforms/Utils/InstructionWorklist.h"
using namespace llvm;
using namespace llvm::PatternMatch;
STATISTIC(NumWorklistIterations,
"Number of instruction combining iterations performed");
STATISTIC(NumCombined , "Number of insts combined");
STATISTIC(NumConstProp, "Number of constant folds");
STATISTIC(NumDeadInst , "Number of dead inst eliminated");
STATISTIC(NumSunkInst , "Number of instructions sunk");
STATISTIC(NumExpand, "Number of expansions");
STATISTIC(NumFactor , "Number of factorizations");
STATISTIC(NumReassoc , "Number of reassociations");
DEBUG_COUNTER(VisitCounter, "instcombine-visit",
"Controls which instructions are visited");
// FIXME: these limits eventually should be as low as 2.
static constexpr unsigned InstCombineDefaultMaxIterations = 1000;
#ifndef NDEBUG
static constexpr unsigned InstCombineDefaultInfiniteLoopThreshold = 100;
#else
static constexpr unsigned InstCombineDefaultInfiniteLoopThreshold = 1000;
#endif
static cl::opt<bool>
EnableCodeSinking("instcombine-code-sinking", cl::desc("Enable code sinking"),
cl::init(true));
static cl::opt<unsigned> LimitMaxIterations(
"instcombine-max-iterations",
cl::desc("Limit the maximum number of instruction combining iterations"),
cl::init(InstCombineDefaultMaxIterations));
static cl::opt<unsigned> InfiniteLoopDetectionThreshold(
"instcombine-infinite-loop-threshold",
cl::desc("Number of instruction combining iterations considered an "
"infinite loop"),
cl::init(InstCombineDefaultInfiniteLoopThreshold), cl::Hidden);
static cl::opt<unsigned>
MaxArraySize("instcombine-maxarray-size", cl::init(1024),
cl::desc("Maximum array size considered when doing a combine"));
// FIXME: Remove this flag when it is no longer necessary to convert
// llvm.dbg.declare to avoid inaccurate debug info. Setting this to false
// increases variable availability at the cost of accuracy. Variables that
// cannot be promoted by mem2reg or SROA will be described as living in memory
// for their entire lifetime. However, passes like DSE and instcombine can
// delete stores to the alloca, leading to misleading and inaccurate debug
// information. This flag can be removed when those passes are fixed.
static cl::opt<unsigned> ShouldLowerDbgDeclare("instcombine-lower-dbg-declare",
cl::Hidden, cl::init(true));
Optional<Instruction *>
InstCombiner::targetInstCombineIntrinsic(IntrinsicInst &II) {
// Handle target specific intrinsics
if (II.getCalledFunction()->isTargetIntrinsic()) {
return TTI.instCombineIntrinsic(*this, II);
}
return None;
}
Optional<Value *> InstCombiner::targetSimplifyDemandedUseBitsIntrinsic(
IntrinsicInst &II, APInt DemandedMask, KnownBits &Known,
bool &KnownBitsComputed) {
// Handle target specific intrinsics
if (II.getCalledFunction()->isTargetIntrinsic()) {
return TTI.simplifyDemandedUseBitsIntrinsic(*this, II, DemandedMask, Known,
KnownBitsComputed);
}
return None;
}
Optional<Value *> InstCombiner::targetSimplifyDemandedVectorEltsIntrinsic(
IntrinsicInst &II, APInt DemandedElts, APInt &UndefElts, APInt &UndefElts2,
APInt &UndefElts3,
std::function<void(Instruction *, unsigned, APInt, APInt &)>
SimplifyAndSetOp) {
// Handle target specific intrinsics
if (II.getCalledFunction()->isTargetIntrinsic()) {
return TTI.simplifyDemandedVectorEltsIntrinsic(
*this, II, DemandedElts, UndefElts, UndefElts2, UndefElts3,
SimplifyAndSetOp);
}
return None;
}
Value *InstCombinerImpl::EmitGEPOffset(User *GEP) {
return llvm::EmitGEPOffset(&Builder, DL, GEP);
}
/// Legal integers and common types are considered desirable. This is used to
/// avoid creating instructions with types that may not be supported well by the
/// the backend.
/// NOTE: This treats i8, i16 and i32 specially because they are common
/// types in frontend languages.
bool InstCombinerImpl::isDesirableIntType(unsigned BitWidth) const {
switch (BitWidth) {
case 8:
case 16:
case 32:
return true;
default:
return DL.isLegalInteger(BitWidth);
}
}
/// Return true if it is desirable to convert an integer computation from a
/// given bit width to a new bit width.
/// We don't want to convert from a legal to an illegal type or from a smaller
/// to a larger illegal type. A width of '1' is always treated as a desirable
/// type because i1 is a fundamental type in IR, and there are many specialized
/// optimizations for i1 types. Common/desirable widths are equally treated as
/// legal to convert to, in order to open up more combining opportunities.
bool InstCombinerImpl::shouldChangeType(unsigned FromWidth,
unsigned ToWidth) const {
bool FromLegal = FromWidth == 1 || DL.isLegalInteger(FromWidth);
bool ToLegal = ToWidth == 1 || DL.isLegalInteger(ToWidth);
// Convert to desirable widths even if they are not legal types.
// Only shrink types, to prevent infinite loops.
if (ToWidth < FromWidth && isDesirableIntType(ToWidth))
return true;
// If this is a legal integer from type, and the result would be an illegal
// type, don't do the transformation.
if (FromLegal && !ToLegal)
return false;
// Otherwise, if both are illegal, do not increase the size of the result. We
// do allow things like i160 -> i64, but not i64 -> i160.
if (!FromLegal && !ToLegal && ToWidth > FromWidth)
return false;
return true;
}
/// Return true if it is desirable to convert a computation from 'From' to 'To'.
/// We don't want to convert from a legal to an illegal type or from a smaller
/// to a larger illegal type. i1 is always treated as a legal type because it is
/// a fundamental type in IR, and there are many specialized optimizations for
/// i1 types.
bool InstCombinerImpl::shouldChangeType(Type *From, Type *To) const {
// TODO: This could be extended to allow vectors. Datalayout changes might be
// needed to properly support that.
if (!From->isIntegerTy() || !To->isIntegerTy())
return false;
unsigned FromWidth = From->getPrimitiveSizeInBits();
unsigned ToWidth = To->getPrimitiveSizeInBits();
return shouldChangeType(FromWidth, ToWidth);
}
// Return true, if No Signed Wrap should be maintained for I.
// The No Signed Wrap flag can be kept if the operation "B (I.getOpcode) C",
// where both B and C should be ConstantInts, results in a constant that does
// not overflow. This function only handles the Add and Sub opcodes. For
// all other opcodes, the function conservatively returns false.
static bool maintainNoSignedWrap(BinaryOperator &I, Value *B, Value *C) {
auto *OBO = dyn_cast<OverflowingBinaryOperator>(&I);
if (!OBO || !OBO->hasNoSignedWrap())
return false;
// We reason about Add and Sub Only.
Instruction::BinaryOps Opcode = I.getOpcode();
if (Opcode != Instruction::Add && Opcode != Instruction::Sub)
return false;
const APInt *BVal, *CVal;
if (!match(B, m_APInt(BVal)) || !match(C, m_APInt(CVal)))
return false;
bool Overflow = false;
if (Opcode == Instruction::Add)
(void)BVal->sadd_ov(*CVal, Overflow);
else
(void)BVal->ssub_ov(*CVal, Overflow);
return !Overflow;
}
static bool hasNoUnsignedWrap(BinaryOperator &I) {
auto *OBO = dyn_cast<OverflowingBinaryOperator>(&I);
return OBO && OBO->hasNoUnsignedWrap();
}
static bool hasNoSignedWrap(BinaryOperator &I) {
auto *OBO = dyn_cast<OverflowingBinaryOperator>(&I);
return OBO && OBO->hasNoSignedWrap();
}
/// Conservatively clears subclassOptionalData after a reassociation or
/// commutation. We preserve fast-math flags when applicable as they can be
/// preserved.
static void ClearSubclassDataAfterReassociation(BinaryOperator &I) {
FPMathOperator *FPMO = dyn_cast<FPMathOperator>(&I);
if (!FPMO) {
I.clearSubclassOptionalData();
return;
}
FastMathFlags FMF = I.getFastMathFlags();
I.clearSubclassOptionalData();
I.setFastMathFlags(FMF);
}
/// Combine constant operands of associative operations either before or after a
/// cast to eliminate one of the associative operations:
/// (op (cast (op X, C2)), C1) --> (cast (op X, op (C1, C2)))
/// (op (cast (op X, C2)), C1) --> (op (cast X), op (C1, C2))
static bool simplifyAssocCastAssoc(BinaryOperator *BinOp1,
InstCombinerImpl &IC) {
auto *Cast = dyn_cast<CastInst>(BinOp1->getOperand(0));
if (!Cast || !Cast->hasOneUse())
return false;
// TODO: Enhance logic for other casts and remove this check.
auto CastOpcode = Cast->getOpcode();
if (CastOpcode != Instruction::ZExt)
return false;
// TODO: Enhance logic for other BinOps and remove this check.
if (!BinOp1->isBitwiseLogicOp())
return false;
auto AssocOpcode = BinOp1->getOpcode();
auto *BinOp2 = dyn_cast<BinaryOperator>(Cast->getOperand(0));
if (!BinOp2 || !BinOp2->hasOneUse() || BinOp2->getOpcode() != AssocOpcode)
return false;
Constant *C1, *C2;
if (!match(BinOp1->getOperand(1), m_Constant(C1)) ||
!match(BinOp2->getOperand(1), m_Constant(C2)))
return false;
// TODO: This assumes a zext cast.
// Eg, if it was a trunc, we'd cast C1 to the source type because casting C2
// to the destination type might lose bits.
// Fold the constants together in the destination type:
// (op (cast (op X, C2)), C1) --> (op (cast X), FoldedC)
Type *DestTy = C1->getType();
Constant *CastC2 = ConstantExpr::getCast(CastOpcode, C2, DestTy);
Constant *FoldedC = ConstantExpr::get(AssocOpcode, C1, CastC2);
IC.replaceOperand(*Cast, 0, BinOp2->getOperand(0));
IC.replaceOperand(*BinOp1, 1, FoldedC);
return true;
}
// Simplifies IntToPtr/PtrToInt RoundTrip Cast To BitCast.
// inttoptr ( ptrtoint (x) ) --> x
Value *InstCombinerImpl::simplifyIntToPtrRoundTripCast(Value *Val) {
auto *IntToPtr = dyn_cast<IntToPtrInst>(Val);
if (IntToPtr && DL.getPointerTypeSizeInBits(IntToPtr->getDestTy()) ==
DL.getTypeSizeInBits(IntToPtr->getSrcTy())) {
auto *PtrToInt = dyn_cast<PtrToIntInst>(IntToPtr->getOperand(0));
Type *CastTy = IntToPtr->getDestTy();
if (PtrToInt &&
CastTy->getPointerAddressSpace() ==
PtrToInt->getSrcTy()->getPointerAddressSpace() &&
DL.getPointerTypeSizeInBits(PtrToInt->getSrcTy()) ==
DL.getTypeSizeInBits(PtrToInt->getDestTy())) {
return CastInst::CreateBitOrPointerCast(PtrToInt->getOperand(0), CastTy,
"", PtrToInt);
}
}
return nullptr;
}
/// This performs a few simplifications for operators that are associative or
/// commutative:
///
/// Commutative operators:
///
/// 1. Order operands such that they are listed from right (least complex) to
/// left (most complex). This puts constants before unary operators before
/// binary operators.
///
/// Associative operators:
///
/// 2. Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
/// 3. Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
///
/// Associative and commutative operators:
///
/// 4. Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
/// 5. Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
/// 6. Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
/// if C1 and C2 are constants.
bool InstCombinerImpl::SimplifyAssociativeOrCommutative(BinaryOperator &I) {
Instruction::BinaryOps Opcode = I.getOpcode();
bool Changed = false;
do {
// Order operands such that they are listed from right (least complex) to
// left (most complex). This puts constants before unary operators before
// binary operators.
if (I.isCommutative() && getComplexity(I.getOperand(0)) <
getComplexity(I.getOperand(1)))
Changed = !I.swapOperands();
BinaryOperator *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
BinaryOperator *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
if (I.isAssociative()) {
// Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
if (Op0 && Op0->getOpcode() == Opcode) {
Value *A = Op0->getOperand(0);
Value *B = Op0->getOperand(1);
Value *C = I.getOperand(1);
// Does "B op C" simplify?
if (Value *V = SimplifyBinOp(Opcode, B, C, SQ.getWithInstruction(&I))) {
// It simplifies to V. Form "A op V".
replaceOperand(I, 0, A);
replaceOperand(I, 1, V);
bool IsNUW = hasNoUnsignedWrap(I) && hasNoUnsignedWrap(*Op0);
bool IsNSW = maintainNoSignedWrap(I, B, C) && hasNoSignedWrap(*Op0);
// Conservatively clear all optional flags since they may not be
// preserved by the reassociation. Reset nsw/nuw based on the above
// analysis.
ClearSubclassDataAfterReassociation(I);
// Note: this is only valid because SimplifyBinOp doesn't look at
// the operands to Op0.
if (IsNUW)
I.setHasNoUnsignedWrap(true);
if (IsNSW)
I.setHasNoSignedWrap(true);
Changed = true;
++NumReassoc;
continue;
}
}
// Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
if (Op1 && Op1->getOpcode() == Opcode) {
Value *A = I.getOperand(0);
Value *B = Op1->getOperand(0);
Value *C = Op1->getOperand(1);
// Does "A op B" simplify?
if (Value *V = SimplifyBinOp(Opcode, A, B, SQ.getWithInstruction(&I))) {
// It simplifies to V. Form "V op C".
replaceOperand(I, 0, V);
replaceOperand(I, 1, C);
// Conservatively clear the optional flags, since they may not be
// preserved by the reassociation.
ClearSubclassDataAfterReassociation(I);
Changed = true;
++NumReassoc;
continue;
}
}
}
if (I.isAssociative() && I.isCommutative()) {
if (simplifyAssocCastAssoc(&I, *this)) {
Changed = true;
++NumReassoc;
continue;
}
// Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
if (Op0 && Op0->getOpcode() == Opcode) {
Value *A = Op0->getOperand(0);
Value *B = Op0->getOperand(1);
Value *C = I.getOperand(1);
// Does "C op A" simplify?
if (Value *V = SimplifyBinOp(Opcode, C, A, SQ.getWithInstruction(&I))) {
// It simplifies to V. Form "V op B".
replaceOperand(I, 0, V);
replaceOperand(I, 1, B);
// Conservatively clear the optional flags, since they may not be
// preserved by the reassociation.
ClearSubclassDataAfterReassociation(I);
Changed = true;
++NumReassoc;
continue;
}
}
// Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
if (Op1 && Op1->getOpcode() == Opcode) {
Value *A = I.getOperand(0);
Value *B = Op1->getOperand(0);
Value *C = Op1->getOperand(1);
// Does "C op A" simplify?
if (Value *V = SimplifyBinOp(Opcode, C, A, SQ.getWithInstruction(&I))) {
// It simplifies to V. Form "B op V".
replaceOperand(I, 0, B);
replaceOperand(I, 1, V);
// Conservatively clear the optional flags, since they may not be
// preserved by the reassociation.
ClearSubclassDataAfterReassociation(I);
Changed = true;
++NumReassoc;
continue;
}
}
// Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
// if C1 and C2 are constants.
Value *A, *B;
Constant *C1, *C2;
if (Op0 && Op1 &&
Op0->getOpcode() == Opcode && Op1->getOpcode() == Opcode &&
match(Op0, m_OneUse(m_BinOp(m_Value(A), m_Constant(C1)))) &&
match(Op1, m_OneUse(m_BinOp(m_Value(B), m_Constant(C2))))) {
bool IsNUW = hasNoUnsignedWrap(I) &&
hasNoUnsignedWrap(*Op0) &&
hasNoUnsignedWrap(*Op1);
BinaryOperator *NewBO = (IsNUW && Opcode == Instruction::Add) ?
BinaryOperator::CreateNUW(Opcode, A, B) :
BinaryOperator::Create(Opcode, A, B);
if (isa<FPMathOperator>(NewBO)) {
FastMathFlags Flags = I.getFastMathFlags();
Flags &= Op0->getFastMathFlags();
Flags &= Op1->getFastMathFlags();
NewBO->setFastMathFlags(Flags);
}
InsertNewInstWith(NewBO, I);
NewBO->takeName(Op1);
replaceOperand(I, 0, NewBO);
replaceOperand(I, 1, ConstantExpr::get(Opcode, C1, C2));
// Conservatively clear the optional flags, since they may not be
// preserved by the reassociation.
ClearSubclassDataAfterReassociation(I);
if (IsNUW)
I.setHasNoUnsignedWrap(true);
Changed = true;
continue;
}
}
// No further simplifications.
return Changed;
} while (true);
}
/// Return whether "X LOp (Y ROp Z)" is always equal to
/// "(X LOp Y) ROp (X LOp Z)".
static bool leftDistributesOverRight(Instruction::BinaryOps LOp,
Instruction::BinaryOps ROp) {
// X & (Y | Z) <--> (X & Y) | (X & Z)
// X & (Y ^ Z) <--> (X & Y) ^ (X & Z)
if (LOp == Instruction::And)
return ROp == Instruction::Or || ROp == Instruction::Xor;
// X | (Y & Z) <--> (X | Y) & (X | Z)
if (LOp == Instruction::Or)
return ROp == Instruction::And;
// X * (Y + Z) <--> (X * Y) + (X * Z)
// X * (Y - Z) <--> (X * Y) - (X * Z)
if (LOp == Instruction::Mul)
return ROp == Instruction::Add || ROp == Instruction::Sub;
return false;
}
/// Return whether "(X LOp Y) ROp Z" is always equal to
/// "(X ROp Z) LOp (Y ROp Z)".
static bool rightDistributesOverLeft(Instruction::BinaryOps LOp,
Instruction::BinaryOps ROp) {
if (Instruction::isCommutative(ROp))
return leftDistributesOverRight(ROp, LOp);
// (X {&|^} Y) >> Z <--> (X >> Z) {&|^} (Y >> Z) for all shifts.
return Instruction::isBitwiseLogicOp(LOp) && Instruction::isShift(ROp);
// TODO: It would be nice to handle division, aka "(X + Y)/Z = X/Z + Y/Z",
// but this requires knowing that the addition does not overflow and other
// such subtleties.
}
/// This function returns identity value for given opcode, which can be used to
/// factor patterns like (X * 2) + X ==> (X * 2) + (X * 1) ==> X * (2 + 1).
static Value *getIdentityValue(Instruction::BinaryOps Opcode, Value *V) {
if (isa<Constant>(V))
return nullptr;
return ConstantExpr::getBinOpIdentity(Opcode, V->getType());
}
/// This function predicates factorization using distributive laws. By default,
/// it just returns the 'Op' inputs. But for special-cases like
/// 'add(shl(X, 5), ...)', this function will have TopOpcode == Instruction::Add
/// and Op = shl(X, 5). The 'shl' is treated as the more general 'mul X, 32' to
/// allow more factorization opportunities.
static Instruction::BinaryOps
getBinOpsForFactorization(Instruction::BinaryOps TopOpcode, BinaryOperator *Op,
Value *&LHS, Value *&RHS) {
assert(Op && "Expected a binary operator");
LHS = Op->getOperand(0);
RHS = Op->getOperand(1);
if (TopOpcode == Instruction::Add || TopOpcode == Instruction::Sub) {
Constant *C;
if (match(Op, m_Shl(m_Value(), m_Constant(C)))) {
// X << C --> X * (1 << C)
RHS = ConstantExpr::getShl(ConstantInt::get(Op->getType(), 1), C);
return Instruction::Mul;
}
// TODO: We can add other conversions e.g. shr => div etc.
}
return Op->getOpcode();
}
/// This tries to simplify binary operations by factorizing out common terms
/// (e. g. "(A*B)+(A*C)" -> "A*(B+C)").
Value *InstCombinerImpl::tryFactorization(BinaryOperator &I,
Instruction::BinaryOps InnerOpcode,
Value *A, Value *B, Value *C,
Value *D) {
assert(A && B && C && D && "All values must be provided");
Value *V = nullptr;
Value *SimplifiedInst = nullptr;
Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
Instruction::BinaryOps TopLevelOpcode = I.getOpcode();
// Does "X op' Y" always equal "Y op' X"?
bool InnerCommutative = Instruction::isCommutative(InnerOpcode);
// Does "X op' (Y op Z)" always equal "(X op' Y) op (X op' Z)"?
if (leftDistributesOverRight(InnerOpcode, TopLevelOpcode))
// Does the instruction have the form "(A op' B) op (A op' D)" or, in the
// commutative case, "(A op' B) op (C op' A)"?
if (A == C || (InnerCommutative && A == D)) {
if (A != C)
std::swap(C, D);
// Consider forming "A op' (B op D)".
// If "B op D" simplifies then it can be formed with no cost.
V = SimplifyBinOp(TopLevelOpcode, B, D, SQ.getWithInstruction(&I));
// If "B op D" doesn't simplify then only go on if both of the existing
// operations "A op' B" and "C op' D" will be zapped as no longer used.
if (!V && LHS->hasOneUse() && RHS->hasOneUse())
V = Builder.CreateBinOp(TopLevelOpcode, B, D, RHS->getName());
if (V) {
SimplifiedInst = Builder.CreateBinOp(InnerOpcode, A, V);
}
}
// Does "(X op Y) op' Z" always equal "(X op' Z) op (Y op' Z)"?
if (!SimplifiedInst && rightDistributesOverLeft(TopLevelOpcode, InnerOpcode))
// Does the instruction have the form "(A op' B) op (C op' B)" or, in the
// commutative case, "(A op' B) op (B op' D)"?
if (B == D || (InnerCommutative && B == C)) {
if (B != D)
std::swap(C, D);
// Consider forming "(A op C) op' B".
// If "A op C" simplifies then it can be formed with no cost.
V = SimplifyBinOp(TopLevelOpcode, A, C, SQ.getWithInstruction(&I));
// If "A op C" doesn't simplify then only go on if both of the existing
// operations "A op' B" and "C op' D" will be zapped as no longer used.
if (!V && LHS->hasOneUse() && RHS->hasOneUse())
V = Builder.CreateBinOp(TopLevelOpcode, A, C, LHS->getName());
if (V) {
SimplifiedInst = Builder.CreateBinOp(InnerOpcode, V, B);
}
}
if (SimplifiedInst) {
++NumFactor;
SimplifiedInst->takeName(&I);
// Check if we can add NSW/NUW flags to SimplifiedInst. If so, set them.
if (BinaryOperator *BO = dyn_cast<BinaryOperator>(SimplifiedInst)) {
if (isa<OverflowingBinaryOperator>(SimplifiedInst)) {
bool HasNSW = false;
bool HasNUW = false;
if (isa<OverflowingBinaryOperator>(&I)) {
HasNSW = I.hasNoSignedWrap();
HasNUW = I.hasNoUnsignedWrap();
}
if (auto *LOBO = dyn_cast<OverflowingBinaryOperator>(LHS)) {
HasNSW &= LOBO->hasNoSignedWrap();
HasNUW &= LOBO->hasNoUnsignedWrap();
}
if (auto *ROBO = dyn_cast<OverflowingBinaryOperator>(RHS)) {
HasNSW &= ROBO->hasNoSignedWrap();
HasNUW &= ROBO->hasNoUnsignedWrap();
}
if (TopLevelOpcode == Instruction::Add &&
InnerOpcode == Instruction::Mul) {
// We can propagate 'nsw' if we know that
// %Y = mul nsw i16 %X, C
// %Z = add nsw i16 %Y, %X
// =>
// %Z = mul nsw i16 %X, C+1
//
// iff C+1 isn't INT_MIN
const APInt *CInt;
if (match(V, m_APInt(CInt))) {
if (!CInt->isMinSignedValue())
BO->setHasNoSignedWrap(HasNSW);
}
// nuw can be propagated with any constant or nuw value.
BO->setHasNoUnsignedWrap(HasNUW);
}
}
}
}
return SimplifiedInst;
}
/// This tries to simplify binary operations which some other binary operation
/// distributes over either by factorizing out common terms
/// (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this results in
/// simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is a win).
/// Returns the simplified value, or null if it didn't simplify.
Value *InstCombinerImpl::SimplifyUsingDistributiveLaws(BinaryOperator &I) {
Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
Instruction::BinaryOps TopLevelOpcode = I.getOpcode();
{
// Factorization.
Value *A, *B, *C, *D;
Instruction::BinaryOps LHSOpcode, RHSOpcode;
if (Op0)
LHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op0, A, B);
if (Op1)
RHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op1, C, D);
// The instruction has the form "(A op' B) op (C op' D)". Try to factorize
// a common term.
if (Op0 && Op1 && LHSOpcode == RHSOpcode)
if (Value *V = tryFactorization(I, LHSOpcode, A, B, C, D))
return V;
// The instruction has the form "(A op' B) op (C)". Try to factorize common
// term.
if (Op0)
if (Value *Ident = getIdentityValue(LHSOpcode, RHS))
if (Value *V = tryFactorization(I, LHSOpcode, A, B, RHS, Ident))
return V;
// The instruction has the form "(B) op (C op' D)". Try to factorize common
// term.
if (Op1)
if (Value *Ident = getIdentityValue(RHSOpcode, LHS))
if (Value *V = tryFactorization(I, RHSOpcode, LHS, Ident, C, D))
return V;
}
// Expansion.
if (Op0 && rightDistributesOverLeft(Op0->getOpcode(), TopLevelOpcode)) {
// The instruction has the form "(A op' B) op C". See if expanding it out
// to "(A op C) op' (B op C)" results in simplifications.
Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op'
// Disable the use of undef because it's not safe to distribute undef.
auto SQDistributive = SQ.getWithInstruction(&I).getWithoutUndef();
Value *L = SimplifyBinOp(TopLevelOpcode, A, C, SQDistributive);
Value *R = SimplifyBinOp(TopLevelOpcode, B, C, SQDistributive);
// Do "A op C" and "B op C" both simplify?
if (L && R) {
// They do! Return "L op' R".
++NumExpand;
C = Builder.CreateBinOp(InnerOpcode, L, R);
C->takeName(&I);
return C;
}
// Does "A op C" simplify to the identity value for the inner opcode?
if (L && L == ConstantExpr::getBinOpIdentity(InnerOpcode, L->getType())) {
// They do! Return "B op C".
++NumExpand;
C = Builder.CreateBinOp(TopLevelOpcode, B, C);
C->takeName(&I);
return C;
}
// Does "B op C" simplify to the identity value for the inner opcode?
if (R && R == ConstantExpr::getBinOpIdentity(InnerOpcode, R->getType())) {
// They do! Return "A op C".
++NumExpand;
C = Builder.CreateBinOp(TopLevelOpcode, A, C);
C->takeName(&I);
return C;
}
}
if (Op1 && leftDistributesOverRight(TopLevelOpcode, Op1->getOpcode())) {
// The instruction has the form "A op (B op' C)". See if expanding it out
// to "(A op B) op' (A op C)" results in simplifications.
Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
Instruction::BinaryOps InnerOpcode = Op1->getOpcode(); // op'
// Disable the use of undef because it's not safe to distribute undef.
auto SQDistributive = SQ.getWithInstruction(&I).getWithoutUndef();
Value *L = SimplifyBinOp(TopLevelOpcode, A, B, SQDistributive);
Value *R = SimplifyBinOp(TopLevelOpcode, A, C, SQDistributive);
// Do "A op B" and "A op C" both simplify?
if (L && R) {
// They do! Return "L op' R".
++NumExpand;
A = Builder.CreateBinOp(InnerOpcode, L, R);
A->takeName(&I);
return A;
}
// Does "A op B" simplify to the identity value for the inner opcode?
if (L && L == ConstantExpr::getBinOpIdentity(InnerOpcode, L->getType())) {
// They do! Return "A op C".
++NumExpand;
A = Builder.CreateBinOp(TopLevelOpcode, A, C);
A->takeName(&I);
return A;
}
// Does "A op C" simplify to the identity value for the inner opcode?
if (R && R == ConstantExpr::getBinOpIdentity(InnerOpcode, R->getType())) {
// They do! Return "A op B".
++NumExpand;
A = Builder.CreateBinOp(TopLevelOpcode, A, B);
A->takeName(&I);
return A;
}
}
return SimplifySelectsFeedingBinaryOp(I, LHS, RHS);
}
Value *InstCombinerImpl::SimplifySelectsFeedingBinaryOp(BinaryOperator &I,
Value *LHS,
Value *RHS) {
Value *A, *B, *C, *D, *E, *F;
bool LHSIsSelect = match(LHS, m_Select(m_Value(A), m_Value(B), m_Value(C)));
bool RHSIsSelect = match(RHS, m_Select(m_Value(D), m_Value(E), m_Value(F)));
if (!LHSIsSelect && !RHSIsSelect)
return nullptr;
FastMathFlags FMF;
BuilderTy::FastMathFlagGuard Guard(Builder);
if (isa<FPMathOperator>(&I)) {
FMF = I.getFastMathFlags();
Builder.setFastMathFlags(FMF);
}
Instruction::BinaryOps Opcode = I.getOpcode();
SimplifyQuery Q = SQ.getWithInstruction(&I);
Value *Cond, *True = nullptr, *False = nullptr;
if (LHSIsSelect && RHSIsSelect && A == D) {
// (A ? B : C) op (A ? E : F) -> A ? (B op E) : (C op F)
Cond = A;
True = SimplifyBinOp(Opcode, B, E, FMF, Q);
False = SimplifyBinOp(Opcode, C, F, FMF, Q);
if (LHS->hasOneUse() && RHS->hasOneUse()) {
if (False && !True)
True = Builder.CreateBinOp(Opcode, B, E);
else if (True && !False)
False = Builder.CreateBinOp(Opcode, C, F);
}
} else if (LHSIsSelect && LHS->hasOneUse()) {
// (A ? B : C) op Y -> A ? (B op Y) : (C op Y)
Cond = A;
True = SimplifyBinOp(Opcode, B, RHS, FMF, Q);
False = SimplifyBinOp(Opcode, C, RHS, FMF, Q);
} else if (RHSIsSelect && RHS->hasOneUse()) {
// X op (D ? E : F) -> D ? (X op E) : (X op F)
Cond = D;
True = SimplifyBinOp(Opcode, LHS, E, FMF, Q);
False = SimplifyBinOp(Opcode, LHS, F, FMF, Q);
}
if (!True || !False)
return nullptr;
Value *SI = Builder.CreateSelect(Cond, True, False);
SI->takeName(&I);
return SI;
}
/// Freely adapt every user of V as-if V was changed to !V.
/// WARNING: only if canFreelyInvertAllUsersOf() said this can be done.
void InstCombinerImpl::freelyInvertAllUsersOf(Value *I) {
for (User *U : I->users()) {
switch (cast<Instruction>(U)->getOpcode()) {
case Instruction::Select: {
auto *SI = cast<SelectInst>(U);
SI->swapValues();
SI->swapProfMetadata();
break;
}
case Instruction::Br:
cast<BranchInst>(U)->swapSuccessors(); // swaps prof metadata too
break;
case Instruction::Xor:
replaceInstUsesWith(cast<Instruction>(*U), I);
break;
default:
llvm_unreachable("Got unexpected user - out of sync with "
"canFreelyInvertAllUsersOf() ?");
}
}
}
/// Given a 'sub' instruction, return the RHS of the instruction if the LHS is a
/// constant zero (which is the 'negate' form).
Value *InstCombinerImpl::dyn_castNegVal(Value *V) const {
Value *NegV;
if (match(V, m_Neg(m_Value(NegV))))
return NegV;
// Constants can be considered to be negated values if they can be folded.
if (ConstantInt *C = dyn_cast<ConstantInt>(V))
return ConstantExpr::getNeg(C);
if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
if (C->getType()->getElementType()->isIntegerTy())
return ConstantExpr::getNeg(C);
if (ConstantVector *CV = dyn_cast<ConstantVector>(V)) {
for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) {
Constant *Elt = CV->getAggregateElement(i);
if (!Elt)
return nullptr;
if (isa<UndefValue>(Elt))
continue;
if (!isa<ConstantInt>(Elt))
return nullptr;
}
return ConstantExpr::getNeg(CV);
}
// Negate integer vector splats.
if (auto *CV = dyn_cast<Constant>(V))
if (CV->getType()->isVectorTy() &&
CV->getType()->getScalarType()->isIntegerTy() && CV->getSplatValue())
return ConstantExpr::getNeg(CV);
return nullptr;
}
/// A binop with a constant operand and a sign-extended boolean operand may be
/// converted into a select of constants by applying the binary operation to
/// the constant with the two possible values of the extended boolean (0 or -1).
Instruction *InstCombinerImpl::foldBinopOfSextBoolToSelect(BinaryOperator &BO) {
// TODO: Handle non-commutative binop (constant is operand 0).
// TODO: Handle zext.
// TODO: Peek through 'not' of cast.
Value *BO0 = BO.getOperand(0);
Value *BO1 = BO.getOperand(1);
Value *X;
Constant *C;
if (!match(BO0, m_SExt(m_Value(X))) || !match(BO1, m_ImmConstant(C)) ||
!X->getType()->isIntOrIntVectorTy(1))
return nullptr;
// bo (sext i1 X), C --> select X, (bo -1, C), (bo 0, C)
Constant *Ones = ConstantInt::getAllOnesValue(BO.getType());
Constant *Zero = ConstantInt::getNullValue(BO.getType());
Constant *TVal = ConstantExpr::get(BO.getOpcode(), Ones, C);
Constant *FVal = ConstantExpr::get(BO.getOpcode(), Zero, C);
return SelectInst::Create(X, TVal, FVal);
}
static Value *foldOperationIntoSelectOperand(Instruction &I, Value *SO,
InstCombiner::BuilderTy &Builder) {
if (auto *Cast = dyn_cast<CastInst>(&I))
return Builder.CreateCast(Cast->getOpcode(), SO, I.getType());
if (auto *II = dyn_cast<IntrinsicInst>(&I)) {
assert(canConstantFoldCallTo(II, cast<Function>(II->getCalledOperand())) &&
"Expected constant-foldable intrinsic");
Intrinsic::ID IID = II->getIntrinsicID();
if (II->arg_size() == 1)
return Builder.CreateUnaryIntrinsic(IID, SO);
// This works for real binary ops like min/max (where we always expect the
// constant operand to be canonicalized as op1) and unary ops with a bonus
// constant argument like ctlz/cttz.
// TODO: Handle non-commutative binary intrinsics as below for binops.
assert(II->arg_size() == 2 && "Expected binary intrinsic");
assert(isa<Constant>(II->getArgOperand(1)) && "Expected constant operand");
return Builder.CreateBinaryIntrinsic(IID, SO, II->getArgOperand(1));
}
assert(I.isBinaryOp() && "Unexpected opcode for select folding");
// Figure out if the constant is the left or the right argument.
bool ConstIsRHS = isa<Constant>(I.getOperand(1));
Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
if (auto *SOC = dyn_cast<Constant>(SO)) {
if (ConstIsRHS)
return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
}
Value *Op0 = SO, *Op1 = ConstOperand;
if (!ConstIsRHS)
std::swap(Op0, Op1);
auto *BO = cast<BinaryOperator>(&I);
Value *RI = Builder.CreateBinOp(BO->getOpcode(), Op0, Op1,
SO->getName() + ".op");
auto *FPInst = dyn_cast<Instruction>(RI);
if (FPInst && isa<FPMathOperator>(FPInst))
FPInst->copyFastMathFlags(BO);
return RI;
}
Instruction *InstCombinerImpl::FoldOpIntoSelect(Instruction &Op,
SelectInst *SI) {
// Don't modify shared select instructions.
if (!SI->hasOneUse())
return nullptr;
Value *TV = SI->getTrueValue();
Value *FV = SI->getFalseValue();
if (!(isa<Constant>(TV) || isa<Constant>(FV)))
return nullptr;
// Bool selects with constant operands can be folded to logical ops.
if (SI->getType()->isIntOrIntVectorTy(1))
return nullptr;
// If it's a bitcast involving vectors, make sure it has the same number of
// elements on both sides.
if (auto *BC = dyn_cast<BitCastInst>(&Op)) {
VectorType *DestTy = dyn_cast<VectorType>(BC->getDestTy());
VectorType *SrcTy = dyn_cast<VectorType>(BC->getSrcTy());
// Verify that either both or neither are vectors.
if ((SrcTy == nullptr) != (DestTy == nullptr))
return nullptr;
// If vectors, verify that they have the same number of elements.
if (SrcTy && SrcTy->getElementCount() != DestTy->getElementCount())
return nullptr;
}
// Test if a CmpInst instruction is used exclusively by a select as
// part of a minimum or maximum operation. If so, refrain from doing
// any other folding. This helps out other analyses which understand
// non-obfuscated minimum and maximum idioms, such as ScalarEvolution
// and CodeGen. And in this case, at least one of the comparison
// operands has at least one user besides the compare (the select),
// which would often largely negate the benefit of folding anyway.
if (auto *CI = dyn_cast<CmpInst>(SI->getCondition())) {
if (CI->hasOneUse()) {
Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1);
// FIXME: This is a hack to avoid infinite looping with min/max patterns.
// We have to ensure that vector constants that only differ with
// undef elements are treated as equivalent.
auto areLooselyEqual = [](Value *A, Value *B) {
if (A == B)
return true;
// Test for vector constants.
Constant *ConstA, *ConstB;
if (!match(A, m_Constant(ConstA)) || !match(B, m_Constant(ConstB)))
return false;
// TODO: Deal with FP constants?
if (!A->getType()->isIntOrIntVectorTy() || A->getType() != B->getType())
return false;
// Compare for equality including undefs as equal.
auto *Cmp = ConstantExpr::getCompare(ICmpInst::ICMP_EQ, ConstA, ConstB);
const APInt *C;
return match(Cmp, m_APIntAllowUndef(C)) && C->isOne();
};
if ((areLooselyEqual(TV, Op0) && areLooselyEqual(FV, Op1)) ||
(areLooselyEqual(FV, Op0) && areLooselyEqual(TV, Op1)))
return nullptr;
}
}
Value *NewTV = foldOperationIntoSelectOperand(Op, TV, Builder);
Value *NewFV = foldOperationIntoSelectOperand(Op, FV, Builder);
return SelectInst::Create(SI->getCondition(), NewTV, NewFV, "", nullptr, SI);
}
static Value *foldOperationIntoPhiValue(BinaryOperator *I, Value *InV,
InstCombiner::BuilderTy &Builder) {
bool ConstIsRHS = isa<Constant>(I->getOperand(1));
Constant *C = cast<Constant>(I->getOperand(ConstIsRHS));
if (auto *InC = dyn_cast<Constant>(InV)) {
if (ConstIsRHS)
return ConstantExpr::get(I->getOpcode(), InC, C);
return ConstantExpr::get(I->getOpcode(), C, InC);
}
Value *Op0 = InV, *Op1 = C;
if (!ConstIsRHS)
std::swap(Op0, Op1);
Value *RI = Builder.CreateBinOp(I->getOpcode(), Op0, Op1, "phi.bo");
auto *FPInst = dyn_cast<Instruction>(RI);
if (FPInst && isa<FPMathOperator>(FPInst))
FPInst->copyFastMathFlags(I);
return RI;
}
Instruction *InstCombinerImpl::foldOpIntoPhi(Instruction &I, PHINode *PN) {
unsigned NumPHIValues = PN->getNumIncomingValues();
if (NumPHIValues == 0)
return nullptr;
// We normally only transform phis with a single use. However, if a PHI has
// multiple uses and they are all the same operation, we can fold *all* of the
// uses into the PHI.
if (!PN->hasOneUse()) {
// Walk the use list for the instruction, comparing them to I.
for (User *U : PN->users()) {
Instruction *UI = cast<Instruction>(U);
if (UI != &I && !I.isIdenticalTo(UI))
return nullptr;
}
// Otherwise, we can replace *all* users with the new PHI we form.
}
// Check to see if all of the operands of the PHI are simple constants
// (constantint/constantfp/undef). If there is one non-constant value,
// remember the BB it is in. If there is more than one or if *it* is a PHI,
// bail out. We don't do arbitrary constant expressions here because moving
// their computation can be expensive without a cost model.
BasicBlock *NonConstBB = nullptr;
for (unsigned i = 0; i != NumPHIValues; ++i) {
Value *InVal = PN->getIncomingValue(i);
// For non-freeze, require constant operand
// For freeze, require non-undef, non-poison operand
if (!isa<FreezeInst>(I) && match(InVal, m_ImmConstant()))
continue;
if (isa<FreezeInst>(I) && isGuaranteedNotToBeUndefOrPoison(InVal))
continue;
if (isa<PHINode>(InVal)) return nullptr; // Itself a phi.
if (NonConstBB) return nullptr; // More than one non-const value.
NonConstBB = PN->getIncomingBlock(i);
// If the InVal is an invoke at the end of the pred block, then we can't
// insert a computation after it without breaking the edge.
if (isa<InvokeInst>(InVal))
if (cast<Instruction>(InVal)->getParent() == NonConstBB)
return nullptr;
// If the incoming non-constant value is in I's block, we will remove one
// instruction, but insert another equivalent one, leading to infinite
// instcombine.
if (isPotentiallyReachable(I.getParent(), NonConstBB, nullptr, &DT, LI))
return nullptr;
}
// If there is exactly one non-constant value, we can insert a copy of the
// operation in that block. However, if this is a critical edge, we would be
// inserting the computation on some other paths (e.g. inside a loop). Only
// do this if the pred block is unconditionally branching into the phi block.
// Also, make sure that the pred block is not dead code.
if (NonConstBB != nullptr) {
BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
if (!BI || !BI->isUnconditional() || !DT.isReachableFromEntry(NonConstBB))
return nullptr;
}
// Okay, we can do the transformation: create the new PHI node.
PHINode *NewPN = PHINode::Create(I.getType(), PN->getNumIncomingValues());
InsertNewInstBefore(NewPN, *PN);
NewPN->takeName(PN);
// If we are going to have to insert a new computation, do so right before the
// predecessor's terminator.
if (NonConstBB)
Builder.SetInsertPoint(NonConstBB->getTerminator());
// Next, add all of the operands to the PHI.
if (SelectInst *SI = dyn_cast<SelectInst>(&I)) {
// We only currently try to fold the condition of a select when it is a phi,
// not the true/false values.
Value *TrueV = SI->getTrueValue();
Value *FalseV = SI->getFalseValue();
BasicBlock *PhiTransBB = PN->getParent();
for (unsigned i = 0; i != NumPHIValues; ++i) {
BasicBlock *ThisBB = PN->getIncomingBlock(i);
Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB);
Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB);
Value *InV = nullptr;
// Beware of ConstantExpr: it may eventually evaluate to getNullValue,
// even if currently isNullValue gives false.
Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i));
// For vector constants, we cannot use isNullValue to fold into
// FalseVInPred versus TrueVInPred. When we have individual nonzero
// elements in the vector, we will incorrectly fold InC to
// `TrueVInPred`.
if (InC && isa<ConstantInt>(InC))
InV = InC->isNullValue() ? FalseVInPred : TrueVInPred;
else {
// Generate the select in the same block as PN's current incoming block.
// Note: ThisBB need not be the NonConstBB because vector constants
// which are constants by definition are handled here.
// FIXME: This can lead to an increase in IR generation because we might
// generate selects for vector constant phi operand, that could not be
// folded to TrueVInPred or FalseVInPred as done for ConstantInt. For
// non-vector phis, this transformation was always profitable because
// the select would be generated exactly once in the NonConstBB.
Builder.SetInsertPoint(ThisBB->getTerminator());
InV = Builder.CreateSelect(PN->getIncomingValue(i), TrueVInPred,
FalseVInPred, "phi.sel");
}
NewPN->addIncoming(InV, ThisBB);
}
} else if (CmpInst *CI = dyn_cast<CmpInst>(&I)) {
Constant *C = cast<Constant>(I.getOperand(1));
for (unsigned i = 0; i != NumPHIValues; ++i) {
Value *InV = nullptr;
if (auto *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
else
InV = Builder.CreateCmp(CI->getPredicate(), PN->getIncomingValue(i),
C, "phi.cmp");
NewPN->addIncoming(InV, PN->getIncomingBlock(i));
}
} else if (auto *BO = dyn_cast<BinaryOperator>(&I)) {
for (unsigned i = 0; i != NumPHIValues; ++i) {
Value *InV = foldOperationIntoPhiValue(BO, PN->getIncomingValue(i),
Builder);
NewPN->addIncoming(InV, PN->getIncomingBlock(i));
}
} else if (isa<FreezeInst>(&I)) {
for (unsigned i = 0; i != NumPHIValues; ++i) {
Value *InV;
if (NonConstBB == PN->getIncomingBlock(i))
InV = Builder.CreateFreeze(PN->getIncomingValue(i), "phi.fr");
else
InV = PN->getIncomingValue(i);
NewPN->addIncoming(InV, PN->getIncomingBlock(i));
}
} else {
CastInst *CI = cast<CastInst>(&I);
Type *RetTy = CI->getType();
for (unsigned i = 0; i != NumPHIValues; ++i) {
Value *InV;
if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
else
InV = Builder.CreateCast(CI->getOpcode(), PN->getIncomingValue(i),
I.getType(), "phi.cast");
NewPN->addIncoming(InV, PN->getIncomingBlock(i));
}
}
for (User *U : make_early_inc_range(PN->users())) {
Instruction *User = cast<Instruction>(U);
if (User == &I) continue;
replaceInstUsesWith(*User, NewPN);
eraseInstFromFunction(*User);
}
return replaceInstUsesWith(I, NewPN);
}
Instruction *InstCombinerImpl::foldBinOpIntoSelectOrPhi(BinaryOperator &I) {
if (!isa<Constant>(I.getOperand(1)))
return nullptr;
if (auto *Sel = dyn_cast<SelectInst>(I.getOperand(0))) {
if (Instruction *NewSel = FoldOpIntoSelect(I, Sel))
return NewSel;
} else if (auto *PN = dyn_cast<PHINode>(I.getOperand(0))) {
if (Instruction *NewPhi = foldOpIntoPhi(I, PN))
return NewPhi;
}
return nullptr;
}
/// Given a pointer type and a constant offset, determine whether or not there
/// is a sequence of GEP indices into the pointed type that will land us at the
/// specified offset. If so, fill them into NewIndices and return the resultant
/// element type, otherwise return null.
Type *
InstCombinerImpl::FindElementAtOffset(PointerType *PtrTy, int64_t IntOffset,
SmallVectorImpl<Value *> &NewIndices) {
Type *Ty = PtrTy->getElementType();
if (!Ty->isSized())
return nullptr;
APInt Offset(DL.getIndexTypeSizeInBits(PtrTy), IntOffset);
SmallVector<APInt> Indices = DL.getGEPIndicesForOffset(Ty, Offset);
if (!Offset.isZero())
return nullptr;
for (const APInt &Index : Indices)
NewIndices.push_back(Builder.getInt(Index));
return Ty;
}
static bool shouldMergeGEPs(GEPOperator &GEP, GEPOperator &Src) {
// If this GEP has only 0 indices, it is the same pointer as
// Src. If Src is not a trivial GEP too, don't combine
// the indices.
if (GEP.hasAllZeroIndices() && !Src.hasAllZeroIndices() &&
!Src.hasOneUse())
return false;
return true;
}
/// Return a value X such that Val = X * Scale, or null if none.
/// If the multiplication is known not to overflow, then NoSignedWrap is set.
Value *InstCombinerImpl::Descale(Value *Val, APInt Scale, bool &NoSignedWrap) {
assert(isa<IntegerType>(Val->getType()) && "Can only descale integers!");
assert(cast<IntegerType>(Val->getType())->getBitWidth() ==
Scale.getBitWidth() && "Scale not compatible with value!");
// If Val is zero or Scale is one then Val = Val * Scale.
if (match(Val, m_Zero()) || Scale == 1) {
NoSignedWrap = true;
return Val;
}
// If Scale is zero then it does not divide Val.
if (Scale.isMinValue())
return nullptr;
// Look through chains of multiplications, searching for a constant that is
// divisible by Scale. For example, descaling X*(Y*(Z*4)) by a factor of 4
// will find the constant factor 4 and produce X*(Y*Z). Descaling X*(Y*8) by
// a factor of 4 will produce X*(Y*2). The principle of operation is to bore
// down from Val:
//
// Val = M1 * X || Analysis starts here and works down
// M1 = M2 * Y || Doesn't descend into terms with more
// M2 = Z * 4 \/ than one use
//
// Then to modify a term at the bottom:
//
// Val = M1 * X
// M1 = Z * Y || Replaced M2 with Z
//
// Then to work back up correcting nsw flags.
// Op - the term we are currently analyzing. Starts at Val then drills down.
// Replaced with its descaled value before exiting from the drill down loop.
Value *Op = Val;
// Parent - initially null, but after drilling down notes where Op came from.
// In the example above, Parent is (Val, 0) when Op is M1, because M1 is the
// 0'th operand of Val.
std::pair<Instruction *, unsigned> Parent;
// Set if the transform requires a descaling at deeper levels that doesn't
// overflow.
bool RequireNoSignedWrap = false;
// Log base 2 of the scale. Negative if not a power of 2.
int32_t logScale = Scale.exactLogBase2();
for (;; Op = Parent.first->getOperand(Parent.second)) { // Drill down
if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
// If Op is a constant divisible by Scale then descale to the quotient.
APInt Quotient(Scale), Remainder(Scale); // Init ensures right bitwidth.
APInt::sdivrem(CI->getValue(), Scale, Quotient, Remainder);
if (!Remainder.isMinValue())
// Not divisible by Scale.
return nullptr;
// Replace with the quotient in the parent.
Op = ConstantInt::get(CI->getType(), Quotient);
NoSignedWrap = true;
break;
}
if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op)) {
if (BO->getOpcode() == Instruction::Mul) {
// Multiplication.
NoSignedWrap = BO->hasNoSignedWrap();
if (RequireNoSignedWrap && !NoSignedWrap)
return nullptr;
// There are three cases for multiplication: multiplication by exactly
// the scale, multiplication by a constant different to the scale, and
// multiplication by something else.
Value *LHS = BO->getOperand(0);
Value *RHS = BO->getOperand(1);
if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
// Multiplication by a constant.
if (CI->getValue() == Scale) {
// Multiplication by exactly the scale, replace the multiplication
// by its left-hand side in the parent.
Op = LHS;
break;
}
// Otherwise drill down into the constant.
if (!Op->hasOneUse())
return nullptr;
Parent = std::make_pair(BO, 1);
continue;
}
// Multiplication by something else. Drill down into the left-hand side
// since that's where the reassociate pass puts the good stuff.
if (!Op->hasOneUse())
return nullptr;
Parent = std::make_pair(BO, 0);
continue;
}
if (logScale > 0 && BO->getOpcode() == Instruction::Shl &&
isa<ConstantInt>(BO->getOperand(1))) {
// Multiplication by a power of 2.
NoSignedWrap = BO->hasNoSignedWrap();
if (RequireNoSignedWrap && !NoSignedWrap)
return nullptr;
Value *LHS = BO->getOperand(0);
int32_t Amt = cast<ConstantInt>(BO->getOperand(1))->
getLimitedValue(Scale.getBitWidth());
// Op = LHS << Amt.
if (Amt == logScale) {
// Multiplication by exactly the scale, replace the multiplication
// by its left-hand side in the parent.
Op = LHS;
break;
}
if (Amt < logScale || !Op->hasOneUse())
return nullptr;
// Multiplication by more than the scale. Reduce the multiplying amount
// by the scale in the parent.
Parent = std::make_pair(BO, 1);
Op = ConstantInt::get(BO->getType(), Amt - logScale);
break;
}
}
if (!Op->hasOneUse())
return nullptr;
if (CastInst *Cast = dyn_cast<CastInst>(Op)) {
if (Cast->getOpcode() == Instruction::SExt) {
// Op is sign-extended from a smaller type, descale in the smaller type.
unsigned SmallSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
APInt SmallScale = Scale.trunc(SmallSize);
// Suppose Op = sext X, and we descale X as Y * SmallScale. We want to
// descale Op as (sext Y) * Scale. In order to have
// sext (Y * SmallScale) = (sext Y) * Scale
// some conditions need to hold however: SmallScale must sign-extend to
// Scale and the multiplication Y * SmallScale should not overflow.
if (SmallScale.sext(Scale.getBitWidth()) != Scale)
// SmallScale does not sign-extend to Scale.
return nullptr;
assert(SmallScale.exactLogBase2() == logScale);
// Require that Y * SmallScale must not overflow.
RequireNoSignedWrap = true;
// Drill down through the cast.
Parent = std::make_pair(Cast, 0);
Scale = SmallScale;
continue;
}
if (Cast->getOpcode() == Instruction::Trunc) {
// Op is truncated from a larger type, descale in the larger type.
// Suppose Op = trunc X, and we descale X as Y * sext Scale. Then
// trunc (Y * sext Scale) = (trunc Y) * Scale
// always holds. However (trunc Y) * Scale may overflow even if
// trunc (Y * sext Scale) does not, so nsw flags need to be cleared
// from this point up in the expression (see later).
if (RequireNoSignedWrap)
return nullptr;
// Drill down through the cast.
unsigned LargeSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
Parent = std::make_pair(Cast, 0);
Scale = Scale.sext(LargeSize);
if (logScale + 1 == (int32_t)Cast->getType()->getPrimitiveSizeInBits())
logScale = -1;
assert(Scale.exactLogBase2() == logScale);
continue;
}
}
// Unsupported expression, bail out.
return nullptr;
}
// If Op is zero then Val = Op * Scale.
if (match(Op, m_Zero())) {
NoSignedWrap = true;
return Op;
}
// We know that we can successfully descale, so from here on we can safely
// modify the IR. Op holds the descaled version of the deepest term in the
// expression. NoSignedWrap is 'true' if multiplying Op by Scale is known
// not to overflow.
if (!Parent.first)
// The expression only had one term.
return Op;
// Rewrite the parent using the descaled version of its operand.
assert(Parent.first->hasOneUse() && "Drilled down when more than one use!");
assert(Op != Parent.first->getOperand(Parent.second) &&
"Descaling was a no-op?");
replaceOperand(*Parent.first, Parent.second, Op);
Worklist.push(Parent.first);
// Now work back up the expression correcting nsw flags. The logic is based
// on the following observation: if X * Y is known not to overflow as a signed
// multiplication, and Y is replaced by a value Z with smaller absolute value,
// then X * Z will not overflow as a signed multiplication either. As we work
// our way up, having NoSignedWrap 'true' means that the descaled value at the
// current level has strictly smaller absolute value than the original.
Instruction *Ancestor = Parent.first;
do {
if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Ancestor)) {
// If the multiplication wasn't nsw then we can't say anything about the
// value of the descaled multiplication, and we have to clear nsw flags
// from this point on up.
bool OpNoSignedWrap = BO->hasNoSignedWrap();
NoSignedWrap &= OpNoSignedWrap;
if (NoSignedWrap != OpNoSignedWrap) {
BO->setHasNoSignedWrap(NoSignedWrap);
Worklist.push(Ancestor);
}
} else if (Ancestor->getOpcode() == Instruction::Trunc) {
// The fact that the descaled input to the trunc has smaller absolute
// value than the original input doesn't tell us anything useful about
// the absolute values of the truncations.
NoSignedWrap = false;
}
assert((Ancestor->getOpcode() != Instruction::SExt || NoSignedWrap) &&
"Failed to keep proper track of nsw flags while drilling down?");
if (Ancestor == Val)
// Got to the top, all done!
return Val;
// Move up one level in the expression.
assert(Ancestor->hasOneUse() && "Drilled down when more than one use!");
Ancestor = Ancestor->user_back();
} while (true);
}
Instruction *InstCombinerImpl::foldVectorBinop(BinaryOperator &Inst) {
if (!isa<VectorType>(Inst.getType()))
return nullptr;
BinaryOperator::BinaryOps Opcode = Inst.getOpcode();
Value *LHS = Inst.getOperand(0), *RHS = Inst.getOperand(1);
assert(cast<VectorType>(LHS->getType())->getElementCount() ==
cast<VectorType>(Inst.getType())->getElementCount());
assert(cast<VectorType>(RHS->getType())->getElementCount() ==
cast<VectorType>(Inst.getType())->getElementCount());
// If both operands of the binop are vector concatenations, then perform the
// narrow binop on each pair of the source operands followed by concatenation
// of the results.
Value *L0, *L1, *R0, *R1;
ArrayRef<int> Mask;
if (match(LHS, m_Shuffle(m_Value(L0), m_Value(L1), m_Mask(Mask))) &&
match(RHS, m_Shuffle(m_Value(R0), m_Value(R1), m_SpecificMask(Mask))) &&
LHS->hasOneUse() && RHS->hasOneUse() &&
cast<ShuffleVectorInst>(LHS)->isConcat() &&
cast<ShuffleVectorInst>(RHS)->isConcat()) {
// This transform does not have the speculative execution constraint as
// below because the shuffle is a concatenation. The new binops are
// operating on exactly the same elements as the existing binop.
// TODO: We could ease the mask requirement to allow different undef lanes,
// but that requires an analysis of the binop-with-undef output value.
Value *NewBO0 = Builder.CreateBinOp(Opcode, L0, R0);
if (auto *BO = dyn_cast<BinaryOperator>(NewBO0))
BO->copyIRFlags(&Inst);
Value *NewBO1 = Builder.CreateBinOp(Opcode, L1, R1);
if (auto *BO = dyn_cast<BinaryOperator>(NewBO1))
BO->copyIRFlags(&Inst);
return new ShuffleVectorInst(NewBO0, NewBO1, Mask);
}
// It may not be safe to reorder shuffles and things like div, urem, etc.
// because we may trap when executing those ops on unknown vector elements.
// See PR20059.
if (!isSafeToSpeculativelyExecute(&Inst))
return nullptr;
auto createBinOpShuffle = [&](Value *X, Value *Y, ArrayRef<int> M) {
Value *XY = Builder.CreateBinOp(Opcode, X, Y);
if (auto *BO = dyn_cast<BinaryOperator>(XY))
BO->copyIRFlags(&Inst);
return new ShuffleVectorInst(XY, M);
};
// If both arguments of the binary operation are shuffles that use the same
// mask and shuffle within a single vector, move the shuffle after the binop.
Value *V1, *V2;
if (match(LHS, m_Shuffle(m_Value(V1), m_Undef(), m_Mask(Mask))) &&
match(RHS, m_Shuffle(m_Value(V2), m_Undef(), m_SpecificMask(Mask))) &&
V1->getType() == V2->getType() &&
(LHS->hasOneUse() || RHS->hasOneUse() || LHS == RHS)) {
// Op(shuffle(V1, Mask), shuffle(V2, Mask)) -> shuffle(Op(V1, V2), Mask)
return createBinOpShuffle(V1, V2, Mask);
}
// If both arguments of a commutative binop are select-shuffles that use the
// same mask with commuted operands, the shuffles are unnecessary.
if (Inst.isCommutative() &&
match(LHS, m_Shuffle(m_Value(V1), m_Value(V2), m_Mask(Mask))) &&
match(RHS,
m_Shuffle(m_Specific(V2), m_Specific(V1), m_SpecificMask(Mask)))) {
auto *LShuf = cast<ShuffleVectorInst>(LHS);
auto *RShuf = cast<ShuffleVectorInst>(RHS);
// TODO: Allow shuffles that contain undefs in the mask?
// That is legal, but it reduces undef knowledge.
// TODO: Allow arbitrary shuffles by shuffling after binop?
// That might be legal, but we have to deal with poison.
if (LShuf->isSelect() &&
!is_contained(LShuf->getShuffleMask(), UndefMaskElem) &&
RShuf->isSelect() &&
!is_contained(RShuf->getShuffleMask(), UndefMaskElem)) {
// Example:
// LHS = shuffle V1, V2, <0, 5, 6, 3>
// RHS = shuffle V2, V1, <0, 5, 6, 3>
// LHS + RHS --> (V10+V20, V21+V11, V22+V12, V13+V23) --> V1 + V2
Instruction *NewBO = BinaryOperator::Create(Opcode, V1, V2);
NewBO->copyIRFlags(&Inst);
return NewBO;
}
}
// If one argument is a shuffle within one vector and the other is a constant,
// try moving the shuffle after the binary operation. This canonicalization
// intends to move shuffles closer to other shuffles and binops closer to
// other binops, so they can be folded. It may also enable demanded elements
// transforms.
Constant *C;
auto *InstVTy = dyn_cast<FixedVectorType>(Inst.getType());
if (InstVTy &&
match(&Inst,
m_c_BinOp(m_OneUse(m_Shuffle(m_Value(V1), m_Undef(), m_Mask(Mask))),
m_ImmConstant(C))) &&
cast<FixedVectorType>(V1->getType())->getNumElements() <=
InstVTy->getNumElements()) {
assert(InstVTy->getScalarType() == V1->getType()->getScalarType() &&
"Shuffle should not change scalar type");
// Find constant NewC that has property:
// shuffle(NewC, ShMask) = C
// If such constant does not exist (example: ShMask=<0,0> and C=<1,2>)
// reorder is not possible. A 1-to-1 mapping is not required. Example:
// ShMask = <1,1,2,2> and C = <5,5,6,6> --> NewC = <undef,5,6,undef>
bool ConstOp1 = isa<Constant>(RHS);
ArrayRef<int> ShMask = Mask;
unsigned SrcVecNumElts =
cast<FixedVectorType>(V1->getType())->getNumElements();
UndefValue *UndefScalar = UndefValue::get(C->getType()->getScalarType());
SmallVector<Constant *, 16> NewVecC(SrcVecNumElts, UndefScalar);
bool MayChange = true;
unsigned NumElts = InstVTy->getNumElements();
for (unsigned I = 0; I < NumElts; ++I) {
Constant *CElt = C->getAggregateElement(I);
if (ShMask[I] >= 0) {
assert(ShMask[I] < (int)NumElts && "Not expecting narrowing shuffle");
Constant *NewCElt = NewVecC[ShMask[I]];
// Bail out if:
// 1. The constant vector contains a constant expression.
// 2. The shuffle needs an element of the constant vector that can't
// be mapped to a new constant vector.
// 3. This is a widening shuffle that copies elements of V1 into the
// extended elements (extending with undef is allowed).
if (!CElt || (!isa<UndefValue>(NewCElt) && NewCElt != CElt) ||
I >= SrcVecNumElts) {
MayChange = false;
break;
}
NewVecC[ShMask[I]] = CElt;
}
// If this is a widening shuffle, we must be able to extend with undef
// elements. If the original binop does not produce an undef in the high
// lanes, then this transform is not safe.
// Similarly for undef lanes due to the shuffle mask, we can only
// transform binops that preserve undef.
// TODO: We could shuffle those non-undef constant values into the
// result by using a constant vector (rather than an undef vector)
// as operand 1 of the new binop, but that might be too aggressive
// for target-independent shuffle creation.
if (I >= SrcVecNumElts || ShMask[I] < 0) {
Constant *MaybeUndef =
ConstOp1 ? ConstantExpr::get(Opcode, UndefScalar, CElt)
: ConstantExpr::get(Opcode, CElt, UndefScalar);
if (!match(MaybeUndef, m_Undef())) {
MayChange = false;
break;
}
}
}
if (MayChange) {
Constant *NewC = ConstantVector::get(NewVecC);
// It may not be safe to execute a binop on a vector with undef elements
// because the entire instruction can be folded to undef or create poison
// that did not exist in the original code.
if (Inst.isIntDivRem() || (Inst.isShift() && ConstOp1))
NewC = getSafeVectorConstantForBinop(Opcode, NewC, ConstOp1);
// Op(shuffle(V1, Mask), C) -> shuffle(Op(V1, NewC), Mask)
// Op(C, shuffle(V1, Mask)) -> shuffle(Op(NewC, V1), Mask)
Value *NewLHS = ConstOp1 ? V1 : NewC;
Value *NewRHS = ConstOp1 ? NewC : V1;
return createBinOpShuffle(NewLHS, NewRHS, Mask);
}
}
// Try to reassociate to sink a splat shuffle after a binary operation.
if (Inst.isAssociative() && Inst.isCommutative()) {
// Canonicalize shuffle operand as LHS.
if (isa<ShuffleVectorInst>(RHS))
std::swap(LHS, RHS);
Value *X;
ArrayRef<int> MaskC;
int SplatIndex;
Value *Y, *OtherOp;
if (!match(LHS,
m_OneUse(m_Shuffle(m_Value(X), m_Undef(), m_Mask(MaskC)))) ||
!match(MaskC, m_SplatOrUndefMask(SplatIndex)) ||
X->getType() != Inst.getType() ||
!match(RHS, m_OneUse(m_BinOp(Opcode, m_Value(Y), m_Value(OtherOp)))))
return nullptr;
// FIXME: This may not be safe if the analysis allows undef elements. By
// moving 'Y' before the splat shuffle, we are implicitly assuming
// that it is not undef/poison at the splat index.
if (isSplatValue(OtherOp, SplatIndex)) {
std::swap(Y, OtherOp);
} else if (!isSplatValue(Y, SplatIndex)) {
return nullptr;
}
// X and Y are splatted values, so perform the binary operation on those
// values followed by a splat followed by the 2nd binary operation:
// bo (splat X), (bo Y, OtherOp) --> bo (splat (bo X, Y)), OtherOp
Value *NewBO = Builder.CreateBinOp(Opcode, X, Y);
SmallVector<int, 8> NewMask(MaskC.size(), SplatIndex);
Value *NewSplat = Builder.CreateShuffleVector(NewBO, NewMask);
Instruction *R = BinaryOperator::Create(Opcode, NewSplat, OtherOp);
// Intersect FMF on both new binops. Other (poison-generating) flags are
// dropped to be safe.
if (isa<FPMathOperator>(R)) {
R->copyFastMathFlags(&Inst);
R->andIRFlags(RHS);
}
if (auto *NewInstBO = dyn_cast<BinaryOperator>(NewBO))
NewInstBO->copyIRFlags(R);
return R;
}
return nullptr;
}
/// Try to narrow the width of a binop if at least 1 operand is an extend of
/// of a value. This requires a potentially expensive known bits check to make
/// sure the narrow op does not overflow.
Instruction *InstCombinerImpl::narrowMathIfNoOverflow(BinaryOperator &BO) {
// We need at least one extended operand.
Value *Op0 = BO.getOperand(0), *Op1 = BO.getOperand(1);
// If this is a sub, we swap the operands since we always want an extension
// on the RHS. The LHS can be an extension or a constant.
if (BO.getOpcode() == Instruction::Sub)
std::swap(Op0, Op1);
Value *X;
bool IsSext = match(Op0, m_SExt(m_Value(X)));
if (!IsSext && !match(Op0, m_ZExt(m_Value(X))))
return nullptr;
// If both operands are the same extension from the same source type and we
// can eliminate at least one (hasOneUse), this might work.
CastInst::CastOps CastOpc = IsSext ? Instruction::SExt : Instruction::ZExt;
Value *Y;
if (!(match(Op1, m_ZExtOrSExt(m_Value(Y))) && X->getType() == Y->getType() &&
cast<Operator>(Op1)->getOpcode() == CastOpc &&
(Op0->hasOneUse() || Op1->hasOneUse()))) {
// If that did not match, see if we have a suitable constant operand.
// Truncating and extending must produce the same constant.
Constant *WideC;
if (!Op0->hasOneUse() || !match(Op1, m_Constant(WideC)))
return nullptr;
Constant *NarrowC = ConstantExpr::getTrunc(WideC, X->getType());
if (ConstantExpr::getCast(CastOpc, NarrowC, BO.getType()) != WideC)
return nullptr;
Y = NarrowC;
}
// Swap back now that we found our operands.
if (BO.getOpcode() == Instruction::Sub)
std::swap(X, Y);
// Both operands have narrow versions. Last step: the math must not overflow
// in the narrow width.
if (!willNotOverflow(BO.getOpcode(), X, Y, BO, IsSext))
return nullptr;
// bo (ext X), (ext Y) --> ext (bo X, Y)
// bo (ext X), C --> ext (bo X, C')
Value *NarrowBO = Builder.CreateBinOp(BO.getOpcode(), X, Y, "narrow");
if (auto *NewBinOp = dyn_cast<BinaryOperator>(NarrowBO)) {
if (IsSext)
NewBinOp->setHasNoSignedWrap();
else
NewBinOp->setHasNoUnsignedWrap();
}
return CastInst::Create(CastOpc, NarrowBO, BO.getType());
}
static bool isMergedGEPInBounds(GEPOperator &GEP1, GEPOperator &GEP2) {
// At least one GEP must be inbounds.
if (!GEP1.isInBounds() && !GEP2.isInBounds())
return false;
return (GEP1.isInBounds() || GEP1.hasAllZeroIndices()) &&
(GEP2.isInBounds() || GEP2.hasAllZeroIndices());
}
/// Thread a GEP operation with constant indices through the constant true/false
/// arms of a select.
static Instruction *foldSelectGEP(GetElementPtrInst &GEP,
InstCombiner::BuilderTy &Builder) {
if (!GEP.hasAllConstantIndices())
return nullptr;
Instruction *Sel;
Value *Cond;
Constant *TrueC, *FalseC;
if (!match(GEP.getPointerOperand(), m_Instruction(Sel)) ||
!match(Sel,
m_Select(m_Value(Cond), m_Constant(TrueC), m_Constant(FalseC))))
return nullptr;
// gep (select Cond, TrueC, FalseC), IndexC --> select Cond, TrueC', FalseC'
// Propagate 'inbounds' and metadata from existing instructions.
// Note: using IRBuilder to create the constants for efficiency.
SmallVector<Value *, 4> IndexC(GEP.indices());
bool IsInBounds = GEP.isInBounds();
Type *Ty = GEP.getSourceElementType();
Value *NewTrueC = IsInBounds ? Builder.CreateInBoundsGEP(Ty, TrueC, IndexC)
: Builder.CreateGEP(Ty, TrueC, IndexC);
Value *NewFalseC = IsInBounds ? Builder.CreateInBoundsGEP(Ty, FalseC, IndexC)
: Builder.CreateGEP(Ty, FalseC, IndexC);
return SelectInst::Create(Cond, NewTrueC, NewFalseC, "", nullptr, Sel);
}
Instruction *InstCombinerImpl::visitGetElementPtrInst(GetElementPtrInst &GEP) {
SmallVector<Value *, 8> Ops(GEP.operands());
Type *GEPType = GEP.getType();
Type *GEPEltType = GEP.getSourceElementType();
bool IsGEPSrcEleScalable = isa<ScalableVectorType>(GEPEltType);
if (Value *V = SimplifyGEPInst(GEPEltType, Ops, GEP.isInBounds(),
SQ.getWithInstruction(&GEP)))
return replaceInstUsesWith(GEP, V);
// For vector geps, use the generic demanded vector support.
// Skip if GEP return type is scalable. The number of elements is unknown at
// compile-time.
if (auto *GEPFVTy = dyn_cast<FixedVectorType>(GEPType)) {
auto VWidth = GEPFVTy->getNumElements();
APInt UndefElts(VWidth, 0);
APInt AllOnesEltMask(APInt::getAllOnes(VWidth));
if (Value *V = SimplifyDemandedVectorElts(&GEP, AllOnesEltMask,
UndefElts)) {
if (V != &GEP)
return replaceInstUsesWith(GEP, V);
return &GEP;
}
// TODO: 1) Scalarize splat operands, 2) scalarize entire instruction if
// possible (decide on canonical form for pointer broadcast), 3) exploit
// undef elements to decrease demanded bits
}
Value *PtrOp = GEP.getOperand(0);
// Eliminate unneeded casts for indices, and replace indices which displace
// by multiples of a zero size type with zero.
bool MadeChange = false;
// Index width may not be the same width as pointer width.
// Data layout chooses the right type based on supported integer types.
Type *NewScalarIndexTy =
DL.getIndexType(GEP.getPointerOperandType()->getScalarType());
gep_type_iterator GTI = gep_type_begin(GEP);
for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end(); I != E;
++I, ++GTI) {
// Skip indices into struct types.
if (GTI.isStruct())
continue;
Type *IndexTy = (*I)->getType();
Type *NewIndexType =
IndexTy->isVectorTy()
? VectorType::get(NewScalarIndexTy,
cast<VectorType>(IndexTy)->getElementCount())
: NewScalarIndexTy;
// If the element type has zero size then any index over it is equivalent
// to an index of zero, so replace it with zero if it is not zero already.
Type *EltTy = GTI.getIndexedType();
if (EltTy->isSized() && DL.getTypeAllocSize(EltTy).isZero())
if (!isa<Constant>(*I) || !match(I->get(), m_Zero())) {
*I = Constant::getNullValue(NewIndexType);
MadeChange = true;
}
if (IndexTy != NewIndexType) {
// If we are using a wider index than needed for this platform, shrink
// it to what we need. If narrower, sign-extend it to what we need.
// This explicit cast can make subsequent optimizations more obvious.
*I = Builder.CreateIntCast(*I, NewIndexType, true);
MadeChange = true;
}
}
if (MadeChange)
return &GEP;
// Check to see if the inputs to the PHI node are getelementptr instructions.
if (auto *PN = dyn_cast<PHINode>(PtrOp)) {
auto *Op1 = dyn_cast<GetElementPtrInst>(PN->getOperand(0));
if (!Op1)
return nullptr;
// Don't fold a GEP into itself through a PHI node. This can only happen
// through the back-edge of a loop. Folding a GEP into itself means that
// the value of the previous iteration needs to be stored in the meantime,
// thus requiring an additional register variable to be live, but not
// actually achieving anything (the GEP still needs to be executed once per
// loop iteration).
if (Op1 == &GEP)
return nullptr;
int DI = -1;
for (auto I = PN->op_begin()+1, E = PN->op_end(); I !=E; ++I) {
auto *Op2 = dyn_cast<GetElementPtrInst>(*I);
if (!Op2 || Op1->getNumOperands() != Op2->getNumOperands())
return nullptr;
// As for Op1 above, don't try to fold a GEP into itself.
if (Op2 == &GEP)
return nullptr;
// Keep track of the type as we walk the GEP.
Type *CurTy = nullptr;
for (unsigned J = 0, F = Op1->getNumOperands(); J != F; ++J) {
if (Op1->getOperand(J)->getType() != Op2->getOperand(J)->getType())
return nullptr;
if (Op1->getOperand(J) != Op2->getOperand(J)) {
if (DI == -1) {
// We have not seen any differences yet in the GEPs feeding the
// PHI yet, so we record this one if it is allowed to be a
// variable.
// The first two arguments can vary for any GEP, the rest have to be
// static for struct slots
if (J > 1) {
assert(CurTy && "No current type?");
if (CurTy->isStructTy())
return nullptr;
}
DI = J;
} else {
// The GEP is different by more than one input. While this could be
// extended to support GEPs that vary by more than one variable it
// doesn't make sense since it greatly increases the complexity and
// would result in an R+R+R addressing mode which no backend
// directly supports and would need to be broken into several
// simpler instructions anyway.
return nullptr;
}
}
// Sink down a layer of the type for the next iteration.
if (J > 0) {
if (J == 1) {
CurTy = Op1->getSourceElementType();
} else {
CurTy =
GetElementPtrInst::getTypeAtIndex(CurTy, Op1->getOperand(J));
}
}
}
}
// If not all GEPs are identical we'll have to create a new PHI node.
// Check that the old PHI node has only one use so that it will get
// removed.
if (DI != -1 && !PN->hasOneUse())
return nullptr;
auto *NewGEP = cast<GetElementPtrInst>(Op1->clone());
if (DI == -1) {
// All the GEPs feeding the PHI are identical. Clone one down into our
// BB so that it can be merged with the current GEP.
} else {
// All the GEPs feeding the PHI differ at a single offset. Clone a GEP
// into the current block so it can be merged, and create a new PHI to
// set that index.
PHINode *NewPN;
{
IRBuilderBase::InsertPointGuard Guard(Builder);
Builder.SetInsertPoint(PN);
NewPN = Builder.CreatePHI(Op1->getOperand(DI)->getType(),
PN->getNumOperands());
}
for (auto &I : PN->operands())
NewPN->addIncoming(cast<GEPOperator>(I)->getOperand(DI),
PN->getIncomingBlock(I));
NewGEP->setOperand(DI, NewPN);
}
GEP.getParent()->getInstList().insert(
GEP.getParent()->getFirstInsertionPt(), NewGEP);
replaceOperand(GEP, 0, NewGEP);
PtrOp = NewGEP;
}
// Combine Indices - If the source pointer to this getelementptr instruction
// is a getelementptr instruction, combine the indices of the two
// getelementptr instructions into a single instruction.
if (auto *Src = dyn_cast<GEPOperator>(PtrOp)) {
if (!shouldMergeGEPs(*cast<GEPOperator>(&GEP), *Src))
return nullptr;
if (Src->getNumOperands() == 2 && GEP.getNumOperands() == 2 &&
Src->hasOneUse()) {
Value *GO1 = GEP.getOperand(1);
Value *SO1 = Src->getOperand(1);
if (LI) {
// Try to reassociate loop invariant GEP chains to enable LICM.
if (Loop *L = LI->getLoopFor(GEP.getParent())) {
// Reassociate the two GEPs if SO1 is variant in the loop and GO1 is
// invariant: this breaks the dependence between GEPs and allows LICM
// to hoist the invariant part out of the loop.
if (L->isLoopInvariant(GO1) && !L->isLoopInvariant(SO1)) {
// We have to be careful here.
// We have something like:
// %src = getelementptr <ty>, <ty>* %base, <ty> %idx
// %gep = getelementptr <ty>, <ty>* %src, <ty> %idx2
// If we just swap idx & idx2 then we could inadvertantly
// change %src from a vector to a scalar, or vice versa.
// Cases:
// 1) %base a scalar & idx a scalar & idx2 a vector
// => Swapping idx & idx2 turns %src into a vector type.
// 2) %base a scalar & idx a vector & idx2 a scalar
// => Swapping idx & idx2 turns %src in a scalar type
// 3) %base, %idx, and %idx2 are scalars
// => %src & %gep are scalars
// => swapping idx & idx2 is safe
// 4) %base a vector
// => %src is a vector
// => swapping idx & idx2 is safe.
auto *SO0 = Src->getOperand(0);
auto *SO0Ty = SO0->getType();
if (!isa<VectorType>(GEPType) || // case 3
isa<VectorType>(SO0Ty)) { // case 4
Src->setOperand(1, GO1);
GEP.setOperand(1, SO1);
return &GEP;
} else {
// Case 1 or 2
// -- have to recreate %src & %gep
// put NewSrc at same location as %src
Builder.SetInsertPoint(cast<Instruction>(PtrOp));
Value *NewSrc =
Builder.CreateGEP(GEPEltType, SO0, GO1, Src->getName());
// Propagate 'inbounds' if the new source was not constant-folded.
if (auto *NewSrcGEPI = dyn_cast<GetElementPtrInst>(NewSrc))
NewSrcGEPI->setIsInBounds(Src->isInBounds());
GetElementPtrInst *NewGEP =
GetElementPtrInst::Create(GEPEltType, NewSrc, {SO1});
NewGEP->setIsInBounds(GEP.isInBounds());
return NewGEP;
}
}
}
}
}
// Note that if our source is a gep chain itself then we wait for that
// chain to be resolved before we perform this transformation. This
// avoids us creating a TON of code in some cases.
if (auto *SrcGEP = dyn_cast<GEPOperator>(Src->getOperand(0)))
if (SrcGEP->getNumOperands() == 2 && shouldMergeGEPs(*Src, *SrcGEP))
return nullptr; // Wait until our source is folded to completion.
SmallVector<Value*, 8> Indices;
// Find out whether the last index in the source GEP is a sequential idx.
bool EndsWithSequential = false;
for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src);
I != E; ++I)
EndsWithSequential = I.isSequential();
// Can we combine the two pointer arithmetics offsets?
if (EndsWithSequential) {
// Replace: gep (gep %P, long B), long A, ...
// With: T = long A+B; gep %P, T, ...
Value *SO1 = Src->getOperand(Src->getNumOperands()-1);
Value *GO1 = GEP.getOperand(1);
// If they aren't the same type, then the input hasn't been processed
// by the loop above yet (which canonicalizes sequential index types to
// intptr_t). Just avoid transforming this until the input has been
// normalized.
if (SO1->getType() != GO1->getType())
return nullptr;
Value *Sum =
SimplifyAddInst(GO1, SO1, false, false, SQ.getWithInstruction(&GEP));
// Only do the combine when we are sure the cost after the
// merge is never more than that before the merge.
if (Sum == nullptr)
return nullptr;
// Update the GEP in place if possible.
if (Src->getNumOperands() == 2) {
GEP.setIsInBounds(isMergedGEPInBounds(*Src, *cast<GEPOperator>(&GEP)));
replaceOperand(GEP, 0, Src->getOperand(0));
replaceOperand(GEP, 1, Sum);
return &GEP;
}
Indices.append(Src->op_begin()+1, Src->op_end()-1);
Indices.push_back(Sum);
Indices.append(GEP.op_begin()+2, GEP.op_end());
} else if (isa<Constant>(*GEP.idx_begin()) &&
cast<Constant>(*GEP.idx_begin())->isNullValue() &&
Src->getNumOperands() != 1) {
// Otherwise we can do the fold if the first index of the GEP is a zero
Indices.append(Src->op_begin()+1, Src->op_end());
Indices.append(GEP.idx_begin()+1, GEP.idx_end());
}
if (!Indices.empty())
return isMergedGEPInBounds(*Src, *cast<GEPOperator>(&GEP))
? GetElementPtrInst::CreateInBounds(
Src->getSourceElementType(), Src->getOperand(0), Indices,
GEP.getName())
: GetElementPtrInst::Create(Src->getSourceElementType(),
Src->getOperand(0), Indices,
GEP.getName());
}
// Skip if GEP source element type is scalable. The type alloc size is unknown
// at compile-time.
if (GEP.getNumIndices() == 1 && !IsGEPSrcEleScalable) {
unsigned AS = GEP.getPointerAddressSpace();
if (GEP.getOperand(1)->getType()->getScalarSizeInBits() ==
DL.getIndexSizeInBits(AS)) {
uint64_t TyAllocSize = DL.getTypeAllocSize(GEPEltType).getFixedSize();
bool Matched = false;
uint64_t C;
Value *V = nullptr;
if (TyAllocSize == 1) {
V = GEP.getOperand(1);
Matched = true;
} else if (match(GEP.getOperand(1),
m_AShr(m_Value(V), m_ConstantInt(C)))) {
if (TyAllocSize == 1ULL << C)
Matched = true;
} else if (match(GEP.getOperand(1),
m_SDiv(m_Value(V), m_ConstantInt(C)))) {
if (TyAllocSize == C)
Matched = true;
}
// Canonicalize (gep i8* X, (ptrtoint Y)-(ptrtoint X)) to (bitcast Y), but
// only if both point to the same underlying object (otherwise provenance
// is not necessarily retained).
Value *Y;
Value *X = GEP.getOperand(0);
if (Matched &&
match(V, m_Sub(m_PtrToInt(m_Value(Y)), m_PtrToInt(m_Specific(X)))) &&
getUnderlyingObject(X) == getUnderlyingObject(Y))
return CastInst::CreatePointerBitCastOrAddrSpaceCast(Y, GEPType);
}
}
// We do not handle pointer-vector geps here.
if (GEPType->isVectorTy())
return nullptr;
// Handle gep(bitcast x) and gep(gep x, 0, 0, 0).
Value *StrippedPtr = PtrOp->stripPointerCasts();
PointerType *StrippedPtrTy = cast<PointerType>(StrippedPtr->getType());
if (StrippedPtr != PtrOp) {
bool HasZeroPointerIndex = false;
Type *StrippedPtrEltTy = StrippedPtrTy->getElementType();
if (auto *C = dyn_cast<ConstantInt>(GEP.getOperand(1)))
HasZeroPointerIndex = C->isZero();
// Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
// into : GEP [10 x i8]* X, i32 0, ...
//
// Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
// into : GEP i8* X, ...
//
// This occurs when the program declares an array extern like "int X[];"
if (HasZeroPointerIndex) {
if (auto *CATy = dyn_cast<ArrayType>(GEPEltType)) {
// GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
if (CATy->getElementType() == StrippedPtrEltTy) {
// -> GEP i8* X, ...
SmallVector<Value *, 8> Idx(drop_begin(GEP.indices()));
GetElementPtrInst *Res = GetElementPtrInst::Create(
StrippedPtrEltTy, StrippedPtr, Idx, GEP.getName());
Res->setIsInBounds(GEP.isInBounds());
if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace())
return Res;
// Insert Res, and create an addrspacecast.
// e.g.,
// GEP (addrspacecast i8 addrspace(1)* X to [0 x i8]*), i32 0, ...
// ->
// %0 = GEP i8 addrspace(1)* X, ...
// addrspacecast i8 addrspace(1)* %0 to i8*
return new AddrSpaceCastInst(Builder.Insert(Res), GEPType);
}
if (auto *XATy = dyn_cast<ArrayType>(StrippedPtrEltTy)) {
// GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
if (CATy->getElementType() == XATy->getElementType()) {
// -> GEP [10 x i8]* X, i32 0, ...
// At this point, we know that the cast source type is a pointer
// to an array of the same type as the destination pointer
// array. Because the array type is never stepped over (there
// is a leading zero) we can fold the cast into this GEP.
if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace()) {
GEP.setSourceElementType(XATy);
return replaceOperand(GEP, 0, StrippedPtr);
}
// Cannot replace the base pointer directly because StrippedPtr's
// address space is different. Instead, create a new GEP followed by
// an addrspacecast.
// e.g.,
// GEP (addrspacecast [10 x i8] addrspace(1)* X to [0 x i8]*),
// i32 0, ...
// ->
// %0 = GEP [10 x i8] addrspace(1)* X, ...
// addrspacecast i8 addrspace(1)* %0 to i8*
SmallVector<Value *, 8> Idx(GEP.indices());
Value *NewGEP =
GEP.isInBounds()
? Builder.CreateInBoundsGEP(StrippedPtrEltTy, StrippedPtr,
Idx, GEP.getName())
: Builder.CreateGEP(StrippedPtrEltTy, StrippedPtr, Idx,
GEP.getName());
return new AddrSpaceCastInst(NewGEP, GEPType);
}
}
}
} else if (GEP.getNumOperands() == 2 && !IsGEPSrcEleScalable) {
// Skip if GEP source element type is scalable. The type alloc size is
// unknown at compile-time.
// Transform things like: %t = getelementptr i32*
// bitcast ([2 x i32]* %str to i32*), i32 %V into: %t1 = getelementptr [2
// x i32]* %str, i32 0, i32 %V; bitcast
if (StrippedPtrEltTy->isArrayTy() &&
DL.getTypeAllocSize(StrippedPtrEltTy->getArrayElementType()) ==
DL.getTypeAllocSize(GEPEltType)) {
Type *IdxType = DL.getIndexType(GEPType);
Value *Idx[2] = { Constant::getNullValue(IdxType), GEP.getOperand(1) };
Value *NewGEP =
GEP.isInBounds()
? Builder.CreateInBoundsGEP(StrippedPtrEltTy, StrippedPtr, Idx,
GEP.getName())
: Builder.CreateGEP(StrippedPtrEltTy, StrippedPtr, Idx,
GEP.getName());
// V and GEP are both pointer types --> BitCast
return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP, GEPType);
}
// Transform things like:
// %V = mul i64 %N, 4
// %t = getelementptr i8* bitcast (i32* %arr to i8*), i32 %V
// into: %t1 = getelementptr i32* %arr, i32 %N; bitcast
if (GEPEltType->isSized() && StrippedPtrEltTy->isSized()) {
// Check that changing the type amounts to dividing the index by a scale
// factor.
uint64_t ResSize = DL.getTypeAllocSize(GEPEltType).getFixedSize();
uint64_t SrcSize = DL.getTypeAllocSize(StrippedPtrEltTy).getFixedSize();
if (ResSize && SrcSize % ResSize == 0) {
Value *Idx = GEP.getOperand(1);
unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
uint64_t Scale = SrcSize / ResSize;
// Earlier transforms ensure that the index has the right type
// according to Data Layout, which considerably simplifies the
// logic by eliminating implicit casts.
assert(Idx->getType() == DL.getIndexType(GEPType) &&
"Index type does not match the Data Layout preferences");
bool NSW;
if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
// Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
// If the multiplication NewIdx * Scale may overflow then the new
// GEP may not be "inbounds".
Value *NewGEP =
GEP.isInBounds() && NSW
? Builder.CreateInBoundsGEP(StrippedPtrEltTy, StrippedPtr,
NewIdx, GEP.getName())
: Builder.CreateGEP(StrippedPtrEltTy, StrippedPtr, NewIdx,
GEP.getName());
// The NewGEP must be pointer typed, so must the old one -> BitCast
return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
GEPType);
}
}
}
// Similarly, transform things like:
// getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
// (where tmp = 8*tmp2) into:
// getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
if (GEPEltType->isSized() && StrippedPtrEltTy->isSized() &&
StrippedPtrEltTy->isArrayTy()) {
// Check that changing to the array element type amounts to dividing the
// index by a scale factor.
uint64_t ResSize = DL.getTypeAllocSize(GEPEltType).getFixedSize();
uint64_t ArrayEltSize =
DL.getTypeAllocSize(StrippedPtrEltTy->getArrayElementType())
.getFixedSize();
if (ResSize && ArrayEltSize % ResSize == 0) {
Value *Idx = GEP.getOperand(1);
unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
uint64_t Scale = ArrayEltSize / ResSize;
// Earlier transforms ensure that the index has the right type
// according to the Data Layout, which considerably simplifies
// the logic by eliminating implicit casts.
assert(Idx->getType() == DL.getIndexType(GEPType) &&
"Index type does not match the Data Layout preferences");
bool NSW;
if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
// Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
// If the multiplication NewIdx * Scale may overflow then the new
// GEP may not be "inbounds".
Type *IndTy = DL.getIndexType(GEPType);
Value *Off[2] = {Constant::getNullValue(IndTy), NewIdx};
Value *NewGEP =
GEP.isInBounds() && NSW
? Builder.CreateInBoundsGEP(StrippedPtrEltTy, StrippedPtr,
Off, GEP.getName())
: Builder.CreateGEP(StrippedPtrEltTy, StrippedPtr, Off,
GEP.getName());
// The NewGEP must be pointer typed, so must the old one -> BitCast
return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
GEPType);
}
}
}
}
}
// addrspacecast between types is canonicalized as a bitcast, then an
// addrspacecast. To take advantage of the below bitcast + struct GEP, look
// through the addrspacecast.
Value *ASCStrippedPtrOp = PtrOp;
if (auto *ASC = dyn_cast<AddrSpaceCastInst>(PtrOp)) {
// X = bitcast A addrspace(1)* to B addrspace(1)*
// Y = addrspacecast A addrspace(1)* to B addrspace(2)*
// Z = gep Y, <...constant indices...>
// Into an addrspacecasted GEP of the struct.
if (auto *BC = dyn_cast<BitCastInst>(ASC->getOperand(0)))
ASCStrippedPtrOp = BC;
}
if (auto *BCI = dyn_cast<BitCastInst>(ASCStrippedPtrOp)) {
Value *SrcOp = BCI->getOperand(0);
PointerType *SrcType = cast<PointerType>(BCI->getSrcTy());
Type *SrcEltType = SrcType->getElementType();
// GEP directly using the source operand if this GEP is accessing an element
// of a bitcasted pointer to vector or array of the same dimensions:
// gep (bitcast <c x ty>* X to [c x ty]*), Y, Z --> gep X, Y, Z
// gep (bitcast [c x ty]* X to <c x ty>*), Y, Z --> gep X, Y, Z
auto areMatchingArrayAndVecTypes = [](Type *ArrTy, Type *VecTy,
const DataLayout &DL) {
auto *VecVTy = cast<FixedVectorType>(VecTy);
return ArrTy->getArrayElementType() == VecVTy->getElementType() &&
ArrTy->getArrayNumElements() == VecVTy->getNumElements() &&
DL.getTypeAllocSize(ArrTy) == DL.getTypeAllocSize(VecTy);
};
if (GEP.getNumOperands() == 3 &&
((GEPEltType->isArrayTy() && isa<FixedVectorType>(SrcEltType) &&
areMatchingArrayAndVecTypes(GEPEltType, SrcEltType, DL)) ||
(isa<FixedVectorType>(GEPEltType) && SrcEltType->isArrayTy() &&
areMatchingArrayAndVecTypes(SrcEltType, GEPEltType, DL)))) {
// Create a new GEP here, as using `setOperand()` followed by
// `setSourceElementType()` won't actually update the type of the
// existing GEP Value. Causing issues if this Value is accessed when
// constructing an AddrSpaceCastInst
Value *NGEP =
GEP.isInBounds()
? Builder.CreateInBoundsGEP(SrcEltType, SrcOp, {Ops[1], Ops[2]})
: Builder.CreateGEP(SrcEltType, SrcOp, {Ops[1], Ops[2]});
NGEP->takeName(&GEP);
// Preserve GEP address space to satisfy users
if (NGEP->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
return new AddrSpaceCastInst(NGEP, GEPType);
return replaceInstUsesWith(GEP, NGEP);
}
// See if we can simplify:
// X = bitcast A* to B*
// Y = gep X, <...constant indices...>
// into a gep of the original struct. This is important for SROA and alias
// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
unsigned OffsetBits = DL.getIndexTypeSizeInBits(GEPType);
APInt Offset(OffsetBits, 0);
// If the bitcast argument is an allocation, The bitcast is for convertion
// to actual type of allocation. Removing such bitcasts, results in having
// GEPs with i8* base and pure byte offsets. That means GEP is not aware of
// struct or array hierarchy.
// By avoiding such GEPs, phi translation and MemoryDependencyAnalysis have
// a better chance to succeed.
if (!isa<BitCastInst>(SrcOp) && GEP.accumulateConstantOffset(DL, Offset) &&
!isAllocationFn(SrcOp, &TLI)) {
// If this GEP instruction doesn't move the pointer, just replace the GEP
// with a bitcast of the real input to the dest type.
if (!Offset) {
// If the bitcast is of an allocation, and the allocation will be
// converted to match the type of the cast, don't touch this.
if (isa<AllocaInst>(SrcOp)) {
// See if the bitcast simplifies, if so, don't nuke this GEP yet.
if (Instruction *I = visitBitCast(*BCI)) {
if (I != BCI) {
I->takeName(BCI);
BCI->getParent()->getInstList().insert(BCI->getIterator(), I);
replaceInstUsesWith(*BCI, I);
}
return &GEP;
}
}
if (SrcType->getPointerAddressSpace() != GEP.getAddressSpace())
return new AddrSpaceCastInst(SrcOp, GEPType);
return new BitCastInst(SrcOp, GEPType);
}
// Otherwise, if the offset is non-zero, we need to find out if there is a
// field at Offset in 'A's type. If so, we can pull the cast through the
// GEP.
SmallVector<Value*, 8> NewIndices;
if (FindElementAtOffset(SrcType, Offset.getSExtValue(), NewIndices)) {
Value *NGEP =
GEP.isInBounds()
? Builder.CreateInBoundsGEP(SrcEltType, SrcOp, NewIndices)
: Builder.CreateGEP(SrcEltType, SrcOp, NewIndices);
if (NGEP->getType() == GEPType)
return replaceInstUsesWith(GEP, NGEP);
NGEP->takeName(&GEP);
if (NGEP->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
return new AddrSpaceCastInst(NGEP, GEPType);
return new BitCastInst(NGEP, GEPType);
}
}
}
if (!GEP.isInBounds()) {
unsigned IdxWidth =
DL.getIndexSizeInBits(PtrOp->getType()->getPointerAddressSpace());
APInt BasePtrOffset(IdxWidth, 0);
Value *UnderlyingPtrOp =
PtrOp->stripAndAccumulateInBoundsConstantOffsets(DL,
BasePtrOffset);
if (auto *AI = dyn_cast<AllocaInst>(UnderlyingPtrOp)) {
if (GEP.accumulateConstantOffset(DL, BasePtrOffset) &&
BasePtrOffset.isNonNegative()) {
APInt AllocSize(
IdxWidth,
DL.getTypeAllocSize(AI->getAllocatedType()).getKnownMinSize());
if (BasePtrOffset.ule(AllocSize)) {
return GetElementPtrInst::CreateInBounds(
GEP.getSourceElementType(), PtrOp, makeArrayRef(Ops).slice(1),
GEP.getName());
}
}
}
}
if (Instruction *R = foldSelectGEP(GEP, Builder))
return R;
return nullptr;
}
static bool isNeverEqualToUnescapedAlloc(Value *V, const TargetLibraryInfo *TLI,
Instruction *AI) {
if (isa<ConstantPointerNull>(V))
return true;
if (auto *LI = dyn_cast<LoadInst>(V))
return isa<GlobalVariable>(LI->getPointerOperand());
// Two distinct allocations will never be equal.
// We rely on LookThroughBitCast in isAllocLikeFn being false, since looking
// through bitcasts of V can cause
// the result statement below to be true, even when AI and V (ex:
// i8* ->i32* ->i8* of AI) are the same allocations.
return isAllocLikeFn(V, TLI) && V != AI;
}
static bool isAllocSiteRemovable(Instruction *AI,
SmallVectorImpl<WeakTrackingVH> &Users,
const TargetLibraryInfo *TLI) {
SmallVector<Instruction*, 4> Worklist;
Worklist.push_back(AI);
do {
Instruction *PI = Worklist.pop_back_val();
for (User *U : PI->users()) {
Instruction *I = cast<Instruction>(U);
switch (I->getOpcode()) {
default:
// Give up the moment we see something we can't handle.
return false;
case Instruction::AddrSpaceCast:
case Instruction::BitCast:
case Instruction::GetElementPtr:
Users.emplace_back(I);
Worklist.push_back(I);
continue;
case Instruction::ICmp: {
ICmpInst *ICI = cast<ICmpInst>(I);
// We can fold eq/ne comparisons with null to false/true, respectively.
// We also fold comparisons in some conditions provided the alloc has
// not escaped (see isNeverEqualToUnescapedAlloc).
if (!ICI->isEquality())
return false;
unsigned OtherIndex = (ICI->getOperand(0) == PI) ? 1 : 0;
if (!isNeverEqualToUnescapedAlloc(ICI->getOperand(OtherIndex), TLI, AI))
return false;
Users.emplace_back(I);
continue;
}
case Instruction::Call:
// Ignore no-op and store intrinsics.
if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
switch (II->getIntrinsicID()) {
default:
return false;
case Intrinsic::memmove:
case Intrinsic::memcpy:
case Intrinsic::memset: {
MemIntrinsic *MI = cast<MemIntrinsic>(II);
if (MI->isVolatile() || MI->getRawDest() != PI)
return false;
LLVM_FALLTHROUGH;
}
case Intrinsic::assume:
case Intrinsic::invariant_start:
case Intrinsic::invariant_end:
case Intrinsic::lifetime_start:
case Intrinsic::lifetime_end:
case Intrinsic::objectsize:
Users.emplace_back(I);
continue;
case Intrinsic::launder_invariant_group:
case Intrinsic::strip_invariant_group:
Users.emplace_back(I);
Worklist.push_back(I);
continue;
}
}
if (isFreeCall(I, TLI)) {
Users.emplace_back(I);
continue;
}
if (isReallocLikeFn(I, TLI, true)) {
Users.emplace_back(I);
Worklist.push_back(I);
continue;
}
return false;
case Instruction::Store: {
StoreInst *SI = cast<StoreInst>(I);
if (SI->isVolatile() || SI->getPointerOperand() != PI)
return false;
Users.emplace_back(I);
continue;
}
}
llvm_unreachable("missing a return?");
}
} while (!Worklist.empty());