blob: 723bb4c0d68988b83bf1f0224e490bcc9ae0050f [file] [log] [blame]
//===- InstructionCombining.cpp - Combine multiple instructions -----------===//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See 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/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/LoopInfo.h"
#include "llvm/Analysis/MemoryBuiltins.h"
#include "llvm/Analysis/OptimizationRemarkEmitter.h"
#include "llvm/Analysis/TargetFolder.h"
#include "llvm/Analysis/TargetLibraryInfo.h"
#include "llvm/Analysis/ValueTracking.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/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/InstCombine/InstCombineWorklist.h"
#include "llvm/Transforms/Utils/Local.h"
#include <algorithm>
#include <cassert>
#include <cstdint>
#include <memory>
#include <string>
#include <utility>
using namespace llvm;
using namespace llvm::PatternMatch;
#define DEBUG_TYPE "instcombine"
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");
static cl::opt<bool>
EnableCodeSinking("instcombine-code-sinking", cl::desc("Enable code sinking"),
static cl::opt<bool>
cl::desc("Enable expensive instruction combines"));
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));
Value *InstCombiner::EmitGEPOffset(User *GEP) {
return llvm::EmitGEPOffset(&Builder, DL, GEP);
/// 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 legal type
/// because i1 is a fundamental type in IR, and there are many specialized
/// optimizations for i1 types. Widths of 8, 16 or 32 are equally treated as
/// legal to convert to, in order to open up more combining opportunities.
/// NOTE: this treats i8, i16 and i32 specially, due to them being so common
/// from frontend languages.
bool InstCombiner::shouldChangeType(unsigned FromWidth,
unsigned ToWidth) const {
bool FromLegal = FromWidth == 1 || DL.isLegalInteger(FromWidth);
bool ToLegal = ToWidth == 1 || DL.isLegalInteger(ToWidth);
// Convert to widths of 8, 16 or 32 even if they are not legal types. Only
// shrink types, to prevent infinite loops.
if (ToWidth < FromWidth && (ToWidth == 8 || ToWidth == 16 || ToWidth == 32))
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 InstCombiner::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) {
OverflowingBinaryOperator *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);
(void)BVal->ssub_ov(*CVal, Overflow);
return !Overflow;
/// 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) {
FastMathFlags FMF = I.getFastMathFlags();
/// 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) {
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);
Cast->setOperand(0, BinOp2->getOperand(0));
BinOp1->setOperand(1, FoldedC);
return true;
/// 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 InstCombiner::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)) <
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".
I.setOperand(0, A);
I.setOperand(1, V);
// Conservatively clear the optional flags, since they may not be
// preserved by the reassociation.
if (MaintainNoSignedWrap(I, B, C) &&
(!Op0 || (isa<BinaryOperator>(Op0) && Op0->hasNoSignedWrap()))) {
// Note: this is only valid because SimplifyBinOp doesn't look at
// the operands to Op0.
} else {
Changed = true;
// 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".
I.setOperand(0, V);
I.setOperand(1, C);
// Conservatively clear the optional flags, since they may not be
// preserved by the reassociation.
Changed = true;
if (I.isAssociative() && I.isCommutative()) {
if (simplifyAssocCastAssoc(&I)) {
Changed = true;
// 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".
I.setOperand(0, V);
I.setOperand(1, B);
// Conservatively clear the optional flags, since they may not be
// preserved by the reassociation.
Changed = true;
// 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".
I.setOperand(0, B);
I.setOperand(1, V);
// Conservatively clear the optional flags, since they may not be
// preserved by the reassociation.
Changed = true;
// 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))))) {
BinaryOperator *NewBO = BinaryOperator::Create(Opcode, A, B);
if (isa<FPMathOperator>(NewBO)) {
FastMathFlags Flags = I.getFastMathFlags();
Flags &= Op0->getFastMathFlags();
Flags &= Op1->getFastMathFlags();
InsertNewInstWith(NewBO, I);
I.setOperand(0, NewBO);
I.setOperand(1, ConstantExpr::get(Opcode, C1, C2));
// Conservatively clear the optional flags, since they may not be
// preserved by the reassociation.
Changed = true;
// 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 *InstCombiner::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) {
// Check if we can add NSW flag to SimplifiedInst. If so, set NSW flag.
// TODO: Check for NUW.
if (BinaryOperator *BO = dyn_cast<BinaryOperator>(SimplifiedInst)) {
if (isa<OverflowingBinaryOperator>(SimplifiedInst)) {
bool HasNSW = false;
if (isa<OverflowingBinaryOperator>(&I))
HasNSW = I.hasNoSignedWrap();
if (auto *LOBO = dyn_cast<OverflowingBinaryOperator>(LHS))
HasNSW &= LOBO->hasNoSignedWrap();
if (auto *ROBO = dyn_cast<OverflowingBinaryOperator>(RHS))
HasNSW &= ROBO->hasNoSignedWrap();
// 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 (TopLevelOpcode == Instruction::Add &&
InnerOpcode == Instruction::Mul)
if (match(V, m_APInt(CInt)) && !CInt->isMinSignedValue())
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 *InstCombiner::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'
Value *L = SimplifyBinOp(TopLevelOpcode, A, C, SQ.getWithInstruction(&I));
Value *R = SimplifyBinOp(TopLevelOpcode, B, C, SQ.getWithInstruction(&I));
// Do "A op C" and "B op C" both simplify?
if (L && R) {
// They do! Return "L op' R".
C = Builder.CreateBinOp(InnerOpcode, L, R);
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".
C = Builder.CreateBinOp(TopLevelOpcode, B, C);
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".
C = Builder.CreateBinOp(TopLevelOpcode, A, C);
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'
Value *L = SimplifyBinOp(TopLevelOpcode, A, B, SQ.getWithInstruction(&I));
Value *R = SimplifyBinOp(TopLevelOpcode, A, C, SQ.getWithInstruction(&I));
// Do "A op B" and "A op C" both simplify?
if (L && R) {
// They do! Return "L op' R".
A = Builder.CreateBinOp(InnerOpcode, L, R);
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".
A = Builder.CreateBinOp(TopLevelOpcode, A, C);
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".
A = Builder.CreateBinOp(TopLevelOpcode, A, B);
return A;
return SimplifySelectsFeedingBinaryOp(I, LHS, RHS);
Value *InstCombiner::SimplifySelectsFeedingBinaryOp(BinaryOperator &I,
Value *LHS, Value *RHS) {
Instruction::BinaryOps Opcode = I.getOpcode();
// (op (select (a, b, c)), (select (a, d, e))) -> (select (a, (op b, d), (op
// c, e)))
Value *A, *B, *C, *D, *E;
Value *SI = nullptr;
if (match(LHS, m_Select(m_Value(A), m_Value(B), m_Value(C))) &&
match(RHS, m_Select(m_Specific(A), m_Value(D), m_Value(E)))) {
bool SelectsHaveOneUse = LHS->hasOneUse() && RHS->hasOneUse();
BuilderTy::FastMathFlagGuard Guard(Builder);
if (isa<FPMathOperator>(&I))
Value *V1 = SimplifyBinOp(Opcode, C, E, SQ.getWithInstruction(&I));
Value *V2 = SimplifyBinOp(Opcode, B, D, SQ.getWithInstruction(&I));
if (V1 && V2)
SI = Builder.CreateSelect(A, V2, V1);
else if (V2 && SelectsHaveOneUse)
SI = Builder.CreateSelect(A, V2, Builder.CreateBinOp(Opcode, C, E));
else if (V1 && SelectsHaveOneUse)
SI = Builder.CreateSelect(A, Builder.CreateBinOp(Opcode, B, D), V1);
if (SI)
return SI;
/// Given a 'sub' instruction, return the RHS of the instruction if the LHS is a
/// constant zero (which is the 'negate' form).
Value *InstCombiner::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))
if (!isa<ConstantInt>(Elt))
return nullptr;
return ConstantExpr::getNeg(CV);
return nullptr;
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());
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))
return RI;
Instruction *InstCombiner::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->getNumElements() != DestTy->getNumElements())
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);
if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
(SI->getOperand(2) == Op0 && SI->getOperand(1) == 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, "phitmp");
auto *FPInst = dyn_cast<Instruction>(RI);
if (FPInst && isa<FPMathOperator>(FPInst))
return RI;
Instruction *InstCombiner::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);
if (isa<Constant>(InVal) && !isa<ConstantExpr>(InVal))
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, &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.
if (NonConstBB != nullptr) {
BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
if (!BI || !BI->isUnconditional()) return nullptr;
// Okay, we can do the transformation: create the new PHI node.
PHINode *NewPN = PHINode::Create(I.getType(), PN->getNumIncomingValues());
InsertNewInstBefore(NewPN, *PN);
// If we are going to have to insert a new computation, do so right before the
// predecessor's terminator.
if (NonConstBB)
// 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<ConstantExpr>(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.
InV = Builder.CreateSelect(PN->getIncomingValue(i), TrueVInPred,
FalseVInPred, "phitmp");
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 (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
else if (isa<ICmpInst>(CI))
InV = Builder.CreateICmp(CI->getPredicate(), PN->getIncomingValue(i),
C, "phitmp");
InV = Builder.CreateFCmp(CI->getPredicate(), PN->getIncomingValue(i),
C, "phitmp");
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),
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);
InV = Builder.CreateCast(CI->getOpcode(), PN->getIncomingValue(i),
I.getType(), "phitmp");
NewPN->addIncoming(InV, PN->getIncomingBlock(i));
for (auto UI = PN->user_begin(), E = PN->user_end(); UI != E;) {
Instruction *User = cast<Instruction>(*UI++);
if (User == &I) continue;
replaceInstUsesWith(*User, NewPN);
return replaceInstUsesWith(I, NewPN);
Instruction *InstCombiner::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 *InstCombiner::FindElementAtOffset(PointerType *PtrTy, int64_t Offset,
SmallVectorImpl<Value *> &NewIndices) {
Type *Ty = PtrTy->getElementType();
if (!Ty->isSized())
return nullptr;
// Start with the index over the outer type. Note that the type size
// might be zero (even if the offset isn't zero) if the indexed type
// is something like [0 x {int, int}]
Type *IndexTy = DL.getIndexType(PtrTy);
int64_t FirstIdx = 0;
if (int64_t TySize = DL.getTypeAllocSize(Ty)) {
FirstIdx = Offset/TySize;
Offset -= FirstIdx*TySize;
// Handle hosts where % returns negative instead of values [0..TySize).
if (Offset < 0) {
Offset += TySize;
assert(Offset >= 0);
assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
NewIndices.push_back(ConstantInt::get(IndexTy, FirstIdx));
// Index into the types. If we fail, set OrigBase to null.
while (Offset) {
// Indexing into tail padding between struct/array elements.
if (uint64_t(Offset * 8) >= DL.getTypeSizeInBits(Ty))
return nullptr;
if (StructType *STy = dyn_cast<StructType>(Ty)) {
const StructLayout *SL = DL.getStructLayout(STy);
assert(Offset < (int64_t)SL->getSizeInBytes() &&
"Offset must stay within the indexed type");
unsigned Elt = SL->getElementContainingOffset(Offset);
Offset -= SL->getElementOffset(Elt);
Ty = STy->getElementType(Elt);
} else if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
uint64_t EltSize = DL.getTypeAllocSize(AT->getElementType());
assert(EltSize && "Cannot index into a zero-sized array");
Offset %= EltSize;
Ty = AT->getElementType();
} else {
// Otherwise, we can't index into the middle of this atomic type, bail.
return nullptr;
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() &&
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 *InstCombiner::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;
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;
// Otherwise drill down into the constant.
if (!Op->hasOneUse())
return nullptr;
Parent = std::make_pair(BO, 1);
// 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);
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))->
// Op = LHS << Amt.
if (Amt == logScale) {
// Multiplication by exactly the scale, replace the multiplication
// by its left-hand side in the parent.
Op = LHS;
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);
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;
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);
// 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?");
Parent.first->setOperand(Parent.second, Op);
// 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) {
} 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 *InstCombiner::foldVectorBinop(BinaryOperator &Inst) {
if (!Inst.getType()->isVectorTy()) return nullptr;
BinaryOperator::BinaryOps Opcode = Inst.getOpcode();
unsigned NumElts = cast<VectorType>(Inst.getType())->getNumElements();
Value *LHS = Inst.getOperand(0), *RHS = Inst.getOperand(1);
assert(cast<VectorType>(LHS->getType())->getNumElements() == NumElts);
assert(cast<VectorType>(RHS->getType())->getNumElements() == NumElts);
// 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;
Constant *Mask;
if (match(LHS, m_ShuffleVector(m_Value(L0), m_Value(L1), m_Constant(Mask))) &&
match(RHS, m_ShuffleVector(m_Value(R0), m_Value(R1), m_Specific(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))
Value *NewBO1 = Builder.CreateBinOp(Opcode, L1, R1);
if (auto *BO = dyn_cast<BinaryOperator>(NewBO1))
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, Constant *M) {
Value *XY = Builder.CreateBinOp(Opcode, X, Y);
if (auto *BO = dyn_cast<BinaryOperator>(XY))
return new ShuffleVectorInst(XY, UndefValue::get(XY->getType()), 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_ShuffleVector(m_Value(V1), m_Undef(), m_Constant(Mask))) &&
match(RHS, m_ShuffleVector(m_Value(V2), m_Undef(), m_Specific(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 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;
if (match(&Inst, m_c_BinOp(
m_OneUse(m_ShuffleVector(m_Value(V1), m_Undef(), m_Constant(Mask))),
m_Constant(C))) &&
V1->getType()->getVectorNumElements() <= NumElts) {
assert(Inst.getType()->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);
SmallVector<int, 16> ShMask;
ShuffleVectorInst::getShuffleMask(Mask, ShMask);
unsigned SrcVecNumElts = V1->getType()->getVectorNumElements();
UndefValue *UndefScalar = UndefValue::get(C->getType()->getScalarType());
SmallVector<Constant *, 16> NewVecC(SrcVecNumElts, UndefScalar);
bool MayChange = true;
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;
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.
// 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) {
Constant *MaybeUndef =
ConstOp1 ? ConstantExpr::get(Opcode, UndefScalar, CElt)
: ConstantExpr::get(Opcode, CElt, UndefScalar);
if (!isa<UndefValue>(MaybeUndef)) {
MayChange = false;
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);
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 *InstCombiner::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)
return CastInst::Create(CastOpc, NarrowBO, BO.getType());
Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end());
Type *GEPType = GEP.getType();
Type *GEPEltType = GEP.getSourceElementType();
if (Value *V = SimplifyGEPInst(GEPEltType, Ops, SQ.getWithInstruction(&GEP)))
return replaceInstUsesWith(GEP, V);
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 =
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())
Type *IndexTy = (*I)->getType();
Type *NewIndexType =
? VectorType::get(NewScalarIndexTy, IndexTy->getVectorNumElements())
: 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) == 0)
if (!isa<Constant>(*I) || !cast<Constant>(*I)->isNullValue()) {
*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 && 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 if (auto *CT = dyn_cast<CompositeType>(CurTy)) {
CurTy = CT->getTypeAtIndex(Op1->getOperand(J));
} else {
CurTy = nullptr;
// 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.
GEP.getParent()->getFirstInsertionPt(), NewGEP);
} 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);
NewPN = Builder.CreatePHI(Op1->getOperand(DI)->getType(),
for (auto &I : PN->operands())
NewGEP->setOperand(DI, NewPN);
GEP.getParent()->getFirstInsertionPt(), NewGEP);
NewGEP->setOperand(DI, NewPN);
GEP.setOperand(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;
// Try to reassociate loop invariant GEP chains to enable LICM.
if (LI && Src->getNumOperands() == 2 && GEP.getNumOperands() == 2 &&
Src->hasOneUse()) {
if (Loop *L = LI->getLoopFor(GEP.getParent())) {
Value *GO1 = GEP.getOperand(1);
Value *SO1 = Src->getOperand(1);
// 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
auto *NewSrc = cast<GetElementPtrInst>(
Builder.CreateGEP(GEPEltType, SO0, GO1, Src->getName()));
auto *NewGEP = GetElementPtrInst::Create(GEPEltType, NewSrc, {SO1});
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.setOperand(0, Src->getOperand(0));
GEP.setOperand(1, Sum);
return &GEP;
Indices.append(Src->op_begin()+1, Src->op_end()-1);
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 GEP.isInBounds() && Src->isInBounds()
? GetElementPtrInst::CreateInBounds(
Src->getSourceElementType(), Src->getOperand(0), Indices,
: GetElementPtrInst::Create(Src->getSourceElementType(),
Src->getOperand(0), Indices,
if (GEP.getNumIndices() == 1) {
unsigned AS = GEP.getPointerAddressSpace();
if (GEP.getOperand(1)->getType()->getScalarSizeInBits() ==
DL.getIndexSizeInBits(AS)) {
uint64_t TyAllocSize = DL.getTypeAllocSize(GEPEltType);
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;
if (Matched) {
// Canonicalize (gep i8* X, -(ptrtoint Y))
// to (inttoptr (sub (ptrtoint X), (ptrtoint Y)))
// The GEP pattern is emitted by the SCEV expander for certain kinds of
// pointer arithmetic.
if (match(V, m_Neg(m_PtrToInt(m_Value())))) {
Operator *Index = cast<Operator>(V);
Value *PtrToInt = Builder.CreatePtrToInt(PtrOp, Index->getType());
Value *NewSub = Builder.CreateSub(PtrToInt, Index->getOperand(1));
return CastInst::Create(Instruction::IntToPtr, NewSub, GEPType);
// Canonicalize (gep i8* X, (ptrtoint Y)-(ptrtoint X))
// to (bitcast Y)
Value *Y;
if (match(V, m_Sub(m_PtrToInt(m_Value(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(GEP.idx_begin()+1, GEP.idx_end());
GetElementPtrInst *Res = GetElementPtrInst::Create(
StrippedPtrEltTy, StrippedPtr, Idx, GEP.getName());
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.setOperand(0, StrippedPtr);
return &GEP;
// 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.idx_begin(), GEP.idx_end());
Value *NewGEP =
? Builder.CreateInBoundsGEP(StrippedPtrEltTy, StrippedPtr,
Idx, GEP.getName())
: Builder.CreateGEP(StrippedPtrEltTy, StrippedPtr, Idx,
return new AddrSpaceCastInst(NewGEP, GEPType);
} else if (GEP.getNumOperands() == 2) {
// 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 =
? Builder.CreateInBoundsGEP(StrippedPtrEltTy, StrippedPtr, Idx,
: Builder.CreateGEP(StrippedPtrEltTy, StrippedPtr, Idx,
// 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);
uint64_t SrcSize = DL.getTypeAllocSize(StrippedPtrEltTy);
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,
// The NewGEP must be pointer typed, so must the old one -> BitCast
return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
// 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);
uint64_t ArrayEltSize =
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,
// The NewGEP must be pointer typed, so must the old one -> BitCast
return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
// 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) {
return ArrTy->getArrayElementType() == VecTy->getVectorElementType() &&
ArrTy->getArrayNumElements() == VecTy->getVectorNumElements();
if (GEP.getNumOperands() == 3 &&
((GEPEltType->isArrayTy() && SrcEltType->isVectorTy() &&
areMatchingArrayAndVecTypes(GEPEltType, SrcEltType)) ||
(GEPEltType->isVectorTy() && SrcEltType->isArrayTy() &&
areMatchingArrayAndVecTypes(SrcEltType, GEPEltType)))) {
// 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 =
? Builder.CreateInBoundsGEP(SrcEltType, SrcOp, {Ops[1], Ops[2]})
: Builder.CreateGEP(SrcEltType, SrcOp, {Ops[1], Ops[2]});
// 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 (!isa<BitCastInst>(SrcOp) && GEP.accumulateConstantOffset(DL, Offset)) {
// 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) || isAllocationFn(SrcOp, &TLI)) {
// See if the bitcast simplifies, if so, don't nuke this GEP yet.
if (Instruction *I = visitBitCast(*BCI)) {
if (I != 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 =
? Builder.CreateInBoundsGEP(SrcEltType, SrcOp, NewIndices)
: Builder.CreateGEP(SrcEltType, SrcOp, NewIndices);
if (NGEP->getType() == GEPType)
return replaceInstUsesWith(GEP, NGEP);
if (NGEP->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
return new AddrSpaceCastInst(NGEP, GEPType);
return new BitCastInst(NGEP, GEPType);
if (!GEP.isInBounds()) {
unsigned IdxWidth =
APInt BasePtrOffset(IdxWidth, 0);
Value *UnderlyingPtrOp =
if (auto *AI = dyn_cast<AllocaInst>(UnderlyingPtrOp)) {
if (GEP.accumulateConstantOffset(DL, BasePtrOffset) &&
BasePtrOffset.isNonNegative()) {
APInt AllocSize(IdxWidth, DL.getTypeAllocSize(AI->getAllocatedType()));
if (BasePtrOffset.ule(AllocSize)) {
return GetElementPtrInst::CreateInBounds(
GEP.getSourceElementType(), PtrOp, makeArrayRef(Ops).slice(1),
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;
do {
Instruction *PI = Worklist.pop_back_val();
for (User *U : PI->users()) {
Instruction *I = cast<Instruction>(U);
switch (I->getOpcode()) {
// Give up the moment we see something we can't handle.
return false;
case Instruction::AddrSpaceCast:
case Instruction::BitCast:
case Instruction::GetElementPtr:
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;
case Instruction::Call:
// Ignore no-op and store intrinsics.
if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
switch (II->getIntrinsicID()) {
return false;
case Intrinsic::memmove:
case Intrinsic::memcpy:
case Intrinsic::memset: {
MemIntrinsic *MI = cast<MemIntrinsic>(II);
if (MI->isVolatile() || MI->getRawDest() != PI)
return false;
case Intrinsic::invariant_start:
case Intrinsic::invariant_end:
case Intrinsic::lifetime_start:
case Intrinsic::lifetime_end:
case Intrinsic::objectsize:
if (isFreeCall(I, TLI)) {
return false;
case Instruction::Store: {
StoreInst *SI = cast<StoreInst>(I);
if (SI->isVolatile() || SI->getPointerOperand() != PI)
return false;
llvm_unreachable("missing a return?");
} while (!Worklist.empty());
return true;
Instruction *InstCombiner::visitAllocSite(Instruction &MI) {
// If we have a malloc call which is only used in any amount of comparisons to
// null and free calls, delete the calls and replace the comparisons with true
// or false as appropriate.
// This is based on the principle that we can substitute our own allocation
// function (which will never return null) rather than knowledge of the
// specific function being called. In some sense this can change the permitted
// outputs of a program (when we convert a malloc to an alloca, the fact that
// the allocation is now on the stack is potentially visible, for example),
// but we believe in a permissible manner.
SmallVector<WeakTrackingVH, 64> Users;
// If we are removing an alloca with a dbg.declare, insert dbg.value calls
// before each store.
TinyPtrVector<DbgVariableIntrinsic *> DIIs;
std::unique_ptr<DIBuilder> DIB;
if (isa<AllocaInst>(MI)) {
DIIs = FindDbgAddrUses(&MI);
DIB.reset(new DIBuilder(*MI.getModule(), /*AllowUnresolved=*/false));
if (isAllocSiteRemovable(&MI, Users, &TLI)) {
for (unsigned i = 0, e = Users.size(); i != e; ++i) {
// Lowering all @llvm.objectsize calls first because they may
// use a bitcast/GEP of the alloca we are removing.
if (!Users[i])
Instruction *I = cast<Instruction>(&*Users[i]);
if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
if (II->getIntrinsicID() == Intrinsic::objectsize) {
Value *Result =
lowerObjectSizeCall(II, DL, &TLI, /*MustSucceed=*/true);
replaceInstUsesWith(*I, Result);
Users[i] = nullptr; // Skip examining in the next loop.
for (unsigned i = 0, e = Users.size(); i != e; ++i) {
if (!Users[i])
Instruction *I = cast<Instruction>(&*Users[i]);
if (ICmpInst *C = dyn_cast<ICmpInst>(I)) {
} else if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I) ||
isa<AddrSpaceCastInst>(I)) {
replaceInstUsesWith(*I, UndefValue::get(I->getType()));
} else if (auto *SI = dyn_cast<StoreInst>(I)) {
for (auto *DII : DIIs)
ConvertDebugDeclareToDebugValue(DII, SI, *DIB);
if (InvokeInst *II = dyn_cast<InvokeInst>(&MI)) {
// Replace invoke with a NOP intrinsic to maintain the original CFG
Module *M = II->getModule();
Function *F = Intrinsic::getDeclaration(M, Intrinsic::donothing);
InvokeInst::Create(F, II->getNormalDest(), II->getUnwindDest(),
None, "", II->getParent());
for (auto *DII : DIIs)
return eraseInstFromFunction(MI);
return nullptr;
/// Move the call to free before a NULL test.
/// Check if this free is accessed after its argument has been test
/// against NULL (property 0).
/// If yes, it is legal to move this call in its predecessor block.
/// The move is performed only if the block containing the call to free
/// will be removed, i.e.:
/// 1. it has only one predecessor P, and P has two successors
/// 2. it contains the call, noops, and an unconditional branch
/// 3. its successor is the same as its predecessor's successor
/// The profitability is out-of concern here and this function should
/// be called only if the caller knows this transformation would be
/// profitable (e.g., for code size).
static Instruction *tryToMoveFreeBeforeNullTest(CallInst &FI,
const DataLayout &DL) {
Value *Op = FI.getArgOperand(0);
BasicBlock *FreeInstrBB = FI.getParent();
BasicBlock *PredBB = FreeInstrBB->getSinglePredecessor();
// Validate part of constraint #1: Only one predecessor
// FIXME: We can extend the number of predecessor, but in that case, we
// would duplicate the call to free in each predecessor and it may
// not be profitable even for code size.
if (!PredBB)
return nullptr;
// Validate constraint #2: Does this block contains only the call to
// free, noops, and an unconditional branch?
BasicBlock *SuccBB;
Instruction *FreeInstrBBTerminator = FreeInstrBB->getTerminator();
if (!match(FreeInstrBBTerminator, m_UnconditionalBr(SuccBB)))
return nullptr;
// If there are only 2 instructions in the block, at this point,
// this is the call to free and unconditional.
// If there are more than 2 instructions, check that they are noops
// i.e., they won't hurt the performance of the generated code.
if (FreeInstrBB->size() != 2) {
for (const Instruction &Inst : *FreeInstrBB) {
if (&Inst == &FI || &Inst == FreeInstrBBTerminator)
auto *Cast = dyn_cast<CastInst>(&Inst);
if (!Cast || !Cast->isNoopCast(DL))
return nullptr;
// Validate the rest of constraint #1 by matching on the pred branch.
Instruction *TI = PredBB->getTerminator();
BasicBlock *TrueBB, *FalseBB;
ICmpInst::Predicate Pred;
if (!match(TI, m_Br(m_ICmp(Pred,
TrueBB, FalseBB)))
return nullptr;
if (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE)
return nullptr;
// Validate constraint #3: Ensure the null case just falls through.
if (SuccBB != (Pred == ICmpInst::ICMP_EQ ? TrueBB : FalseBB))
return nullptr;
assert(FreeInstrBB == (Pred == ICmpInst::ICMP_EQ ? FalseBB : TrueBB) &&
"Broken CFG: missing edge from predecessor to successor");
// At this point, we know that everything in FreeInstrBB can be moved
// before TI.