#include "llvm/Transforms/Scalar.h"
#include "InstCombine.h"
#include "llvm/IntrinsicInst.h"
-#include "llvm/LLVMContext.h"
-#include "llvm/DerivedTypes.h"
-#include "llvm/GlobalVariable.h"
-#include "llvm/Operator.h"
#include "llvm/Analysis/ConstantFolding.h"
#include "llvm/Analysis/InstructionSimplify.h"
#include "llvm/Analysis/MemoryBuiltins.h"
#include "llvm/Target/TargetData.h"
-#include "llvm/Transforms/Utils/BasicBlockUtils.h"
+#include "llvm/Target/TargetLibraryInfo.h"
#include "llvm/Transforms/Utils/Local.h"
+#include "llvm/Support/CFG.h"
#include "llvm/Support/Debug.h"
-#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/GetElementPtrTypeIterator.h"
-#include "llvm/Support/MathExtras.h"
#include "llvm/Support/PatternMatch.h"
+#include "llvm/Support/ValueHandle.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/Statistic.h"
-#include "llvm/ADT/STLExtras.h"
+#include "llvm/ADT/StringSwitch.h"
+#include "llvm-c/Initialization.h"
#include <algorithm>
#include <climits>
using namespace llvm;
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");
+// Initialization Routines
+void llvm::initializeInstCombine(PassRegistry &Registry) {
+ initializeInstCombinerPass(Registry);
+}
+
+void LLVMInitializeInstCombine(LLVMPassRegistryRef R) {
+ initializeInstCombine(*unwrap(R));
+}
char InstCombiner::ID = 0;
-static RegisterPass<InstCombiner>
-X("instcombine", "Combine redundant instructions");
+INITIALIZE_PASS_BEGIN(InstCombiner, "instcombine",
+ "Combine redundant instructions", false, false)
+INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
+INITIALIZE_PASS_END(InstCombiner, "instcombine",
+ "Combine redundant instructions", false, false)
void InstCombiner::getAnalysisUsage(AnalysisUsage &AU) const {
- AU.addPreservedID(LCSSAID);
AU.setPreservesCFG();
+ AU.addRequired<TargetLibraryInfo>();
}
/// ShouldChangeType - 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 for example, or from a smaller to a larger illegal type.
-bool InstCombiner::ShouldChangeType(const Type *From, const Type *To) const {
- assert(isa<IntegerType>(From) && isa<IntegerType>(To));
+bool InstCombiner::ShouldChangeType(Type *From, Type *To) const {
+ assert(From->isIntegerTy() && To->isIntegerTy());
// If we don't have TD, we don't know if the source/dest are legal.
if (!TD) return false;
return true;
}
+// 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;
+ }
+
+ ConstantInt *CB = dyn_cast<ConstantInt>(B);
+ ConstantInt *CC = dyn_cast<ConstantInt>(C);
+
+ if (!CB || !CC) {
+ return false;
+ }
+
+ const APInt &BVal = CB->getValue();
+ const APInt &CVal = CC->getValue();
+ bool Overflow = false;
+
+ if (Opcode == Instruction::Add) {
+ BVal.sadd_ov(CVal, Overflow);
+ } else {
+ BVal.ssub_ov(CVal, Overflow);
+ }
+
+ return !Overflow;
+}
-// SimplifyCommutative - This performs a few simplifications for commutative
-// operators:
+/// SimplifyAssociativeOrCommutative - This performs a few simplifications for
+/// operators which 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.
//
-// 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
-// 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
+// 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:
//
-bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
+// 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;
- if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
- Changed = !I.swapOperands();
- if (!I.isAssociative()) return Changed;
-
- Instruction::BinaryOps Opcode = I.getOpcode();
- if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
- if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
- if (isa<Constant>(I.getOperand(1))) {
- Constant *Folded = ConstantExpr::get(I.getOpcode(),
- cast<Constant>(I.getOperand(1)),
- cast<Constant>(Op->getOperand(1)));
- I.setOperand(0, Op->getOperand(0));
- I.setOperand(1, Folded);
- return true;
+ 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, TD)) {
+ // 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.
+ I.clearSubclassOptionalData();
+ I.setHasNoSignedWrap(true);
+ } else {
+ I.clearSubclassOptionalData();
+ }
+
+ Changed = true;
+ ++NumReassoc;
+ continue;
+ }
}
-
- if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1)))
- if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
- Op->hasOneUse() && Op1->hasOneUse()) {
- Constant *C1 = cast<Constant>(Op->getOperand(1));
- Constant *C2 = cast<Constant>(Op1->getOperand(1));
-
- // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
- Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
- Instruction *New = BinaryOperator::Create(Opcode, Op->getOperand(0),
- Op1->getOperand(0),
- Op1->getName(), &I);
- Worklist.Add(New);
- I.setOperand(0, New);
- I.setOperand(1, Folded);
- return 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, TD)) {
+ // 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.
+ I.clearSubclassOptionalData();
+ Changed = true;
+ ++NumReassoc;
+ continue;
}
+ }
}
- return Changed;
+
+ if (I.isAssociative() && I.isCommutative()) {
+ // 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, TD)) {
+ // 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.
+ I.clearSubclassOptionalData();
+ 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, TD)) {
+ // 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.
+ I.clearSubclassOptionalData();
+ 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.
+ if (Op0 && Op1 &&
+ Op0->getOpcode() == Opcode && Op1->getOpcode() == Opcode &&
+ isa<Constant>(Op0->getOperand(1)) &&
+ isa<Constant>(Op1->getOperand(1)) &&
+ Op0->hasOneUse() && Op1->hasOneUse()) {
+ Value *A = Op0->getOperand(0);
+ Constant *C1 = cast<Constant>(Op0->getOperand(1));
+ Value *B = Op1->getOperand(0);
+ Constant *C2 = cast<Constant>(Op1->getOperand(1));
+
+ Constant *Folded = ConstantExpr::get(Opcode, C1, C2);
+ BinaryOperator *New = BinaryOperator::Create(Opcode, A, B);
+ InsertNewInstWith(New, I);
+ New->takeName(Op1);
+ I.setOperand(0, New);
+ I.setOperand(1, Folded);
+ // Conservatively clear the optional flags, since they may not be
+ // preserved by the reassociation.
+ I.clearSubclassOptionalData();
+
+ Changed = true;
+ continue;
+ }
+ }
+
+ // No further simplifications.
+ return Changed;
+ } while (1);
+}
+
+/// LeftDistributesOverRight - 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) {
+ switch (LOp) {
+ default:
+ return false;
+
+ case Instruction::And:
+ // And distributes over Or and Xor.
+ switch (ROp) {
+ default:
+ return false;
+ case Instruction::Or:
+ case Instruction::Xor:
+ return true;
+ }
+
+ case Instruction::Mul:
+ // Multiplication distributes over addition and subtraction.
+ switch (ROp) {
+ default:
+ return false;
+ case Instruction::Add:
+ case Instruction::Sub:
+ return true;
+ }
+
+ case Instruction::Or:
+ // Or distributes over And.
+ switch (ROp) {
+ default:
+ return false;
+ case Instruction::And:
+ return true;
+ }
+ }
+}
+
+/// RightDistributesOverLeft - 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);
+ // 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.
+ return false;
+}
+
+/// SimplifyUsingDistributiveLaws - 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(); // op
+
+ // Factorization.
+ if (Op0 && Op1 && Op0->getOpcode() == Op1->getOpcode()) {
+ // The instruction has the form "(A op' B) op (C op' D)". Try to factorize
+ // a common term.
+ Value *A = Op0->getOperand(0), *B = Op0->getOperand(1);
+ Value *C = Op1->getOperand(0), *D = Op1->getOperand(1);
+ Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op'
+
+ // 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.
+ Value *V = SimplifyBinOp(TopLevelOpcode, B, D, TD);
+ // 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 && Op0->hasOneUse() && Op1->hasOneUse())
+ V = Builder->CreateBinOp(TopLevelOpcode, B, D, Op1->getName());
+ if (V) {
+ ++NumFactor;
+ V = Builder->CreateBinOp(InnerOpcode, A, V);
+ V->takeName(&I);
+ return V;
+ }
+ }
+
+ // Does "(X op Y) op' Z" always equal "(X op' Z) op (Y op' Z)"?
+ if (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.
+ Value *V = SimplifyBinOp(TopLevelOpcode, A, C, TD);
+ // 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 && Op0->hasOneUse() && Op1->hasOneUse())
+ V = Builder->CreateBinOp(TopLevelOpcode, A, C, Op0->getName());
+ if (V) {
+ ++NumFactor;
+ V = Builder->CreateBinOp(InnerOpcode, V, B);
+ V->takeName(&I);
+ 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'
+
+ // Do "A op C" and "B op C" both simplify?
+ if (Value *L = SimplifyBinOp(TopLevelOpcode, A, C, TD))
+ if (Value *R = SimplifyBinOp(TopLevelOpcode, B, C, TD)) {
+ // They do! Return "L op' R".
+ ++NumExpand;
+ // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
+ if ((L == A && R == B) ||
+ (Instruction::isCommutative(InnerOpcode) && L == B && R == A))
+ return Op0;
+ // Otherwise return "L op' R" if it simplifies.
+ if (Value *V = SimplifyBinOp(InnerOpcode, L, R, TD))
+ return V;
+ // Otherwise, create a new instruction.
+ C = Builder->CreateBinOp(InnerOpcode, L, R);
+ 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'
+
+ // Do "A op B" and "A op C" both simplify?
+ if (Value *L = SimplifyBinOp(TopLevelOpcode, A, B, TD))
+ if (Value *R = SimplifyBinOp(TopLevelOpcode, A, C, TD)) {
+ // They do! Return "L op' R".
+ ++NumExpand;
+ // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
+ if ((L == B && R == C) ||
+ (Instruction::isCommutative(InnerOpcode) && L == C && R == B))
+ return Op1;
+ // Otherwise return "L op' R" if it simplifies.
+ if (Value *V = SimplifyBinOp(InnerOpcode, L, R, TD))
+ return V;
+ // Otherwise, create a new instruction.
+ A = Builder->CreateBinOp(InnerOpcode, L, R);
+ A->takeName(&I);
+ return A;
+ }
+ }
+
+ return 0;
}
// dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
return ConstantExpr::getNeg(C);
if (ConstantVector *C = dyn_cast<ConstantVector>(V))
- if (C->getType()->getElementType()->isInteger())
+ if (C->getType()->getElementType()->isIntegerTy())
return ConstantExpr::getNeg(C);
return 0;
return ConstantExpr::getFNeg(C);
if (ConstantVector *C = dyn_cast<ConstantVector>(V))
- if (C->getType()->getElementType()->isFloatingPoint())
+ if (C->getType()->getElementType()->isFloatingPointTy())
return ConstantExpr::getFNeg(C);
return 0;
static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
InstCombiner *IC) {
- if (CastInst *CI = dyn_cast<CastInst>(&I))
+ if (CastInst *CI = dyn_cast<CastInst>(&I)) {
return IC->Builder->CreateCast(CI->getOpcode(), SO, I.getType());
+ }
// Figure out if the constant is the left or the right argument.
bool ConstIsRHS = isa<Constant>(I.getOperand(1));
if (isa<Constant>(TV) || isa<Constant>(FV)) {
// Bool selects with constant operands can be folded to logical ops.
- if (SI->getType() == Type::getInt1Ty(SI->getContext())) return 0;
-
+ if (SI->getType()->isIntegerTy(1)) return 0;
+
+ // If it's a bitcast involving vectors, make sure it has the same number of
+ // elements on both sides.
+ if (BitCastInst *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 == NULL) != (DestTy == NULL)) return 0;
+ // If vectors, verify that they have the same number of elements.
+ if (SrcTy && SrcTy->getNumElements() != DestTy->getNumElements())
+ return 0;
+ }
+
Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, this);
Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, this);
- return SelectInst::Create(SI->getCondition(), SelectTrueVal,
- SelectFalseVal);
+ return SelectInst::Create(SI->getCondition(),
+ SelectTrueVal, SelectFalseVal);
}
return 0;
}
/// has a PHI node as operand #0, see if we can fold the instruction into the
/// PHI (which is only possible if all operands to the PHI are constants).
///
-/// If AllowAggressive is true, FoldOpIntoPhi will allow certain transforms
-/// that would normally be unprofitable because they strongly encourage jump
-/// threading.
-Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I,
- bool AllowAggressive) {
- AllowAggressive = false;
+Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
PHINode *PN = cast<PHINode>(I.getOperand(0));
unsigned NumPHIValues = PN->getNumIncomingValues();
- if (NumPHIValues == 0 ||
- // We normally only transform phis with a single use, unless we're trying
- // hard to make jump threading happen.
- (!PN->hasOneUse() && !AllowAggressive))
+ if (NumPHIValues == 0)
return 0;
+ // 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 (Value::use_iterator UI = PN->use_begin(), E = PN->use_end();
+ UI != E; ++UI) {
+ Instruction *User = cast<Instruction>(*UI);
+ if (User != &I && !I.isIdenticalTo(User))
+ return 0;
+ }
+ // 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,
// bail out. We don't do arbitrary constant expressions here because moving
// their computation can be expensive without a cost model.
BasicBlock *NonConstBB = 0;
- for (unsigned i = 0; i != NumPHIValues; ++i)
- if (!isa<Constant>(PN->getIncomingValue(i)) ||
- isa<ConstantExpr>(PN->getIncomingValue(i))) {
- if (NonConstBB) return 0; // More than one non-const value.
- if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
- NonConstBB = PN->getIncomingBlock(i);
-
- // If the incoming non-constant value is in I's block, we have an infinite
- // loop.
- if (NonConstBB == I.getParent())
+ for (unsigned i = 0; i != NumPHIValues; ++i) {
+ Value *InVal = PN->getIncomingValue(i);
+ if (isa<Constant>(InVal) && !isa<ConstantExpr>(InVal))
+ continue;
+
+ if (isa<PHINode>(InVal)) return 0; // Itself a phi.
+ if (NonConstBB) return 0; // 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 (InvokeInst *II = dyn_cast<InvokeInst>(InVal))
+ if (II->getParent() == NonConstBB)
return 0;
- }
+
+ // 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 (NonConstBB == I.getParent())
+ return 0;
+ }
// 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 one some other paths (e.g. inside a loop). Only
// do this if the pred block is unconditionally branching into the phi block.
- if (NonConstBB != 0 && !AllowAggressive) {
+ if (NonConstBB != 0) {
BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
if (!BI || !BI->isUnconditional()) return 0;
}
// Okay, we can do the transformation: create the new PHI node.
- PHINode *NewPN = PHINode::Create(I.getType(), "");
- NewPN->reserveOperandSpace(PN->getNumOperands()/2);
+ 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
+ // predecessors 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,
Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB);
Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB);
Value *InV = 0;
- if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
+ if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
InV = InC->isNullValue() ? FalseVInPred : TrueVInPred;
- } else {
- assert(PN->getIncomingBlock(i) == NonConstBB);
- InV = SelectInst::Create(PN->getIncomingValue(i), TrueVInPred,
- FalseVInPred,
- "phitmp", NonConstBB->getTerminator());
- Worklist.Add(cast<Instruction>(InV));
- }
+ else
+ 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 = 0;
+ 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");
+ else
+ InV = Builder->CreateFCmp(CI->getPredicate(), PN->getIncomingValue(i),
+ C, "phitmp");
+ NewPN->addIncoming(InV, PN->getIncomingBlock(i));
+ }
} else if (I.getNumOperands() == 2) {
Constant *C = cast<Constant>(I.getOperand(1));
for (unsigned i = 0; i != NumPHIValues; ++i) {
Value *InV = 0;
- if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
- if (CmpInst *CI = dyn_cast<CmpInst>(&I))
- InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
- else
- InV = ConstantExpr::get(I.getOpcode(), InC, C);
- } else {
- assert(PN->getIncomingBlock(i) == NonConstBB);
- if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
- InV = BinaryOperator::Create(BO->getOpcode(),
- PN->getIncomingValue(i), C, "phitmp",
- NonConstBB->getTerminator());
- else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
- InV = CmpInst::Create(CI->getOpcode(),
- CI->getPredicate(),
- PN->getIncomingValue(i), C, "phitmp",
- NonConstBB->getTerminator());
- else
- llvm_unreachable("Unknown binop!");
-
- Worklist.Add(cast<Instruction>(InV));
- }
+ if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
+ InV = ConstantExpr::get(I.getOpcode(), InC, C);
+ else
+ InV = Builder->CreateBinOp(cast<BinaryOperator>(I).getOpcode(),
+ PN->getIncomingValue(i), C, "phitmp");
NewPN->addIncoming(InV, PN->getIncomingBlock(i));
}
} else {
CastInst *CI = cast<CastInst>(&I);
- const Type *RetTy = CI->getType();
+ Type *RetTy = CI->getType();
for (unsigned i = 0; i != NumPHIValues; ++i) {
Value *InV;
- if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
+ if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
- } else {
- assert(PN->getIncomingBlock(i) == NonConstBB);
- InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
- I.getType(), "phitmp",
- NonConstBB->getTerminator());
- Worklist.Add(cast<Instruction>(InV));
- }
+ else
+ InV = Builder->CreateCast(CI->getOpcode(),
+ PN->getIncomingValue(i), I.getType(), "phitmp");
NewPN->addIncoming(InV, PN->getIncomingBlock(i));
}
}
+
+ for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end();
+ UI != E; ) {
+ Instruction *User = cast<Instruction>(*UI++);
+ if (User == &I) continue;
+ ReplaceInstUsesWith(*User, NewPN);
+ EraseInstFromFunction(*User);
+ }
return ReplaceInstUsesWith(I, NewPN);
}
/// or not there is a sequence of GEP indices into the type that will land us at
/// the specified offset. If so, fill them into NewIndices and return the
/// resultant element type, otherwise return null.
-const Type *InstCombiner::FindElementAtOffset(const Type *Ty, int64_t Offset,
+Type *InstCombiner::FindElementAtOffset(Type *Ty, int64_t Offset,
SmallVectorImpl<Value*> &NewIndices) {
if (!TD) return 0;
if (!Ty->isSized()) return 0;
// 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}]
- const Type *IntPtrTy = TD->getIntPtrType(Ty->getContext());
+ Type *IntPtrTy = TD->getIntPtrType(Ty->getContext());
int64_t FirstIdx = 0;
if (int64_t TySize = TD->getTypeAllocSize(Ty)) {
FirstIdx = Offset/TySize;
if (uint64_t(Offset*8) >= TD->getTypeSizeInBits(Ty))
return 0;
- if (const StructType *STy = dyn_cast<StructType>(Ty)) {
+ if (StructType *STy = dyn_cast<StructType>(Ty)) {
const StructLayout *SL = TD->getStructLayout(STy);
assert(Offset < (int64_t)SL->getSizeInBytes() &&
"Offset must stay within the indexed type");
Offset -= SL->getElementOffset(Elt);
Ty = STy->getElementType(Elt);
- } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
+ } else if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
uint64_t EltSize = TD->getTypeAllocSize(AT->getElementType());
assert(EltSize && "Cannot index into a zero-sized array");
NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
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;
+}
Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end());
- if (Value *V = SimplifyGEPInst(&Ops[0], Ops.size(), TD))
+ if (Value *V = SimplifyGEPInst(Ops, TD))
return ReplaceInstUsesWith(GEP, V);
Value *PtrOp = GEP.getOperand(0);
- if (isa<UndefValue>(GEP.getOperand(0)))
- return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
-
- // Eliminate unneeded casts for indices.
+ // Eliminate unneeded casts for indices, and replace indices which displace
+ // by multiples of a zero size type with zero.
if (TD) {
bool MadeChange = false;
- unsigned PtrSize = TD->getPointerSizeInBits();
-
+ Type *IntPtrTy = TD->getIntPtrType(GEP.getContext());
+
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) {
- if (!isa<SequentialType>(*GTI)) continue;
-
- // 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.
- unsigned OpBits = cast<IntegerType>((*I)->getType())->getBitWidth();
- if (OpBits == PtrSize)
- continue;
-
- *I = Builder->CreateIntCast(*I, TD->getIntPtrType(GEP.getContext()),true);
- MadeChange = true;
+ // Skip indices into struct types.
+ SequentialType *SeqTy = dyn_cast<SequentialType>(*GTI);
+ if (!SeqTy) continue;
+
+ // 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.
+ if (SeqTy->getElementType()->isSized() &&
+ TD->getTypeAllocSize(SeqTy->getElementType()) == 0)
+ if (!isa<Constant>(*I) || !cast<Constant>(*I)->isNullValue()) {
+ *I = Constant::getNullValue(IntPtrTy);
+ MadeChange = true;
+ }
+
+ Type *IndexTy = (*I)->getType();
+ if (IndexTy != IntPtrTy && !IndexTy->isVectorTy()) {
+ // 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, IntPtrTy, true);
+ MadeChange = true;
+ }
}
if (MadeChange) return &GEP;
}
// getelementptr instructions into a single instruction.
//
if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) {
+ if (!shouldMergeGEPs(*cast<GEPOperator>(&GEP), *Src))
+ return 0;
+
// Note that if our source is a gep chain itself that 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 (GetElementPtrInst *SrcGEP =
- dyn_cast<GetElementPtrInst>(Src->getOperand(0)))
- if (SrcGEP->getNumOperands() == 2)
+ if (GEPOperator *SrcGEP =
+ dyn_cast<GEPOperator>(Src->getOperand(0)))
+ if (SrcGEP->getNumOperands() == 2 && shouldMergeGEPs(*Src, *SrcGEP))
return 0; // Wait until our source is folded to completion.
SmallVector<Value*, 8> Indices;
bool EndsWithSequential = false;
for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src);
I != E; ++I)
- EndsWithSequential = !isa<StructType>(*I);
+ EndsWithSequential = !(*I)->isStructTy();
// Can we combine the two pointer arithmetics offsets?
if (EndsWithSequential) {
if (!Indices.empty())
return (GEP.isInBounds() && Src->isInBounds()) ?
- GetElementPtrInst::CreateInBounds(Src->getOperand(0), Indices.begin(),
- Indices.end(), GEP.getName()) :
- GetElementPtrInst::Create(Src->getOperand(0), Indices.begin(),
- Indices.end(), GEP.getName());
+ GetElementPtrInst::CreateInBounds(Src->getOperand(0), Indices,
+ GEP.getName()) :
+ GetElementPtrInst::Create(Src->getOperand(0), Indices, GEP.getName());
}
-
+
// Handle gep(bitcast x) and gep(gep x, 0, 0, 0).
Value *StrippedPtr = PtrOp->stripPointerCasts();
- if (StrippedPtr != PtrOp) {
- const PointerType *StrippedPtrTy =cast<PointerType>(StrippedPtr->getType());
+ PointerType *StrippedPtrTy = dyn_cast<PointerType>(StrippedPtr->getType());
+ // We do not handle pointer-vector geps here
+ if (!StrippedPtr)
+ return 0;
+
+ if (StrippedPtr != PtrOp &&
+ StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace()) {
bool HasZeroPointerIndex = false;
if (ConstantInt *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) {
- const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
- if (const ArrayType *CATy =
+ PointerType *CPTy = cast<PointerType>(PtrOp->getType());
+ if (ArrayType *CATy =
dyn_cast<ArrayType>(CPTy->getElementType())) {
// GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
if (CATy->getElementType() == StrippedPtrTy->getElementType()) {
// -> GEP i8* X, ...
SmallVector<Value*, 8> Idx(GEP.idx_begin()+1, GEP.idx_end());
GetElementPtrInst *Res =
- GetElementPtrInst::Create(StrippedPtr, Idx.begin(),
- Idx.end(), GEP.getName());
+ GetElementPtrInst::Create(StrippedPtr, Idx, GEP.getName());
Res->setIsInBounds(GEP.isInBounds());
return Res;
}
- if (const ArrayType *XATy =
+ if (ArrayType *XATy =
dyn_cast<ArrayType>(StrippedPtrTy->getElementType())){
// GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
if (CATy->getElementType() == XATy->getElementType()) {
// 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
- const Type *SrcElTy = StrippedPtrTy->getElementType();
- const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
- if (TD && isa<ArrayType>(SrcElTy) &&
+ Type *SrcElTy = StrippedPtrTy->getElementType();
+ Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
+ if (TD && SrcElTy->isArrayTy() &&
TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
TD->getTypeAllocSize(ResElTy)) {
Value *Idx[2];
Idx[0] = Constant::getNullValue(Type::getInt32Ty(GEP.getContext()));
Idx[1] = GEP.getOperand(1);
Value *NewGEP = GEP.isInBounds() ?
- Builder->CreateInBoundsGEP(StrippedPtr, Idx, Idx + 2, GEP.getName()) :
- Builder->CreateGEP(StrippedPtr, Idx, Idx + 2, GEP.getName());
+ Builder->CreateInBoundsGEP(StrippedPtr, Idx, GEP.getName()) :
+ Builder->CreateGEP(StrippedPtr, Idx, GEP.getName());
// V and GEP are both pointer types --> BitCast
return new BitCastInst(NewGEP, GEP.getType());
}
// (where tmp = 8*tmp2) into:
// getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
- if (TD && isa<ArrayType>(SrcElTy) &&
- ResElTy == Type::getInt8Ty(GEP.getContext())) {
+ if (TD && SrcElTy->isArrayTy() && ResElTy->isIntegerTy(8)) {
uint64_t ArrayEltSize =
TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType());
Idx[0] = Constant::getNullValue(Type::getInt32Ty(GEP.getContext()));
Idx[1] = NewIdx;
Value *NewGEP = GEP.isInBounds() ?
- Builder->CreateInBoundsGEP(StrippedPtr, Idx, Idx + 2,GEP.getName()):
- Builder->CreateGEP(StrippedPtr, Idx, Idx + 2, GEP.getName());
+ Builder->CreateInBoundsGEP(StrippedPtr, Idx, GEP.getName()):
+ Builder->CreateGEP(StrippedPtr, Idx, GEP.getName());
// The NewGEP must be pointer typed, so must the old one -> BitCast
return new BitCastInst(NewGEP, GEP.getType());
}
}
}
}
-
+
/// See if we can simplify:
/// X = bitcast A* to B*
/// Y = gep X, <...constant indices...>
/// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
if (TD &&
- !isa<BitCastInst>(BCI->getOperand(0)) && GEP.hasAllConstantIndices()) {
+ !isa<BitCastInst>(BCI->getOperand(0)) && GEP.hasAllConstantIndices() &&
+ StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace()) {
+
// Determine how much the GEP moves the pointer. We are guaranteed to get
// a constant back from EmitGEPOffset.
ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(&GEP));
int64_t Offset = OffsetV->getSExtValue();
-
+
// 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 == 0) {
// field at Offset in 'A's type. If so, we can pull the cast through the
// GEP.
SmallVector<Value*, 8> NewIndices;
- const Type *InTy =
+ Type *InTy =
cast<PointerType>(BCI->getOperand(0)->getType())->getElementType();
if (FindElementAtOffset(InTy, Offset, NewIndices)) {
Value *NGEP = GEP.isInBounds() ?
- Builder->CreateInBoundsGEP(BCI->getOperand(0), NewIndices.begin(),
- NewIndices.end()) :
- Builder->CreateGEP(BCI->getOperand(0), NewIndices.begin(),
- NewIndices.end());
+ Builder->CreateInBoundsGEP(BCI->getOperand(0), NewIndices) :
+ Builder->CreateGEP(BCI->getOperand(0), NewIndices);
if (NGEP->getType() == GEP.getType())
return ReplaceInstUsesWith(GEP, NGEP);
return 0;
}
-Instruction *InstCombiner::visitFree(Instruction &FI) {
- Value *Op = FI.getOperand(1);
+
+
+static bool IsOnlyNullComparedAndFreed(Value *V, SmallVectorImpl<WeakVH> &Users,
+ int Depth = 0) {
+ if (Depth == 8)
+ return false;
+
+ for (Value::use_iterator UI = V->use_begin(), UE = V->use_end();
+ UI != UE; ++UI) {
+ User *U = *UI;
+ if (isFreeCall(U)) {
+ Users.push_back(U);
+ continue;
+ }
+ if (ICmpInst *ICI = dyn_cast<ICmpInst>(U)) {
+ if (ICI->isEquality() && isa<ConstantPointerNull>(ICI->getOperand(1))) {
+ Users.push_back(ICI);
+ continue;
+ }
+ }
+ if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
+ if (IsOnlyNullComparedAndFreed(BCI, Users, Depth+1)) {
+ Users.push_back(BCI);
+ continue;
+ }
+ }
+ if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
+ if (IsOnlyNullComparedAndFreed(GEPI, Users, Depth+1)) {
+ Users.push_back(GEPI);
+ continue;
+ }
+ }
+ if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) {
+ if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
+ II->getIntrinsicID() == Intrinsic::lifetime_end) {
+ Users.push_back(II);
+ continue;
+ }
+ }
+ return false;
+ }
+ return true;
+}
+
+Instruction *InstCombiner::visitMalloc(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.
+ SmallVector<WeakVH, 64> Users;
+ if (IsOnlyNullComparedAndFreed(&MI, Users)) {
+ for (unsigned i = 0, e = Users.size(); i != e; ++i) {
+ Instruction *I = cast_or_null<Instruction>(&*Users[i]);
+ if (!I) continue;
+
+ if (ICmpInst *C = dyn_cast<ICmpInst>(I)) {
+ ReplaceInstUsesWith(*C,
+ ConstantInt::get(Type::getInt1Ty(C->getContext()),
+ C->isFalseWhenEqual()));
+ } else if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) {
+ ReplaceInstUsesWith(*I, UndefValue::get(I->getType()));
+ }
+ EraseInstFromFunction(*I);
+ }
+ return EraseInstFromFunction(MI);
+ }
+ return 0;
+}
+
+
+
+Instruction *InstCombiner::visitFree(CallInst &FI) {
+ Value *Op = FI.getArgOperand(0);
// free undef -> unreachable.
if (isa<UndefValue>(Op)) {
// Insert a new store to null because we cannot modify the CFG here.
- new StoreInst(ConstantInt::getTrue(FI.getContext()),
- UndefValue::get(Type::getInt1PtrTy(FI.getContext())), &FI);
+ Builder->CreateStore(ConstantInt::getTrue(FI.getContext()),
+ UndefValue::get(Type::getInt1PtrTy(FI.getContext())));
return EraseInstFromFunction(FI);
}
if (isa<ConstantPointerNull>(Op))
return EraseInstFromFunction(FI);
- // If we have a malloc call whose only use is a free call, delete both.
- if (isMalloc(Op)) {
- if (CallInst* CI = extractMallocCallFromBitCast(Op)) {
- if (Op->hasOneUse() && CI->hasOneUse()) {
- EraseInstFromFunction(FI);
- EraseInstFromFunction(*CI);
- return EraseInstFromFunction(*cast<Instruction>(Op));
- }
- } else {
- // Op is a call to malloc
- if (Op->hasOneUse()) {
- EraseInstFromFunction(FI);
- return EraseInstFromFunction(*cast<Instruction>(Op));
- }
- }
- }
-
return 0;
}
!isa<Constant>(X)) {
// Swap Destinations and condition...
BI.setCondition(X);
- BI.setSuccessor(0, FalseDest);
- BI.setSuccessor(1, TrueDest);
+ BI.swapSuccessors();
return &BI;
}
Cond->setPredicate(FCmpInst::getInversePredicate(FPred));
// Swap Destinations and condition.
- BI.setSuccessor(0, FalseDest);
- BI.setSuccessor(1, TrueDest);
+ BI.swapSuccessors();
Worklist.Add(Cond);
return &BI;
}
ICmpInst *Cond = cast<ICmpInst>(BI.getCondition());
Cond->setPredicate(ICmpInst::getInversePredicate(IPred));
// Swap Destinations and condition.
- BI.setSuccessor(0, FalseDest);
- BI.setSuccessor(1, TrueDest);
+ BI.swapSuccessors();
Worklist.Add(Cond);
return &BI;
}
if (I->getOpcode() == Instruction::Add)
if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
// change 'switch (X+4) case 1:' into 'switch (X) case -3'
- for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
- SI.setOperand(i,
- ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
- AddRHS));
- SI.setOperand(0, I->getOperand(0));
+ unsigned NumCases = SI.getNumCases();
+ // Skip the first item since that's the default case.
+ for (unsigned i = 1; i < NumCases; ++i) {
+ ConstantInt* CaseVal = SI.getCaseValue(i);
+ Constant* NewCaseVal = ConstantExpr::getSub(cast<Constant>(CaseVal),
+ AddRHS);
+ assert(isa<ConstantInt>(NewCaseVal) &&
+ "Result of expression should be constant");
+ SI.setSuccessorValue(i, cast<ConstantInt>(NewCaseVal));
+ }
+ SI.setCondition(I->getOperand(0));
Worklist.Add(I);
return &SI;
}
if (EV.getNumIndices() > 1)
// Extract the remaining indices out of the constant indexed by the
// first index
- return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end());
+ return ExtractValueInst::Create(V, EV.getIndices().slice(1));
else
return ReplaceInstUsesWith(EV, V);
}
// with
// %E = extractvalue { i32, { i32 } } %A, 0
return ExtractValueInst::Create(IV->getAggregateOperand(),
- EV.idx_begin(), EV.idx_end());
+ EV.getIndices());
}
if (exti == exte && insi == inse)
// Both iterators are at the end: Index lists are identical. Replace
// by switching the order of the insert and extract (though the
// insertvalue should be left in, since it may have other uses).
Value *NewEV = Builder->CreateExtractValue(IV->getAggregateOperand(),
- EV.idx_begin(), EV.idx_end());
+ EV.getIndices());
return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
- insi, inse);
+ makeArrayRef(insi, inse));
}
if (insi == inse)
// The insert list is a prefix of the extract list
// with
// %E extractvalue { i32 } { i32 42 }, 0
return ExtractValueInst::Create(IV->getInsertedValueOperand(),
- exti, exte);
+ makeArrayRef(exti, exte));
}
if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) {
// We're extracting from an intrinsic, see if we're the only user, which
// allows us to simplify multiple result intrinsics to simpler things that
- // just get one value..
+ // just get one value.
if (II->hasOneUse()) {
// Check if we're grabbing the overflow bit or the result of a 'with
// overflow' intrinsic. If it's the latter we can remove the intrinsic
case Intrinsic::uadd_with_overflow:
case Intrinsic::sadd_with_overflow:
if (*EV.idx_begin() == 0) { // Normal result.
- Value *LHS = II->getOperand(1), *RHS = II->getOperand(2);
- II->replaceAllUsesWith(UndefValue::get(II->getType()));
+ Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
+ ReplaceInstUsesWith(*II, UndefValue::get(II->getType()));
EraseInstFromFunction(*II);
return BinaryOperator::CreateAdd(LHS, RHS);
}
+
+ // If the normal result of the add is dead, and the RHS is a constant,
+ // we can transform this into a range comparison.
+ // overflow = uadd a, -4 --> overflow = icmp ugt a, 3
+ if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow)
+ if (ConstantInt *CI = dyn_cast<ConstantInt>(II->getArgOperand(1)))
+ return new ICmpInst(ICmpInst::ICMP_UGT, II->getArgOperand(0),
+ ConstantExpr::getNot(CI));
break;
case Intrinsic::usub_with_overflow:
case Intrinsic::ssub_with_overflow:
if (*EV.idx_begin() == 0) { // Normal result.
- Value *LHS = II->getOperand(1), *RHS = II->getOperand(2);
- II->replaceAllUsesWith(UndefValue::get(II->getType()));
+ Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
+ ReplaceInstUsesWith(*II, UndefValue::get(II->getType()));
EraseInstFromFunction(*II);
return BinaryOperator::CreateSub(LHS, RHS);
}
case Intrinsic::umul_with_overflow:
case Intrinsic::smul_with_overflow:
if (*EV.idx_begin() == 0) { // Normal result.
- Value *LHS = II->getOperand(1), *RHS = II->getOperand(2);
- II->replaceAllUsesWith(UndefValue::get(II->getType()));
+ Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
+ ReplaceInstUsesWith(*II, UndefValue::get(II->getType()));
EraseInstFromFunction(*II);
return BinaryOperator::CreateMul(LHS, RHS);
}
}
}
}
- // Can't simplify extracts from other values. Note that nested extracts are
- // already simplified implicitely by the above (extract ( extract (insert) )
+ if (LoadInst *L = dyn_cast<LoadInst>(Agg))
+ // If the (non-volatile) load only has one use, we can rewrite this to a
+ // load from a GEP. This reduces the size of the load.
+ // FIXME: If a load is used only by extractvalue instructions then this
+ // could be done regardless of having multiple uses.
+ if (L->isSimple() && L->hasOneUse()) {
+ // extractvalue has integer indices, getelementptr has Value*s. Convert.
+ SmallVector<Value*, 4> Indices;
+ // Prefix an i32 0 since we need the first element.
+ Indices.push_back(Builder->getInt32(0));
+ for (ExtractValueInst::idx_iterator I = EV.idx_begin(), E = EV.idx_end();
+ I != E; ++I)
+ Indices.push_back(Builder->getInt32(*I));
+
+ // We need to insert these at the location of the old load, not at that of
+ // the extractvalue.
+ Builder->SetInsertPoint(L->getParent(), L);
+ Value *GEP = Builder->CreateInBoundsGEP(L->getPointerOperand(), Indices);
+ // Returning the load directly will cause the main loop to insert it in
+ // the wrong spot, so use ReplaceInstUsesWith().
+ return ReplaceInstUsesWith(EV, Builder->CreateLoad(GEP));
+ }
+ // We could simplify extracts from other values. Note that nested extracts may
+ // already be simplified implicitly by the above: extract (extract (insert) )
// will be translated into extract ( insert ( extract ) ) first and then just
- // the value inserted, if appropriate).
+ // the value inserted, if appropriate. Similarly for extracts from single-use
+ // loads: extract (extract (load)) will be translated to extract (load (gep))
+ // and if again single-use then via load (gep (gep)) to load (gep).
+ // However, double extracts from e.g. function arguments or return values
+ // aren't handled yet.
+ return 0;
+}
+
+enum Personality_Type {
+ Unknown_Personality,
+ GNU_Ada_Personality,
+ GNU_CXX_Personality,
+ GNU_ObjC_Personality
+};
+
+/// RecognizePersonality - See if the given exception handling personality
+/// function is one that we understand. If so, return a description of it;
+/// otherwise return Unknown_Personality.
+static Personality_Type RecognizePersonality(Value *Pers) {
+ Function *F = dyn_cast<Function>(Pers->stripPointerCasts());
+ if (!F)
+ return Unknown_Personality;
+ return StringSwitch<Personality_Type>(F->getName())
+ .Case("__gnat_eh_personality", GNU_Ada_Personality)
+ .Case("__gxx_personality_v0", GNU_CXX_Personality)
+ .Case("__objc_personality_v0", GNU_ObjC_Personality)
+ .Default(Unknown_Personality);
+}
+
+/// isCatchAll - Return 'true' if the given typeinfo will match anything.
+static bool isCatchAll(Personality_Type Personality, Constant *TypeInfo) {
+ switch (Personality) {
+ case Unknown_Personality:
+ return false;
+ case GNU_Ada_Personality:
+ // While __gnat_all_others_value will match any Ada exception, it doesn't
+ // match foreign exceptions (or didn't, before gcc-4.7).
+ return false;
+ case GNU_CXX_Personality:
+ case GNU_ObjC_Personality:
+ return TypeInfo->isNullValue();
+ }
+ llvm_unreachable("Unknown personality!");
+}
+
+static bool shorter_filter(const Value *LHS, const Value *RHS) {
+ return
+ cast<ArrayType>(LHS->getType())->getNumElements()
+ <
+ cast<ArrayType>(RHS->getType())->getNumElements();
+}
+
+Instruction *InstCombiner::visitLandingPadInst(LandingPadInst &LI) {
+ // The logic here should be correct for any real-world personality function.
+ // However if that turns out not to be true, the offending logic can always
+ // be conditioned on the personality function, like the catch-all logic is.
+ Personality_Type Personality = RecognizePersonality(LI.getPersonalityFn());
+
+ // Simplify the list of clauses, eg by removing repeated catch clauses
+ // (these are often created by inlining).
+ bool MakeNewInstruction = false; // If true, recreate using the following:
+ SmallVector<Value *, 16> NewClauses; // - Clauses for the new instruction;
+ bool CleanupFlag = LI.isCleanup(); // - The new instruction is a cleanup.
+
+ SmallPtrSet<Value *, 16> AlreadyCaught; // Typeinfos known caught already.
+ for (unsigned i = 0, e = LI.getNumClauses(); i != e; ++i) {
+ bool isLastClause = i + 1 == e;
+ if (LI.isCatch(i)) {
+ // A catch clause.
+ Value *CatchClause = LI.getClause(i);
+ Constant *TypeInfo = cast<Constant>(CatchClause->stripPointerCasts());
+
+ // If we already saw this clause, there is no point in having a second
+ // copy of it.
+ if (AlreadyCaught.insert(TypeInfo)) {
+ // This catch clause was not already seen.
+ NewClauses.push_back(CatchClause);
+ } else {
+ // Repeated catch clause - drop the redundant copy.
+ MakeNewInstruction = true;
+ }
+
+ // If this is a catch-all then there is no point in keeping any following
+ // clauses or marking the landingpad as having a cleanup.
+ if (isCatchAll(Personality, TypeInfo)) {
+ if (!isLastClause)
+ MakeNewInstruction = true;
+ CleanupFlag = false;
+ break;
+ }
+ } else {
+ // A filter clause. If any of the filter elements were already caught
+ // then they can be dropped from the filter. It is tempting to try to
+ // exploit the filter further by saying that any typeinfo that does not
+ // occur in the filter can't be caught later (and thus can be dropped).
+ // However this would be wrong, since typeinfos can match without being
+ // equal (for example if one represents a C++ class, and the other some
+ // class derived from it).
+ assert(LI.isFilter(i) && "Unsupported landingpad clause!");
+ Value *FilterClause = LI.getClause(i);
+ ArrayType *FilterType = cast<ArrayType>(FilterClause->getType());
+ unsigned NumTypeInfos = FilterType->getNumElements();
+
+ // An empty filter catches everything, so there is no point in keeping any
+ // following clauses or marking the landingpad as having a cleanup. By
+ // dealing with this case here the following code is made a bit simpler.
+ if (!NumTypeInfos) {
+ NewClauses.push_back(FilterClause);
+ if (!isLastClause)
+ MakeNewInstruction = true;
+ CleanupFlag = false;
+ break;
+ }
+
+ bool MakeNewFilter = false; // If true, make a new filter.
+ SmallVector<Constant *, 16> NewFilterElts; // New elements.
+ if (isa<ConstantAggregateZero>(FilterClause)) {
+ // Not an empty filter - it contains at least one null typeinfo.
+ assert(NumTypeInfos > 0 && "Should have handled empty filter already!");
+ Constant *TypeInfo =
+ Constant::getNullValue(FilterType->getElementType());
+ // If this typeinfo is a catch-all then the filter can never match.
+ if (isCatchAll(Personality, TypeInfo)) {
+ // Throw the filter away.
+ MakeNewInstruction = true;
+ continue;
+ }
+
+ // There is no point in having multiple copies of this typeinfo, so
+ // discard all but the first copy if there is more than one.
+ NewFilterElts.push_back(TypeInfo);
+ if (NumTypeInfos > 1)
+ MakeNewFilter = true;
+ } else {
+ ConstantArray *Filter = cast<ConstantArray>(FilterClause);
+ SmallPtrSet<Value *, 16> SeenInFilter; // For uniquing the elements.
+ NewFilterElts.reserve(NumTypeInfos);
+
+ // Remove any filter elements that were already caught or that already
+ // occurred in the filter. While there, see if any of the elements are
+ // catch-alls. If so, the filter can be discarded.
+ bool SawCatchAll = false;
+ for (unsigned j = 0; j != NumTypeInfos; ++j) {
+ Value *Elt = Filter->getOperand(j);
+ Constant *TypeInfo = cast<Constant>(Elt->stripPointerCasts());
+ if (isCatchAll(Personality, TypeInfo)) {
+ // This element is a catch-all. Bail out, noting this fact.
+ SawCatchAll = true;
+ break;
+ }
+ if (AlreadyCaught.count(TypeInfo))
+ // Already caught by an earlier clause, so having it in the filter
+ // is pointless.
+ continue;
+ // There is no point in having multiple copies of the same typeinfo in
+ // a filter, so only add it if we didn't already.
+ if (SeenInFilter.insert(TypeInfo))
+ NewFilterElts.push_back(cast<Constant>(Elt));
+ }
+ // A filter containing a catch-all cannot match anything by definition.
+ if (SawCatchAll) {
+ // Throw the filter away.
+ MakeNewInstruction = true;
+ continue;
+ }
+
+ // If we dropped something from the filter, make a new one.
+ if (NewFilterElts.size() < NumTypeInfos)
+ MakeNewFilter = true;
+ }
+ if (MakeNewFilter) {
+ FilterType = ArrayType::get(FilterType->getElementType(),
+ NewFilterElts.size());
+ FilterClause = ConstantArray::get(FilterType, NewFilterElts);
+ MakeNewInstruction = true;
+ }
+
+ NewClauses.push_back(FilterClause);
+
+ // If the new filter is empty then it will catch everything so there is
+ // no point in keeping any following clauses or marking the landingpad
+ // as having a cleanup. The case of the original filter being empty was
+ // already handled above.
+ if (MakeNewFilter && !NewFilterElts.size()) {
+ assert(MakeNewInstruction && "New filter but not a new instruction!");
+ CleanupFlag = false;
+ break;
+ }
+ }
+ }
+
+ // If several filters occur in a row then reorder them so that the shortest
+ // filters come first (those with the smallest number of elements). This is
+ // advantageous because shorter filters are more likely to match, speeding up
+ // unwinding, but mostly because it increases the effectiveness of the other
+ // filter optimizations below.
+ for (unsigned i = 0, e = NewClauses.size(); i + 1 < e; ) {
+ unsigned j;
+ // Find the maximal 'j' s.t. the range [i, j) consists entirely of filters.
+ for (j = i; j != e; ++j)
+ if (!isa<ArrayType>(NewClauses[j]->getType()))
+ break;
+
+ // Check whether the filters are already sorted by length. We need to know
+ // if sorting them is actually going to do anything so that we only make a
+ // new landingpad instruction if it does.
+ for (unsigned k = i; k + 1 < j; ++k)
+ if (shorter_filter(NewClauses[k+1], NewClauses[k])) {
+ // Not sorted, so sort the filters now. Doing an unstable sort would be
+ // correct too but reordering filters pointlessly might confuse users.
+ std::stable_sort(NewClauses.begin() + i, NewClauses.begin() + j,
+ shorter_filter);
+ MakeNewInstruction = true;
+ break;
+ }
+
+ // Look for the next batch of filters.
+ i = j + 1;
+ }
+
+ // If typeinfos matched if and only if equal, then the elements of a filter L
+ // that occurs later than a filter F could be replaced by the intersection of
+ // the elements of F and L. In reality two typeinfos can match without being
+ // equal (for example if one represents a C++ class, and the other some class
+ // derived from it) so it would be wrong to perform this transform in general.
+ // However the transform is correct and useful if F is a subset of L. In that
+ // case L can be replaced by F, and thus removed altogether since repeating a
+ // filter is pointless. So here we look at all pairs of filters F and L where
+ // L follows F in the list of clauses, and remove L if every element of F is
+ // an element of L. This can occur when inlining C++ functions with exception
+ // specifications.
+ for (unsigned i = 0; i + 1 < NewClauses.size(); ++i) {
+ // Examine each filter in turn.
+ Value *Filter = NewClauses[i];
+ ArrayType *FTy = dyn_cast<ArrayType>(Filter->getType());
+ if (!FTy)
+ // Not a filter - skip it.
+ continue;
+ unsigned FElts = FTy->getNumElements();
+ // Examine each filter following this one. Doing this backwards means that
+ // we don't have to worry about filters disappearing under us when removed.
+ for (unsigned j = NewClauses.size() - 1; j != i; --j) {
+ Value *LFilter = NewClauses[j];
+ ArrayType *LTy = dyn_cast<ArrayType>(LFilter->getType());
+ if (!LTy)
+ // Not a filter - skip it.
+ continue;
+ // If Filter is a subset of LFilter, i.e. every element of Filter is also
+ // an element of LFilter, then discard LFilter.
+ SmallVector<Value *, 16>::iterator J = NewClauses.begin() + j;
+ // If Filter is empty then it is a subset of LFilter.
+ if (!FElts) {
+ // Discard LFilter.
+ NewClauses.erase(J);
+ MakeNewInstruction = true;
+ // Move on to the next filter.
+ continue;
+ }
+ unsigned LElts = LTy->getNumElements();
+ // If Filter is longer than LFilter then it cannot be a subset of it.
+ if (FElts > LElts)
+ // Move on to the next filter.
+ continue;
+ // At this point we know that LFilter has at least one element.
+ if (isa<ConstantAggregateZero>(LFilter)) { // LFilter only contains zeros.
+ // Filter is a subset of LFilter iff Filter contains only zeros (as we
+ // already know that Filter is not longer than LFilter).
+ if (isa<ConstantAggregateZero>(Filter)) {
+ assert(FElts <= LElts && "Should have handled this case earlier!");
+ // Discard LFilter.
+ NewClauses.erase(J);
+ MakeNewInstruction = true;
+ }
+ // Move on to the next filter.
+ continue;
+ }
+ ConstantArray *LArray = cast<ConstantArray>(LFilter);
+ if (isa<ConstantAggregateZero>(Filter)) { // Filter only contains zeros.
+ // Since Filter is non-empty and contains only zeros, it is a subset of
+ // LFilter iff LFilter contains a zero.
+ assert(FElts > 0 && "Should have eliminated the empty filter earlier!");
+ for (unsigned l = 0; l != LElts; ++l)
+ if (LArray->getOperand(l)->isNullValue()) {
+ // LFilter contains a zero - discard it.
+ NewClauses.erase(J);
+ MakeNewInstruction = true;
+ break;
+ }
+ // Move on to the next filter.
+ continue;
+ }
+ // At this point we know that both filters are ConstantArrays. Loop over
+ // operands to see whether every element of Filter is also an element of
+ // LFilter. Since filters tend to be short this is probably faster than
+ // using a method that scales nicely.
+ ConstantArray *FArray = cast<ConstantArray>(Filter);
+ bool AllFound = true;
+ for (unsigned f = 0; f != FElts; ++f) {
+ Value *FTypeInfo = FArray->getOperand(f)->stripPointerCasts();
+ AllFound = false;
+ for (unsigned l = 0; l != LElts; ++l) {
+ Value *LTypeInfo = LArray->getOperand(l)->stripPointerCasts();
+ if (LTypeInfo == FTypeInfo) {
+ AllFound = true;
+ break;
+ }
+ }
+ if (!AllFound)
+ break;
+ }
+ if (AllFound) {
+ // Discard LFilter.
+ NewClauses.erase(J);
+ MakeNewInstruction = true;
+ }
+ // Move on to the next filter.
+ }
+ }
+
+ // If we changed any of the clauses, replace the old landingpad instruction
+ // with a new one.
+ if (MakeNewInstruction) {
+ LandingPadInst *NLI = LandingPadInst::Create(LI.getType(),
+ LI.getPersonalityFn(),
+ NewClauses.size());
+ for (unsigned i = 0, e = NewClauses.size(); i != e; ++i)
+ NLI->addClause(NewClauses[i]);
+ // A landing pad with no clauses must have the cleanup flag set. It is
+ // theoretically possible, though highly unlikely, that we eliminated all
+ // clauses. If so, force the cleanup flag to true.
+ if (NewClauses.empty())
+ CleanupFlag = true;
+ NLI->setCleanup(CleanupFlag);
+ return NLI;
+ }
+
+ // Even if none of the clauses changed, we may nonetheless have understood
+ // that the cleanup flag is pointless. Clear it if so.
+ if (LI.isCleanup() != CleanupFlag) {
+ assert(!CleanupFlag && "Adding a cleanup, not removing one?!");
+ LI.setCleanup(CleanupFlag);
+ return &LI;
+ }
+
return 0;
}
assert(I->hasOneUse() && "Invariants didn't hold!");
// Cannot move control-flow-involving, volatile loads, vaarg, etc.
- if (isa<PHINode>(I) || I->mayHaveSideEffects() || isa<TerminatorInst>(I))
+ if (isa<PHINode>(I) || isa<LandingPadInst>(I) || I->mayHaveSideEffects() ||
+ isa<TerminatorInst>(I))
return false;
// Do not sink alloca instructions out of the entry block.
return false;
}
- BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
-
+ BasicBlock::iterator InsertPos = DestBlock->getFirstInsertionPt();
I->moveBefore(InsertPos);
++NumSunkInst;
return true;
static bool AddReachableCodeToWorklist(BasicBlock *BB,
SmallPtrSet<BasicBlock*, 64> &Visited,
InstCombiner &IC,
- const TargetData *TD) {
+ const TargetData *TD,
+ const TargetLibraryInfo *TLI) {
bool MadeIRChange = false;
SmallVector<BasicBlock*, 256> Worklist;
Worklist.push_back(BB);
-
- std::vector<Instruction*> InstrsForInstCombineWorklist;
- InstrsForInstCombineWorklist.reserve(128);
- SmallPtrSet<ConstantExpr*, 64> FoldedConstants;
-
- while (!Worklist.empty()) {
- BB = Worklist.back();
- Worklist.pop_back();
+ SmallVector<Instruction*, 128> InstrsForInstCombineWorklist;
+ DenseMap<ConstantExpr*, Constant*> FoldedConstants;
+
+ do {
+ BB = Worklist.pop_back_val();
// We have now visited this block! If we've already been here, ignore it.
if (!Visited.insert(BB)) continue;
// ConstantProp instruction if trivially constant.
if (!Inst->use_empty() && isa<Constant>(Inst->getOperand(0)))
- if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
+ if (Constant *C = ConstantFoldInstruction(Inst, TD, TLI)) {
DEBUG(errs() << "IC: ConstFold to: " << *C << " from: "
<< *Inst << '\n');
Inst->replaceAllUsesWith(C);
continue;
}
-
-
if (TD) {
// See if we can constant fold its operands.
for (User::op_iterator i = Inst->op_begin(), e = Inst->op_end();
i != e; ++i) {
ConstantExpr *CE = dyn_cast<ConstantExpr>(i);
if (CE == 0) continue;
-
- // If we already folded this constant, don't try again.
- if (!FoldedConstants.insert(CE))
- continue;
-
- Constant *NewC = ConstantFoldConstantExpression(CE, TD);
- if (NewC && NewC != CE) {
- *i = NewC;
+
+ Constant*& FoldRes = FoldedConstants[CE];
+ if (!FoldRes)
+ FoldRes = ConstantFoldConstantExpression(CE, TD, TLI);
+ if (!FoldRes)
+ FoldRes = CE;
+
+ if (FoldRes != CE) {
+ *i = FoldRes;
MadeIRChange = true;
}
}
}
-
InstrsForInstCombineWorklist.push_back(Inst);
}
for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
Worklist.push_back(TI->getSuccessor(i));
- }
+ } while (!Worklist.empty());
// Once we've found all of the instructions to add to instcombine's worklist,
// add them in reverse order. This way instcombine will visit from the top
MadeIRChange = false;
DEBUG(errs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
- << F.getNameStr() << "\n");
+ << F.getName() << "\n");
{
// Do a depth-first traversal of the function, populate the worklist with
// the reachable instructions. Ignore blocks that are not reachable. Keep
// track of which blocks we visit.
SmallPtrSet<BasicBlock*, 64> Visited;
- MadeIRChange |= AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
+ MadeIRChange |= AddReachableCodeToWorklist(F.begin(), Visited, *this, TD,
+ TLI);
// Do a quick scan over the function. If we find any blocks that are
// unreachable, remove any instructions inside of them. This prevents
// the instcombine code from having to deal with some bad special cases.
- for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
- if (!Visited.count(BB)) {
- Instruction *Term = BB->getTerminator();
- while (Term != BB->begin()) { // Remove instrs bottom-up
- BasicBlock::iterator I = Term; --I;
-
- DEBUG(errs() << "IC: DCE: " << *I << '\n');
- // A debug intrinsic shouldn't force another iteration if we weren't
- // going to do one without it.
- if (!isa<DbgInfoIntrinsic>(I)) {
- ++NumDeadInst;
- MadeIRChange = true;
- }
-
- // If I is not void type then replaceAllUsesWith undef.
- // This allows ValueHandlers and custom metadata to adjust itself.
- if (!I->getType()->isVoidTy())
- I->replaceAllUsesWith(UndefValue::get(I->getType()));
- I->eraseFromParent();
+ for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
+ if (Visited.count(BB)) continue;
+
+ // Delete the instructions backwards, as it has a reduced likelihood of
+ // having to update as many def-use and use-def chains.
+ Instruction *EndInst = BB->getTerminator(); // Last not to be deleted.
+ while (EndInst != BB->begin()) {
+ // Delete the next to last instruction.
+ BasicBlock::iterator I = EndInst;
+ Instruction *Inst = --I;
+ if (!Inst->use_empty())
+ Inst->replaceAllUsesWith(UndefValue::get(Inst->getType()));
+ if (isa<LandingPadInst>(Inst)) {
+ EndInst = Inst;
+ continue;
+ }
+ if (!isa<DbgInfoIntrinsic>(Inst)) {
+ ++NumDeadInst;
+ MadeIRChange = true;
}
+ Inst->eraseFromParent();
}
+ }
}
while (!Worklist.isEmpty()) {
// Instruction isn't dead, see if we can constant propagate it.
if (!I->use_empty() && isa<Constant>(I->getOperand(0)))
- if (Constant *C = ConstantFoldInstruction(I, TD)) {
+ if (Constant *C = ConstantFoldInstruction(I, TD, TLI)) {
DEBUG(errs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n');
// Add operands to the worklist.
// Now that we have an instruction, try combining it to simplify it.
Builder->SetInsertPoint(I->getParent(), I);
+ Builder->SetCurrentDebugLocation(I->getDebugLoc());
#ifndef NDEBUG
std::string OrigI;
DEBUG(errs() << "IC: Old = " << *I << '\n'
<< " New = " << *Result << '\n');
+ if (!I->getDebugLoc().isUnknown())
+ Result->setDebugLoc(I->getDebugLoc());
// Everything uses the new instruction now.
I->replaceAllUsesWith(Result);
+ // Move the name to the new instruction first.
+ Result->takeName(I);
+
// Push the new instruction and any users onto the worklist.
Worklist.Add(Result);
Worklist.AddUsersToWorkList(*Result);
- // Move the name to the new instruction first.
- Result->takeName(I);
-
// Insert the new instruction into the basic block...
BasicBlock *InstParent = I->getParent();
BasicBlock::iterator InsertPos = I;
- if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
- while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
- ++InsertPos;
+ // If we replace a PHI with something that isn't a PHI, fix up the
+ // insertion point.
+ if (!isa<PHINode>(Result) && isa<PHINode>(InsertPos))
+ InsertPos = InstParent->getFirstInsertionPt();
InstParent->getInstList().insert(InsertPos, Result);
bool InstCombiner::runOnFunction(Function &F) {
- MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
TD = getAnalysisIfAvailable<TargetData>();
-
+ TLI = &getAnalysis<TargetLibraryInfo>();
/// Builder - This is an IRBuilder that automatically inserts new
/// instructions into the worklist when they are created.
bool EverMadeChange = false;
+ // Lower dbg.declare intrinsics otherwise their value may be clobbered
+ // by instcombiner.
+ EverMadeChange = LowerDbgDeclare(F);
+
// Iterate while there is work to do.
unsigned Iteration = 0;
while (DoOneIteration(F, Iteration++))