#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:
//
-bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
+// 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;
- 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;
+ }
+ }
+ }
+
+ 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;
}
- return Changed;
+ }
+}
+
+/// 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;
}
-/// isFreeToInvert - Return true if the specified value is free to invert (apply
-/// ~ to). This happens in cases where the ~ can be eliminated.
-static inline bool isFreeToInvert(Value *V) {
- // ~(~(X)) -> X.
- if (BinaryOperator::isNot(V))
- return true;
-
- // Constants can be considered to be not'ed values.
- if (isa<ConstantInt>(V))
- return true;
-
- // Compares can be inverted if they have a single use.
- if (CmpInst *CI = dyn_cast<CmpInst>(V))
- return CI->hasOneUse();
-
- return false;
-}
-
-static inline Value *dyn_castNotVal(Value *V) {
- // If this is not(not(x)) don't return that this is a not: we want the two
- // not's to be folded first.
- if (BinaryOperator::isNot(V)) {
- Value *Operand = BinaryOperator::getNotArgument(V);
- if (!isFreeToInvert(Operand))
- return Operand;
- }
-
- // Constants can be considered to be not'ed values...
- if (ConstantInt *C = dyn_cast<ConstantInt>(V))
- return ConstantInt::get(C->getType(), ~C->getValue());
- return 0;
-}
-
-
-
-/// AddOne - Add one to a ConstantInt.
-static Constant *AddOne(Constant *C) {
- return ConstantExpr::getAdd(C, ConstantInt::get(C->getType(), 1));
-}
-/// SubOne - Subtract one from a ConstantInt.
-static Constant *SubOne(ConstantInt *C) {
- return ConstantInt::get(C->getContext(), C->getValue()-1);
-}
-
-
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;
-
- Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, this);
- Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, this);
-
- return SelectInst::Create(SI->getCondition(), SelectTrueVal,
- SelectFalseVal);
- }
- return 0;
-}
-
-
-/// FoldOpIntoPhi - Given a binary operator, cast instruction, or select which
-/// 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;
- 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))
- return 0;
-
-
- // 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 = 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())
+ 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;
}
-
- // 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) {
- 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);
- InsertNewInstBefore(NewPN, *PN);
- NewPN->takeName(PN);
-
- // 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 = 0;
- 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));
- }
- NewPN->addIncoming(InV, ThisBB);
- }
- } 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));
- }
- NewPN->addIncoming(InV, PN->getIncomingBlock(i));
- }
- } else {
- CastInst *CI = cast<CastInst>(&I);
- const Type *RetTy = CI->getType();
- for (unsigned i = 0; i != NumPHIValues; ++i) {
- Value *InV;
- if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
- InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
- } else {
- assert(PN->getIncomingBlock(i) == NonConstBB);
- InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
- I.getType(), "phitmp",
- NonConstBB->getTerminator());
- Worklist.Add(cast<Instruction>(InV));
- }
- NewPN->addIncoming(InV, PN->getIncomingBlock(i));
- }
- }
- return ReplaceInstUsesWith(I, NewPN);
-}
-
-
-/// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
-/// are carefully arranged to allow folding of expressions such as:
-///
-/// (A < B) | (A > B) --> (A != B)
-///
-/// Note that this is only valid if the first and second predicates have the
-/// same sign. Is illegal to do: (A u< B) | (A s> B)
-///
-/// Three bits are used to represent the condition, as follows:
-/// 0 A > B
-/// 1 A == B
-/// 2 A < B
-///
-/// <=> Value Definition
-/// 000 0 Always false
-/// 001 1 A > B
-/// 010 2 A == B
-/// 011 3 A >= B
-/// 100 4 A < B
-/// 101 5 A != B
-/// 110 6 A <= B
-/// 111 7 Always true
-///
-static unsigned getICmpCode(const ICmpInst *ICI) {
- switch (ICI->getPredicate()) {
- // False -> 0
- case ICmpInst::ICMP_UGT: return 1; // 001
- case ICmpInst::ICMP_SGT: return 1; // 001
- case ICmpInst::ICMP_EQ: return 2; // 010
- case ICmpInst::ICMP_UGE: return 3; // 011
- case ICmpInst::ICMP_SGE: return 3; // 011
- case ICmpInst::ICMP_ULT: return 4; // 100
- case ICmpInst::ICMP_SLT: return 4; // 100
- case ICmpInst::ICMP_NE: return 5; // 101
- case ICmpInst::ICMP_ULE: return 6; // 110
- case ICmpInst::ICMP_SLE: return 6; // 110
- // True -> 7
- default:
- llvm_unreachable("Invalid ICmp predicate!");
- return 0;
- }
-}
-
-/// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
-/// predicate into a three bit mask. It also returns whether it is an ordered
-/// predicate by reference.
-static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
- isOrdered = false;
- switch (CC) {
- case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000
- case FCmpInst::FCMP_UNO: return 0; // 000
- case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001
- case FCmpInst::FCMP_UGT: return 1; // 001
- case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010
- case FCmpInst::FCMP_UEQ: return 2; // 010
- case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011
- case FCmpInst::FCMP_UGE: return 3; // 011
- case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100
- case FCmpInst::FCMP_ULT: return 4; // 100
- case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101
- case FCmpInst::FCMP_UNE: return 5; // 101
- case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110
- case FCmpInst::FCMP_ULE: return 6; // 110
- // True -> 7
- default:
- // Not expecting FCMP_FALSE and FCMP_TRUE;
- llvm_unreachable("Unexpected FCmp predicate!");
- return 0;
- }
-}
-
-/// getICmpValue - This is the complement of getICmpCode, which turns an
-/// opcode and two operands into either a constant true or false, or a brand
-/// new ICmp instruction. The sign is passed in to determine which kind
-/// of predicate to use in the new icmp instruction.
-static Value *getICmpValue(bool Sign, unsigned Code, Value *LHS, Value *RHS) {
- switch (Code) {
- default: assert(0 && "Illegal ICmp code!");
- case 0:
- return ConstantInt::getFalse(LHS->getContext());
- case 1:
- if (Sign)
- return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
- return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
- case 2:
- return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
- case 3:
- if (Sign)
- return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
- return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
- case 4:
- if (Sign)
- return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
- return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
- case 5:
- return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
- case 6:
- if (Sign)
- return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
- return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
- case 7:
- return ConstantInt::getTrue(LHS->getContext());
- }
-}
-
-/// getFCmpValue - This is the complement of getFCmpCode, which turns an
-/// opcode and two operands into either a FCmp instruction. isordered is passed
-/// in to determine which kind of predicate to use in the new fcmp instruction.
-static Value *getFCmpValue(bool isordered, unsigned code,
- Value *LHS, Value *RHS) {
- switch (code) {
- default: llvm_unreachable("Illegal FCmp code!");
- case 0:
- if (isordered)
- return new FCmpInst(FCmpInst::FCMP_ORD, LHS, RHS);
- else
- return new FCmpInst(FCmpInst::FCMP_UNO, LHS, RHS);
- case 1:
- if (isordered)
- return new FCmpInst(FCmpInst::FCMP_OGT, LHS, RHS);
- else
- return new FCmpInst(FCmpInst::FCMP_UGT, LHS, RHS);
- case 2:
- if (isordered)
- return new FCmpInst(FCmpInst::FCMP_OEQ, LHS, RHS);
- else
- return new FCmpInst(FCmpInst::FCMP_UEQ, LHS, RHS);
- case 3:
- if (isordered)
- return new FCmpInst(FCmpInst::FCMP_OGE, LHS, RHS);
- else
- return new FCmpInst(FCmpInst::FCMP_UGE, LHS, RHS);
- case 4:
- if (isordered)
- return new FCmpInst(FCmpInst::FCMP_OLT, LHS, RHS);
- else
- return new FCmpInst(FCmpInst::FCMP_ULT, LHS, RHS);
- case 5:
- if (isordered)
- return new FCmpInst(FCmpInst::FCMP_ONE, LHS, RHS);
- else
- return new FCmpInst(FCmpInst::FCMP_UNE, LHS, RHS);
- case 6:
- if (isordered)
- return new FCmpInst(FCmpInst::FCMP_OLE, LHS, RHS);
- else
- return new FCmpInst(FCmpInst::FCMP_ULE, LHS, RHS);
- case 7: return ConstantInt::getTrue(LHS->getContext());
- }
-}
-
-/// PredicatesFoldable - Return true if both predicates match sign or if at
-/// least one of them is an equality comparison (which is signless).
-static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
- return (CmpInst::isSigned(p1) == CmpInst::isSigned(p2)) ||
- (CmpInst::isSigned(p1) && ICmpInst::isEquality(p2)) ||
- (CmpInst::isSigned(p2) && ICmpInst::isEquality(p1));
-}
-
-// OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
-// the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
-// guaranteed to be a binary operator.
-Instruction *InstCombiner::OptAndOp(Instruction *Op,
- ConstantInt *OpRHS,
- ConstantInt *AndRHS,
- BinaryOperator &TheAnd) {
- Value *X = Op->getOperand(0);
- Constant *Together = 0;
- if (!Op->isShift())
- Together = ConstantExpr::getAnd(AndRHS, OpRHS);
-
- switch (Op->getOpcode()) {
- case Instruction::Xor:
- if (Op->hasOneUse()) {
- // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
- Value *And = Builder->CreateAnd(X, AndRHS);
- And->takeName(Op);
- return BinaryOperator::CreateXor(And, Together);
- }
- break;
- case Instruction::Or:
- if (Together == AndRHS) // (X | C) & C --> C
- return ReplaceInstUsesWith(TheAnd, AndRHS);
-
- if (Op->hasOneUse() && Together != OpRHS) {
- // (X | C1) & C2 --> (X | (C1&C2)) & C2
- Value *Or = Builder->CreateOr(X, Together);
- Or->takeName(Op);
- return BinaryOperator::CreateAnd(Or, AndRHS);
- }
- break;
- case Instruction::Add:
- if (Op->hasOneUse()) {
- // Adding a one to a single bit bit-field should be turned into an XOR
- // of the bit. First thing to check is to see if this AND is with a
- // single bit constant.
- const APInt &AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
-
- // If there is only one bit set.
- if (AndRHSV.isPowerOf2()) {
- // Ok, at this point, we know that we are masking the result of the
- // ADD down to exactly one bit. If the constant we are adding has
- // no bits set below this bit, then we can eliminate the ADD.
- const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
-
- // Check to see if any bits below the one bit set in AndRHSV are set.
- if ((AddRHS & (AndRHSV-1)) == 0) {
- // If not, the only thing that can effect the output of the AND is
- // the bit specified by AndRHSV. If that bit is set, the effect of
- // the XOR is to toggle the bit. If it is clear, then the ADD has
- // no effect.
- if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
- TheAnd.setOperand(0, X);
- return &TheAnd;
- } else {
- // Pull the XOR out of the AND.
- Value *NewAnd = Builder->CreateAnd(X, AndRHS);
- NewAnd->takeName(Op);
- return BinaryOperator::CreateXor(NewAnd, AndRHS);
- }
- }
- }
- }
- break;
-
- case Instruction::Shl: {
- // We know that the AND will not produce any of the bits shifted in, so if
- // the anded constant includes them, clear them now!
- //
- uint32_t BitWidth = AndRHS->getType()->getBitWidth();
- uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
- APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
- ConstantInt *CI = ConstantInt::get(AndRHS->getContext(),
- AndRHS->getValue() & ShlMask);
-
- if (CI->getValue() == ShlMask) {
- // Masking out bits that the shift already masks
- return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
- } else if (CI != AndRHS) { // Reducing bits set in and.
- TheAnd.setOperand(1, CI);
- return &TheAnd;
- }
- break;
- }
- case Instruction::LShr: {
- // We know that the AND will not produce any of the bits shifted in, so if
- // the anded constant includes them, clear them now! This only applies to
- // unsigned shifts, because a signed shr may bring in set bits!
- //
- uint32_t BitWidth = AndRHS->getType()->getBitWidth();
- uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
- APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
- ConstantInt *CI = ConstantInt::get(Op->getContext(),
- AndRHS->getValue() & ShrMask);
-
- if (CI->getValue() == ShrMask) {
- // Masking out bits that the shift already masks.
- return ReplaceInstUsesWith(TheAnd, Op);
- } else if (CI != AndRHS) {
- TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
- return &TheAnd;
- }
- break;
- }
- case Instruction::AShr:
- // Signed shr.
- // See if this is shifting in some sign extension, then masking it out
- // with an and.
- if (Op->hasOneUse()) {
- uint32_t BitWidth = AndRHS->getType()->getBitWidth();
- uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
- APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
- Constant *C = ConstantInt::get(Op->getContext(),
- AndRHS->getValue() & ShrMask);
- if (C == AndRHS) { // Masking out bits shifted in.
- // (Val ashr C1) & C2 -> (Val lshr C1) & C2
- // Make the argument unsigned.
- Value *ShVal = Op->getOperand(0);
- ShVal = Builder->CreateLShr(ShVal, OpRHS, Op->getName());
- return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
- }
- }
- break;
- }
- return 0;
-}
-
-
-/// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
-/// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
-/// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
-/// whether to treat the V, Lo and HI as signed or not. IB is the location to
-/// insert new instructions.
-Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
- bool isSigned, bool Inside,
- Instruction &IB) {
- assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
- ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
- "Lo is not <= Hi in range emission code!");
- if (Inside) {
- if (Lo == Hi) // Trivially false.
- return new ICmpInst(ICmpInst::ICMP_NE, V, V);
-
- // V >= Min && V < Hi --> V < Hi
- if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
- ICmpInst::Predicate pred = (isSigned ?
- ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
- return new ICmpInst(pred, V, Hi);
- }
-
- // Emit V-Lo <u Hi-Lo
- Constant *NegLo = ConstantExpr::getNeg(Lo);
- Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
- Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
- return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
- }
-
- if (Lo == Hi) // Trivially true.
- return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
-
- // V < Min || V >= Hi -> V > Hi-1
- Hi = SubOne(cast<ConstantInt>(Hi));
- if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
- ICmpInst::Predicate pred = (isSigned ?
- ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
- return new ICmpInst(pred, V, Hi);
- }
-
- // Emit V-Lo >u Hi-1-Lo
- // Note that Hi has already had one subtracted from it, above.
- ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
- Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
- Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
- return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
-}
-
-// isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
-// any number of 0s on either side. The 1s are allowed to wrap from LSB to
-// MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
-// not, since all 1s are not contiguous.
-static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
- const APInt& V = Val->getValue();
- uint32_t BitWidth = Val->getType()->getBitWidth();
- if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
-
- // look for the first zero bit after the run of ones
- MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
- // look for the first non-zero bit
- ME = V.getActiveBits();
- return true;
-}
-
-/// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
-/// where isSub determines whether the operator is a sub. If we can fold one of
-/// the following xforms:
-///
-/// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
-/// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
-/// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
-///
-/// return (A +/- B).
-///
-Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
- ConstantInt *Mask, bool isSub,
- Instruction &I) {
- Instruction *LHSI = dyn_cast<Instruction>(LHS);
- if (!LHSI || LHSI->getNumOperands() != 2 ||
- !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
-
- ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
-
- switch (LHSI->getOpcode()) {
- default: return 0;
- case Instruction::And:
- if (ConstantExpr::getAnd(N, Mask) == Mask) {
- // If the AndRHS is a power of two minus one (0+1+), this is simple.
- if ((Mask->getValue().countLeadingZeros() +
- Mask->getValue().countPopulation()) ==
- Mask->getValue().getBitWidth())
- break;
+ Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, this);
+ Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, this);
- // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
- // part, we don't need any explicit masks to take them out of A. If that
- // is all N is, ignore it.
- uint32_t MB = 0, ME = 0;
- if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
- uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
- APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
- if (MaskedValueIsZero(RHS, Mask))
- break;
- }
- }
- return 0;
- case Instruction::Or:
- case Instruction::Xor:
- // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
- if ((Mask->getValue().countLeadingZeros() +
- Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
- && ConstantExpr::getAnd(N, Mask)->isNullValue())
- break;
- return 0;
+ return SelectInst::Create(SI->getCondition(),
+ SelectTrueVal, SelectFalseVal);
}
-
- if (isSub)
- return Builder->CreateSub(LHSI->getOperand(0), RHS, "fold");
- return Builder->CreateAdd(LHSI->getOperand(0), RHS, "fold");
-}
-
-/// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
-Instruction *InstCombiner::FoldAndOfICmps(Instruction &I,
- ICmpInst *LHS, ICmpInst *RHS) {
- ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
-
- // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
- if (PredicatesFoldable(LHSCC, RHSCC)) {
- if (LHS->getOperand(0) == RHS->getOperand(1) &&
- LHS->getOperand(1) == RHS->getOperand(0))
- LHS->swapOperands();
- if (LHS->getOperand(0) == RHS->getOperand(0) &&
- LHS->getOperand(1) == RHS->getOperand(1)) {
- Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
- unsigned Code = getICmpCode(LHS) & getICmpCode(RHS);
- bool isSigned = LHS->isSigned() || RHS->isSigned();
- Value *RV = getICmpValue(isSigned, Code, Op0, Op1);
- if (Instruction *I = dyn_cast<Instruction>(RV))
- return I;
- // Otherwise, it's a constant boolean value.
- return ReplaceInstUsesWith(I, RV);
- }
- }
-
- // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
- Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0);
- ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1));
- ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1));
- if (LHSCst == 0 || RHSCst == 0) return 0;
-
- if (LHSCst == RHSCst && LHSCC == RHSCC) {
- // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
- // where C is a power of 2
- if (LHSCC == ICmpInst::ICMP_ULT &&
- LHSCst->getValue().isPowerOf2()) {
- Value *NewOr = Builder->CreateOr(Val, Val2);
- return new ICmpInst(LHSCC, NewOr, LHSCst);
- }
-
- // (icmp eq A, 0) & (icmp eq B, 0) --> (icmp eq (A|B), 0)
- if (LHSCC == ICmpInst::ICMP_EQ && LHSCst->isZero()) {
- Value *NewOr = Builder->CreateOr(Val, Val2);
- return new ICmpInst(LHSCC, NewOr, LHSCst);
- }
- }
-
- // From here on, we only handle:
- // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
- if (Val != Val2) return 0;
-
- // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
- if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
- RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
- LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
- RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
- return 0;
-
- // We can't fold (ugt x, C) & (sgt x, C2).
- if (!PredicatesFoldable(LHSCC, RHSCC))
- return 0;
-
- // Ensure that the larger constant is on the RHS.
- bool ShouldSwap;
- if (CmpInst::isSigned(LHSCC) ||
- (ICmpInst::isEquality(LHSCC) &&
- CmpInst::isSigned(RHSCC)))
- ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
- else
- ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
-
- if (ShouldSwap) {
- std::swap(LHS, RHS);
- std::swap(LHSCst, RHSCst);
- std::swap(LHSCC, RHSCC);
- }
-
- // At this point, we know we have have two icmp instructions
- // comparing a value against two constants and and'ing the result
- // together. Because of the above check, we know that we only have
- // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
- // (from the icmp folding check above), that the two constants
- // are not equal and that the larger constant is on the RHS
- assert(LHSCst != RHSCst && "Compares not folded above?");
-
- switch (LHSCC) {
- default: llvm_unreachable("Unknown integer condition code!");
- case ICmpInst::ICMP_EQ:
- switch (RHSCC) {
- default: llvm_unreachable("Unknown integer condition code!");
- case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
- case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
- case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
- return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
- case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
- case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
- case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
- return ReplaceInstUsesWith(I, LHS);
- }
- case ICmpInst::ICMP_NE:
- switch (RHSCC) {
- default: llvm_unreachable("Unknown integer condition code!");
- case ICmpInst::ICMP_ULT:
- if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
- return new ICmpInst(ICmpInst::ICMP_ULT, Val, LHSCst);
- break; // (X != 13 & X u< 15) -> no change
- case ICmpInst::ICMP_SLT:
- if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
- return new ICmpInst(ICmpInst::ICMP_SLT, Val, LHSCst);
- break; // (X != 13 & X s< 15) -> no change
- case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
- case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
- case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
- return ReplaceInstUsesWith(I, RHS);
- case ICmpInst::ICMP_NE:
- if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
- Constant *AddCST = ConstantExpr::getNeg(LHSCst);
- Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
- return new ICmpInst(ICmpInst::ICMP_UGT, Add,
- ConstantInt::get(Add->getType(), 1));
- }
- break; // (X != 13 & X != 15) -> no change
- }
- break;
- case ICmpInst::ICMP_ULT:
- switch (RHSCC) {
- default: llvm_unreachable("Unknown integer condition code!");
- case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
- case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
- return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
- case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
- break;
- case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
- case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
- return ReplaceInstUsesWith(I, LHS);
- case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
- break;
- }
- break;
- case ICmpInst::ICMP_SLT:
- switch (RHSCC) {
- default: llvm_unreachable("Unknown integer condition code!");
- case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
- case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
- return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
- case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
- break;
- case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
- case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
- return ReplaceInstUsesWith(I, LHS);
- case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
- break;
- }
- break;
- case ICmpInst::ICMP_UGT:
- switch (RHSCC) {
- default: llvm_unreachable("Unknown integer condition code!");
- case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
- case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
- return ReplaceInstUsesWith(I, RHS);
- case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
- break;
- case ICmpInst::ICMP_NE:
- if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
- return new ICmpInst(LHSCC, Val, RHSCst);
- break; // (X u> 13 & X != 15) -> no change
- case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
- return InsertRangeTest(Val, AddOne(LHSCst),
- RHSCst, false, true, I);
- case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
- break;
- }
- break;
- case ICmpInst::ICMP_SGT:
- switch (RHSCC) {
- default: llvm_unreachable("Unknown integer condition code!");
- case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
- case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
- return ReplaceInstUsesWith(I, RHS);
- case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
- break;
- case ICmpInst::ICMP_NE:
- if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
- return new ICmpInst(LHSCC, Val, RHSCst);
- break; // (X s> 13 & X != 15) -> no change
- case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
- return InsertRangeTest(Val, AddOne(LHSCst),
- RHSCst, true, true, I);
- case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
- break;
- }
- break;
- }
-
- return 0;
-}
-
-Instruction *InstCombiner::FoldAndOfFCmps(Instruction &I, FCmpInst *LHS,
- FCmpInst *RHS) {
-
- if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
- RHS->getPredicate() == FCmpInst::FCMP_ORD) {
- // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
- if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
- if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
- // If either of the constants are nans, then the whole thing returns
- // false.
- if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
- return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
- return new FCmpInst(FCmpInst::FCMP_ORD,
- LHS->getOperand(0), RHS->getOperand(0));
- }
-
- // Handle vector zeros. This occurs because the canonical form of
- // "fcmp ord x,x" is "fcmp ord x, 0".
- if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
- isa<ConstantAggregateZero>(RHS->getOperand(1)))
- return new FCmpInst(FCmpInst::FCMP_ORD,
- LHS->getOperand(0), RHS->getOperand(0));
- return 0;
- }
-
- Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
- Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
- FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
-
-
- if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
- // Swap RHS operands to match LHS.
- Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
- std::swap(Op1LHS, Op1RHS);
- }
-
- if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
- // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
- if (Op0CC == Op1CC)
- return new FCmpInst((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
-
- if (Op0CC == FCmpInst::FCMP_FALSE || Op1CC == FCmpInst::FCMP_FALSE)
- return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
- if (Op0CC == FCmpInst::FCMP_TRUE)
- return ReplaceInstUsesWith(I, RHS);
- if (Op1CC == FCmpInst::FCMP_TRUE)
- return ReplaceInstUsesWith(I, LHS);
-
- bool Op0Ordered;
- bool Op1Ordered;
- unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
- unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
- if (Op1Pred == 0) {
- std::swap(LHS, RHS);
- std::swap(Op0Pred, Op1Pred);
- std::swap(Op0Ordered, Op1Ordered);
- }
- if (Op0Pred == 0) {
- // uno && ueq -> uno && (uno || eq) -> ueq
- // ord && olt -> ord && (ord && lt) -> olt
- if (Op0Ordered == Op1Ordered)
- return ReplaceInstUsesWith(I, RHS);
-
- // uno && oeq -> uno && (ord && eq) -> false
- // uno && ord -> false
- if (!Op0Ordered)
- return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
- // ord && ueq -> ord && (uno || eq) -> oeq
- return cast<Instruction>(getFCmpValue(true, Op1Pred, Op0LHS, Op0RHS));
- }
- }
-
- return 0;
-}
-
-
-Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
- bool Changed = SimplifyCommutative(I);
- Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
-
- if (Value *V = SimplifyAndInst(Op0, Op1, TD))
- return ReplaceInstUsesWith(I, V);
-
- // See if we can simplify any instructions used by the instruction whose sole
- // purpose is to compute bits we don't care about.
- if (SimplifyDemandedInstructionBits(I))
- return &I;
-
- if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
- const APInt &AndRHSMask = AndRHS->getValue();
- APInt NotAndRHS(~AndRHSMask);
-
- // Optimize a variety of ((val OP C1) & C2) combinations...
- if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
- Value *Op0LHS = Op0I->getOperand(0);
- Value *Op0RHS = Op0I->getOperand(1);
- switch (Op0I->getOpcode()) {
- default: break;
- case Instruction::Xor:
- case Instruction::Or:
- // If the mask is only needed on one incoming arm, push it up.
- if (!Op0I->hasOneUse()) break;
-
- if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
- // Not masking anything out for the LHS, move to RHS.
- Value *NewRHS = Builder->CreateAnd(Op0RHS, AndRHS,
- Op0RHS->getName()+".masked");
- return BinaryOperator::Create(Op0I->getOpcode(), Op0LHS, NewRHS);
- }
- if (!isa<Constant>(Op0RHS) &&
- MaskedValueIsZero(Op0RHS, NotAndRHS)) {
- // Not masking anything out for the RHS, move to LHS.
- Value *NewLHS = Builder->CreateAnd(Op0LHS, AndRHS,
- Op0LHS->getName()+".masked");
- return BinaryOperator::Create(Op0I->getOpcode(), NewLHS, Op0RHS);
- }
-
- break;
- case Instruction::Add:
- // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
- // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
- // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
- if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
- return BinaryOperator::CreateAnd(V, AndRHS);
- if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
- return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
- break;
-
- case Instruction::Sub:
- // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
- // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
- // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
- if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
- return BinaryOperator::CreateAnd(V, AndRHS);
-
- // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
- // has 1's for all bits that the subtraction with A might affect.
- if (Op0I->hasOneUse()) {
- uint32_t BitWidth = AndRHSMask.getBitWidth();
- uint32_t Zeros = AndRHSMask.countLeadingZeros();
- APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
-
- ConstantInt *A = dyn_cast<ConstantInt>(Op0LHS);
- if (!(A && A->isZero()) && // avoid infinite recursion.
- MaskedValueIsZero(Op0LHS, Mask)) {
- Value *NewNeg = Builder->CreateNeg(Op0RHS);
- return BinaryOperator::CreateAnd(NewNeg, AndRHS);
- }
- }
- break;
-
- case Instruction::Shl:
- case Instruction::LShr:
- // (1 << x) & 1 --> zext(x == 0)
- // (1 >> x) & 1 --> zext(x == 0)
- if (AndRHSMask == 1 && Op0LHS == AndRHS) {
- Value *NewICmp =
- Builder->CreateICmpEQ(Op0RHS, Constant::getNullValue(I.getType()));
- return new ZExtInst(NewICmp, I.getType());
- }
- break;
- }
-
- if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
- if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
- return Res;
- } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
- // If this is an integer truncation or change from signed-to-unsigned, and
- // if the source is an and/or with immediate, transform it. This
- // frequently occurs for bitfield accesses.
- if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
- if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
- CastOp->getNumOperands() == 2)
- if (ConstantInt *AndCI =dyn_cast<ConstantInt>(CastOp->getOperand(1))){
- if (CastOp->getOpcode() == Instruction::And) {
- // Change: and (cast (and X, C1) to T), C2
- // into : and (cast X to T), trunc_or_bitcast(C1)&C2
- // This will fold the two constants together, which may allow
- // other simplifications.
- Value *NewCast = Builder->CreateTruncOrBitCast(
- CastOp->getOperand(0), I.getType(),
- CastOp->getName()+".shrunk");
- // trunc_or_bitcast(C1)&C2
- Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
- C3 = ConstantExpr::getAnd(C3, AndRHS);
- return BinaryOperator::CreateAnd(NewCast, C3);
- } else if (CastOp->getOpcode() == Instruction::Or) {
- // Change: and (cast (or X, C1) to T), C2
- // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
- Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
- if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS)
- // trunc(C1)&C2
- return ReplaceInstUsesWith(I, AndRHS);
- }
- }
- }
- }
-
- // Try to fold constant and into select arguments.
- if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
- if (Instruction *R = FoldOpIntoSelect(I, SI))
- return R;
- if (isa<PHINode>(Op0))
- if (Instruction *NV = FoldOpIntoPhi(I))
- return NV;
- }
-
-
- // (~A & ~B) == (~(A | B)) - De Morgan's Law
- if (Value *Op0NotVal = dyn_castNotVal(Op0))
- if (Value *Op1NotVal = dyn_castNotVal(Op1))
- if (Op0->hasOneUse() && Op1->hasOneUse()) {
- Value *Or = Builder->CreateOr(Op0NotVal, Op1NotVal,
- I.getName()+".demorgan");
- return BinaryOperator::CreateNot(Or);
- }
-
- {
- Value *A = 0, *B = 0, *C = 0, *D = 0;
- // (A|B) & ~(A&B) -> A^B
- if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
- match(Op1, m_Not(m_And(m_Value(C), m_Value(D)))) &&
- ((A == C && B == D) || (A == D && B == C)))
- return BinaryOperator::CreateXor(A, B);
-
- // ~(A&B) & (A|B) -> A^B
- if (match(Op1, m_Or(m_Value(A), m_Value(B))) &&
- match(Op0, m_Not(m_And(m_Value(C), m_Value(D)))) &&
- ((A == C && B == D) || (A == D && B == C)))
- return BinaryOperator::CreateXor(A, B);
-
- if (Op0->hasOneUse() &&
- match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
- if (A == Op1) { // (A^B)&A -> A&(A^B)
- I.swapOperands(); // Simplify below
- std::swap(Op0, Op1);
- } else if (B == Op1) { // (A^B)&B -> B&(B^A)
- cast<BinaryOperator>(Op0)->swapOperands();
- I.swapOperands(); // Simplify below
- std::swap(Op0, Op1);
- }
- }
-
- if (Op1->hasOneUse() &&
- match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
- if (B == Op0) { // B&(A^B) -> B&(B^A)
- cast<BinaryOperator>(Op1)->swapOperands();
- std::swap(A, B);
- }
- if (A == Op0) // A&(A^B) -> A & ~B
- return BinaryOperator::CreateAnd(A, Builder->CreateNot(B, "tmp"));
- }
-
- // (A&((~A)|B)) -> A&B
- if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A))) ||
- match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1)))))
- return BinaryOperator::CreateAnd(A, Op1);
- if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A))) ||
- match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0)))))
- return BinaryOperator::CreateAnd(A, Op0);
- }
-
- if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1))
- if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
- if (Instruction *Res = FoldAndOfICmps(I, LHS, RHS))
- return Res;
-
- // fold (and (cast A), (cast B)) -> (cast (and A, B))
- if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
- if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
- if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
- const Type *SrcTy = Op0C->getOperand(0)->getType();
- if (SrcTy == Op1C->getOperand(0)->getType() &&
- SrcTy->isIntOrIntVector() &&
- // Only do this if the casts both really cause code to be generated.
- ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
- I.getType()) &&
- ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
- I.getType())) {
- Value *NewOp = Builder->CreateAnd(Op0C->getOperand(0),
- Op1C->getOperand(0), I.getName());
- return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
- }
- }
-
- // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
- if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
- if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
- if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
- SI0->getOperand(1) == SI1->getOperand(1) &&
- (SI0->hasOneUse() || SI1->hasOneUse())) {
- Value *NewOp =
- Builder->CreateAnd(SI0->getOperand(0), SI1->getOperand(0),
- SI0->getName());
- return BinaryOperator::Create(SI1->getOpcode(), NewOp,
- SI1->getOperand(1));
- }
- }
-
- // If and'ing two fcmp, try combine them into one.
- if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
- if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
- if (Instruction *Res = FoldAndOfFCmps(I, LHS, RHS))
- return Res;
- }
-
- return Changed ? &I : 0;
-}
-
-/// CollectBSwapParts - Analyze the specified subexpression and see if it is
-/// capable of providing pieces of a bswap. The subexpression provides pieces
-/// of a bswap if it is proven that each of the non-zero bytes in the output of
-/// the expression came from the corresponding "byte swapped" byte in some other
-/// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
-/// we know that the expression deposits the low byte of %X into the high byte
-/// of the bswap result and that all other bytes are zero. This expression is
-/// accepted, the high byte of ByteValues is set to X to indicate a correct
-/// match.
-///
-/// This function returns true if the match was unsuccessful and false if so.
-/// On entry to the function the "OverallLeftShift" is a signed integer value
-/// indicating the number of bytes that the subexpression is later shifted. For
-/// example, if the expression is later right shifted by 16 bits, the
-/// OverallLeftShift value would be -2 on entry. This is used to specify which
-/// byte of ByteValues is actually being set.
-///
-/// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
-/// byte is masked to zero by a user. For example, in (X & 255), X will be
-/// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
-/// this function to working on up to 32-byte (256 bit) values. ByteMask is
-/// always in the local (OverallLeftShift) coordinate space.
-///
-static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
- SmallVector<Value*, 8> &ByteValues) {
- if (Instruction *I = dyn_cast<Instruction>(V)) {
- // If this is an or instruction, it may be an inner node of the bswap.
- if (I->getOpcode() == Instruction::Or) {
- return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
- ByteValues) ||
- CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
- ByteValues);
- }
-
- // If this is a logical shift by a constant multiple of 8, recurse with
- // OverallLeftShift and ByteMask adjusted.
- if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
- unsigned ShAmt =
- cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
- // Ensure the shift amount is defined and of a byte value.
- if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
- return true;
-
- unsigned ByteShift = ShAmt >> 3;
- if (I->getOpcode() == Instruction::Shl) {
- // X << 2 -> collect(X, +2)
- OverallLeftShift += ByteShift;
- ByteMask >>= ByteShift;
- } else {
- // X >>u 2 -> collect(X, -2)
- OverallLeftShift -= ByteShift;
- ByteMask <<= ByteShift;
- ByteMask &= (~0U >> (32-ByteValues.size()));
- }
-
- if (OverallLeftShift >= (int)ByteValues.size()) return true;
- if (OverallLeftShift <= -(int)ByteValues.size()) return true;
-
- return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
- ByteValues);
- }
-
- // If this is a logical 'and' with a mask that clears bytes, clear the
- // corresponding bytes in ByteMask.
- if (I->getOpcode() == Instruction::And &&
- isa<ConstantInt>(I->getOperand(1))) {
- // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
- unsigned NumBytes = ByteValues.size();
- APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
- const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
-
- for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
- // If this byte is masked out by a later operation, we don't care what
- // the and mask is.
- if ((ByteMask & (1 << i)) == 0)
- continue;
-
- // If the AndMask is all zeros for this byte, clear the bit.
- APInt MaskB = AndMask & Byte;
- if (MaskB == 0) {
- ByteMask &= ~(1U << i);
- continue;
- }
-
- // If the AndMask is not all ones for this byte, it's not a bytezap.
- if (MaskB != Byte)
- return true;
-
- // Otherwise, this byte is kept.
- }
-
- return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
- ByteValues);
- }
- }
-
- // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
- // the input value to the bswap. Some observations: 1) if more than one byte
- // is demanded from this input, then it could not be successfully assembled
- // into a byteswap. At least one of the two bytes would not be aligned with
- // their ultimate destination.
- if (!isPowerOf2_32(ByteMask)) return true;
- unsigned InputByteNo = CountTrailingZeros_32(ByteMask);
-
- // 2) The input and ultimate destinations must line up: if byte 3 of an i32
- // is demanded, it needs to go into byte 0 of the result. This means that the
- // byte needs to be shifted until it lands in the right byte bucket. The
- // shift amount depends on the position: if the byte is coming from the high
- // part of the value (e.g. byte 3) then it must be shifted right. If from the
- // low part, it must be shifted left.
- unsigned DestByteNo = InputByteNo + OverallLeftShift;
- if (InputByteNo < ByteValues.size()/2) {
- if (ByteValues.size()-1-DestByteNo != InputByteNo)
- return true;
- } else {
- if (ByteValues.size()-1-DestByteNo != InputByteNo)
- return true;
- }
-
- // If the destination byte value is already defined, the values are or'd
- // together, which isn't a bswap (unless it's an or of the same bits).
- if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
- return true;
- ByteValues[DestByteNo] = V;
- return false;
-}
-
-/// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
-/// If so, insert the new bswap intrinsic and return it.
-Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
- const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
- if (!ITy || ITy->getBitWidth() % 16 ||
- // ByteMask only allows up to 32-byte values.
- ITy->getBitWidth() > 32*8)
- return 0; // Can only bswap pairs of bytes. Can't do vectors.
-
- /// ByteValues - For each byte of the result, we keep track of which value
- /// defines each byte.
- SmallVector<Value*, 8> ByteValues;
- ByteValues.resize(ITy->getBitWidth()/8);
-
- // Try to find all the pieces corresponding to the bswap.
- uint32_t ByteMask = ~0U >> (32-ByteValues.size());
- if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
- return 0;
-
- // Check to see if all of the bytes come from the same value.
- Value *V = ByteValues[0];
- if (V == 0) return 0; // Didn't find a byte? Must be zero.
-
- // Check to make sure that all of the bytes come from the same value.
- for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
- if (ByteValues[i] != V)
- return 0;
- const Type *Tys[] = { ITy };
- Module *M = I.getParent()->getParent()->getParent();
- Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
- return CallInst::Create(F, V);
-}
-
-/// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
-/// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
-/// we can simplify this expression to "cond ? C : D or B".
-static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
- Value *C, Value *D) {
- // If A is not a select of -1/0, this cannot match.
- Value *Cond = 0;
- if (!match(A, m_SelectCst<-1, 0>(m_Value(Cond))))
- return 0;
-
- // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
- if (match(D, m_SelectCst<0, -1>(m_Specific(Cond))))
- return SelectInst::Create(Cond, C, B);
- if (match(D, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond)))))
- return SelectInst::Create(Cond, C, B);
- // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
- if (match(B, m_SelectCst<0, -1>(m_Specific(Cond))))
- return SelectInst::Create(Cond, C, D);
- if (match(B, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond)))))
- return SelectInst::Create(Cond, C, D);
- return 0;
-}
-
-/// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
-Instruction *InstCombiner::FoldOrOfICmps(Instruction &I,
- ICmpInst *LHS, ICmpInst *RHS) {
- ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
-
- // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
- if (PredicatesFoldable(LHSCC, RHSCC)) {
- if (LHS->getOperand(0) == RHS->getOperand(1) &&
- LHS->getOperand(1) == RHS->getOperand(0))
- LHS->swapOperands();
- if (LHS->getOperand(0) == RHS->getOperand(0) &&
- LHS->getOperand(1) == RHS->getOperand(1)) {
- Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
- unsigned Code = getICmpCode(LHS) | getICmpCode(RHS);
- bool isSigned = LHS->isSigned() || RHS->isSigned();
- Value *RV = getICmpValue(isSigned, Code, Op0, Op1);
- if (Instruction *I = dyn_cast<Instruction>(RV))
- return I;
- // Otherwise, it's a constant boolean value.
- return ReplaceInstUsesWith(I, RV);
- }
- }
-
- // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
- Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0);
- ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1));
- ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1));
- if (LHSCst == 0 || RHSCst == 0) return 0;
-
- // (icmp ne A, 0) | (icmp ne B, 0) --> (icmp ne (A|B), 0)
- if (LHSCst == RHSCst && LHSCC == RHSCC &&
- LHSCC == ICmpInst::ICMP_NE && LHSCst->isZero()) {
- Value *NewOr = Builder->CreateOr(Val, Val2);
- return new ICmpInst(LHSCC, NewOr, LHSCst);
- }
-
- // From here on, we only handle:
- // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
- if (Val != Val2) return 0;
-
- // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
- if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
- RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
- LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
- RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
- return 0;
-
- // We can't fold (ugt x, C) | (sgt x, C2).
- if (!PredicatesFoldable(LHSCC, RHSCC))
- return 0;
-
- // Ensure that the larger constant is on the RHS.
- bool ShouldSwap;
- if (CmpInst::isSigned(LHSCC) ||
- (ICmpInst::isEquality(LHSCC) &&
- CmpInst::isSigned(RHSCC)))
- ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
- else
- ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
-
- if (ShouldSwap) {
- std::swap(LHS, RHS);
- std::swap(LHSCst, RHSCst);
- std::swap(LHSCC, RHSCC);
- }
-
- // At this point, we know we have have two icmp instructions
- // comparing a value against two constants and or'ing the result
- // together. Because of the above check, we know that we only have
- // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
- // icmp folding check above), that the two constants are not
- // equal.
- assert(LHSCst != RHSCst && "Compares not folded above?");
-
- switch (LHSCC) {
- default: llvm_unreachable("Unknown integer condition code!");
- case ICmpInst::ICMP_EQ:
- switch (RHSCC) {
- default: llvm_unreachable("Unknown integer condition code!");
- case ICmpInst::ICMP_EQ:
- if (LHSCst == SubOne(RHSCst)) {
- // (X == 13 | X == 14) -> X-13 <u 2
- Constant *AddCST = ConstantExpr::getNeg(LHSCst);
- Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
- AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
- return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
- }
- break; // (X == 13 | X == 15) -> no change
- case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
- case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
- break;
- case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
- case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
- case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
- return ReplaceInstUsesWith(I, RHS);
- }
- break;
- case ICmpInst::ICMP_NE:
- switch (RHSCC) {
- default: llvm_unreachable("Unknown integer condition code!");
- case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
- case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
- case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
- return ReplaceInstUsesWith(I, LHS);
- case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
- case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
- case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
- return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
- }
- break;
- case ICmpInst::ICMP_ULT:
- switch (RHSCC) {
- default: llvm_unreachable("Unknown integer condition code!");
- case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
- break;
- case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
- // If RHSCst is [us]MAXINT, it is always false. Not handling
- // this can cause overflow.
- if (RHSCst->isMaxValue(false))
- return ReplaceInstUsesWith(I, LHS);
- return InsertRangeTest(Val, LHSCst, AddOne(RHSCst),
- false, false, I);
- case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
- break;
- case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
- case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
- return ReplaceInstUsesWith(I, RHS);
- case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
- break;
- }
- break;
- case ICmpInst::ICMP_SLT:
- switch (RHSCC) {
- default: llvm_unreachable("Unknown integer condition code!");
- case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
- break;
- case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
- // If RHSCst is [us]MAXINT, it is always false. Not handling
- // this can cause overflow.
- if (RHSCst->isMaxValue(true))
- return ReplaceInstUsesWith(I, LHS);
- return InsertRangeTest(Val, LHSCst, AddOne(RHSCst),
- true, false, I);
- case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
- break;
- case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
- case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
- return ReplaceInstUsesWith(I, RHS);
- case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
- break;
- }
- break;
- case ICmpInst::ICMP_UGT:
- switch (RHSCC) {
- default: llvm_unreachable("Unknown integer condition code!");
- case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
- case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
- return ReplaceInstUsesWith(I, LHS);
- case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
- break;
- case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
- case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
- return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
- case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
- break;
- }
- break;
- case ICmpInst::ICMP_SGT:
- switch (RHSCC) {
- default: llvm_unreachable("Unknown integer condition code!");
- case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
- case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
- return ReplaceInstUsesWith(I, LHS);
- case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
- break;
- case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
- case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
- return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
- case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
- break;
- }
- break;
- }
- return 0;
-}
-
-Instruction *InstCombiner::FoldOrOfFCmps(Instruction &I, FCmpInst *LHS,
- FCmpInst *RHS) {
- if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
- RHS->getPredicate() == FCmpInst::FCMP_UNO &&
- LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
- if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
- if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
- // If either of the constants are nans, then the whole thing returns
- // true.
- if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
- return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
-
- // Otherwise, no need to compare the two constants, compare the
- // rest.
- return new FCmpInst(FCmpInst::FCMP_UNO,
- LHS->getOperand(0), RHS->getOperand(0));
- }
-
- // Handle vector zeros. This occurs because the canonical form of
- // "fcmp uno x,x" is "fcmp uno x, 0".
- if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
- isa<ConstantAggregateZero>(RHS->getOperand(1)))
- return new FCmpInst(FCmpInst::FCMP_UNO,
- LHS->getOperand(0), RHS->getOperand(0));
-
- return 0;
- }
-
- Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
- Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
- FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
-
- if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
- // Swap RHS operands to match LHS.
- Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
- std::swap(Op1LHS, Op1RHS);
- }
- if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
- // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
- if (Op0CC == Op1CC)
- return new FCmpInst((FCmpInst::Predicate)Op0CC,
- Op0LHS, Op0RHS);
- if (Op0CC == FCmpInst::FCMP_TRUE || Op1CC == FCmpInst::FCMP_TRUE)
- return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
- if (Op0CC == FCmpInst::FCMP_FALSE)
- return ReplaceInstUsesWith(I, RHS);
- if (Op1CC == FCmpInst::FCMP_FALSE)
- return ReplaceInstUsesWith(I, LHS);
- bool Op0Ordered;
- bool Op1Ordered;
- unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
- unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
- if (Op0Ordered == Op1Ordered) {
- // If both are ordered or unordered, return a new fcmp with
- // or'ed predicates.
- Value *RV = getFCmpValue(Op0Ordered, Op0Pred|Op1Pred, Op0LHS, Op0RHS);
- if (Instruction *I = dyn_cast<Instruction>(RV))
- return I;
- // Otherwise, it's a constant boolean value...
- return ReplaceInstUsesWith(I, RV);
- }
- }
- return 0;
-}
-
-/// FoldOrWithConstants - This helper function folds:
-///
-/// ((A | B) & C1) | (B & C2)
-///
-/// into:
-///
-/// (A & C1) | B
-///
-/// when the XOR of the two constants is "all ones" (-1).
-Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
- Value *A, Value *B, Value *C) {
- ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
- if (!CI1) return 0;
-
- Value *V1 = 0;
- ConstantInt *CI2 = 0;
- if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)))) return 0;
-
- APInt Xor = CI1->getValue() ^ CI2->getValue();
- if (!Xor.isAllOnesValue()) return 0;
-
- if (V1 == A || V1 == B) {
- Value *NewOp = Builder->CreateAnd((V1 == A) ? B : A, CI1);
- return BinaryOperator::CreateOr(NewOp, V1);
- }
-
- return 0;
-}
-
-Instruction *InstCombiner::visitOr(BinaryOperator &I) {
- bool Changed = SimplifyCommutative(I);
- Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
-
- if (Value *V = SimplifyOrInst(Op0, Op1, TD))
- return ReplaceInstUsesWith(I, V);
-
-
- // See if we can simplify any instructions used by the instruction whose sole
- // purpose is to compute bits we don't care about.
- if (SimplifyDemandedInstructionBits(I))
- return &I;
-
- if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
- ConstantInt *C1 = 0; Value *X = 0;
- // (X & C1) | C2 --> (X | C2) & (C1|C2)
- if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) &&
- Op0->hasOneUse()) {
- Value *Or = Builder->CreateOr(X, RHS);
- Or->takeName(Op0);
- return BinaryOperator::CreateAnd(Or,
- ConstantInt::get(I.getContext(),
- RHS->getValue() | C1->getValue()));
- }
-
- // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
- if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) &&
- Op0->hasOneUse()) {
- Value *Or = Builder->CreateOr(X, RHS);
- Or->takeName(Op0);
- return BinaryOperator::CreateXor(Or,
- ConstantInt::get(I.getContext(),
- C1->getValue() & ~RHS->getValue()));
- }
-
- // Try to fold constant and into select arguments.
- if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
- if (Instruction *R = FoldOpIntoSelect(I, SI))
- return R;
- if (isa<PHINode>(Op0))
- if (Instruction *NV = FoldOpIntoPhi(I))
- return NV;
- }
-
- Value *A = 0, *B = 0;
- ConstantInt *C1 = 0, *C2 = 0;
-
- // (A | B) | C and A | (B | C) -> bswap if possible.
- // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
- if (match(Op0, m_Or(m_Value(), m_Value())) ||
- match(Op1, m_Or(m_Value(), m_Value())) ||
- (match(Op0, m_Shift(m_Value(), m_Value())) &&
- match(Op1, m_Shift(m_Value(), m_Value())))) {
- if (Instruction *BSwap = MatchBSwap(I))
- return BSwap;
- }
-
- // (X^C)|Y -> (X|Y)^C iff Y&C == 0
- if (Op0->hasOneUse() &&
- match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
- MaskedValueIsZero(Op1, C1->getValue())) {
- Value *NOr = Builder->CreateOr(A, Op1);
- NOr->takeName(Op0);
- return BinaryOperator::CreateXor(NOr, C1);
- }
-
- // Y|(X^C) -> (X|Y)^C iff Y&C == 0
- if (Op1->hasOneUse() &&
- match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
- MaskedValueIsZero(Op0, C1->getValue())) {
- Value *NOr = Builder->CreateOr(A, Op0);
- NOr->takeName(Op0);
- return BinaryOperator::CreateXor(NOr, C1);
- }
-
- // (A & C)|(B & D)
- Value *C = 0, *D = 0;
- if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
- match(Op1, m_And(m_Value(B), m_Value(D)))) {
- Value *V1 = 0, *V2 = 0, *V3 = 0;
- C1 = dyn_cast<ConstantInt>(C);
- C2 = dyn_cast<ConstantInt>(D);
- if (C1 && C2) { // (A & C1)|(B & C2)
- // If we have: ((V + N) & C1) | (V & C2)
- // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
- // replace with V+N.
- if (C1->getValue() == ~C2->getValue()) {
- if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
- match(A, m_Add(m_Value(V1), m_Value(V2)))) {
- // Add commutes, try both ways.
- if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
- return ReplaceInstUsesWith(I, A);
- if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
- return ReplaceInstUsesWith(I, A);
- }
- // Or commutes, try both ways.
- if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
- match(B, m_Add(m_Value(V1), m_Value(V2)))) {
- // Add commutes, try both ways.
- if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
- return ReplaceInstUsesWith(I, B);
- if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
- return ReplaceInstUsesWith(I, B);
- }
- }
-
- // ((V | N) & C1) | (V & C2) --> (V|N) & (C1|C2)
- // iff (C1&C2) == 0 and (N&~C1) == 0
- if ((C1->getValue() & C2->getValue()) == 0) {
- if (match(A, m_Or(m_Value(V1), m_Value(V2))) &&
- ((V1 == B && MaskedValueIsZero(V2, ~C1->getValue())) || // (V|N)
- (V2 == B && MaskedValueIsZero(V1, ~C1->getValue())))) // (N|V)
- return BinaryOperator::CreateAnd(A,
- ConstantInt::get(A->getContext(),
- C1->getValue()|C2->getValue()));
- // Or commutes, try both ways.
- if (match(B, m_Or(m_Value(V1), m_Value(V2))) &&
- ((V1 == A && MaskedValueIsZero(V2, ~C2->getValue())) || // (V|N)
- (V2 == A && MaskedValueIsZero(V1, ~C2->getValue())))) // (N|V)
- return BinaryOperator::CreateAnd(B,
- ConstantInt::get(B->getContext(),
- C1->getValue()|C2->getValue()));
- }
- }
-
- // Check to see if we have any common things being and'ed. If so, find the
- // terms for V1 & (V2|V3).
- if (Op0->hasOneUse() || Op1->hasOneUse()) {
- V1 = 0;
- if (A == B) // (A & C)|(A & D) == A & (C|D)
- V1 = A, V2 = C, V3 = D;
- else if (A == D) // (A & C)|(B & A) == A & (B|C)
- V1 = A, V2 = B, V3 = C;
- else if (C == B) // (A & C)|(C & D) == C & (A|D)
- V1 = C, V2 = A, V3 = D;
- else if (C == D) // (A & C)|(B & C) == C & (A|B)
- V1 = C, V2 = A, V3 = B;
-
- if (V1) {
- Value *Or = Builder->CreateOr(V2, V3, "tmp");
- return BinaryOperator::CreateAnd(V1, Or);
- }
- }
-
- // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants
- if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D))
- return Match;
- if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C))
- return Match;
- if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D))
- return Match;
- if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C))
- return Match;
-
- // ((A&~B)|(~A&B)) -> A^B
- if ((match(C, m_Not(m_Specific(D))) &&
- match(B, m_Not(m_Specific(A)))))
- return BinaryOperator::CreateXor(A, D);
- // ((~B&A)|(~A&B)) -> A^B
- if ((match(A, m_Not(m_Specific(D))) &&
- match(B, m_Not(m_Specific(C)))))
- return BinaryOperator::CreateXor(C, D);
- // ((A&~B)|(B&~A)) -> A^B
- if ((match(C, m_Not(m_Specific(B))) &&
- match(D, m_Not(m_Specific(A)))))
- return BinaryOperator::CreateXor(A, B);
- // ((~B&A)|(B&~A)) -> A^B
- if ((match(A, m_Not(m_Specific(B))) &&
- match(D, m_Not(m_Specific(C)))))
- return BinaryOperator::CreateXor(C, B);
- }
-
- // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
- if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
- if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
- if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
- SI0->getOperand(1) == SI1->getOperand(1) &&
- (SI0->hasOneUse() || SI1->hasOneUse())) {
- Value *NewOp = Builder->CreateOr(SI0->getOperand(0), SI1->getOperand(0),
- SI0->getName());
- return BinaryOperator::Create(SI1->getOpcode(), NewOp,
- SI1->getOperand(1));
- }
- }
-
- // ((A|B)&1)|(B&-2) -> (A&1) | B
- if (match(Op0, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
- match(Op0, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
- Instruction *Ret = FoldOrWithConstants(I, Op1, A, B, C);
- if (Ret) return Ret;
- }
- // (B&-2)|((A|B)&1) -> (A&1) | B
- if (match(Op1, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
- match(Op1, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
- Instruction *Ret = FoldOrWithConstants(I, Op0, A, B, C);
- if (Ret) return Ret;
- }
-
- // (~A | ~B) == (~(A & B)) - De Morgan's Law
- if (Value *Op0NotVal = dyn_castNotVal(Op0))
- if (Value *Op1NotVal = dyn_castNotVal(Op1))
- if (Op0->hasOneUse() && Op1->hasOneUse()) {
- Value *And = Builder->CreateAnd(Op0NotVal, Op1NotVal,
- I.getName()+".demorgan");
- return BinaryOperator::CreateNot(And);
- }
-
- if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
- if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
- if (Instruction *Res = FoldOrOfICmps(I, LHS, RHS))
- return Res;
-
- // fold (or (cast A), (cast B)) -> (cast (or A, B))
- if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
- if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
- if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
- if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
- !isa<ICmpInst>(Op1C->getOperand(0))) {
- const Type *SrcTy = Op0C->getOperand(0)->getType();
- if (SrcTy == Op1C->getOperand(0)->getType() &&
- SrcTy->isIntOrIntVector() &&
- // Only do this if the casts both really cause code to be
- // generated.
- ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
- I.getType()) &&
- ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
- I.getType())) {
- Value *NewOp = Builder->CreateOr(Op0C->getOperand(0),
- Op1C->getOperand(0), I.getName());
- return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
- }
- }
- }
- }
-
-
- // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
- if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
- if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
- if (Instruction *Res = FoldOrOfFCmps(I, LHS, RHS))
- return Res;
- }
-
- return Changed ? &I : 0;
-}
-
-Instruction *InstCombiner::visitXor(BinaryOperator &I) {
- bool Changed = SimplifyCommutative(I);
- Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
-
- if (isa<UndefValue>(Op1)) {
- if (isa<UndefValue>(Op0))
- // Handle undef ^ undef -> 0 special case. This is a common
- // idiom (misuse).
- return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
- return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
- }
-
- // xor X, X = 0
- if (Op0 == Op1)
- return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
-
- // See if we can simplify any instructions used by the instruction whose sole
- // purpose is to compute bits we don't care about.
- if (SimplifyDemandedInstructionBits(I))
- return &I;
- if (isa<VectorType>(I.getType()))
- if (isa<ConstantAggregateZero>(Op1))
- return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
-
- // Is this a ~ operation?
- if (Value *NotOp = dyn_castNotVal(&I)) {
- if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
- if (Op0I->getOpcode() == Instruction::And ||
- Op0I->getOpcode() == Instruction::Or) {
- // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
- // ~(~X | Y) === (X & ~Y) - De Morgan's Law
- if (dyn_castNotVal(Op0I->getOperand(1)))
- Op0I->swapOperands();
- if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
- Value *NotY =
- Builder->CreateNot(Op0I->getOperand(1),
- Op0I->getOperand(1)->getName()+".not");
- if (Op0I->getOpcode() == Instruction::And)
- return BinaryOperator::CreateOr(Op0NotVal, NotY);
- return BinaryOperator::CreateAnd(Op0NotVal, NotY);
- }
-
- // ~(X & Y) --> (~X | ~Y) - De Morgan's Law
- // ~(X | Y) === (~X & ~Y) - De Morgan's Law
- if (isFreeToInvert(Op0I->getOperand(0)) &&
- isFreeToInvert(Op0I->getOperand(1))) {
- Value *NotX =
- Builder->CreateNot(Op0I->getOperand(0), "notlhs");
- Value *NotY =
- Builder->CreateNot(Op0I->getOperand(1), "notrhs");
- if (Op0I->getOpcode() == Instruction::And)
- return BinaryOperator::CreateOr(NotX, NotY);
- return BinaryOperator::CreateAnd(NotX, NotY);
- }
- }
- }
- }
-
-
- if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
- if (RHS->isOne() && Op0->hasOneUse()) {
- // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
- if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
- return new ICmpInst(ICI->getInversePredicate(),
- ICI->getOperand(0), ICI->getOperand(1));
-
- if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
- return new FCmpInst(FCI->getInversePredicate(),
- FCI->getOperand(0), FCI->getOperand(1));
- }
-
- // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
- if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
- if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
- if (CI->hasOneUse() && Op0C->hasOneUse()) {
- Instruction::CastOps Opcode = Op0C->getOpcode();
- if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) &&
- (RHS == ConstantExpr::getCast(Opcode,
- ConstantInt::getTrue(I.getContext()),
- Op0C->getDestTy()))) {
- CI->setPredicate(CI->getInversePredicate());
- return CastInst::Create(Opcode, CI, Op0C->getType());
- }
- }
- }
- }
-
- if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
- // ~(c-X) == X-c-1 == X+(-c-1)
- if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
- if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
- Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
- Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
- ConstantInt::get(I.getType(), 1));
- return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
- }
-
- if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
- if (Op0I->getOpcode() == Instruction::Add) {
- // ~(X-c) --> (-c-1)-X
- if (RHS->isAllOnesValue()) {
- Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
- return BinaryOperator::CreateSub(
- ConstantExpr::getSub(NegOp0CI,
- ConstantInt::get(I.getType(), 1)),
- Op0I->getOperand(0));
- } else if (RHS->getValue().isSignBit()) {
- // (X + C) ^ signbit -> (X + C + signbit)
- Constant *C = ConstantInt::get(I.getContext(),
- RHS->getValue() + Op0CI->getValue());
- return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
-
- }
- } else if (Op0I->getOpcode() == Instruction::Or) {
- // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
- if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
- Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
- // Anything in both C1 and C2 is known to be zero, remove it from
- // NewRHS.
- Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
- NewRHS = ConstantExpr::getAnd(NewRHS,
- ConstantExpr::getNot(CommonBits));
- Worklist.Add(Op0I);
- I.setOperand(0, Op0I->getOperand(0));
- I.setOperand(1, NewRHS);
- return &I;
- }
- }
- }
- }
-
- // Try to fold constant and into select arguments.
- if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
- if (Instruction *R = FoldOpIntoSelect(I, SI))
- return R;
- if (isa<PHINode>(Op0))
- if (Instruction *NV = FoldOpIntoPhi(I))
- return NV;
- }
-
- if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
- if (X == Op1)
- return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
+ return 0;
+}
- if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
- if (X == Op0)
- return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
+/// FoldOpIntoPhi - Given a binary operator, cast instruction, or select which
+/// 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).
+///
+Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
+ PHINode *PN = cast<PHINode>(I.getOperand(0));
+ unsigned NumPHIValues = PN->getNumIncomingValues();
+ if (NumPHIValues == 0)
+ return 0;
- BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
- if (Op1I) {
- Value *A, *B;
- if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
- if (A == Op0) { // B^(B|A) == (A|B)^B
- Op1I->swapOperands();
- I.swapOperands();
- std::swap(Op0, Op1);
- } else if (B == Op0) { // B^(A|B) == (A|B)^B
- I.swapOperands(); // Simplified below.
- std::swap(Op0, Op1);
- }
- } else if (match(Op1I, m_Xor(m_Specific(Op0), m_Value(B)))) {
- return ReplaceInstUsesWith(I, B); // A^(A^B) == B
- } else if (match(Op1I, m_Xor(m_Value(A), m_Specific(Op0)))) {
- return ReplaceInstUsesWith(I, A); // A^(B^A) == B
- } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) &&
- Op1I->hasOneUse()){
- if (A == Op0) { // A^(A&B) -> A^(B&A)
- Op1I->swapOperands();
- std::swap(A, B);
- }
- if (B == Op0) { // A^(B&A) -> (B&A)^A
- I.swapOperands(); // Simplified below.
- std::swap(Op0, Op1);
- }
+ // 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.
}
- BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
- if (Op0I) {
- Value *A, *B;
- if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
- Op0I->hasOneUse()) {
- if (A == Op1) // (B|A)^B == (A|B)^B
- std::swap(A, B);
- if (B == Op1) // (A|B)^B == A & ~B
- return BinaryOperator::CreateAnd(A, Builder->CreateNot(Op1, "tmp"));
- } else if (match(Op0I, m_Xor(m_Specific(Op1), m_Value(B)))) {
- return ReplaceInstUsesWith(I, B); // (A^B)^A == B
- } else if (match(Op0I, m_Xor(m_Value(A), m_Specific(Op1)))) {
- return ReplaceInstUsesWith(I, A); // (B^A)^A == B
- } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
- Op0I->hasOneUse()){
- if (A == Op1) // (A&B)^A -> (B&A)^A
- std::swap(A, B);
- if (B == Op1 && // (B&A)^A == ~B & A
- !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
- return BinaryOperator::CreateAnd(Builder->CreateNot(A, "tmp"), Op1);
- }
- }
+ // 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 = 0;
+ 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;
}
- // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
- if (Op0I && Op1I && Op0I->isShift() &&
- Op0I->getOpcode() == Op1I->getOpcode() &&
- Op0I->getOperand(1) == Op1I->getOperand(1) &&
- (Op1I->hasOneUse() || Op1I->hasOneUse())) {
- Value *NewOp =
- Builder->CreateXor(Op0I->getOperand(0), Op1I->getOperand(0),
- Op0I->getName());
- return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
- Op1I->getOperand(1));
+ // 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) {
+ BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
+ if (!BI || !BI->isUnconditional()) return 0;
}
-
- if (Op0I && Op1I) {
- Value *A, *B, *C, *D;
- // (A & B)^(A | B) -> A ^ B
- if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
- match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
- if ((A == C && B == D) || (A == D && B == C))
- return BinaryOperator::CreateXor(A, B);
+
+ // Okay, we can do the transformation: create the new PHI node.
+ PHINode *NewPN = PHINode::Create(I.getType(), PN->getNumIncomingValues());
+ InsertNewInstBefore(NewPN, *PN);
+ NewPN->takeName(PN);
+
+ // If we are going to have to insert a new computation, do so right before the
+ // 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,
+ // 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 = 0;
+ if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
+ InV = InC->isNullValue() ? FalseVInPred : TrueVInPred;
+ else
+ InV = Builder->CreateSelect(PN->getIncomingValue(i),
+ TrueVInPred, FalseVInPred, "phitmp");
+ NewPN->addIncoming(InV, ThisBB);
}
- // (A | B)^(A & B) -> A ^ B
- if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
- match(Op1I, m_And(m_Value(C), m_Value(D)))) {
- if ((A == C && B == D) || (A == D && B == C))
- return BinaryOperator::CreateXor(A, B);
+ } 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));
}
-
- // (A & B)^(C & D)
- if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
- match(Op0I, m_And(m_Value(A), m_Value(B))) &&
- match(Op1I, m_And(m_Value(C), m_Value(D)))) {
- // (X & Y)^(X & Y) -> (Y^Z) & X
- Value *X = 0, *Y = 0, *Z = 0;
- if (A == C)
- X = A, Y = B, Z = D;
- else if (A == D)
- X = A, Y = B, Z = C;
- else if (B == C)
- X = B, Y = A, Z = D;
- else if (B == D)
- X = B, Y = A, Z = C;
-
- if (X) {
- Value *NewOp = Builder->CreateXor(Y, Z, Op0->getName());
- return BinaryOperator::CreateAnd(NewOp, X);
- }
+ } 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)))
+ 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);
+ Type *RetTy = CI->getType();
+ for (unsigned i = 0; i != NumPHIValues; ++i) {
+ Value *InV;
+ if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
+ InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
+ else
+ InV = Builder->CreateCast(CI->getOpcode(),
+ PN->getIncomingValue(i), I.getType(), "phitmp");
+ NewPN->addIncoming(InV, PN->getIncomingBlock(i));
}
}
-
- // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
- if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
- if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
- if (PredicatesFoldable(LHS->getPredicate(), RHS->getPredicate())) {
- if (LHS->getOperand(0) == RHS->getOperand(1) &&
- LHS->getOperand(1) == RHS->getOperand(0))
- LHS->swapOperands();
- if (LHS->getOperand(0) == RHS->getOperand(0) &&
- LHS->getOperand(1) == RHS->getOperand(1)) {
- Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
- unsigned Code = getICmpCode(LHS) ^ getICmpCode(RHS);
- bool isSigned = LHS->isSigned() || RHS->isSigned();
- Value *RV = getICmpValue(isSigned, Code, Op0, Op1);
- if (Instruction *I = dyn_cast<Instruction>(RV))
- return I;
- // Otherwise, it's a constant boolean value.
- return ReplaceInstUsesWith(I, RV);
- }
- }
-
- // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
- if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
- if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
- if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
- const Type *SrcTy = Op0C->getOperand(0)->getType();
- if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
- // Only do this if the casts both really cause code to be generated.
- ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
- I.getType()) &&
- ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
- I.getType())) {
- Value *NewOp = Builder->CreateXor(Op0C->getOperand(0),
- Op1C->getOperand(0), I.getName());
- return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
- }
- }
+
+ 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 Changed ? &I : 0;
+ return ReplaceInstUsesWith(I, NewPN);
}
-
-
-
/// FindElementAtOffset - Given a type and a constant offset, determine whether
/// 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++))