1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // InstructionCombining - Combine instructions to form fewer, simple
11 // instructions. This pass does not modify the CFG. This pass is where
12 // algebraic simplification happens.
14 // This pass combines things like:
20 // This is a simple worklist driven algorithm.
22 // This pass guarantees that the following canonicalizations are performed on
24 // 1. If a binary operator has a constant operand, it is moved to the RHS
25 // 2. Bitwise operators with constant operands are always grouped so that
26 // shifts are performed first, then or's, then and's, then xor's.
27 // 3. Compare instructions are converted from <,>,<=,>= to ==,!= if possible
28 // 4. All cmp instructions on boolean values are replaced with logical ops
29 // 5. add X, X is represented as (X*2) => (X << 1)
30 // 6. Multiplies with a power-of-two constant argument are transformed into
34 //===----------------------------------------------------------------------===//
36 #include "llvm/Transforms/Scalar.h"
37 #include "InstCombine.h"
38 #include "llvm-c/Initialization.h"
39 #include "llvm/ADT/SmallPtrSet.h"
40 #include "llvm/ADT/Statistic.h"
41 #include "llvm/ADT/StringSwitch.h"
42 #include "llvm/Analysis/AssumptionTracker.h"
43 #include "llvm/Analysis/CFG.h"
44 #include "llvm/Analysis/ConstantFolding.h"
45 #include "llvm/Analysis/InstructionSimplify.h"
46 #include "llvm/Analysis/LoopInfo.h"
47 #include "llvm/Analysis/MemoryBuiltins.h"
48 #include "llvm/Analysis/ValueTracking.h"
49 #include "llvm/IR/CFG.h"
50 #include "llvm/IR/DataLayout.h"
51 #include "llvm/IR/Dominators.h"
52 #include "llvm/IR/GetElementPtrTypeIterator.h"
53 #include "llvm/IR/IntrinsicInst.h"
54 #include "llvm/IR/PatternMatch.h"
55 #include "llvm/IR/ValueHandle.h"
56 #include "llvm/Support/CommandLine.h"
57 #include "llvm/Support/Debug.h"
58 #include "llvm/Target/TargetLibraryInfo.h"
59 #include "llvm/Transforms/Utils/Local.h"
63 using namespace llvm::PatternMatch;
65 #define DEBUG_TYPE "instcombine"
67 STATISTIC(NumCombined , "Number of insts combined");
68 STATISTIC(NumConstProp, "Number of constant folds");
69 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
70 STATISTIC(NumSunkInst , "Number of instructions sunk");
71 STATISTIC(NumExpand, "Number of expansions");
72 STATISTIC(NumFactor , "Number of factorizations");
73 STATISTIC(NumReassoc , "Number of reassociations");
75 // Initialization Routines
76 void llvm::initializeInstCombine(PassRegistry &Registry) {
77 initializeInstCombinerPass(Registry);
80 void LLVMInitializeInstCombine(LLVMPassRegistryRef R) {
81 initializeInstCombine(*unwrap(R));
84 char InstCombiner::ID = 0;
85 INITIALIZE_PASS_BEGIN(InstCombiner, "instcombine",
86 "Combine redundant instructions", false, false)
87 INITIALIZE_PASS_DEPENDENCY(AssumptionTracker)
88 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
89 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
90 INITIALIZE_PASS_END(InstCombiner, "instcombine",
91 "Combine redundant instructions", false, false)
93 void InstCombiner::getAnalysisUsage(AnalysisUsage &AU) const {
95 AU.addRequired<AssumptionTracker>();
96 AU.addRequired<TargetLibraryInfo>();
97 AU.addRequired<DominatorTreeWrapperPass>();
98 AU.addPreserved<DominatorTreeWrapperPass>();
102 Value *InstCombiner::EmitGEPOffset(User *GEP) {
103 return llvm::EmitGEPOffset(Builder, *getDataLayout(), GEP);
106 /// ShouldChangeType - Return true if it is desirable to convert a computation
107 /// from 'From' to 'To'. We don't want to convert from a legal to an illegal
108 /// type for example, or from a smaller to a larger illegal type.
109 bool InstCombiner::ShouldChangeType(Type *From, Type *To) const {
110 assert(From->isIntegerTy() && To->isIntegerTy());
112 // If we don't have DL, we don't know if the source/dest are legal.
113 if (!DL) return false;
115 unsigned FromWidth = From->getPrimitiveSizeInBits();
116 unsigned ToWidth = To->getPrimitiveSizeInBits();
117 bool FromLegal = DL->isLegalInteger(FromWidth);
118 bool ToLegal = DL->isLegalInteger(ToWidth);
120 // If this is a legal integer from type, and the result would be an illegal
121 // type, don't do the transformation.
122 if (FromLegal && !ToLegal)
125 // Otherwise, if both are illegal, do not increase the size of the result. We
126 // do allow things like i160 -> i64, but not i64 -> i160.
127 if (!FromLegal && !ToLegal && ToWidth > FromWidth)
133 // Return true, if No Signed Wrap should be maintained for I.
134 // The No Signed Wrap flag can be kept if the operation "B (I.getOpcode) C",
135 // where both B and C should be ConstantInts, results in a constant that does
136 // not overflow. This function only handles the Add and Sub opcodes. For
137 // all other opcodes, the function conservatively returns false.
138 static bool MaintainNoSignedWrap(BinaryOperator &I, Value *B, Value *C) {
139 OverflowingBinaryOperator *OBO = dyn_cast<OverflowingBinaryOperator>(&I);
140 if (!OBO || !OBO->hasNoSignedWrap()) {
144 // We reason about Add and Sub Only.
145 Instruction::BinaryOps Opcode = I.getOpcode();
146 if (Opcode != Instruction::Add &&
147 Opcode != Instruction::Sub) {
151 ConstantInt *CB = dyn_cast<ConstantInt>(B);
152 ConstantInt *CC = dyn_cast<ConstantInt>(C);
158 const APInt &BVal = CB->getValue();
159 const APInt &CVal = CC->getValue();
160 bool Overflow = false;
162 if (Opcode == Instruction::Add) {
163 BVal.sadd_ov(CVal, Overflow);
165 BVal.ssub_ov(CVal, Overflow);
171 /// Conservatively clears subclassOptionalData after a reassociation or
172 /// commutation. We preserve fast-math flags when applicable as they can be
174 static void ClearSubclassDataAfterReassociation(BinaryOperator &I) {
175 FPMathOperator *FPMO = dyn_cast<FPMathOperator>(&I);
177 I.clearSubclassOptionalData();
181 FastMathFlags FMF = I.getFastMathFlags();
182 I.clearSubclassOptionalData();
183 I.setFastMathFlags(FMF);
186 /// SimplifyAssociativeOrCommutative - This performs a few simplifications for
187 /// operators which are associative or commutative:
189 // Commutative operators:
191 // 1. Order operands such that they are listed from right (least complex) to
192 // left (most complex). This puts constants before unary operators before
195 // Associative operators:
197 // 2. Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
198 // 3. Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
200 // Associative and commutative operators:
202 // 4. Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
203 // 5. Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
204 // 6. Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
205 // if C1 and C2 are constants.
207 bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) {
208 Instruction::BinaryOps Opcode = I.getOpcode();
209 bool Changed = false;
212 // Order operands such that they are listed from right (least complex) to
213 // left (most complex). This puts constants before unary operators before
215 if (I.isCommutative() && getComplexity(I.getOperand(0)) <
216 getComplexity(I.getOperand(1)))
217 Changed = !I.swapOperands();
219 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
220 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
222 if (I.isAssociative()) {
223 // Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
224 if (Op0 && Op0->getOpcode() == Opcode) {
225 Value *A = Op0->getOperand(0);
226 Value *B = Op0->getOperand(1);
227 Value *C = I.getOperand(1);
229 // Does "B op C" simplify?
230 if (Value *V = SimplifyBinOp(Opcode, B, C, DL)) {
231 // It simplifies to V. Form "A op V".
234 // Conservatively clear the optional flags, since they may not be
235 // preserved by the reassociation.
236 if (MaintainNoSignedWrap(I, B, C) &&
237 (!Op0 || (isa<BinaryOperator>(Op0) && Op0->hasNoSignedWrap()))) {
238 // Note: this is only valid because SimplifyBinOp doesn't look at
239 // the operands to Op0.
240 I.clearSubclassOptionalData();
241 I.setHasNoSignedWrap(true);
243 ClearSubclassDataAfterReassociation(I);
252 // Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
253 if (Op1 && Op1->getOpcode() == Opcode) {
254 Value *A = I.getOperand(0);
255 Value *B = Op1->getOperand(0);
256 Value *C = Op1->getOperand(1);
258 // Does "A op B" simplify?
259 if (Value *V = SimplifyBinOp(Opcode, A, B, DL)) {
260 // It simplifies to V. Form "V op C".
263 // Conservatively clear the optional flags, since they may not be
264 // preserved by the reassociation.
265 ClearSubclassDataAfterReassociation(I);
273 if (I.isAssociative() && I.isCommutative()) {
274 // Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
275 if (Op0 && Op0->getOpcode() == Opcode) {
276 Value *A = Op0->getOperand(0);
277 Value *B = Op0->getOperand(1);
278 Value *C = I.getOperand(1);
280 // Does "C op A" simplify?
281 if (Value *V = SimplifyBinOp(Opcode, C, A, DL)) {
282 // It simplifies to V. Form "V op B".
285 // Conservatively clear the optional flags, since they may not be
286 // preserved by the reassociation.
287 ClearSubclassDataAfterReassociation(I);
294 // Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
295 if (Op1 && Op1->getOpcode() == Opcode) {
296 Value *A = I.getOperand(0);
297 Value *B = Op1->getOperand(0);
298 Value *C = Op1->getOperand(1);
300 // Does "C op A" simplify?
301 if (Value *V = SimplifyBinOp(Opcode, C, A, DL)) {
302 // It simplifies to V. Form "B op V".
305 // Conservatively clear the optional flags, since they may not be
306 // preserved by the reassociation.
307 ClearSubclassDataAfterReassociation(I);
314 // Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
315 // if C1 and C2 are constants.
317 Op0->getOpcode() == Opcode && Op1->getOpcode() == Opcode &&
318 isa<Constant>(Op0->getOperand(1)) &&
319 isa<Constant>(Op1->getOperand(1)) &&
320 Op0->hasOneUse() && Op1->hasOneUse()) {
321 Value *A = Op0->getOperand(0);
322 Constant *C1 = cast<Constant>(Op0->getOperand(1));
323 Value *B = Op1->getOperand(0);
324 Constant *C2 = cast<Constant>(Op1->getOperand(1));
326 Constant *Folded = ConstantExpr::get(Opcode, C1, C2);
327 BinaryOperator *New = BinaryOperator::Create(Opcode, A, B);
328 if (isa<FPMathOperator>(New)) {
329 FastMathFlags Flags = I.getFastMathFlags();
330 Flags &= Op0->getFastMathFlags();
331 Flags &= Op1->getFastMathFlags();
332 New->setFastMathFlags(Flags);
334 InsertNewInstWith(New, I);
336 I.setOperand(0, New);
337 I.setOperand(1, Folded);
338 // Conservatively clear the optional flags, since they may not be
339 // preserved by the reassociation.
340 ClearSubclassDataAfterReassociation(I);
347 // No further simplifications.
352 /// LeftDistributesOverRight - Whether "X LOp (Y ROp Z)" is always equal to
353 /// "(X LOp Y) ROp (X LOp Z)".
354 static bool LeftDistributesOverRight(Instruction::BinaryOps LOp,
355 Instruction::BinaryOps ROp) {
360 case Instruction::And:
361 // And distributes over Or and Xor.
365 case Instruction::Or:
366 case Instruction::Xor:
370 case Instruction::Mul:
371 // Multiplication distributes over addition and subtraction.
375 case Instruction::Add:
376 case Instruction::Sub:
380 case Instruction::Or:
381 // Or distributes over And.
385 case Instruction::And:
391 /// RightDistributesOverLeft - Whether "(X LOp Y) ROp Z" is always equal to
392 /// "(X ROp Z) LOp (Y ROp Z)".
393 static bool RightDistributesOverLeft(Instruction::BinaryOps LOp,
394 Instruction::BinaryOps ROp) {
395 if (Instruction::isCommutative(ROp))
396 return LeftDistributesOverRight(ROp, LOp);
401 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
402 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
403 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
404 case Instruction::And:
405 case Instruction::Or:
406 case Instruction::Xor:
410 case Instruction::Shl:
411 case Instruction::LShr:
412 case Instruction::AShr:
416 // TODO: It would be nice to handle division, aka "(X + Y)/Z = X/Z + Y/Z",
417 // but this requires knowing that the addition does not overflow and other
422 /// This function returns identity value for given opcode, which can be used to
423 /// factor patterns like (X * 2) + X ==> (X * 2) + (X * 1) ==> X * (2 + 1).
424 static Value *getIdentityValue(Instruction::BinaryOps OpCode, Value *V) {
425 if (isa<Constant>(V))
428 if (OpCode == Instruction::Mul)
429 return ConstantInt::get(V->getType(), 1);
431 // TODO: We can handle other cases e.g. Instruction::And, Instruction::Or etc.
436 /// This function factors binary ops which can be combined using distributive
437 /// laws. This function tries to transform 'Op' based TopLevelOpcode to enable
438 /// factorization e.g for ADD(SHL(X , 2), MUL(X, 5)), When this function called
439 /// with TopLevelOpcode == Instruction::Add and Op = SHL(X, 2), transforms
440 /// SHL(X, 2) to MUL(X, 4) i.e. returns Instruction::Mul with LHS set to 'X' and
442 static Instruction::BinaryOps
443 getBinOpsForFactorization(Instruction::BinaryOps TopLevelOpcode,
444 BinaryOperator *Op, Value *&LHS, Value *&RHS) {
446 return Instruction::BinaryOpsEnd;
448 LHS = Op->getOperand(0);
449 RHS = Op->getOperand(1);
451 switch (TopLevelOpcode) {
453 return Op->getOpcode();
455 case Instruction::Add:
456 case Instruction::Sub:
457 if (Op->getOpcode() == Instruction::Shl) {
458 if (Constant *CST = dyn_cast<Constant>(Op->getOperand(1))) {
459 // The multiplier is really 1 << CST.
460 RHS = ConstantExpr::getShl(ConstantInt::get(Op->getType(), 1), CST);
461 return Instruction::Mul;
464 return Op->getOpcode();
467 // TODO: We can add other conversions e.g. shr => div etc.
470 /// This tries to simplify binary operations by factorizing out common terms
471 /// (e. g. "(A*B)+(A*C)" -> "A*(B+C)").
472 static Value *tryFactorization(InstCombiner::BuilderTy *Builder,
473 const DataLayout *DL, BinaryOperator &I,
474 Instruction::BinaryOps InnerOpcode, Value *A,
475 Value *B, Value *C, Value *D) {
477 // If any of A, B, C, D are null, we can not factor I, return early.
478 // Checking A and C should be enough.
479 if (!A || !C || !B || !D)
482 Value *SimplifiedInst = nullptr;
483 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
484 Instruction::BinaryOps TopLevelOpcode = I.getOpcode();
486 // Does "X op' Y" always equal "Y op' X"?
487 bool InnerCommutative = Instruction::isCommutative(InnerOpcode);
489 // Does "X op' (Y op Z)" always equal "(X op' Y) op (X op' Z)"?
490 if (LeftDistributesOverRight(InnerOpcode, TopLevelOpcode))
491 // Does the instruction have the form "(A op' B) op (A op' D)" or, in the
492 // commutative case, "(A op' B) op (C op' A)"?
493 if (A == C || (InnerCommutative && A == D)) {
496 // Consider forming "A op' (B op D)".
497 // If "B op D" simplifies then it can be formed with no cost.
498 Value *V = SimplifyBinOp(TopLevelOpcode, B, D, DL);
499 // If "B op D" doesn't simplify then only go on if both of the existing
500 // operations "A op' B" and "C op' D" will be zapped as no longer used.
501 if (!V && LHS->hasOneUse() && RHS->hasOneUse())
502 V = Builder->CreateBinOp(TopLevelOpcode, B, D, RHS->getName());
504 SimplifiedInst = Builder->CreateBinOp(InnerOpcode, A, V);
508 // Does "(X op Y) op' Z" always equal "(X op' Z) op (Y op' Z)"?
509 if (!SimplifiedInst && RightDistributesOverLeft(TopLevelOpcode, InnerOpcode))
510 // Does the instruction have the form "(A op' B) op (C op' B)" or, in the
511 // commutative case, "(A op' B) op (B op' D)"?
512 if (B == D || (InnerCommutative && B == C)) {
515 // Consider forming "(A op C) op' B".
516 // If "A op C" simplifies then it can be formed with no cost.
517 Value *V = SimplifyBinOp(TopLevelOpcode, A, C, DL);
519 // If "A op C" doesn't simplify then only go on if both of the existing
520 // operations "A op' B" and "C op' D" will be zapped as no longer used.
521 if (!V && LHS->hasOneUse() && RHS->hasOneUse())
522 V = Builder->CreateBinOp(TopLevelOpcode, A, C, LHS->getName());
524 SimplifiedInst = Builder->CreateBinOp(InnerOpcode, V, B);
528 if (SimplifiedInst) {
530 SimplifiedInst->takeName(&I);
532 // Check if we can add NSW flag to SimplifiedInst. If so, set NSW flag.
533 // TODO: Check for NUW.
534 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(SimplifiedInst)) {
535 if (isa<OverflowingBinaryOperator>(SimplifiedInst)) {
537 if (isa<OverflowingBinaryOperator>(&I))
538 HasNSW = I.hasNoSignedWrap();
540 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS))
541 if (isa<OverflowingBinaryOperator>(Op0))
542 HasNSW &= Op0->hasNoSignedWrap();
544 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS))
545 if (isa<OverflowingBinaryOperator>(Op1))
546 HasNSW &= Op1->hasNoSignedWrap();
547 BO->setHasNoSignedWrap(HasNSW);
551 return SimplifiedInst;
554 /// SimplifyUsingDistributiveLaws - This tries to simplify binary operations
555 /// which some other binary operation distributes over either by factorizing
556 /// out common terms (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this
557 /// results in simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is
558 /// a win). Returns the simplified value, or null if it didn't simplify.
559 Value *InstCombiner::SimplifyUsingDistributiveLaws(BinaryOperator &I) {
560 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
561 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
562 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
565 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
566 auto TopLevelOpcode = I.getOpcode();
567 auto LHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op0, A, B);
568 auto RHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op1, C, D);
570 // The instruction has the form "(A op' B) op (C op' D)". Try to factorize
572 if (LHSOpcode == RHSOpcode) {
573 if (Value *V = tryFactorization(Builder, DL, I, LHSOpcode, A, B, C, D))
577 // The instruction has the form "(A op' B) op (C)". Try to factorize common
579 if (Value *V = tryFactorization(Builder, DL, I, LHSOpcode, A, B, RHS,
580 getIdentityValue(LHSOpcode, RHS)))
583 // The instruction has the form "(B) op (C op' D)". Try to factorize common
585 if (Value *V = tryFactorization(Builder, DL, I, RHSOpcode, LHS,
586 getIdentityValue(RHSOpcode, LHS), C, D))
590 if (Op0 && RightDistributesOverLeft(Op0->getOpcode(), TopLevelOpcode)) {
591 // The instruction has the form "(A op' B) op C". See if expanding it out
592 // to "(A op C) op' (B op C)" results in simplifications.
593 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
594 Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op'
596 // Do "A op C" and "B op C" both simplify?
597 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, C, DL))
598 if (Value *R = SimplifyBinOp(TopLevelOpcode, B, C, DL)) {
599 // They do! Return "L op' R".
601 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
602 if ((L == A && R == B) ||
603 (Instruction::isCommutative(InnerOpcode) && L == B && R == A))
605 // Otherwise return "L op' R" if it simplifies.
606 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, DL))
608 // Otherwise, create a new instruction.
609 C = Builder->CreateBinOp(InnerOpcode, L, R);
615 if (Op1 && LeftDistributesOverRight(TopLevelOpcode, Op1->getOpcode())) {
616 // The instruction has the form "A op (B op' C)". See if expanding it out
617 // to "(A op B) op' (A op C)" results in simplifications.
618 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
619 Instruction::BinaryOps InnerOpcode = Op1->getOpcode(); // op'
621 // Do "A op B" and "A op C" both simplify?
622 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, B, DL))
623 if (Value *R = SimplifyBinOp(TopLevelOpcode, A, C, DL)) {
624 // They do! Return "L op' R".
626 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
627 if ((L == B && R == C) ||
628 (Instruction::isCommutative(InnerOpcode) && L == C && R == B))
630 // Otherwise return "L op' R" if it simplifies.
631 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, DL))
633 // Otherwise, create a new instruction.
634 A = Builder->CreateBinOp(InnerOpcode, L, R);
643 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
644 // if the LHS is a constant zero (which is the 'negate' form).
646 Value *InstCombiner::dyn_castNegVal(Value *V) const {
647 if (BinaryOperator::isNeg(V))
648 return BinaryOperator::getNegArgument(V);
650 // Constants can be considered to be negated values if they can be folded.
651 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
652 return ConstantExpr::getNeg(C);
654 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
655 if (C->getType()->getElementType()->isIntegerTy())
656 return ConstantExpr::getNeg(C);
661 // dyn_castFNegVal - Given a 'fsub' instruction, return the RHS of the
662 // instruction if the LHS is a constant negative zero (which is the 'negate'
665 Value *InstCombiner::dyn_castFNegVal(Value *V, bool IgnoreZeroSign) const {
666 if (BinaryOperator::isFNeg(V, IgnoreZeroSign))
667 return BinaryOperator::getFNegArgument(V);
669 // Constants can be considered to be negated values if they can be folded.
670 if (ConstantFP *C = dyn_cast<ConstantFP>(V))
671 return ConstantExpr::getFNeg(C);
673 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
674 if (C->getType()->getElementType()->isFloatingPointTy())
675 return ConstantExpr::getFNeg(C);
680 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
682 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
683 return IC->Builder->CreateCast(CI->getOpcode(), SO, I.getType());
686 // Figure out if the constant is the left or the right argument.
687 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
688 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
690 if (Constant *SOC = dyn_cast<Constant>(SO)) {
692 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
693 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
696 Value *Op0 = SO, *Op1 = ConstOperand;
700 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I)) {
701 Value *RI = IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1,
702 SO->getName()+".op");
703 Instruction *FPInst = dyn_cast<Instruction>(RI);
704 if (FPInst && isa<FPMathOperator>(FPInst))
705 FPInst->copyFastMathFlags(BO);
708 if (ICmpInst *CI = dyn_cast<ICmpInst>(&I))
709 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
710 SO->getName()+".cmp");
711 if (FCmpInst *CI = dyn_cast<FCmpInst>(&I))
712 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
713 SO->getName()+".cmp");
714 llvm_unreachable("Unknown binary instruction type!");
717 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
718 // constant as the other operand, try to fold the binary operator into the
719 // select arguments. This also works for Cast instructions, which obviously do
720 // not have a second operand.
721 Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) {
722 // Don't modify shared select instructions
723 if (!SI->hasOneUse()) return nullptr;
724 Value *TV = SI->getOperand(1);
725 Value *FV = SI->getOperand(2);
727 if (isa<Constant>(TV) || isa<Constant>(FV)) {
728 // Bool selects with constant operands can be folded to logical ops.
729 if (SI->getType()->isIntegerTy(1)) return nullptr;
731 // If it's a bitcast involving vectors, make sure it has the same number of
732 // elements on both sides.
733 if (BitCastInst *BC = dyn_cast<BitCastInst>(&Op)) {
734 VectorType *DestTy = dyn_cast<VectorType>(BC->getDestTy());
735 VectorType *SrcTy = dyn_cast<VectorType>(BC->getSrcTy());
737 // Verify that either both or neither are vectors.
738 if ((SrcTy == nullptr) != (DestTy == nullptr)) return nullptr;
739 // If vectors, verify that they have the same number of elements.
740 if (SrcTy && SrcTy->getNumElements() != DestTy->getNumElements())
744 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, this);
745 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, this);
747 return SelectInst::Create(SI->getCondition(),
748 SelectTrueVal, SelectFalseVal);
754 /// FoldOpIntoPhi - Given a binary operator, cast instruction, or select which
755 /// has a PHI node as operand #0, see if we can fold the instruction into the
756 /// PHI (which is only possible if all operands to the PHI are constants).
758 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
759 PHINode *PN = cast<PHINode>(I.getOperand(0));
760 unsigned NumPHIValues = PN->getNumIncomingValues();
761 if (NumPHIValues == 0)
764 // We normally only transform phis with a single use. However, if a PHI has
765 // multiple uses and they are all the same operation, we can fold *all* of the
766 // uses into the PHI.
767 if (!PN->hasOneUse()) {
768 // Walk the use list for the instruction, comparing them to I.
769 for (User *U : PN->users()) {
770 Instruction *UI = cast<Instruction>(U);
771 if (UI != &I && !I.isIdenticalTo(UI))
774 // Otherwise, we can replace *all* users with the new PHI we form.
777 // Check to see if all of the operands of the PHI are simple constants
778 // (constantint/constantfp/undef). If there is one non-constant value,
779 // remember the BB it is in. If there is more than one or if *it* is a PHI,
780 // bail out. We don't do arbitrary constant expressions here because moving
781 // their computation can be expensive without a cost model.
782 BasicBlock *NonConstBB = nullptr;
783 for (unsigned i = 0; i != NumPHIValues; ++i) {
784 Value *InVal = PN->getIncomingValue(i);
785 if (isa<Constant>(InVal) && !isa<ConstantExpr>(InVal))
788 if (isa<PHINode>(InVal)) return nullptr; // Itself a phi.
789 if (NonConstBB) return nullptr; // More than one non-const value.
791 NonConstBB = PN->getIncomingBlock(i);
793 // If the InVal is an invoke at the end of the pred block, then we can't
794 // insert a computation after it without breaking the edge.
795 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal))
796 if (II->getParent() == NonConstBB)
799 // If the incoming non-constant value is in I's block, we will remove one
800 // instruction, but insert another equivalent one, leading to infinite
802 if (isPotentiallyReachable(I.getParent(), NonConstBB, DT,
803 getAnalysisIfAvailable<LoopInfo>()))
807 // If there is exactly one non-constant value, we can insert a copy of the
808 // operation in that block. However, if this is a critical edge, we would be
809 // inserting the computation on some other paths (e.g. inside a loop). Only
810 // do this if the pred block is unconditionally branching into the phi block.
811 if (NonConstBB != nullptr) {
812 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
813 if (!BI || !BI->isUnconditional()) return nullptr;
816 // Okay, we can do the transformation: create the new PHI node.
817 PHINode *NewPN = PHINode::Create(I.getType(), PN->getNumIncomingValues());
818 InsertNewInstBefore(NewPN, *PN);
821 // If we are going to have to insert a new computation, do so right before the
822 // predecessors terminator.
824 Builder->SetInsertPoint(NonConstBB->getTerminator());
826 // Next, add all of the operands to the PHI.
827 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) {
828 // We only currently try to fold the condition of a select when it is a phi,
829 // not the true/false values.
830 Value *TrueV = SI->getTrueValue();
831 Value *FalseV = SI->getFalseValue();
832 BasicBlock *PhiTransBB = PN->getParent();
833 for (unsigned i = 0; i != NumPHIValues; ++i) {
834 BasicBlock *ThisBB = PN->getIncomingBlock(i);
835 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB);
836 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB);
837 Value *InV = nullptr;
838 // Beware of ConstantExpr: it may eventually evaluate to getNullValue,
839 // even if currently isNullValue gives false.
840 Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i));
841 if (InC && !isa<ConstantExpr>(InC))
842 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred;
844 InV = Builder->CreateSelect(PN->getIncomingValue(i),
845 TrueVInPred, FalseVInPred, "phitmp");
846 NewPN->addIncoming(InV, ThisBB);
848 } else if (CmpInst *CI = dyn_cast<CmpInst>(&I)) {
849 Constant *C = cast<Constant>(I.getOperand(1));
850 for (unsigned i = 0; i != NumPHIValues; ++i) {
851 Value *InV = nullptr;
852 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
853 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
854 else if (isa<ICmpInst>(CI))
855 InV = Builder->CreateICmp(CI->getPredicate(), PN->getIncomingValue(i),
858 InV = Builder->CreateFCmp(CI->getPredicate(), PN->getIncomingValue(i),
860 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
862 } else if (I.getNumOperands() == 2) {
863 Constant *C = cast<Constant>(I.getOperand(1));
864 for (unsigned i = 0; i != NumPHIValues; ++i) {
865 Value *InV = nullptr;
866 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
867 InV = ConstantExpr::get(I.getOpcode(), InC, C);
869 InV = Builder->CreateBinOp(cast<BinaryOperator>(I).getOpcode(),
870 PN->getIncomingValue(i), C, "phitmp");
871 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
874 CastInst *CI = cast<CastInst>(&I);
875 Type *RetTy = CI->getType();
876 for (unsigned i = 0; i != NumPHIValues; ++i) {
878 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
879 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
881 InV = Builder->CreateCast(CI->getOpcode(),
882 PN->getIncomingValue(i), I.getType(), "phitmp");
883 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
887 for (auto UI = PN->user_begin(), E = PN->user_end(); UI != E;) {
888 Instruction *User = cast<Instruction>(*UI++);
889 if (User == &I) continue;
890 ReplaceInstUsesWith(*User, NewPN);
891 EraseInstFromFunction(*User);
893 return ReplaceInstUsesWith(I, NewPN);
896 /// FindElementAtOffset - Given a pointer type and a constant offset, determine
897 /// whether or not there is a sequence of GEP indices into the pointed type that
898 /// will land us at the specified offset. If so, fill them into NewIndices and
899 /// return the resultant element type, otherwise return null.
900 Type *InstCombiner::FindElementAtOffset(Type *PtrTy, int64_t Offset,
901 SmallVectorImpl<Value*> &NewIndices) {
902 assert(PtrTy->isPtrOrPtrVectorTy());
907 Type *Ty = PtrTy->getPointerElementType();
911 // Start with the index over the outer type. Note that the type size
912 // might be zero (even if the offset isn't zero) if the indexed type
913 // is something like [0 x {int, int}]
914 Type *IntPtrTy = DL->getIntPtrType(PtrTy);
915 int64_t FirstIdx = 0;
916 if (int64_t TySize = DL->getTypeAllocSize(Ty)) {
917 FirstIdx = Offset/TySize;
918 Offset -= FirstIdx*TySize;
920 // Handle hosts where % returns negative instead of values [0..TySize).
926 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
929 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
931 // Index into the types. If we fail, set OrigBase to null.
933 // Indexing into tail padding between struct/array elements.
934 if (uint64_t(Offset*8) >= DL->getTypeSizeInBits(Ty))
937 if (StructType *STy = dyn_cast<StructType>(Ty)) {
938 const StructLayout *SL = DL->getStructLayout(STy);
939 assert(Offset < (int64_t)SL->getSizeInBytes() &&
940 "Offset must stay within the indexed type");
942 unsigned Elt = SL->getElementContainingOffset(Offset);
943 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
946 Offset -= SL->getElementOffset(Elt);
947 Ty = STy->getElementType(Elt);
948 } else if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
949 uint64_t EltSize = DL->getTypeAllocSize(AT->getElementType());
950 assert(EltSize && "Cannot index into a zero-sized array");
951 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
953 Ty = AT->getElementType();
955 // Otherwise, we can't index into the middle of this atomic type, bail.
963 static bool shouldMergeGEPs(GEPOperator &GEP, GEPOperator &Src) {
964 // If this GEP has only 0 indices, it is the same pointer as
965 // Src. If Src is not a trivial GEP too, don't combine
967 if (GEP.hasAllZeroIndices() && !Src.hasAllZeroIndices() &&
973 /// Descale - Return a value X such that Val = X * Scale, or null if none. If
974 /// the multiplication is known not to overflow then NoSignedWrap is set.
975 Value *InstCombiner::Descale(Value *Val, APInt Scale, bool &NoSignedWrap) {
976 assert(isa<IntegerType>(Val->getType()) && "Can only descale integers!");
977 assert(cast<IntegerType>(Val->getType())->getBitWidth() ==
978 Scale.getBitWidth() && "Scale not compatible with value!");
980 // If Val is zero or Scale is one then Val = Val * Scale.
981 if (match(Val, m_Zero()) || Scale == 1) {
986 // If Scale is zero then it does not divide Val.
987 if (Scale.isMinValue())
990 // Look through chains of multiplications, searching for a constant that is
991 // divisible by Scale. For example, descaling X*(Y*(Z*4)) by a factor of 4
992 // will find the constant factor 4 and produce X*(Y*Z). Descaling X*(Y*8) by
993 // a factor of 4 will produce X*(Y*2). The principle of operation is to bore
996 // Val = M1 * X || Analysis starts here and works down
997 // M1 = M2 * Y || Doesn't descend into terms with more
998 // M2 = Z * 4 \/ than one use
1000 // Then to modify a term at the bottom:
1003 // M1 = Z * Y || Replaced M2 with Z
1005 // Then to work back up correcting nsw flags.
1007 // Op - the term we are currently analyzing. Starts at Val then drills down.
1008 // Replaced with its descaled value before exiting from the drill down loop.
1011 // Parent - initially null, but after drilling down notes where Op came from.
1012 // In the example above, Parent is (Val, 0) when Op is M1, because M1 is the
1013 // 0'th operand of Val.
1014 std::pair<Instruction*, unsigned> Parent;
1016 // RequireNoSignedWrap - Set if the transform requires a descaling at deeper
1017 // levels that doesn't overflow.
1018 bool RequireNoSignedWrap = false;
1020 // logScale - log base 2 of the scale. Negative if not a power of 2.
1021 int32_t logScale = Scale.exactLogBase2();
1023 for (;; Op = Parent.first->getOperand(Parent.second)) { // Drill down
1025 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
1026 // If Op is a constant divisible by Scale then descale to the quotient.
1027 APInt Quotient(Scale), Remainder(Scale); // Init ensures right bitwidth.
1028 APInt::sdivrem(CI->getValue(), Scale, Quotient, Remainder);
1029 if (!Remainder.isMinValue())
1030 // Not divisible by Scale.
1032 // Replace with the quotient in the parent.
1033 Op = ConstantInt::get(CI->getType(), Quotient);
1034 NoSignedWrap = true;
1038 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op)) {
1040 if (BO->getOpcode() == Instruction::Mul) {
1042 NoSignedWrap = BO->hasNoSignedWrap();
1043 if (RequireNoSignedWrap && !NoSignedWrap)
1046 // There are three cases for multiplication: multiplication by exactly
1047 // the scale, multiplication by a constant different to the scale, and
1048 // multiplication by something else.
1049 Value *LHS = BO->getOperand(0);
1050 Value *RHS = BO->getOperand(1);
1052 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
1053 // Multiplication by a constant.
1054 if (CI->getValue() == Scale) {
1055 // Multiplication by exactly the scale, replace the multiplication
1056 // by its left-hand side in the parent.
1061 // Otherwise drill down into the constant.
1062 if (!Op->hasOneUse())
1065 Parent = std::make_pair(BO, 1);
1069 // Multiplication by something else. Drill down into the left-hand side
1070 // since that's where the reassociate pass puts the good stuff.
1071 if (!Op->hasOneUse())
1074 Parent = std::make_pair(BO, 0);
1078 if (logScale > 0 && BO->getOpcode() == Instruction::Shl &&
1079 isa<ConstantInt>(BO->getOperand(1))) {
1080 // Multiplication by a power of 2.
1081 NoSignedWrap = BO->hasNoSignedWrap();
1082 if (RequireNoSignedWrap && !NoSignedWrap)
1085 Value *LHS = BO->getOperand(0);
1086 int32_t Amt = cast<ConstantInt>(BO->getOperand(1))->
1087 getLimitedValue(Scale.getBitWidth());
1090 if (Amt == logScale) {
1091 // Multiplication by exactly the scale, replace the multiplication
1092 // by its left-hand side in the parent.
1096 if (Amt < logScale || !Op->hasOneUse())
1099 // Multiplication by more than the scale. Reduce the multiplying amount
1100 // by the scale in the parent.
1101 Parent = std::make_pair(BO, 1);
1102 Op = ConstantInt::get(BO->getType(), Amt - logScale);
1107 if (!Op->hasOneUse())
1110 if (CastInst *Cast = dyn_cast<CastInst>(Op)) {
1111 if (Cast->getOpcode() == Instruction::SExt) {
1112 // Op is sign-extended from a smaller type, descale in the smaller type.
1113 unsigned SmallSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1114 APInt SmallScale = Scale.trunc(SmallSize);
1115 // Suppose Op = sext X, and we descale X as Y * SmallScale. We want to
1116 // descale Op as (sext Y) * Scale. In order to have
1117 // sext (Y * SmallScale) = (sext Y) * Scale
1118 // some conditions need to hold however: SmallScale must sign-extend to
1119 // Scale and the multiplication Y * SmallScale should not overflow.
1120 if (SmallScale.sext(Scale.getBitWidth()) != Scale)
1121 // SmallScale does not sign-extend to Scale.
1123 assert(SmallScale.exactLogBase2() == logScale);
1124 // Require that Y * SmallScale must not overflow.
1125 RequireNoSignedWrap = true;
1127 // Drill down through the cast.
1128 Parent = std::make_pair(Cast, 0);
1133 if (Cast->getOpcode() == Instruction::Trunc) {
1134 // Op is truncated from a larger type, descale in the larger type.
1135 // Suppose Op = trunc X, and we descale X as Y * sext Scale. Then
1136 // trunc (Y * sext Scale) = (trunc Y) * Scale
1137 // always holds. However (trunc Y) * Scale may overflow even if
1138 // trunc (Y * sext Scale) does not, so nsw flags need to be cleared
1139 // from this point up in the expression (see later).
1140 if (RequireNoSignedWrap)
1143 // Drill down through the cast.
1144 unsigned LargeSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1145 Parent = std::make_pair(Cast, 0);
1146 Scale = Scale.sext(LargeSize);
1147 if (logScale + 1 == (int32_t)Cast->getType()->getPrimitiveSizeInBits())
1149 assert(Scale.exactLogBase2() == logScale);
1154 // Unsupported expression, bail out.
1158 // If Op is zero then Val = Op * Scale.
1159 if (match(Op, m_Zero())) {
1160 NoSignedWrap = true;
1164 // We know that we can successfully descale, so from here on we can safely
1165 // modify the IR. Op holds the descaled version of the deepest term in the
1166 // expression. NoSignedWrap is 'true' if multiplying Op by Scale is known
1170 // The expression only had one term.
1173 // Rewrite the parent using the descaled version of its operand.
1174 assert(Parent.first->hasOneUse() && "Drilled down when more than one use!");
1175 assert(Op != Parent.first->getOperand(Parent.second) &&
1176 "Descaling was a no-op?");
1177 Parent.first->setOperand(Parent.second, Op);
1178 Worklist.Add(Parent.first);
1180 // Now work back up the expression correcting nsw flags. The logic is based
1181 // on the following observation: if X * Y is known not to overflow as a signed
1182 // multiplication, and Y is replaced by a value Z with smaller absolute value,
1183 // then X * Z will not overflow as a signed multiplication either. As we work
1184 // our way up, having NoSignedWrap 'true' means that the descaled value at the
1185 // current level has strictly smaller absolute value than the original.
1186 Instruction *Ancestor = Parent.first;
1188 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Ancestor)) {
1189 // If the multiplication wasn't nsw then we can't say anything about the
1190 // value of the descaled multiplication, and we have to clear nsw flags
1191 // from this point on up.
1192 bool OpNoSignedWrap = BO->hasNoSignedWrap();
1193 NoSignedWrap &= OpNoSignedWrap;
1194 if (NoSignedWrap != OpNoSignedWrap) {
1195 BO->setHasNoSignedWrap(NoSignedWrap);
1196 Worklist.Add(Ancestor);
1198 } else if (Ancestor->getOpcode() == Instruction::Trunc) {
1199 // The fact that the descaled input to the trunc has smaller absolute
1200 // value than the original input doesn't tell us anything useful about
1201 // the absolute values of the truncations.
1202 NoSignedWrap = false;
1204 assert((Ancestor->getOpcode() != Instruction::SExt || NoSignedWrap) &&
1205 "Failed to keep proper track of nsw flags while drilling down?");
1207 if (Ancestor == Val)
1208 // Got to the top, all done!
1211 // Move up one level in the expression.
1212 assert(Ancestor->hasOneUse() && "Drilled down when more than one use!");
1213 Ancestor = Ancestor->user_back();
1217 /// \brief Creates node of binary operation with the same attributes as the
1218 /// specified one but with other operands.
1219 static Value *CreateBinOpAsGiven(BinaryOperator &Inst, Value *LHS, Value *RHS,
1220 InstCombiner::BuilderTy *B) {
1221 Value *BORes = B->CreateBinOp(Inst.getOpcode(), LHS, RHS);
1222 if (BinaryOperator *NewBO = dyn_cast<BinaryOperator>(BORes)) {
1223 if (isa<OverflowingBinaryOperator>(NewBO)) {
1224 NewBO->setHasNoSignedWrap(Inst.hasNoSignedWrap());
1225 NewBO->setHasNoUnsignedWrap(Inst.hasNoUnsignedWrap());
1227 if (isa<PossiblyExactOperator>(NewBO))
1228 NewBO->setIsExact(Inst.isExact());
1233 /// \brief Makes transformation of binary operation specific for vector types.
1234 /// \param Inst Binary operator to transform.
1235 /// \return Pointer to node that must replace the original binary operator, or
1236 /// null pointer if no transformation was made.
1237 Value *InstCombiner::SimplifyVectorOp(BinaryOperator &Inst) {
1238 if (!Inst.getType()->isVectorTy()) return nullptr;
1240 // It may not be safe to reorder shuffles and things like div, urem, etc.
1241 // because we may trap when executing those ops on unknown vector elements.
1243 if (!isSafeToSpeculativelyExecute(&Inst, DL)) return nullptr;
1245 unsigned VWidth = cast<VectorType>(Inst.getType())->getNumElements();
1246 Value *LHS = Inst.getOperand(0), *RHS = Inst.getOperand(1);
1247 assert(cast<VectorType>(LHS->getType())->getNumElements() == VWidth);
1248 assert(cast<VectorType>(RHS->getType())->getNumElements() == VWidth);
1250 // If both arguments of binary operation are shuffles, which use the same
1251 // mask and shuffle within a single vector, it is worthwhile to move the
1252 // shuffle after binary operation:
1253 // Op(shuffle(v1, m), shuffle(v2, m)) -> shuffle(Op(v1, v2), m)
1254 if (isa<ShuffleVectorInst>(LHS) && isa<ShuffleVectorInst>(RHS)) {
1255 ShuffleVectorInst *LShuf = cast<ShuffleVectorInst>(LHS);
1256 ShuffleVectorInst *RShuf = cast<ShuffleVectorInst>(RHS);
1257 if (isa<UndefValue>(LShuf->getOperand(1)) &&
1258 isa<UndefValue>(RShuf->getOperand(1)) &&
1259 LShuf->getOperand(0)->getType() == RShuf->getOperand(0)->getType() &&
1260 LShuf->getMask() == RShuf->getMask()) {
1261 Value *NewBO = CreateBinOpAsGiven(Inst, LShuf->getOperand(0),
1262 RShuf->getOperand(0), Builder);
1263 Value *Res = Builder->CreateShuffleVector(NewBO,
1264 UndefValue::get(NewBO->getType()), LShuf->getMask());
1269 // If one argument is a shuffle within one vector, the other is a constant,
1270 // try moving the shuffle after the binary operation.
1271 ShuffleVectorInst *Shuffle = nullptr;
1272 Constant *C1 = nullptr;
1273 if (isa<ShuffleVectorInst>(LHS)) Shuffle = cast<ShuffleVectorInst>(LHS);
1274 if (isa<ShuffleVectorInst>(RHS)) Shuffle = cast<ShuffleVectorInst>(RHS);
1275 if (isa<Constant>(LHS)) C1 = cast<Constant>(LHS);
1276 if (isa<Constant>(RHS)) C1 = cast<Constant>(RHS);
1277 if (Shuffle && C1 &&
1278 (isa<ConstantVector>(C1) || isa<ConstantDataVector>(C1)) &&
1279 isa<UndefValue>(Shuffle->getOperand(1)) &&
1280 Shuffle->getType() == Shuffle->getOperand(0)->getType()) {
1281 SmallVector<int, 16> ShMask = Shuffle->getShuffleMask();
1282 // Find constant C2 that has property:
1283 // shuffle(C2, ShMask) = C1
1284 // If such constant does not exist (example: ShMask=<0,0> and C1=<1,2>)
1285 // reorder is not possible.
1286 SmallVector<Constant*, 16> C2M(VWidth,
1287 UndefValue::get(C1->getType()->getScalarType()));
1288 bool MayChange = true;
1289 for (unsigned I = 0; I < VWidth; ++I) {
1290 if (ShMask[I] >= 0) {
1291 assert(ShMask[I] < (int)VWidth);
1292 if (!isa<UndefValue>(C2M[ShMask[I]])) {
1296 C2M[ShMask[I]] = C1->getAggregateElement(I);
1300 Constant *C2 = ConstantVector::get(C2M);
1301 Value *NewLHS, *NewRHS;
1302 if (isa<Constant>(LHS)) {
1304 NewRHS = Shuffle->getOperand(0);
1306 NewLHS = Shuffle->getOperand(0);
1309 Value *NewBO = CreateBinOpAsGiven(Inst, NewLHS, NewRHS, Builder);
1310 Value *Res = Builder->CreateShuffleVector(NewBO,
1311 UndefValue::get(Inst.getType()), Shuffle->getMask());
1319 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
1320 SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end());
1322 if (Value *V = SimplifyGEPInst(Ops, DL, TLI, DT, AT))
1323 return ReplaceInstUsesWith(GEP, V);
1325 Value *PtrOp = GEP.getOperand(0);
1327 // Eliminate unneeded casts for indices, and replace indices which displace
1328 // by multiples of a zero size type with zero.
1330 bool MadeChange = false;
1331 Type *IntPtrTy = DL->getIntPtrType(GEP.getPointerOperandType());
1333 gep_type_iterator GTI = gep_type_begin(GEP);
1334 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end();
1335 I != E; ++I, ++GTI) {
1336 // Skip indices into struct types.
1337 SequentialType *SeqTy = dyn_cast<SequentialType>(*GTI);
1338 if (!SeqTy) continue;
1340 // If the element type has zero size then any index over it is equivalent
1341 // to an index of zero, so replace it with zero if it is not zero already.
1342 if (SeqTy->getElementType()->isSized() &&
1343 DL->getTypeAllocSize(SeqTy->getElementType()) == 0)
1344 if (!isa<Constant>(*I) || !cast<Constant>(*I)->isNullValue()) {
1345 *I = Constant::getNullValue(IntPtrTy);
1349 Type *IndexTy = (*I)->getType();
1350 if (IndexTy != IntPtrTy) {
1351 // If we are using a wider index than needed for this platform, shrink
1352 // it to what we need. If narrower, sign-extend it to what we need.
1353 // This explicit cast can make subsequent optimizations more obvious.
1354 *I = Builder->CreateIntCast(*I, IntPtrTy, true);
1358 if (MadeChange) return &GEP;
1361 // Check to see if the inputs to the PHI node are getelementptr instructions.
1362 if (PHINode *PN = dyn_cast<PHINode>(PtrOp)) {
1363 GetElementPtrInst *Op1 = dyn_cast<GetElementPtrInst>(PN->getOperand(0));
1369 for (auto I = PN->op_begin()+1, E = PN->op_end(); I !=E; ++I) {
1370 GetElementPtrInst *Op2 = dyn_cast<GetElementPtrInst>(*I);
1371 if (!Op2 || Op1->getNumOperands() != Op2->getNumOperands())
1374 // Keep track of the type as we walk the GEP.
1375 Type *CurTy = Op1->getOperand(0)->getType()->getScalarType();
1377 for (unsigned J = 0, F = Op1->getNumOperands(); J != F; ++J) {
1378 if (Op1->getOperand(J)->getType() != Op2->getOperand(J)->getType())
1381 if (Op1->getOperand(J) != Op2->getOperand(J)) {
1383 // We have not seen any differences yet in the GEPs feeding the
1384 // PHI yet, so we record this one if it is allowed to be a
1387 // The first two arguments can vary for any GEP, the rest have to be
1388 // static for struct slots
1389 if (J > 1 && CurTy->isStructTy())
1394 // The GEP is different by more than one input. While this could be
1395 // extended to support GEPs that vary by more than one variable it
1396 // doesn't make sense since it greatly increases the complexity and
1397 // would result in an R+R+R addressing mode which no backend
1398 // directly supports and would need to be broken into several
1399 // simpler instructions anyway.
1404 // Sink down a layer of the type for the next iteration.
1406 if (CompositeType *CT = dyn_cast<CompositeType>(CurTy)) {
1407 CurTy = CT->getTypeAtIndex(Op1->getOperand(J));
1415 GetElementPtrInst *NewGEP = cast<GetElementPtrInst>(Op1->clone());
1418 // All the GEPs feeding the PHI are identical. Clone one down into our
1419 // BB so that it can be merged with the current GEP.
1420 GEP.getParent()->getInstList().insert(GEP.getParent()->getFirstNonPHI(),
1423 // All the GEPs feeding the PHI differ at a single offset. Clone a GEP
1424 // into the current block so it can be merged, and create a new PHI to
1426 Instruction *InsertPt = Builder->GetInsertPoint();
1427 Builder->SetInsertPoint(PN);
1428 PHINode *NewPN = Builder->CreatePHI(Op1->getOperand(DI)->getType(),
1429 PN->getNumOperands());
1430 Builder->SetInsertPoint(InsertPt);
1432 for (auto &I : PN->operands())
1433 NewPN->addIncoming(cast<GEPOperator>(I)->getOperand(DI),
1434 PN->getIncomingBlock(I));
1436 NewGEP->setOperand(DI, NewPN);
1437 GEP.getParent()->getInstList().insert(GEP.getParent()->getFirstNonPHI(),
1439 NewGEP->setOperand(DI, NewPN);
1442 GEP.setOperand(0, NewGEP);
1446 // Combine Indices - If the source pointer to this getelementptr instruction
1447 // is a getelementptr instruction, combine the indices of the two
1448 // getelementptr instructions into a single instruction.
1450 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) {
1451 if (!shouldMergeGEPs(*cast<GEPOperator>(&GEP), *Src))
1454 // Note that if our source is a gep chain itself then we wait for that
1455 // chain to be resolved before we perform this transformation. This
1456 // avoids us creating a TON of code in some cases.
1457 if (GEPOperator *SrcGEP =
1458 dyn_cast<GEPOperator>(Src->getOperand(0)))
1459 if (SrcGEP->getNumOperands() == 2 && shouldMergeGEPs(*Src, *SrcGEP))
1460 return nullptr; // Wait until our source is folded to completion.
1462 SmallVector<Value*, 8> Indices;
1464 // Find out whether the last index in the source GEP is a sequential idx.
1465 bool EndsWithSequential = false;
1466 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src);
1468 EndsWithSequential = !(*I)->isStructTy();
1470 // Can we combine the two pointer arithmetics offsets?
1471 if (EndsWithSequential) {
1472 // Replace: gep (gep %P, long B), long A, ...
1473 // With: T = long A+B; gep %P, T, ...
1476 Value *SO1 = Src->getOperand(Src->getNumOperands()-1);
1477 Value *GO1 = GEP.getOperand(1);
1478 if (SO1 == Constant::getNullValue(SO1->getType())) {
1480 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
1483 // If they aren't the same type, then the input hasn't been processed
1484 // by the loop above yet (which canonicalizes sequential index types to
1485 // intptr_t). Just avoid transforming this until the input has been
1487 if (SO1->getType() != GO1->getType())
1489 Sum = Builder->CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
1492 // Update the GEP in place if possible.
1493 if (Src->getNumOperands() == 2) {
1494 GEP.setOperand(0, Src->getOperand(0));
1495 GEP.setOperand(1, Sum);
1498 Indices.append(Src->op_begin()+1, Src->op_end()-1);
1499 Indices.push_back(Sum);
1500 Indices.append(GEP.op_begin()+2, GEP.op_end());
1501 } else if (isa<Constant>(*GEP.idx_begin()) &&
1502 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
1503 Src->getNumOperands() != 1) {
1504 // Otherwise we can do the fold if the first index of the GEP is a zero
1505 Indices.append(Src->op_begin()+1, Src->op_end());
1506 Indices.append(GEP.idx_begin()+1, GEP.idx_end());
1509 if (!Indices.empty())
1510 return (GEP.isInBounds() && Src->isInBounds()) ?
1511 GetElementPtrInst::CreateInBounds(Src->getOperand(0), Indices,
1513 GetElementPtrInst::Create(Src->getOperand(0), Indices, GEP.getName());
1516 if (DL && GEP.getNumIndices() == 1) {
1517 unsigned AS = GEP.getPointerAddressSpace();
1518 if (GEP.getOperand(1)->getType()->getScalarSizeInBits() ==
1519 DL->getPointerSizeInBits(AS)) {
1520 Type *PtrTy = GEP.getPointerOperandType();
1521 Type *Ty = PtrTy->getPointerElementType();
1522 uint64_t TyAllocSize = DL->getTypeAllocSize(Ty);
1524 bool Matched = false;
1527 if (TyAllocSize == 1) {
1528 V = GEP.getOperand(1);
1530 } else if (match(GEP.getOperand(1),
1531 m_AShr(m_Value(V), m_ConstantInt(C)))) {
1532 if (TyAllocSize == 1ULL << C)
1534 } else if (match(GEP.getOperand(1),
1535 m_SDiv(m_Value(V), m_ConstantInt(C)))) {
1536 if (TyAllocSize == C)
1541 // Canonicalize (gep i8* X, -(ptrtoint Y))
1542 // to (inttoptr (sub (ptrtoint X), (ptrtoint Y)))
1543 // The GEP pattern is emitted by the SCEV expander for certain kinds of
1544 // pointer arithmetic.
1545 if (match(V, m_Neg(m_PtrToInt(m_Value())))) {
1546 Operator *Index = cast<Operator>(V);
1547 Value *PtrToInt = Builder->CreatePtrToInt(PtrOp, Index->getType());
1548 Value *NewSub = Builder->CreateSub(PtrToInt, Index->getOperand(1));
1549 return CastInst::Create(Instruction::IntToPtr, NewSub, GEP.getType());
1551 // Canonicalize (gep i8* X, (ptrtoint Y)-(ptrtoint X))
1554 if (match(V, m_Sub(m_PtrToInt(m_Value(Y)),
1555 m_PtrToInt(m_Specific(GEP.getOperand(0)))))) {
1556 return CastInst::CreatePointerBitCastOrAddrSpaceCast(Y,
1563 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0).
1564 Value *StrippedPtr = PtrOp->stripPointerCasts();
1565 PointerType *StrippedPtrTy = dyn_cast<PointerType>(StrippedPtr->getType());
1567 // We do not handle pointer-vector geps here.
1571 if (StrippedPtr != PtrOp) {
1572 bool HasZeroPointerIndex = false;
1573 if (ConstantInt *C = dyn_cast<ConstantInt>(GEP.getOperand(1)))
1574 HasZeroPointerIndex = C->isZero();
1576 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
1577 // into : GEP [10 x i8]* X, i32 0, ...
1579 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
1580 // into : GEP i8* X, ...
1582 // This occurs when the program declares an array extern like "int X[];"
1583 if (HasZeroPointerIndex) {
1584 PointerType *CPTy = cast<PointerType>(PtrOp->getType());
1585 if (ArrayType *CATy =
1586 dyn_cast<ArrayType>(CPTy->getElementType())) {
1587 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
1588 if (CATy->getElementType() == StrippedPtrTy->getElementType()) {
1589 // -> GEP i8* X, ...
1590 SmallVector<Value*, 8> Idx(GEP.idx_begin()+1, GEP.idx_end());
1591 GetElementPtrInst *Res =
1592 GetElementPtrInst::Create(StrippedPtr, Idx, GEP.getName());
1593 Res->setIsInBounds(GEP.isInBounds());
1594 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace())
1596 // Insert Res, and create an addrspacecast.
1598 // GEP (addrspacecast i8 addrspace(1)* X to [0 x i8]*), i32 0, ...
1600 // %0 = GEP i8 addrspace(1)* X, ...
1601 // addrspacecast i8 addrspace(1)* %0 to i8*
1602 return new AddrSpaceCastInst(Builder->Insert(Res), GEP.getType());
1605 if (ArrayType *XATy =
1606 dyn_cast<ArrayType>(StrippedPtrTy->getElementType())){
1607 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
1608 if (CATy->getElementType() == XATy->getElementType()) {
1609 // -> GEP [10 x i8]* X, i32 0, ...
1610 // At this point, we know that the cast source type is a pointer
1611 // to an array of the same type as the destination pointer
1612 // array. Because the array type is never stepped over (there
1613 // is a leading zero) we can fold the cast into this GEP.
1614 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace()) {
1615 GEP.setOperand(0, StrippedPtr);
1618 // Cannot replace the base pointer directly because StrippedPtr's
1619 // address space is different. Instead, create a new GEP followed by
1620 // an addrspacecast.
1622 // GEP (addrspacecast [10 x i8] addrspace(1)* X to [0 x i8]*),
1625 // %0 = GEP [10 x i8] addrspace(1)* X, ...
1626 // addrspacecast i8 addrspace(1)* %0 to i8*
1627 SmallVector<Value*, 8> Idx(GEP.idx_begin(), GEP.idx_end());
1628 Value *NewGEP = GEP.isInBounds() ?
1629 Builder->CreateInBoundsGEP(StrippedPtr, Idx, GEP.getName()) :
1630 Builder->CreateGEP(StrippedPtr, Idx, GEP.getName());
1631 return new AddrSpaceCastInst(NewGEP, GEP.getType());
1635 } else if (GEP.getNumOperands() == 2) {
1636 // Transform things like:
1637 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
1638 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
1639 Type *SrcElTy = StrippedPtrTy->getElementType();
1640 Type *ResElTy = PtrOp->getType()->getPointerElementType();
1641 if (DL && SrcElTy->isArrayTy() &&
1642 DL->getTypeAllocSize(SrcElTy->getArrayElementType()) ==
1643 DL->getTypeAllocSize(ResElTy)) {
1644 Type *IdxType = DL->getIntPtrType(GEP.getType());
1645 Value *Idx[2] = { Constant::getNullValue(IdxType), GEP.getOperand(1) };
1646 Value *NewGEP = GEP.isInBounds() ?
1647 Builder->CreateInBoundsGEP(StrippedPtr, Idx, GEP.getName()) :
1648 Builder->CreateGEP(StrippedPtr, Idx, GEP.getName());
1650 // V and GEP are both pointer types --> BitCast
1651 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
1655 // Transform things like:
1656 // %V = mul i64 %N, 4
1657 // %t = getelementptr i8* bitcast (i32* %arr to i8*), i32 %V
1658 // into: %t1 = getelementptr i32* %arr, i32 %N; bitcast
1659 if (DL && ResElTy->isSized() && SrcElTy->isSized()) {
1660 // Check that changing the type amounts to dividing the index by a scale
1662 uint64_t ResSize = DL->getTypeAllocSize(ResElTy);
1663 uint64_t SrcSize = DL->getTypeAllocSize(SrcElTy);
1664 if (ResSize && SrcSize % ResSize == 0) {
1665 Value *Idx = GEP.getOperand(1);
1666 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
1667 uint64_t Scale = SrcSize / ResSize;
1669 // Earlier transforms ensure that the index has type IntPtrType, which
1670 // considerably simplifies the logic by eliminating implicit casts.
1671 assert(Idx->getType() == DL->getIntPtrType(GEP.getType()) &&
1672 "Index not cast to pointer width?");
1675 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
1676 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
1677 // If the multiplication NewIdx * Scale may overflow then the new
1678 // GEP may not be "inbounds".
1679 Value *NewGEP = GEP.isInBounds() && NSW ?
1680 Builder->CreateInBoundsGEP(StrippedPtr, NewIdx, GEP.getName()) :
1681 Builder->CreateGEP(StrippedPtr, NewIdx, GEP.getName());
1683 // The NewGEP must be pointer typed, so must the old one -> BitCast
1684 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
1690 // Similarly, transform things like:
1691 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
1692 // (where tmp = 8*tmp2) into:
1693 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
1694 if (DL && ResElTy->isSized() && SrcElTy->isSized() &&
1695 SrcElTy->isArrayTy()) {
1696 // Check that changing to the array element type amounts to dividing the
1697 // index by a scale factor.
1698 uint64_t ResSize = DL->getTypeAllocSize(ResElTy);
1699 uint64_t ArrayEltSize
1700 = DL->getTypeAllocSize(SrcElTy->getArrayElementType());
1701 if (ResSize && ArrayEltSize % ResSize == 0) {
1702 Value *Idx = GEP.getOperand(1);
1703 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
1704 uint64_t Scale = ArrayEltSize / ResSize;
1706 // Earlier transforms ensure that the index has type IntPtrType, which
1707 // considerably simplifies the logic by eliminating implicit casts.
1708 assert(Idx->getType() == DL->getIntPtrType(GEP.getType()) &&
1709 "Index not cast to pointer width?");
1712 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
1713 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
1714 // If the multiplication NewIdx * Scale may overflow then the new
1715 // GEP may not be "inbounds".
1717 Constant::getNullValue(DL->getIntPtrType(GEP.getType())),
1721 Value *NewGEP = GEP.isInBounds() && NSW ?
1722 Builder->CreateInBoundsGEP(StrippedPtr, Off, GEP.getName()) :
1723 Builder->CreateGEP(StrippedPtr, Off, GEP.getName());
1724 // The NewGEP must be pointer typed, so must the old one -> BitCast
1725 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
1736 // addrspacecast between types is canonicalized as a bitcast, then an
1737 // addrspacecast. To take advantage of the below bitcast + struct GEP, look
1738 // through the addrspacecast.
1739 if (AddrSpaceCastInst *ASC = dyn_cast<AddrSpaceCastInst>(PtrOp)) {
1740 // X = bitcast A addrspace(1)* to B addrspace(1)*
1741 // Y = addrspacecast A addrspace(1)* to B addrspace(2)*
1742 // Z = gep Y, <...constant indices...>
1743 // Into an addrspacecasted GEP of the struct.
1744 if (BitCastInst *BC = dyn_cast<BitCastInst>(ASC->getOperand(0)))
1748 /// See if we can simplify:
1749 /// X = bitcast A* to B*
1750 /// Y = gep X, <...constant indices...>
1751 /// into a gep of the original struct. This is important for SROA and alias
1752 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
1753 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
1754 Value *Operand = BCI->getOperand(0);
1755 PointerType *OpType = cast<PointerType>(Operand->getType());
1756 unsigned OffsetBits = DL->getPointerTypeSizeInBits(GEP.getType());
1757 APInt Offset(OffsetBits, 0);
1758 if (!isa<BitCastInst>(Operand) &&
1759 GEP.accumulateConstantOffset(*DL, Offset)) {
1761 // If this GEP instruction doesn't move the pointer, just replace the GEP
1762 // with a bitcast of the real input to the dest type.
1764 // If the bitcast is of an allocation, and the allocation will be
1765 // converted to match the type of the cast, don't touch this.
1766 if (isa<AllocaInst>(Operand) || isAllocationFn(Operand, TLI)) {
1767 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
1768 if (Instruction *I = visitBitCast(*BCI)) {
1771 BCI->getParent()->getInstList().insert(BCI, I);
1772 ReplaceInstUsesWith(*BCI, I);
1778 if (Operand->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
1779 return new AddrSpaceCastInst(Operand, GEP.getType());
1780 return new BitCastInst(Operand, GEP.getType());
1783 // Otherwise, if the offset is non-zero, we need to find out if there is a
1784 // field at Offset in 'A's type. If so, we can pull the cast through the
1786 SmallVector<Value*, 8> NewIndices;
1787 if (FindElementAtOffset(OpType, Offset.getSExtValue(), NewIndices)) {
1788 Value *NGEP = GEP.isInBounds() ?
1789 Builder->CreateInBoundsGEP(Operand, NewIndices) :
1790 Builder->CreateGEP(Operand, NewIndices);
1792 if (NGEP->getType() == GEP.getType())
1793 return ReplaceInstUsesWith(GEP, NGEP);
1794 NGEP->takeName(&GEP);
1796 if (NGEP->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
1797 return new AddrSpaceCastInst(NGEP, GEP.getType());
1798 return new BitCastInst(NGEP, GEP.getType());
1807 isAllocSiteRemovable(Instruction *AI, SmallVectorImpl<WeakVH> &Users,
1808 const TargetLibraryInfo *TLI) {
1809 SmallVector<Instruction*, 4> Worklist;
1810 Worklist.push_back(AI);
1813 Instruction *PI = Worklist.pop_back_val();
1814 for (User *U : PI->users()) {
1815 Instruction *I = cast<Instruction>(U);
1816 switch (I->getOpcode()) {
1818 // Give up the moment we see something we can't handle.
1821 case Instruction::BitCast:
1822 case Instruction::GetElementPtr:
1824 Worklist.push_back(I);
1827 case Instruction::ICmp: {
1828 ICmpInst *ICI = cast<ICmpInst>(I);
1829 // We can fold eq/ne comparisons with null to false/true, respectively.
1830 if (!ICI->isEquality() || !isa<ConstantPointerNull>(ICI->getOperand(1)))
1836 case Instruction::Call:
1837 // Ignore no-op and store intrinsics.
1838 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1839 switch (II->getIntrinsicID()) {
1843 case Intrinsic::memmove:
1844 case Intrinsic::memcpy:
1845 case Intrinsic::memset: {
1846 MemIntrinsic *MI = cast<MemIntrinsic>(II);
1847 if (MI->isVolatile() || MI->getRawDest() != PI)
1851 case Intrinsic::dbg_declare:
1852 case Intrinsic::dbg_value:
1853 case Intrinsic::invariant_start:
1854 case Intrinsic::invariant_end:
1855 case Intrinsic::lifetime_start:
1856 case Intrinsic::lifetime_end:
1857 case Intrinsic::objectsize:
1863 if (isFreeCall(I, TLI)) {
1869 case Instruction::Store: {
1870 StoreInst *SI = cast<StoreInst>(I);
1871 if (SI->isVolatile() || SI->getPointerOperand() != PI)
1877 llvm_unreachable("missing a return?");
1879 } while (!Worklist.empty());
1883 Instruction *InstCombiner::visitAllocSite(Instruction &MI) {
1884 // If we have a malloc call which is only used in any amount of comparisons
1885 // to null and free calls, delete the calls and replace the comparisons with
1886 // true or false as appropriate.
1887 SmallVector<WeakVH, 64> Users;
1888 if (isAllocSiteRemovable(&MI, Users, TLI)) {
1889 for (unsigned i = 0, e = Users.size(); i != e; ++i) {
1890 Instruction *I = cast_or_null<Instruction>(&*Users[i]);
1893 if (ICmpInst *C = dyn_cast<ICmpInst>(I)) {
1894 ReplaceInstUsesWith(*C,
1895 ConstantInt::get(Type::getInt1Ty(C->getContext()),
1896 C->isFalseWhenEqual()));
1897 } else if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) {
1898 ReplaceInstUsesWith(*I, UndefValue::get(I->getType()));
1899 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1900 if (II->getIntrinsicID() == Intrinsic::objectsize) {
1901 ConstantInt *CI = cast<ConstantInt>(II->getArgOperand(1));
1902 uint64_t DontKnow = CI->isZero() ? -1ULL : 0;
1903 ReplaceInstUsesWith(*I, ConstantInt::get(I->getType(), DontKnow));
1906 EraseInstFromFunction(*I);
1909 if (InvokeInst *II = dyn_cast<InvokeInst>(&MI)) {
1910 // Replace invoke with a NOP intrinsic to maintain the original CFG
1911 Module *M = II->getParent()->getParent()->getParent();
1912 Function *F = Intrinsic::getDeclaration(M, Intrinsic::donothing);
1913 InvokeInst::Create(F, II->getNormalDest(), II->getUnwindDest(),
1914 None, "", II->getParent());
1916 return EraseInstFromFunction(MI);
1921 /// \brief Move the call to free before a NULL test.
1923 /// Check if this free is accessed after its argument has been test
1924 /// against NULL (property 0).
1925 /// If yes, it is legal to move this call in its predecessor block.
1927 /// The move is performed only if the block containing the call to free
1928 /// will be removed, i.e.:
1929 /// 1. it has only one predecessor P, and P has two successors
1930 /// 2. it contains the call and an unconditional branch
1931 /// 3. its successor is the same as its predecessor's successor
1933 /// The profitability is out-of concern here and this function should
1934 /// be called only if the caller knows this transformation would be
1935 /// profitable (e.g., for code size).
1936 static Instruction *
1937 tryToMoveFreeBeforeNullTest(CallInst &FI) {
1938 Value *Op = FI.getArgOperand(0);
1939 BasicBlock *FreeInstrBB = FI.getParent();
1940 BasicBlock *PredBB = FreeInstrBB->getSinglePredecessor();
1942 // Validate part of constraint #1: Only one predecessor
1943 // FIXME: We can extend the number of predecessor, but in that case, we
1944 // would duplicate the call to free in each predecessor and it may
1945 // not be profitable even for code size.
1949 // Validate constraint #2: Does this block contains only the call to
1950 // free and an unconditional branch?
1951 // FIXME: We could check if we can speculate everything in the
1952 // predecessor block
1953 if (FreeInstrBB->size() != 2)
1956 if (!match(FreeInstrBB->getTerminator(), m_UnconditionalBr(SuccBB)))
1959 // Validate the rest of constraint #1 by matching on the pred branch.
1960 TerminatorInst *TI = PredBB->getTerminator();
1961 BasicBlock *TrueBB, *FalseBB;
1962 ICmpInst::Predicate Pred;
1963 if (!match(TI, m_Br(m_ICmp(Pred, m_Specific(Op), m_Zero()), TrueBB, FalseBB)))
1965 if (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE)
1968 // Validate constraint #3: Ensure the null case just falls through.
1969 if (SuccBB != (Pred == ICmpInst::ICMP_EQ ? TrueBB : FalseBB))
1971 assert(FreeInstrBB == (Pred == ICmpInst::ICMP_EQ ? FalseBB : TrueBB) &&
1972 "Broken CFG: missing edge from predecessor to successor");
1979 Instruction *InstCombiner::visitFree(CallInst &FI) {
1980 Value *Op = FI.getArgOperand(0);
1982 // free undef -> unreachable.
1983 if (isa<UndefValue>(Op)) {
1984 // Insert a new store to null because we cannot modify the CFG here.
1985 Builder->CreateStore(ConstantInt::getTrue(FI.getContext()),
1986 UndefValue::get(Type::getInt1PtrTy(FI.getContext())));
1987 return EraseInstFromFunction(FI);
1990 // If we have 'free null' delete the instruction. This can happen in stl code
1991 // when lots of inlining happens.
1992 if (isa<ConstantPointerNull>(Op))
1993 return EraseInstFromFunction(FI);
1995 // If we optimize for code size, try to move the call to free before the null
1996 // test so that simplify cfg can remove the empty block and dead code
1997 // elimination the branch. I.e., helps to turn something like:
1998 // if (foo) free(foo);
2002 if (Instruction *I = tryToMoveFreeBeforeNullTest(FI))
2008 Instruction *InstCombiner::visitReturnInst(ReturnInst &RI) {
2009 if (RI.getNumOperands() == 0) // ret void
2012 Value *ResultOp = RI.getOperand(0);
2013 Type *VTy = ResultOp->getType();
2014 if (!VTy->isIntegerTy())
2017 // There might be assume intrinsics dominating this return that completely
2018 // determine the value. If so, constant fold it.
2019 unsigned BitWidth = VTy->getPrimitiveSizeInBits();
2020 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2021 computeKnownBits(ResultOp, KnownZero, KnownOne, 0, &RI);
2022 if ((KnownZero|KnownOne).isAllOnesValue())
2023 RI.setOperand(0, Constant::getIntegerValue(VTy, KnownOne));
2028 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
2029 // Change br (not X), label True, label False to: br X, label False, True
2031 BasicBlock *TrueDest;
2032 BasicBlock *FalseDest;
2033 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
2034 !isa<Constant>(X)) {
2035 // Swap Destinations and condition...
2037 BI.swapSuccessors();
2041 // Canonicalize fcmp_one -> fcmp_oeq
2042 FCmpInst::Predicate FPred; Value *Y;
2043 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
2044 TrueDest, FalseDest)) &&
2045 BI.getCondition()->hasOneUse())
2046 if (FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
2047 FPred == FCmpInst::FCMP_OGE) {
2048 FCmpInst *Cond = cast<FCmpInst>(BI.getCondition());
2049 Cond->setPredicate(FCmpInst::getInversePredicate(FPred));
2051 // Swap Destinations and condition.
2052 BI.swapSuccessors();
2057 // Canonicalize icmp_ne -> icmp_eq
2058 ICmpInst::Predicate IPred;
2059 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
2060 TrueDest, FalseDest)) &&
2061 BI.getCondition()->hasOneUse())
2062 if (IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
2063 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
2064 IPred == ICmpInst::ICMP_SGE) {
2065 ICmpInst *Cond = cast<ICmpInst>(BI.getCondition());
2066 Cond->setPredicate(ICmpInst::getInversePredicate(IPred));
2067 // Swap Destinations and condition.
2068 BI.swapSuccessors();
2076 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
2077 Value *Cond = SI.getCondition();
2078 unsigned BitWidth = cast<IntegerType>(Cond->getType())->getBitWidth();
2079 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2080 computeKnownBits(Cond, KnownZero, KnownOne);
2081 unsigned LeadingKnownZeros = KnownZero.countLeadingOnes();
2082 unsigned LeadingKnownOnes = KnownOne.countLeadingOnes();
2084 // Compute the number of leading bits we can ignore.
2085 for (auto &C : SI.cases()) {
2086 LeadingKnownZeros = std::min(
2087 LeadingKnownZeros, C.getCaseValue()->getValue().countLeadingZeros());
2088 LeadingKnownOnes = std::min(
2089 LeadingKnownOnes, C.getCaseValue()->getValue().countLeadingOnes());
2092 unsigned NewWidth = BitWidth - std::max(LeadingKnownZeros, LeadingKnownOnes);
2094 // Truncate the condition operand if the new type is equal to or larger than
2095 // the largest legal integer type. We need to be conservative here since
2096 // x86 generates redundant zero-extenstion instructions if the operand is
2097 // truncated to i8 or i16.
2098 if (DL && BitWidth > NewWidth &&
2099 NewWidth >= DL->getLargestLegalIntTypeSize()) {
2100 IntegerType *Ty = IntegerType::get(SI.getContext(), NewWidth);
2101 Builder->SetInsertPoint(&SI);
2102 Value *NewCond = Builder->CreateTrunc(SI.getCondition(), Ty, "trunc");
2103 SI.setCondition(NewCond);
2105 for (auto &C : SI.cases())
2106 static_cast<SwitchInst::CaseIt *>(&C)->setValue(ConstantInt::get(
2107 SI.getContext(), C.getCaseValue()->getValue().trunc(NewWidth)));
2110 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
2111 if (I->getOpcode() == Instruction::Add)
2112 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2113 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
2114 // Skip the first item since that's the default case.
2115 for (SwitchInst::CaseIt i = SI.case_begin(), e = SI.case_end();
2117 ConstantInt* CaseVal = i.getCaseValue();
2118 Constant* NewCaseVal = ConstantExpr::getSub(cast<Constant>(CaseVal),
2120 assert(isa<ConstantInt>(NewCaseVal) &&
2121 "Result of expression should be constant");
2122 i.setValue(cast<ConstantInt>(NewCaseVal));
2124 SI.setCondition(I->getOperand(0));
2132 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
2133 Value *Agg = EV.getAggregateOperand();
2135 if (!EV.hasIndices())
2136 return ReplaceInstUsesWith(EV, Agg);
2138 if (Constant *C = dyn_cast<Constant>(Agg)) {
2139 if (Constant *C2 = C->getAggregateElement(*EV.idx_begin())) {
2140 if (EV.getNumIndices() == 0)
2141 return ReplaceInstUsesWith(EV, C2);
2142 // Extract the remaining indices out of the constant indexed by the
2144 return ExtractValueInst::Create(C2, EV.getIndices().slice(1));
2146 return nullptr; // Can't handle other constants
2149 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
2150 // We're extracting from an insertvalue instruction, compare the indices
2151 const unsigned *exti, *exte, *insi, *inse;
2152 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
2153 exte = EV.idx_end(), inse = IV->idx_end();
2154 exti != exte && insi != inse;
2157 // The insert and extract both reference distinctly different elements.
2158 // This means the extract is not influenced by the insert, and we can
2159 // replace the aggregate operand of the extract with the aggregate
2160 // operand of the insert. i.e., replace
2161 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
2162 // %E = extractvalue { i32, { i32 } } %I, 0
2164 // %E = extractvalue { i32, { i32 } } %A, 0
2165 return ExtractValueInst::Create(IV->getAggregateOperand(),
2168 if (exti == exte && insi == inse)
2169 // Both iterators are at the end: Index lists are identical. Replace
2170 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
2171 // %C = extractvalue { i32, { i32 } } %B, 1, 0
2173 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
2175 // The extract list is a prefix of the insert list. i.e. replace
2176 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
2177 // %E = extractvalue { i32, { i32 } } %I, 1
2179 // %X = extractvalue { i32, { i32 } } %A, 1
2180 // %E = insertvalue { i32 } %X, i32 42, 0
2181 // by switching the order of the insert and extract (though the
2182 // insertvalue should be left in, since it may have other uses).
2183 Value *NewEV = Builder->CreateExtractValue(IV->getAggregateOperand(),
2185 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
2186 makeArrayRef(insi, inse));
2189 // The insert list is a prefix of the extract list
2190 // We can simply remove the common indices from the extract and make it
2191 // operate on the inserted value instead of the insertvalue result.
2193 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
2194 // %E = extractvalue { i32, { i32 } } %I, 1, 0
2196 // %E extractvalue { i32 } { i32 42 }, 0
2197 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
2198 makeArrayRef(exti, exte));
2200 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) {
2201 // We're extracting from an intrinsic, see if we're the only user, which
2202 // allows us to simplify multiple result intrinsics to simpler things that
2203 // just get one value.
2204 if (II->hasOneUse()) {
2205 // Check if we're grabbing the overflow bit or the result of a 'with
2206 // overflow' intrinsic. If it's the latter we can remove the intrinsic
2207 // and replace it with a traditional binary instruction.
2208 switch (II->getIntrinsicID()) {
2209 case Intrinsic::uadd_with_overflow:
2210 case Intrinsic::sadd_with_overflow:
2211 if (*EV.idx_begin() == 0) { // Normal result.
2212 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2213 ReplaceInstUsesWith(*II, UndefValue::get(II->getType()));
2214 EraseInstFromFunction(*II);
2215 return BinaryOperator::CreateAdd(LHS, RHS);
2218 // If the normal result of the add is dead, and the RHS is a constant,
2219 // we can transform this into a range comparison.
2220 // overflow = uadd a, -4 --> overflow = icmp ugt a, 3
2221 if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow)
2222 if (ConstantInt *CI = dyn_cast<ConstantInt>(II->getArgOperand(1)))
2223 return new ICmpInst(ICmpInst::ICMP_UGT, II->getArgOperand(0),
2224 ConstantExpr::getNot(CI));
2226 case Intrinsic::usub_with_overflow:
2227 case Intrinsic::ssub_with_overflow:
2228 if (*EV.idx_begin() == 0) { // Normal result.
2229 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2230 ReplaceInstUsesWith(*II, UndefValue::get(II->getType()));
2231 EraseInstFromFunction(*II);
2232 return BinaryOperator::CreateSub(LHS, RHS);
2235 case Intrinsic::umul_with_overflow:
2236 case Intrinsic::smul_with_overflow:
2237 if (*EV.idx_begin() == 0) { // Normal result.
2238 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2239 ReplaceInstUsesWith(*II, UndefValue::get(II->getType()));
2240 EraseInstFromFunction(*II);
2241 return BinaryOperator::CreateMul(LHS, RHS);
2249 if (LoadInst *L = dyn_cast<LoadInst>(Agg))
2250 // If the (non-volatile) load only has one use, we can rewrite this to a
2251 // load from a GEP. This reduces the size of the load.
2252 // FIXME: If a load is used only by extractvalue instructions then this
2253 // could be done regardless of having multiple uses.
2254 if (L->isSimple() && L->hasOneUse()) {
2255 // extractvalue has integer indices, getelementptr has Value*s. Convert.
2256 SmallVector<Value*, 4> Indices;
2257 // Prefix an i32 0 since we need the first element.
2258 Indices.push_back(Builder->getInt32(0));
2259 for (ExtractValueInst::idx_iterator I = EV.idx_begin(), E = EV.idx_end();
2261 Indices.push_back(Builder->getInt32(*I));
2263 // We need to insert these at the location of the old load, not at that of
2264 // the extractvalue.
2265 Builder->SetInsertPoint(L->getParent(), L);
2266 Value *GEP = Builder->CreateInBoundsGEP(L->getPointerOperand(), Indices);
2267 // Returning the load directly will cause the main loop to insert it in
2268 // the wrong spot, so use ReplaceInstUsesWith().
2269 return ReplaceInstUsesWith(EV, Builder->CreateLoad(GEP));
2271 // We could simplify extracts from other values. Note that nested extracts may
2272 // already be simplified implicitly by the above: extract (extract (insert) )
2273 // will be translated into extract ( insert ( extract ) ) first and then just
2274 // the value inserted, if appropriate. Similarly for extracts from single-use
2275 // loads: extract (extract (load)) will be translated to extract (load (gep))
2276 // and if again single-use then via load (gep (gep)) to load (gep).
2277 // However, double extracts from e.g. function arguments or return values
2278 // aren't handled yet.
2282 enum Personality_Type {
2283 Unknown_Personality,
2284 GNU_Ada_Personality,
2285 GNU_CXX_Personality,
2286 GNU_ObjC_Personality
2289 /// RecognizePersonality - See if the given exception handling personality
2290 /// function is one that we understand. If so, return a description of it;
2291 /// otherwise return Unknown_Personality.
2292 static Personality_Type RecognizePersonality(Value *Pers) {
2293 Function *F = dyn_cast<Function>(Pers->stripPointerCasts());
2295 return Unknown_Personality;
2296 return StringSwitch<Personality_Type>(F->getName())
2297 .Case("__gnat_eh_personality", GNU_Ada_Personality)
2298 .Case("__gxx_personality_v0", GNU_CXX_Personality)
2299 .Case("__objc_personality_v0", GNU_ObjC_Personality)
2300 .Default(Unknown_Personality);
2303 /// isCatchAll - Return 'true' if the given typeinfo will match anything.
2304 static bool isCatchAll(Personality_Type Personality, Constant *TypeInfo) {
2305 switch (Personality) {
2306 case Unknown_Personality:
2308 case GNU_Ada_Personality:
2309 // While __gnat_all_others_value will match any Ada exception, it doesn't
2310 // match foreign exceptions (or didn't, before gcc-4.7).
2312 case GNU_CXX_Personality:
2313 case GNU_ObjC_Personality:
2314 return TypeInfo->isNullValue();
2316 llvm_unreachable("Unknown personality!");
2319 static bool shorter_filter(const Value *LHS, const Value *RHS) {
2321 cast<ArrayType>(LHS->getType())->getNumElements()
2323 cast<ArrayType>(RHS->getType())->getNumElements();
2326 Instruction *InstCombiner::visitLandingPadInst(LandingPadInst &LI) {
2327 // The logic here should be correct for any real-world personality function.
2328 // However if that turns out not to be true, the offending logic can always
2329 // be conditioned on the personality function, like the catch-all logic is.
2330 Personality_Type Personality = RecognizePersonality(LI.getPersonalityFn());
2332 // Simplify the list of clauses, eg by removing repeated catch clauses
2333 // (these are often created by inlining).
2334 bool MakeNewInstruction = false; // If true, recreate using the following:
2335 SmallVector<Constant *, 16> NewClauses; // - Clauses for the new instruction;
2336 bool CleanupFlag = LI.isCleanup(); // - The new instruction is a cleanup.
2338 SmallPtrSet<Value *, 16> AlreadyCaught; // Typeinfos known caught already.
2339 for (unsigned i = 0, e = LI.getNumClauses(); i != e; ++i) {
2340 bool isLastClause = i + 1 == e;
2341 if (LI.isCatch(i)) {
2343 Constant *CatchClause = LI.getClause(i);
2344 Constant *TypeInfo = CatchClause->stripPointerCasts();
2346 // If we already saw this clause, there is no point in having a second
2348 if (AlreadyCaught.insert(TypeInfo).second) {
2349 // This catch clause was not already seen.
2350 NewClauses.push_back(CatchClause);
2352 // Repeated catch clause - drop the redundant copy.
2353 MakeNewInstruction = true;
2356 // If this is a catch-all then there is no point in keeping any following
2357 // clauses or marking the landingpad as having a cleanup.
2358 if (isCatchAll(Personality, TypeInfo)) {
2360 MakeNewInstruction = true;
2361 CleanupFlag = false;
2365 // A filter clause. If any of the filter elements were already caught
2366 // then they can be dropped from the filter. It is tempting to try to
2367 // exploit the filter further by saying that any typeinfo that does not
2368 // occur in the filter can't be caught later (and thus can be dropped).
2369 // However this would be wrong, since typeinfos can match without being
2370 // equal (for example if one represents a C++ class, and the other some
2371 // class derived from it).
2372 assert(LI.isFilter(i) && "Unsupported landingpad clause!");
2373 Constant *FilterClause = LI.getClause(i);
2374 ArrayType *FilterType = cast<ArrayType>(FilterClause->getType());
2375 unsigned NumTypeInfos = FilterType->getNumElements();
2377 // An empty filter catches everything, so there is no point in keeping any
2378 // following clauses or marking the landingpad as having a cleanup. By
2379 // dealing with this case here the following code is made a bit simpler.
2380 if (!NumTypeInfos) {
2381 NewClauses.push_back(FilterClause);
2383 MakeNewInstruction = true;
2384 CleanupFlag = false;
2388 bool MakeNewFilter = false; // If true, make a new filter.
2389 SmallVector<Constant *, 16> NewFilterElts; // New elements.
2390 if (isa<ConstantAggregateZero>(FilterClause)) {
2391 // Not an empty filter - it contains at least one null typeinfo.
2392 assert(NumTypeInfos > 0 && "Should have handled empty filter already!");
2393 Constant *TypeInfo =
2394 Constant::getNullValue(FilterType->getElementType());
2395 // If this typeinfo is a catch-all then the filter can never match.
2396 if (isCatchAll(Personality, TypeInfo)) {
2397 // Throw the filter away.
2398 MakeNewInstruction = true;
2402 // There is no point in having multiple copies of this typeinfo, so
2403 // discard all but the first copy if there is more than one.
2404 NewFilterElts.push_back(TypeInfo);
2405 if (NumTypeInfos > 1)
2406 MakeNewFilter = true;
2408 ConstantArray *Filter = cast<ConstantArray>(FilterClause);
2409 SmallPtrSet<Value *, 16> SeenInFilter; // For uniquing the elements.
2410 NewFilterElts.reserve(NumTypeInfos);
2412 // Remove any filter elements that were already caught or that already
2413 // occurred in the filter. While there, see if any of the elements are
2414 // catch-alls. If so, the filter can be discarded.
2415 bool SawCatchAll = false;
2416 for (unsigned j = 0; j != NumTypeInfos; ++j) {
2417 Constant *Elt = Filter->getOperand(j);
2418 Constant *TypeInfo = Elt->stripPointerCasts();
2419 if (isCatchAll(Personality, TypeInfo)) {
2420 // This element is a catch-all. Bail out, noting this fact.
2424 if (AlreadyCaught.count(TypeInfo))
2425 // Already caught by an earlier clause, so having it in the filter
2428 // There is no point in having multiple copies of the same typeinfo in
2429 // a filter, so only add it if we didn't already.
2430 if (SeenInFilter.insert(TypeInfo).second)
2431 NewFilterElts.push_back(cast<Constant>(Elt));
2433 // A filter containing a catch-all cannot match anything by definition.
2435 // Throw the filter away.
2436 MakeNewInstruction = true;
2440 // If we dropped something from the filter, make a new one.
2441 if (NewFilterElts.size() < NumTypeInfos)
2442 MakeNewFilter = true;
2444 if (MakeNewFilter) {
2445 FilterType = ArrayType::get(FilterType->getElementType(),
2446 NewFilterElts.size());
2447 FilterClause = ConstantArray::get(FilterType, NewFilterElts);
2448 MakeNewInstruction = true;
2451 NewClauses.push_back(FilterClause);
2453 // If the new filter is empty then it will catch everything so there is
2454 // no point in keeping any following clauses or marking the landingpad
2455 // as having a cleanup. The case of the original filter being empty was
2456 // already handled above.
2457 if (MakeNewFilter && !NewFilterElts.size()) {
2458 assert(MakeNewInstruction && "New filter but not a new instruction!");
2459 CleanupFlag = false;
2465 // If several filters occur in a row then reorder them so that the shortest
2466 // filters come first (those with the smallest number of elements). This is
2467 // advantageous because shorter filters are more likely to match, speeding up
2468 // unwinding, but mostly because it increases the effectiveness of the other
2469 // filter optimizations below.
2470 for (unsigned i = 0, e = NewClauses.size(); i + 1 < e; ) {
2472 // Find the maximal 'j' s.t. the range [i, j) consists entirely of filters.
2473 for (j = i; j != e; ++j)
2474 if (!isa<ArrayType>(NewClauses[j]->getType()))
2477 // Check whether the filters are already sorted by length. We need to know
2478 // if sorting them is actually going to do anything so that we only make a
2479 // new landingpad instruction if it does.
2480 for (unsigned k = i; k + 1 < j; ++k)
2481 if (shorter_filter(NewClauses[k+1], NewClauses[k])) {
2482 // Not sorted, so sort the filters now. Doing an unstable sort would be
2483 // correct too but reordering filters pointlessly might confuse users.
2484 std::stable_sort(NewClauses.begin() + i, NewClauses.begin() + j,
2486 MakeNewInstruction = true;
2490 // Look for the next batch of filters.
2494 // If typeinfos matched if and only if equal, then the elements of a filter L
2495 // that occurs later than a filter F could be replaced by the intersection of
2496 // the elements of F and L. In reality two typeinfos can match without being
2497 // equal (for example if one represents a C++ class, and the other some class
2498 // derived from it) so it would be wrong to perform this transform in general.
2499 // However the transform is correct and useful if F is a subset of L. In that
2500 // case L can be replaced by F, and thus removed altogether since repeating a
2501 // filter is pointless. So here we look at all pairs of filters F and L where
2502 // L follows F in the list of clauses, and remove L if every element of F is
2503 // an element of L. This can occur when inlining C++ functions with exception
2505 for (unsigned i = 0; i + 1 < NewClauses.size(); ++i) {
2506 // Examine each filter in turn.
2507 Value *Filter = NewClauses[i];
2508 ArrayType *FTy = dyn_cast<ArrayType>(Filter->getType());
2510 // Not a filter - skip it.
2512 unsigned FElts = FTy->getNumElements();
2513 // Examine each filter following this one. Doing this backwards means that
2514 // we don't have to worry about filters disappearing under us when removed.
2515 for (unsigned j = NewClauses.size() - 1; j != i; --j) {
2516 Value *LFilter = NewClauses[j];
2517 ArrayType *LTy = dyn_cast<ArrayType>(LFilter->getType());
2519 // Not a filter - skip it.
2521 // If Filter is a subset of LFilter, i.e. every element of Filter is also
2522 // an element of LFilter, then discard LFilter.
2523 SmallVectorImpl<Constant *>::iterator J = NewClauses.begin() + j;
2524 // If Filter is empty then it is a subset of LFilter.
2527 NewClauses.erase(J);
2528 MakeNewInstruction = true;
2529 // Move on to the next filter.
2532 unsigned LElts = LTy->getNumElements();
2533 // If Filter is longer than LFilter then it cannot be a subset of it.
2535 // Move on to the next filter.
2537 // At this point we know that LFilter has at least one element.
2538 if (isa<ConstantAggregateZero>(LFilter)) { // LFilter only contains zeros.
2539 // Filter is a subset of LFilter iff Filter contains only zeros (as we
2540 // already know that Filter is not longer than LFilter).
2541 if (isa<ConstantAggregateZero>(Filter)) {
2542 assert(FElts <= LElts && "Should have handled this case earlier!");
2544 NewClauses.erase(J);
2545 MakeNewInstruction = true;
2547 // Move on to the next filter.
2550 ConstantArray *LArray = cast<ConstantArray>(LFilter);
2551 if (isa<ConstantAggregateZero>(Filter)) { // Filter only contains zeros.
2552 // Since Filter is non-empty and contains only zeros, it is a subset of
2553 // LFilter iff LFilter contains a zero.
2554 assert(FElts > 0 && "Should have eliminated the empty filter earlier!");
2555 for (unsigned l = 0; l != LElts; ++l)
2556 if (LArray->getOperand(l)->isNullValue()) {
2557 // LFilter contains a zero - discard it.
2558 NewClauses.erase(J);
2559 MakeNewInstruction = true;
2562 // Move on to the next filter.
2565 // At this point we know that both filters are ConstantArrays. Loop over
2566 // operands to see whether every element of Filter is also an element of
2567 // LFilter. Since filters tend to be short this is probably faster than
2568 // using a method that scales nicely.
2569 ConstantArray *FArray = cast<ConstantArray>(Filter);
2570 bool AllFound = true;
2571 for (unsigned f = 0; f != FElts; ++f) {
2572 Value *FTypeInfo = FArray->getOperand(f)->stripPointerCasts();
2574 for (unsigned l = 0; l != LElts; ++l) {
2575 Value *LTypeInfo = LArray->getOperand(l)->stripPointerCasts();
2576 if (LTypeInfo == FTypeInfo) {
2586 NewClauses.erase(J);
2587 MakeNewInstruction = true;
2589 // Move on to the next filter.
2593 // If we changed any of the clauses, replace the old landingpad instruction
2595 if (MakeNewInstruction) {
2596 LandingPadInst *NLI = LandingPadInst::Create(LI.getType(),
2597 LI.getPersonalityFn(),
2599 for (unsigned i = 0, e = NewClauses.size(); i != e; ++i)
2600 NLI->addClause(NewClauses[i]);
2601 // A landing pad with no clauses must have the cleanup flag set. It is
2602 // theoretically possible, though highly unlikely, that we eliminated all
2603 // clauses. If so, force the cleanup flag to true.
2604 if (NewClauses.empty())
2606 NLI->setCleanup(CleanupFlag);
2610 // Even if none of the clauses changed, we may nonetheless have understood
2611 // that the cleanup flag is pointless. Clear it if so.
2612 if (LI.isCleanup() != CleanupFlag) {
2613 assert(!CleanupFlag && "Adding a cleanup, not removing one?!");
2614 LI.setCleanup(CleanupFlag);
2624 /// TryToSinkInstruction - Try to move the specified instruction from its
2625 /// current block into the beginning of DestBlock, which can only happen if it's
2626 /// safe to move the instruction past all of the instructions between it and the
2627 /// end of its block.
2628 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
2629 assert(I->hasOneUse() && "Invariants didn't hold!");
2631 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
2632 if (isa<PHINode>(I) || isa<LandingPadInst>(I) || I->mayHaveSideEffects() ||
2633 isa<TerminatorInst>(I))
2636 // Do not sink alloca instructions out of the entry block.
2637 if (isa<AllocaInst>(I) && I->getParent() ==
2638 &DestBlock->getParent()->getEntryBlock())
2641 // We can only sink load instructions if there is nothing between the load and
2642 // the end of block that could change the value.
2643 if (I->mayReadFromMemory()) {
2644 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
2646 if (Scan->mayWriteToMemory())
2650 BasicBlock::iterator InsertPos = DestBlock->getFirstInsertionPt();
2651 I->moveBefore(InsertPos);
2657 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
2658 /// all reachable code to the worklist.
2660 /// This has a couple of tricks to make the code faster and more powerful. In
2661 /// particular, we constant fold and DCE instructions as we go, to avoid adding
2662 /// them to the worklist (this significantly speeds up instcombine on code where
2663 /// many instructions are dead or constant). Additionally, if we find a branch
2664 /// whose condition is a known constant, we only visit the reachable successors.
2666 static bool AddReachableCodeToWorklist(BasicBlock *BB,
2667 SmallPtrSetImpl<BasicBlock*> &Visited,
2669 const DataLayout *DL,
2670 const TargetLibraryInfo *TLI) {
2671 bool MadeIRChange = false;
2672 SmallVector<BasicBlock*, 256> Worklist;
2673 Worklist.push_back(BB);
2675 SmallVector<Instruction*, 128> InstrsForInstCombineWorklist;
2676 DenseMap<ConstantExpr*, Constant*> FoldedConstants;
2679 BB = Worklist.pop_back_val();
2681 // We have now visited this block! If we've already been here, ignore it.
2682 if (!Visited.insert(BB).second)
2685 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
2686 Instruction *Inst = BBI++;
2688 // DCE instruction if trivially dead.
2689 if (isInstructionTriviallyDead(Inst, TLI)) {
2691 DEBUG(dbgs() << "IC: DCE: " << *Inst << '\n');
2692 Inst->eraseFromParent();
2696 // ConstantProp instruction if trivially constant.
2697 if (!Inst->use_empty() && isa<Constant>(Inst->getOperand(0)))
2698 if (Constant *C = ConstantFoldInstruction(Inst, DL, TLI)) {
2699 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: "
2701 Inst->replaceAllUsesWith(C);
2703 Inst->eraseFromParent();
2708 // See if we can constant fold its operands.
2709 for (User::op_iterator i = Inst->op_begin(), e = Inst->op_end();
2711 ConstantExpr *CE = dyn_cast<ConstantExpr>(i);
2712 if (CE == nullptr) continue;
2714 Constant*& FoldRes = FoldedConstants[CE];
2716 FoldRes = ConstantFoldConstantExpression(CE, DL, TLI);
2720 if (FoldRes != CE) {
2722 MadeIRChange = true;
2727 InstrsForInstCombineWorklist.push_back(Inst);
2730 // Recursively visit successors. If this is a branch or switch on a
2731 // constant, only visit the reachable successor.
2732 TerminatorInst *TI = BB->getTerminator();
2733 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
2734 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
2735 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
2736 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
2737 Worklist.push_back(ReachableBB);
2740 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
2741 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
2742 // See if this is an explicit destination.
2743 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2745 if (i.getCaseValue() == Cond) {
2746 BasicBlock *ReachableBB = i.getCaseSuccessor();
2747 Worklist.push_back(ReachableBB);
2751 // Otherwise it is the default destination.
2752 Worklist.push_back(SI->getDefaultDest());
2757 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
2758 Worklist.push_back(TI->getSuccessor(i));
2759 } while (!Worklist.empty());
2761 // Once we've found all of the instructions to add to instcombine's worklist,
2762 // add them in reverse order. This way instcombine will visit from the top
2763 // of the function down. This jives well with the way that it adds all uses
2764 // of instructions to the worklist after doing a transformation, thus avoiding
2765 // some N^2 behavior in pathological cases.
2766 IC.Worklist.AddInitialGroup(&InstrsForInstCombineWorklist[0],
2767 InstrsForInstCombineWorklist.size());
2769 return MadeIRChange;
2772 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
2773 MadeIRChange = false;
2775 DEBUG(dbgs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
2776 << F.getName() << "\n");
2779 // Do a depth-first traversal of the function, populate the worklist with
2780 // the reachable instructions. Ignore blocks that are not reachable. Keep
2781 // track of which blocks we visit.
2782 SmallPtrSet<BasicBlock*, 64> Visited;
2783 MadeIRChange |= AddReachableCodeToWorklist(F.begin(), Visited, *this, DL,
2786 // Do a quick scan over the function. If we find any blocks that are
2787 // unreachable, remove any instructions inside of them. This prevents
2788 // the instcombine code from having to deal with some bad special cases.
2789 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
2790 if (Visited.count(BB)) continue;
2792 // Delete the instructions backwards, as it has a reduced likelihood of
2793 // having to update as many def-use and use-def chains.
2794 Instruction *EndInst = BB->getTerminator(); // Last not to be deleted.
2795 while (EndInst != BB->begin()) {
2796 // Delete the next to last instruction.
2797 BasicBlock::iterator I = EndInst;
2798 Instruction *Inst = --I;
2799 if (!Inst->use_empty())
2800 Inst->replaceAllUsesWith(UndefValue::get(Inst->getType()));
2801 if (isa<LandingPadInst>(Inst)) {
2805 if (!isa<DbgInfoIntrinsic>(Inst)) {
2807 MadeIRChange = true;
2809 Inst->eraseFromParent();
2814 while (!Worklist.isEmpty()) {
2815 Instruction *I = Worklist.RemoveOne();
2816 if (I == nullptr) continue; // skip null values.
2818 // Check to see if we can DCE the instruction.
2819 if (isInstructionTriviallyDead(I, TLI)) {
2820 DEBUG(dbgs() << "IC: DCE: " << *I << '\n');
2821 EraseInstFromFunction(*I);
2823 MadeIRChange = true;
2827 // Instruction isn't dead, see if we can constant propagate it.
2828 if (!I->use_empty() && isa<Constant>(I->getOperand(0)))
2829 if (Constant *C = ConstantFoldInstruction(I, DL, TLI)) {
2830 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n');
2832 // Add operands to the worklist.
2833 ReplaceInstUsesWith(*I, C);
2835 EraseInstFromFunction(*I);
2836 MadeIRChange = true;
2840 // See if we can trivially sink this instruction to a successor basic block.
2841 if (I->hasOneUse()) {
2842 BasicBlock *BB = I->getParent();
2843 Instruction *UserInst = cast<Instruction>(*I->user_begin());
2844 BasicBlock *UserParent;
2846 // Get the block the use occurs in.
2847 if (PHINode *PN = dyn_cast<PHINode>(UserInst))
2848 UserParent = PN->getIncomingBlock(*I->use_begin());
2850 UserParent = UserInst->getParent();
2852 if (UserParent != BB) {
2853 bool UserIsSuccessor = false;
2854 // See if the user is one of our successors.
2855 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
2856 if (*SI == UserParent) {
2857 UserIsSuccessor = true;
2861 // If the user is one of our immediate successors, and if that successor
2862 // only has us as a predecessors (we'd have to split the critical edge
2863 // otherwise), we can keep going.
2864 if (UserIsSuccessor && UserParent->getSinglePredecessor()) {
2865 // Okay, the CFG is simple enough, try to sink this instruction.
2866 if (TryToSinkInstruction(I, UserParent)) {
2867 MadeIRChange = true;
2868 // We'll add uses of the sunk instruction below, but since sinking
2869 // can expose opportunities for it's *operands* add them to the
2871 for (Use &U : I->operands())
2872 if (Instruction *OpI = dyn_cast<Instruction>(U.get()))
2879 // Now that we have an instruction, try combining it to simplify it.
2880 Builder->SetInsertPoint(I->getParent(), I);
2881 Builder->SetCurrentDebugLocation(I->getDebugLoc());
2886 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str(););
2887 DEBUG(dbgs() << "IC: Visiting: " << OrigI << '\n');
2889 if (Instruction *Result = visit(*I)) {
2891 // Should we replace the old instruction with a new one?
2893 DEBUG(dbgs() << "IC: Old = " << *I << '\n'
2894 << " New = " << *Result << '\n');
2896 if (!I->getDebugLoc().isUnknown())
2897 Result->setDebugLoc(I->getDebugLoc());
2898 // Everything uses the new instruction now.
2899 I->replaceAllUsesWith(Result);
2901 // Move the name to the new instruction first.
2902 Result->takeName(I);
2904 // Push the new instruction and any users onto the worklist.
2905 Worklist.Add(Result);
2906 Worklist.AddUsersToWorkList(*Result);
2908 // Insert the new instruction into the basic block...
2909 BasicBlock *InstParent = I->getParent();
2910 BasicBlock::iterator InsertPos = I;
2912 // If we replace a PHI with something that isn't a PHI, fix up the
2914 if (!isa<PHINode>(Result) && isa<PHINode>(InsertPos))
2915 InsertPos = InstParent->getFirstInsertionPt();
2917 InstParent->getInstList().insert(InsertPos, Result);
2919 EraseInstFromFunction(*I);
2922 DEBUG(dbgs() << "IC: Mod = " << OrigI << '\n'
2923 << " New = " << *I << '\n');
2926 // If the instruction was modified, it's possible that it is now dead.
2927 // if so, remove it.
2928 if (isInstructionTriviallyDead(I, TLI)) {
2929 EraseInstFromFunction(*I);
2932 Worklist.AddUsersToWorkList(*I);
2935 MadeIRChange = true;
2940 return MadeIRChange;
2944 class InstCombinerLibCallSimplifier final : public LibCallSimplifier {
2947 InstCombinerLibCallSimplifier(const DataLayout *DL,
2948 const TargetLibraryInfo *TLI,
2950 : LibCallSimplifier(DL, TLI) {
2954 /// replaceAllUsesWith - override so that instruction replacement
2955 /// can be defined in terms of the instruction combiner framework.
2956 void replaceAllUsesWith(Instruction *I, Value *With) const override {
2957 IC->ReplaceInstUsesWith(*I, With);
2962 bool InstCombiner::runOnFunction(Function &F) {
2963 if (skipOptnoneFunction(F))
2966 AT = &getAnalysis<AssumptionTracker>();
2967 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
2968 DL = DLP ? &DLP->getDataLayout() : nullptr;
2969 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2970 TLI = &getAnalysis<TargetLibraryInfo>();
2973 MinimizeSize = F.getAttributes().hasAttribute(AttributeSet::FunctionIndex,
2974 Attribute::MinSize);
2976 /// Builder - This is an IRBuilder that automatically inserts new
2977 /// instructions into the worklist when they are created.
2978 IRBuilder<true, TargetFolder, InstCombineIRInserter>
2979 TheBuilder(F.getContext(), TargetFolder(DL),
2980 InstCombineIRInserter(Worklist, AT));
2981 Builder = &TheBuilder;
2983 InstCombinerLibCallSimplifier TheSimplifier(DL, TLI, this);
2984 Simplifier = &TheSimplifier;
2986 bool EverMadeChange = false;
2988 // Lower dbg.declare intrinsics otherwise their value may be clobbered
2990 EverMadeChange = LowerDbgDeclare(F);
2992 // Iterate while there is work to do.
2993 unsigned Iteration = 0;
2994 while (DoOneIteration(F, Iteration++))
2995 EverMadeChange = true;
2998 return EverMadeChange;
3001 FunctionPass *llvm::createInstructionCombiningPass() {
3002 return new InstCombiner();