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/ConstantFolding.h"
44 #include "llvm/Analysis/InstructionSimplify.h"
45 #include "llvm/Analysis/MemoryBuiltins.h"
46 #include "llvm/Analysis/ValueTracking.h"
47 #include "llvm/IR/CFG.h"
48 #include "llvm/IR/DataLayout.h"
49 #include "llvm/IR/Dominators.h"
50 #include "llvm/IR/GetElementPtrTypeIterator.h"
51 #include "llvm/IR/IntrinsicInst.h"
52 #include "llvm/IR/PatternMatch.h"
53 #include "llvm/IR/ValueHandle.h"
54 #include "llvm/Support/CommandLine.h"
55 #include "llvm/Support/Debug.h"
56 #include "llvm/Target/TargetLibraryInfo.h"
57 #include "llvm/Transforms/Utils/Local.h"
61 using namespace llvm::PatternMatch;
63 #define DEBUG_TYPE "instcombine"
65 STATISTIC(NumCombined , "Number of insts combined");
66 STATISTIC(NumConstProp, "Number of constant folds");
67 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
68 STATISTIC(NumSunkInst , "Number of instructions sunk");
69 STATISTIC(NumExpand, "Number of expansions");
70 STATISTIC(NumFactor , "Number of factorizations");
71 STATISTIC(NumReassoc , "Number of reassociations");
74 EnableUnsafeFPShrink("enable-double-float-shrink", cl::Hidden,
76 cl::desc("Enable unsafe double to float "
77 "shrinking for math lib calls"));
79 // Initialization Routines
80 void llvm::initializeInstCombine(PassRegistry &Registry) {
81 initializeInstCombinerPass(Registry);
84 void LLVMInitializeInstCombine(LLVMPassRegistryRef R) {
85 initializeInstCombine(*unwrap(R));
88 char InstCombiner::ID = 0;
89 INITIALIZE_PASS_BEGIN(InstCombiner, "instcombine",
90 "Combine redundant instructions", false, false)
91 INITIALIZE_PASS_DEPENDENCY(AssumptionTracker)
92 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
93 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
94 INITIALIZE_PASS_END(InstCombiner, "instcombine",
95 "Combine redundant instructions", false, false)
97 void InstCombiner::getAnalysisUsage(AnalysisUsage &AU) const {
99 AU.addRequired<AssumptionTracker>();
100 AU.addRequired<TargetLibraryInfo>();
101 AU.addRequired<DominatorTreeWrapperPass>();
102 AU.addPreserved<DominatorTreeWrapperPass>();
106 Value *InstCombiner::EmitGEPOffset(User *GEP) {
107 return llvm::EmitGEPOffset(Builder, *getDataLayout(), GEP);
110 /// ShouldChangeType - Return true if it is desirable to convert a computation
111 /// from 'From' to 'To'. We don't want to convert from a legal to an illegal
112 /// type for example, or from a smaller to a larger illegal type.
113 bool InstCombiner::ShouldChangeType(Type *From, Type *To) const {
114 assert(From->isIntegerTy() && To->isIntegerTy());
116 // If we don't have DL, we don't know if the source/dest are legal.
117 if (!DL) return false;
119 unsigned FromWidth = From->getPrimitiveSizeInBits();
120 unsigned ToWidth = To->getPrimitiveSizeInBits();
121 bool FromLegal = DL->isLegalInteger(FromWidth);
122 bool ToLegal = DL->isLegalInteger(ToWidth);
124 // If this is a legal integer from type, and the result would be an illegal
125 // type, don't do the transformation.
126 if (FromLegal && !ToLegal)
129 // Otherwise, if both are illegal, do not increase the size of the result. We
130 // do allow things like i160 -> i64, but not i64 -> i160.
131 if (!FromLegal && !ToLegal && ToWidth > FromWidth)
137 // Return true, if No Signed Wrap should be maintained for I.
138 // The No Signed Wrap flag can be kept if the operation "B (I.getOpcode) C",
139 // where both B and C should be ConstantInts, results in a constant that does
140 // not overflow. This function only handles the Add and Sub opcodes. For
141 // all other opcodes, the function conservatively returns false.
142 static bool MaintainNoSignedWrap(BinaryOperator &I, Value *B, Value *C) {
143 OverflowingBinaryOperator *OBO = dyn_cast<OverflowingBinaryOperator>(&I);
144 if (!OBO || !OBO->hasNoSignedWrap()) {
148 // We reason about Add and Sub Only.
149 Instruction::BinaryOps Opcode = I.getOpcode();
150 if (Opcode != Instruction::Add &&
151 Opcode != Instruction::Sub) {
155 ConstantInt *CB = dyn_cast<ConstantInt>(B);
156 ConstantInt *CC = dyn_cast<ConstantInt>(C);
162 const APInt &BVal = CB->getValue();
163 const APInt &CVal = CC->getValue();
164 bool Overflow = false;
166 if (Opcode == Instruction::Add) {
167 BVal.sadd_ov(CVal, Overflow);
169 BVal.ssub_ov(CVal, Overflow);
175 /// Conservatively clears subclassOptionalData after a reassociation or
176 /// commutation. We preserve fast-math flags when applicable as they can be
178 static void ClearSubclassDataAfterReassociation(BinaryOperator &I) {
179 FPMathOperator *FPMO = dyn_cast<FPMathOperator>(&I);
181 I.clearSubclassOptionalData();
185 FastMathFlags FMF = I.getFastMathFlags();
186 I.clearSubclassOptionalData();
187 I.setFastMathFlags(FMF);
190 /// SimplifyAssociativeOrCommutative - This performs a few simplifications for
191 /// operators which are associative or commutative:
193 // Commutative operators:
195 // 1. Order operands such that they are listed from right (least complex) to
196 // left (most complex). This puts constants before unary operators before
199 // Associative operators:
201 // 2. Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
202 // 3. Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
204 // Associative and commutative operators:
206 // 4. Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
207 // 5. Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
208 // 6. Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
209 // if C1 and C2 are constants.
211 bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) {
212 Instruction::BinaryOps Opcode = I.getOpcode();
213 bool Changed = false;
216 // Order operands such that they are listed from right (least complex) to
217 // left (most complex). This puts constants before unary operators before
219 if (I.isCommutative() && getComplexity(I.getOperand(0)) <
220 getComplexity(I.getOperand(1)))
221 Changed = !I.swapOperands();
223 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
224 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
226 if (I.isAssociative()) {
227 // Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
228 if (Op0 && Op0->getOpcode() == Opcode) {
229 Value *A = Op0->getOperand(0);
230 Value *B = Op0->getOperand(1);
231 Value *C = I.getOperand(1);
233 // Does "B op C" simplify?
234 if (Value *V = SimplifyBinOp(Opcode, B, C, DL)) {
235 // It simplifies to V. Form "A op V".
238 // Conservatively clear the optional flags, since they may not be
239 // preserved by the reassociation.
240 if (MaintainNoSignedWrap(I, B, C) &&
241 (!Op0 || (isa<BinaryOperator>(Op0) && Op0->hasNoSignedWrap()))) {
242 // Note: this is only valid because SimplifyBinOp doesn't look at
243 // the operands to Op0.
244 I.clearSubclassOptionalData();
245 I.setHasNoSignedWrap(true);
247 ClearSubclassDataAfterReassociation(I);
256 // Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
257 if (Op1 && Op1->getOpcode() == Opcode) {
258 Value *A = I.getOperand(0);
259 Value *B = Op1->getOperand(0);
260 Value *C = Op1->getOperand(1);
262 // Does "A op B" simplify?
263 if (Value *V = SimplifyBinOp(Opcode, A, B, DL)) {
264 // It simplifies to V. Form "V op C".
267 // Conservatively clear the optional flags, since they may not be
268 // preserved by the reassociation.
269 ClearSubclassDataAfterReassociation(I);
277 if (I.isAssociative() && I.isCommutative()) {
278 // Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
279 if (Op0 && Op0->getOpcode() == Opcode) {
280 Value *A = Op0->getOperand(0);
281 Value *B = Op0->getOperand(1);
282 Value *C = I.getOperand(1);
284 // Does "C op A" simplify?
285 if (Value *V = SimplifyBinOp(Opcode, C, A, DL)) {
286 // It simplifies to V. Form "V op B".
289 // Conservatively clear the optional flags, since they may not be
290 // preserved by the reassociation.
291 ClearSubclassDataAfterReassociation(I);
298 // Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
299 if (Op1 && Op1->getOpcode() == Opcode) {
300 Value *A = I.getOperand(0);
301 Value *B = Op1->getOperand(0);
302 Value *C = Op1->getOperand(1);
304 // Does "C op A" simplify?
305 if (Value *V = SimplifyBinOp(Opcode, C, A, DL)) {
306 // It simplifies to V. Form "B op V".
309 // Conservatively clear the optional flags, since they may not be
310 // preserved by the reassociation.
311 ClearSubclassDataAfterReassociation(I);
318 // Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
319 // if C1 and C2 are constants.
321 Op0->getOpcode() == Opcode && Op1->getOpcode() == Opcode &&
322 isa<Constant>(Op0->getOperand(1)) &&
323 isa<Constant>(Op1->getOperand(1)) &&
324 Op0->hasOneUse() && Op1->hasOneUse()) {
325 Value *A = Op0->getOperand(0);
326 Constant *C1 = cast<Constant>(Op0->getOperand(1));
327 Value *B = Op1->getOperand(0);
328 Constant *C2 = cast<Constant>(Op1->getOperand(1));
330 Constant *Folded = ConstantExpr::get(Opcode, C1, C2);
331 BinaryOperator *New = BinaryOperator::Create(Opcode, A, B);
332 if (isa<FPMathOperator>(New)) {
333 FastMathFlags Flags = I.getFastMathFlags();
334 Flags &= Op0->getFastMathFlags();
335 Flags &= Op1->getFastMathFlags();
336 New->setFastMathFlags(Flags);
338 InsertNewInstWith(New, I);
340 I.setOperand(0, New);
341 I.setOperand(1, Folded);
342 // Conservatively clear the optional flags, since they may not be
343 // preserved by the reassociation.
344 ClearSubclassDataAfterReassociation(I);
351 // No further simplifications.
356 /// LeftDistributesOverRight - Whether "X LOp (Y ROp Z)" is always equal to
357 /// "(X LOp Y) ROp (X LOp Z)".
358 static bool LeftDistributesOverRight(Instruction::BinaryOps LOp,
359 Instruction::BinaryOps ROp) {
364 case Instruction::And:
365 // And distributes over Or and Xor.
369 case Instruction::Or:
370 case Instruction::Xor:
374 case Instruction::Mul:
375 // Multiplication distributes over addition and subtraction.
379 case Instruction::Add:
380 case Instruction::Sub:
384 case Instruction::Or:
385 // Or distributes over And.
389 case Instruction::And:
395 /// RightDistributesOverLeft - Whether "(X LOp Y) ROp Z" is always equal to
396 /// "(X ROp Z) LOp (Y ROp Z)".
397 static bool RightDistributesOverLeft(Instruction::BinaryOps LOp,
398 Instruction::BinaryOps ROp) {
399 if (Instruction::isCommutative(ROp))
400 return LeftDistributesOverRight(ROp, LOp);
405 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
406 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
407 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
408 case Instruction::And:
409 case Instruction::Or:
410 case Instruction::Xor:
414 case Instruction::Shl:
415 case Instruction::LShr:
416 case Instruction::AShr:
420 // TODO: It would be nice to handle division, aka "(X + Y)/Z = X/Z + Y/Z",
421 // but this requires knowing that the addition does not overflow and other
426 /// This function returns identity value for given opcode, which can be used to
427 /// factor patterns like (X * 2) + X ==> (X * 2) + (X * 1) ==> X * (2 + 1).
428 static Value *getIdentityValue(Instruction::BinaryOps OpCode, Value *V) {
429 if (isa<Constant>(V))
432 if (OpCode == Instruction::Mul)
433 return ConstantInt::get(V->getType(), 1);
435 // TODO: We can handle other cases e.g. Instruction::And, Instruction::Or etc.
440 /// This function factors binary ops which can be combined using distributive
441 /// laws. This function tries to transform 'Op' based TopLevelOpcode to enable
442 /// factorization e.g for ADD(SHL(X , 2), MUL(X, 5)), When this function called
443 /// with TopLevelOpcode == Instruction::Add and Op = SHL(X, 2), transforms
444 /// SHL(X, 2) to MUL(X, 4) i.e. returns Instruction::Mul with LHS set to 'X' and
446 static Instruction::BinaryOps
447 getBinOpsForFactorization(Instruction::BinaryOps TopLevelOpcode,
448 BinaryOperator *Op, Value *&LHS, Value *&RHS) {
450 return Instruction::BinaryOpsEnd;
452 LHS = Op->getOperand(0);
453 RHS = Op->getOperand(1);
455 switch (TopLevelOpcode) {
457 return Op->getOpcode();
459 case Instruction::Add:
460 case Instruction::Sub:
461 if (Op->getOpcode() == Instruction::Shl) {
462 if (Constant *CST = dyn_cast<Constant>(Op->getOperand(1))) {
463 // The multiplier is really 1 << CST.
464 RHS = ConstantExpr::getShl(ConstantInt::get(Op->getType(), 1), CST);
465 return Instruction::Mul;
468 return Op->getOpcode();
471 // TODO: We can add other conversions e.g. shr => div etc.
474 /// This tries to simplify binary operations by factorizing out common terms
475 /// (e. g. "(A*B)+(A*C)" -> "A*(B+C)").
476 static Value *tryFactorization(InstCombiner::BuilderTy *Builder,
477 const DataLayout *DL, BinaryOperator &I,
478 Instruction::BinaryOps InnerOpcode, Value *A,
479 Value *B, Value *C, Value *D) {
481 // If any of A, B, C, D are null, we can not factor I, return early.
482 // Checking A and C should be enough.
483 if (!A || !C || !B || !D)
486 Value *SimplifiedInst = nullptr;
487 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
488 Instruction::BinaryOps TopLevelOpcode = I.getOpcode();
490 // Does "X op' Y" always equal "Y op' X"?
491 bool InnerCommutative = Instruction::isCommutative(InnerOpcode);
493 // Does "X op' (Y op Z)" always equal "(X op' Y) op (X op' Z)"?
494 if (LeftDistributesOverRight(InnerOpcode, TopLevelOpcode))
495 // Does the instruction have the form "(A op' B) op (A op' D)" or, in the
496 // commutative case, "(A op' B) op (C op' A)"?
497 if (A == C || (InnerCommutative && A == D)) {
500 // Consider forming "A op' (B op D)".
501 // If "B op D" simplifies then it can be formed with no cost.
502 Value *V = SimplifyBinOp(TopLevelOpcode, B, D, DL);
503 // If "B op D" doesn't simplify then only go on if both of the existing
504 // operations "A op' B" and "C op' D" will be zapped as no longer used.
505 if (!V && LHS->hasOneUse() && RHS->hasOneUse())
506 V = Builder->CreateBinOp(TopLevelOpcode, B, D, RHS->getName());
508 SimplifiedInst = Builder->CreateBinOp(InnerOpcode, A, V);
512 // Does "(X op Y) op' Z" always equal "(X op' Z) op (Y op' Z)"?
513 if (!SimplifiedInst && RightDistributesOverLeft(TopLevelOpcode, InnerOpcode))
514 // Does the instruction have the form "(A op' B) op (C op' B)" or, in the
515 // commutative case, "(A op' B) op (B op' D)"?
516 if (B == D || (InnerCommutative && B == C)) {
519 // Consider forming "(A op C) op' B".
520 // If "A op C" simplifies then it can be formed with no cost.
521 Value *V = SimplifyBinOp(TopLevelOpcode, A, C, DL);
523 // If "A op C" doesn't simplify then only go on if both of the existing
524 // operations "A op' B" and "C op' D" will be zapped as no longer used.
525 if (!V && LHS->hasOneUse() && RHS->hasOneUse())
526 V = Builder->CreateBinOp(TopLevelOpcode, A, C, LHS->getName());
528 SimplifiedInst = Builder->CreateBinOp(InnerOpcode, V, B);
532 if (SimplifiedInst) {
534 SimplifiedInst->takeName(&I);
536 // Check if we can add NSW flag to SimplifiedInst. If so, set NSW flag.
537 // TODO: Check for NUW.
538 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(SimplifiedInst)) {
539 if (isa<OverflowingBinaryOperator>(SimplifiedInst)) {
541 if (isa<OverflowingBinaryOperator>(&I))
542 HasNSW = I.hasNoSignedWrap();
544 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS))
545 if (isa<OverflowingBinaryOperator>(Op0))
546 HasNSW &= Op0->hasNoSignedWrap();
548 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS))
549 if (isa<OverflowingBinaryOperator>(Op1))
550 HasNSW &= Op1->hasNoSignedWrap();
551 BO->setHasNoSignedWrap(HasNSW);
555 return SimplifiedInst;
558 /// SimplifyUsingDistributiveLaws - This tries to simplify binary operations
559 /// which some other binary operation distributes over either by factorizing
560 /// out common terms (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this
561 /// results in simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is
562 /// a win). Returns the simplified value, or null if it didn't simplify.
563 Value *InstCombiner::SimplifyUsingDistributiveLaws(BinaryOperator &I) {
564 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
565 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
566 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
569 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
570 auto TopLevelOpcode = I.getOpcode();
571 auto LHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op0, A, B);
572 auto RHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op1, C, D);
574 // The instruction has the form "(A op' B) op (C op' D)". Try to factorize
576 if (LHSOpcode == RHSOpcode) {
577 if (Value *V = tryFactorization(Builder, DL, I, LHSOpcode, A, B, C, D))
581 // The instruction has the form "(A op' B) op (C)". Try to factorize common
583 if (Value *V = tryFactorization(Builder, DL, I, LHSOpcode, A, B, RHS,
584 getIdentityValue(LHSOpcode, RHS)))
587 // The instruction has the form "(B) op (C op' D)". Try to factorize common
589 if (Value *V = tryFactorization(Builder, DL, I, RHSOpcode, LHS,
590 getIdentityValue(RHSOpcode, LHS), C, D))
594 if (Op0 && RightDistributesOverLeft(Op0->getOpcode(), TopLevelOpcode)) {
595 // The instruction has the form "(A op' B) op C". See if expanding it out
596 // to "(A op C) op' (B op C)" results in simplifications.
597 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
598 Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op'
600 // Do "A op C" and "B op C" both simplify?
601 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, C, DL))
602 if (Value *R = SimplifyBinOp(TopLevelOpcode, B, C, DL)) {
603 // They do! Return "L op' R".
605 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
606 if ((L == A && R == B) ||
607 (Instruction::isCommutative(InnerOpcode) && L == B && R == A))
609 // Otherwise return "L op' R" if it simplifies.
610 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, DL))
612 // Otherwise, create a new instruction.
613 C = Builder->CreateBinOp(InnerOpcode, L, R);
619 if (Op1 && LeftDistributesOverRight(TopLevelOpcode, Op1->getOpcode())) {
620 // The instruction has the form "A op (B op' C)". See if expanding it out
621 // to "(A op B) op' (A op C)" results in simplifications.
622 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
623 Instruction::BinaryOps InnerOpcode = Op1->getOpcode(); // op'
625 // Do "A op B" and "A op C" both simplify?
626 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, B, DL))
627 if (Value *R = SimplifyBinOp(TopLevelOpcode, A, C, DL)) {
628 // They do! Return "L op' R".
630 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
631 if ((L == B && R == C) ||
632 (Instruction::isCommutative(InnerOpcode) && L == C && R == B))
634 // Otherwise return "L op' R" if it simplifies.
635 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, DL))
637 // Otherwise, create a new instruction.
638 A = Builder->CreateBinOp(InnerOpcode, L, R);
647 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
648 // if the LHS is a constant zero (which is the 'negate' form).
650 Value *InstCombiner::dyn_castNegVal(Value *V) const {
651 if (BinaryOperator::isNeg(V))
652 return BinaryOperator::getNegArgument(V);
654 // Constants can be considered to be negated values if they can be folded.
655 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
656 return ConstantExpr::getNeg(C);
658 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
659 if (C->getType()->getElementType()->isIntegerTy())
660 return ConstantExpr::getNeg(C);
665 // dyn_castFNegVal - Given a 'fsub' instruction, return the RHS of the
666 // instruction if the LHS is a constant negative zero (which is the 'negate'
669 Value *InstCombiner::dyn_castFNegVal(Value *V, bool IgnoreZeroSign) const {
670 if (BinaryOperator::isFNeg(V, IgnoreZeroSign))
671 return BinaryOperator::getFNegArgument(V);
673 // Constants can be considered to be negated values if they can be folded.
674 if (ConstantFP *C = dyn_cast<ConstantFP>(V))
675 return ConstantExpr::getFNeg(C);
677 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
678 if (C->getType()->getElementType()->isFloatingPointTy())
679 return ConstantExpr::getFNeg(C);
684 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
686 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
687 return IC->Builder->CreateCast(CI->getOpcode(), SO, I.getType());
690 // Figure out if the constant is the left or the right argument.
691 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
692 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
694 if (Constant *SOC = dyn_cast<Constant>(SO)) {
696 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
697 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
700 Value *Op0 = SO, *Op1 = ConstOperand;
704 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I)) {
705 Value *RI = IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1,
706 SO->getName()+".op");
707 Instruction *FPInst = dyn_cast<Instruction>(RI);
708 if (FPInst && isa<FPMathOperator>(FPInst))
709 FPInst->copyFastMathFlags(BO);
712 if (ICmpInst *CI = dyn_cast<ICmpInst>(&I))
713 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
714 SO->getName()+".cmp");
715 if (FCmpInst *CI = dyn_cast<FCmpInst>(&I))
716 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
717 SO->getName()+".cmp");
718 llvm_unreachable("Unknown binary instruction type!");
721 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
722 // constant as the other operand, try to fold the binary operator into the
723 // select arguments. This also works for Cast instructions, which obviously do
724 // not have a second operand.
725 Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) {
726 // Don't modify shared select instructions
727 if (!SI->hasOneUse()) return nullptr;
728 Value *TV = SI->getOperand(1);
729 Value *FV = SI->getOperand(2);
731 if (isa<Constant>(TV) || isa<Constant>(FV)) {
732 // Bool selects with constant operands can be folded to logical ops.
733 if (SI->getType()->isIntegerTy(1)) return nullptr;
735 // If it's a bitcast involving vectors, make sure it has the same number of
736 // elements on both sides.
737 if (BitCastInst *BC = dyn_cast<BitCastInst>(&Op)) {
738 VectorType *DestTy = dyn_cast<VectorType>(BC->getDestTy());
739 VectorType *SrcTy = dyn_cast<VectorType>(BC->getSrcTy());
741 // Verify that either both or neither are vectors.
742 if ((SrcTy == nullptr) != (DestTy == nullptr)) return nullptr;
743 // If vectors, verify that they have the same number of elements.
744 if (SrcTy && SrcTy->getNumElements() != DestTy->getNumElements())
748 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, this);
749 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, this);
751 return SelectInst::Create(SI->getCondition(),
752 SelectTrueVal, SelectFalseVal);
758 /// FoldOpIntoPhi - Given a binary operator, cast instruction, or select which
759 /// has a PHI node as operand #0, see if we can fold the instruction into the
760 /// PHI (which is only possible if all operands to the PHI are constants).
762 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
763 PHINode *PN = cast<PHINode>(I.getOperand(0));
764 unsigned NumPHIValues = PN->getNumIncomingValues();
765 if (NumPHIValues == 0)
768 // We normally only transform phis with a single use. However, if a PHI has
769 // multiple uses and they are all the same operation, we can fold *all* of the
770 // uses into the PHI.
771 if (!PN->hasOneUse()) {
772 // Walk the use list for the instruction, comparing them to I.
773 for (User *U : PN->users()) {
774 Instruction *UI = cast<Instruction>(U);
775 if (UI != &I && !I.isIdenticalTo(UI))
778 // Otherwise, we can replace *all* users with the new PHI we form.
781 // Check to see if all of the operands of the PHI are simple constants
782 // (constantint/constantfp/undef). If there is one non-constant value,
783 // remember the BB it is in. If there is more than one or if *it* is a PHI,
784 // bail out. We don't do arbitrary constant expressions here because moving
785 // their computation can be expensive without a cost model.
786 BasicBlock *NonConstBB = nullptr;
787 for (unsigned i = 0; i != NumPHIValues; ++i) {
788 Value *InVal = PN->getIncomingValue(i);
789 if (isa<Constant>(InVal) && !isa<ConstantExpr>(InVal))
792 if (isa<PHINode>(InVal)) return nullptr; // Itself a phi.
793 if (NonConstBB) return nullptr; // More than one non-const value.
795 NonConstBB = PN->getIncomingBlock(i);
797 // If the InVal is an invoke at the end of the pred block, then we can't
798 // insert a computation after it without breaking the edge.
799 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal))
800 if (II->getParent() == NonConstBB)
803 // If the incoming non-constant value is in I's block, we will remove one
804 // instruction, but insert another equivalent one, leading to infinite
806 if (NonConstBB == I.getParent())
810 // If there is exactly one non-constant value, we can insert a copy of the
811 // operation in that block. However, if this is a critical edge, we would be
812 // inserting the computation one some other paths (e.g. inside a loop). Only
813 // do this if the pred block is unconditionally branching into the phi block.
814 if (NonConstBB != nullptr) {
815 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
816 if (!BI || !BI->isUnconditional()) return nullptr;
819 // Okay, we can do the transformation: create the new PHI node.
820 PHINode *NewPN = PHINode::Create(I.getType(), PN->getNumIncomingValues());
821 InsertNewInstBefore(NewPN, *PN);
824 // If we are going to have to insert a new computation, do so right before the
825 // predecessors terminator.
827 Builder->SetInsertPoint(NonConstBB->getTerminator());
829 // Next, add all of the operands to the PHI.
830 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) {
831 // We only currently try to fold the condition of a select when it is a phi,
832 // not the true/false values.
833 Value *TrueV = SI->getTrueValue();
834 Value *FalseV = SI->getFalseValue();
835 BasicBlock *PhiTransBB = PN->getParent();
836 for (unsigned i = 0; i != NumPHIValues; ++i) {
837 BasicBlock *ThisBB = PN->getIncomingBlock(i);
838 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB);
839 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB);
840 Value *InV = nullptr;
841 // Beware of ConstantExpr: it may eventually evaluate to getNullValue,
842 // even if currently isNullValue gives false.
843 Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i));
844 if (InC && !isa<ConstantExpr>(InC))
845 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred;
847 InV = Builder->CreateSelect(PN->getIncomingValue(i),
848 TrueVInPred, FalseVInPred, "phitmp");
849 NewPN->addIncoming(InV, ThisBB);
851 } else if (CmpInst *CI = dyn_cast<CmpInst>(&I)) {
852 Constant *C = cast<Constant>(I.getOperand(1));
853 for (unsigned i = 0; i != NumPHIValues; ++i) {
854 Value *InV = nullptr;
855 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
856 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
857 else if (isa<ICmpInst>(CI))
858 InV = Builder->CreateICmp(CI->getPredicate(), PN->getIncomingValue(i),
861 InV = Builder->CreateFCmp(CI->getPredicate(), PN->getIncomingValue(i),
863 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
865 } else if (I.getNumOperands() == 2) {
866 Constant *C = cast<Constant>(I.getOperand(1));
867 for (unsigned i = 0; i != NumPHIValues; ++i) {
868 Value *InV = nullptr;
869 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
870 InV = ConstantExpr::get(I.getOpcode(), InC, C);
872 InV = Builder->CreateBinOp(cast<BinaryOperator>(I).getOpcode(),
873 PN->getIncomingValue(i), C, "phitmp");
874 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
877 CastInst *CI = cast<CastInst>(&I);
878 Type *RetTy = CI->getType();
879 for (unsigned i = 0; i != NumPHIValues; ++i) {
881 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
882 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
884 InV = Builder->CreateCast(CI->getOpcode(),
885 PN->getIncomingValue(i), I.getType(), "phitmp");
886 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
890 for (auto UI = PN->user_begin(), E = PN->user_end(); UI != E;) {
891 Instruction *User = cast<Instruction>(*UI++);
892 if (User == &I) continue;
893 ReplaceInstUsesWith(*User, NewPN);
894 EraseInstFromFunction(*User);
896 return ReplaceInstUsesWith(I, NewPN);
899 /// FindElementAtOffset - Given a pointer type and a constant offset, determine
900 /// whether or not there is a sequence of GEP indices into the pointed type that
901 /// will land us at the specified offset. If so, fill them into NewIndices and
902 /// return the resultant element type, otherwise return null.
903 Type *InstCombiner::FindElementAtOffset(Type *PtrTy, int64_t Offset,
904 SmallVectorImpl<Value*> &NewIndices) {
905 assert(PtrTy->isPtrOrPtrVectorTy());
910 Type *Ty = PtrTy->getPointerElementType();
914 // Start with the index over the outer type. Note that the type size
915 // might be zero (even if the offset isn't zero) if the indexed type
916 // is something like [0 x {int, int}]
917 Type *IntPtrTy = DL->getIntPtrType(PtrTy);
918 int64_t FirstIdx = 0;
919 if (int64_t TySize = DL->getTypeAllocSize(Ty)) {
920 FirstIdx = Offset/TySize;
921 Offset -= FirstIdx*TySize;
923 // Handle hosts where % returns negative instead of values [0..TySize).
929 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
932 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
934 // Index into the types. If we fail, set OrigBase to null.
936 // Indexing into tail padding between struct/array elements.
937 if (uint64_t(Offset*8) >= DL->getTypeSizeInBits(Ty))
940 if (StructType *STy = dyn_cast<StructType>(Ty)) {
941 const StructLayout *SL = DL->getStructLayout(STy);
942 assert(Offset < (int64_t)SL->getSizeInBytes() &&
943 "Offset must stay within the indexed type");
945 unsigned Elt = SL->getElementContainingOffset(Offset);
946 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
949 Offset -= SL->getElementOffset(Elt);
950 Ty = STy->getElementType(Elt);
951 } else if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
952 uint64_t EltSize = DL->getTypeAllocSize(AT->getElementType());
953 assert(EltSize && "Cannot index into a zero-sized array");
954 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
956 Ty = AT->getElementType();
958 // Otherwise, we can't index into the middle of this atomic type, bail.
966 static bool shouldMergeGEPs(GEPOperator &GEP, GEPOperator &Src) {
967 // If this GEP has only 0 indices, it is the same pointer as
968 // Src. If Src is not a trivial GEP too, don't combine
970 if (GEP.hasAllZeroIndices() && !Src.hasAllZeroIndices() &&
976 /// Descale - Return a value X such that Val = X * Scale, or null if none. If
977 /// the multiplication is known not to overflow then NoSignedWrap is set.
978 Value *InstCombiner::Descale(Value *Val, APInt Scale, bool &NoSignedWrap) {
979 assert(isa<IntegerType>(Val->getType()) && "Can only descale integers!");
980 assert(cast<IntegerType>(Val->getType())->getBitWidth() ==
981 Scale.getBitWidth() && "Scale not compatible with value!");
983 // If Val is zero or Scale is one then Val = Val * Scale.
984 if (match(Val, m_Zero()) || Scale == 1) {
989 // If Scale is zero then it does not divide Val.
990 if (Scale.isMinValue())
993 // Look through chains of multiplications, searching for a constant that is
994 // divisible by Scale. For example, descaling X*(Y*(Z*4)) by a factor of 4
995 // will find the constant factor 4 and produce X*(Y*Z). Descaling X*(Y*8) by
996 // a factor of 4 will produce X*(Y*2). The principle of operation is to bore
999 // Val = M1 * X || Analysis starts here and works down
1000 // M1 = M2 * Y || Doesn't descend into terms with more
1001 // M2 = Z * 4 \/ than one use
1003 // Then to modify a term at the bottom:
1006 // M1 = Z * Y || Replaced M2 with Z
1008 // Then to work back up correcting nsw flags.
1010 // Op - the term we are currently analyzing. Starts at Val then drills down.
1011 // Replaced with its descaled value before exiting from the drill down loop.
1014 // Parent - initially null, but after drilling down notes where Op came from.
1015 // In the example above, Parent is (Val, 0) when Op is M1, because M1 is the
1016 // 0'th operand of Val.
1017 std::pair<Instruction*, unsigned> Parent;
1019 // RequireNoSignedWrap - Set if the transform requires a descaling at deeper
1020 // levels that doesn't overflow.
1021 bool RequireNoSignedWrap = false;
1023 // logScale - log base 2 of the scale. Negative if not a power of 2.
1024 int32_t logScale = Scale.exactLogBase2();
1026 for (;; Op = Parent.first->getOperand(Parent.second)) { // Drill down
1028 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
1029 // If Op is a constant divisible by Scale then descale to the quotient.
1030 APInt Quotient(Scale), Remainder(Scale); // Init ensures right bitwidth.
1031 APInt::sdivrem(CI->getValue(), Scale, Quotient, Remainder);
1032 if (!Remainder.isMinValue())
1033 // Not divisible by Scale.
1035 // Replace with the quotient in the parent.
1036 Op = ConstantInt::get(CI->getType(), Quotient);
1037 NoSignedWrap = true;
1041 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op)) {
1043 if (BO->getOpcode() == Instruction::Mul) {
1045 NoSignedWrap = BO->hasNoSignedWrap();
1046 if (RequireNoSignedWrap && !NoSignedWrap)
1049 // There are three cases for multiplication: multiplication by exactly
1050 // the scale, multiplication by a constant different to the scale, and
1051 // multiplication by something else.
1052 Value *LHS = BO->getOperand(0);
1053 Value *RHS = BO->getOperand(1);
1055 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
1056 // Multiplication by a constant.
1057 if (CI->getValue() == Scale) {
1058 // Multiplication by exactly the scale, replace the multiplication
1059 // by its left-hand side in the parent.
1064 // Otherwise drill down into the constant.
1065 if (!Op->hasOneUse())
1068 Parent = std::make_pair(BO, 1);
1072 // Multiplication by something else. Drill down into the left-hand side
1073 // since that's where the reassociate pass puts the good stuff.
1074 if (!Op->hasOneUse())
1077 Parent = std::make_pair(BO, 0);
1081 if (logScale > 0 && BO->getOpcode() == Instruction::Shl &&
1082 isa<ConstantInt>(BO->getOperand(1))) {
1083 // Multiplication by a power of 2.
1084 NoSignedWrap = BO->hasNoSignedWrap();
1085 if (RequireNoSignedWrap && !NoSignedWrap)
1088 Value *LHS = BO->getOperand(0);
1089 int32_t Amt = cast<ConstantInt>(BO->getOperand(1))->
1090 getLimitedValue(Scale.getBitWidth());
1093 if (Amt == logScale) {
1094 // Multiplication by exactly the scale, replace the multiplication
1095 // by its left-hand side in the parent.
1099 if (Amt < logScale || !Op->hasOneUse())
1102 // Multiplication by more than the scale. Reduce the multiplying amount
1103 // by the scale in the parent.
1104 Parent = std::make_pair(BO, 1);
1105 Op = ConstantInt::get(BO->getType(), Amt - logScale);
1110 if (!Op->hasOneUse())
1113 if (CastInst *Cast = dyn_cast<CastInst>(Op)) {
1114 if (Cast->getOpcode() == Instruction::SExt) {
1115 // Op is sign-extended from a smaller type, descale in the smaller type.
1116 unsigned SmallSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1117 APInt SmallScale = Scale.trunc(SmallSize);
1118 // Suppose Op = sext X, and we descale X as Y * SmallScale. We want to
1119 // descale Op as (sext Y) * Scale. In order to have
1120 // sext (Y * SmallScale) = (sext Y) * Scale
1121 // some conditions need to hold however: SmallScale must sign-extend to
1122 // Scale and the multiplication Y * SmallScale should not overflow.
1123 if (SmallScale.sext(Scale.getBitWidth()) != Scale)
1124 // SmallScale does not sign-extend to Scale.
1126 assert(SmallScale.exactLogBase2() == logScale);
1127 // Require that Y * SmallScale must not overflow.
1128 RequireNoSignedWrap = true;
1130 // Drill down through the cast.
1131 Parent = std::make_pair(Cast, 0);
1136 if (Cast->getOpcode() == Instruction::Trunc) {
1137 // Op is truncated from a larger type, descale in the larger type.
1138 // Suppose Op = trunc X, and we descale X as Y * sext Scale. Then
1139 // trunc (Y * sext Scale) = (trunc Y) * Scale
1140 // always holds. However (trunc Y) * Scale may overflow even if
1141 // trunc (Y * sext Scale) does not, so nsw flags need to be cleared
1142 // from this point up in the expression (see later).
1143 if (RequireNoSignedWrap)
1146 // Drill down through the cast.
1147 unsigned LargeSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1148 Parent = std::make_pair(Cast, 0);
1149 Scale = Scale.sext(LargeSize);
1150 if (logScale + 1 == (int32_t)Cast->getType()->getPrimitiveSizeInBits())
1152 assert(Scale.exactLogBase2() == logScale);
1157 // Unsupported expression, bail out.
1161 // If Op is zero then Val = Op * Scale.
1162 if (match(Op, m_Zero())) {
1163 NoSignedWrap = true;
1167 // We know that we can successfully descale, so from here on we can safely
1168 // modify the IR. Op holds the descaled version of the deepest term in the
1169 // expression. NoSignedWrap is 'true' if multiplying Op by Scale is known
1173 // The expression only had one term.
1176 // Rewrite the parent using the descaled version of its operand.
1177 assert(Parent.first->hasOneUse() && "Drilled down when more than one use!");
1178 assert(Op != Parent.first->getOperand(Parent.second) &&
1179 "Descaling was a no-op?");
1180 Parent.first->setOperand(Parent.second, Op);
1181 Worklist.Add(Parent.first);
1183 // Now work back up the expression correcting nsw flags. The logic is based
1184 // on the following observation: if X * Y is known not to overflow as a signed
1185 // multiplication, and Y is replaced by a value Z with smaller absolute value,
1186 // then X * Z will not overflow as a signed multiplication either. As we work
1187 // our way up, having NoSignedWrap 'true' means that the descaled value at the
1188 // current level has strictly smaller absolute value than the original.
1189 Instruction *Ancestor = Parent.first;
1191 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Ancestor)) {
1192 // If the multiplication wasn't nsw then we can't say anything about the
1193 // value of the descaled multiplication, and we have to clear nsw flags
1194 // from this point on up.
1195 bool OpNoSignedWrap = BO->hasNoSignedWrap();
1196 NoSignedWrap &= OpNoSignedWrap;
1197 if (NoSignedWrap != OpNoSignedWrap) {
1198 BO->setHasNoSignedWrap(NoSignedWrap);
1199 Worklist.Add(Ancestor);
1201 } else if (Ancestor->getOpcode() == Instruction::Trunc) {
1202 // The fact that the descaled input to the trunc has smaller absolute
1203 // value than the original input doesn't tell us anything useful about
1204 // the absolute values of the truncations.
1205 NoSignedWrap = false;
1207 assert((Ancestor->getOpcode() != Instruction::SExt || NoSignedWrap) &&
1208 "Failed to keep proper track of nsw flags while drilling down?");
1210 if (Ancestor == Val)
1211 // Got to the top, all done!
1214 // Move up one level in the expression.
1215 assert(Ancestor->hasOneUse() && "Drilled down when more than one use!");
1216 Ancestor = Ancestor->user_back();
1220 /// \brief Creates node of binary operation with the same attributes as the
1221 /// specified one but with other operands.
1222 static Value *CreateBinOpAsGiven(BinaryOperator &Inst, Value *LHS, Value *RHS,
1223 InstCombiner::BuilderTy *B) {
1224 Value *BORes = B->CreateBinOp(Inst.getOpcode(), LHS, RHS);
1225 if (BinaryOperator *NewBO = dyn_cast<BinaryOperator>(BORes)) {
1226 if (isa<OverflowingBinaryOperator>(NewBO)) {
1227 NewBO->setHasNoSignedWrap(Inst.hasNoSignedWrap());
1228 NewBO->setHasNoUnsignedWrap(Inst.hasNoUnsignedWrap());
1230 if (isa<PossiblyExactOperator>(NewBO))
1231 NewBO->setIsExact(Inst.isExact());
1236 /// \brief Makes transformation of binary operation specific for vector types.
1237 /// \param Inst Binary operator to transform.
1238 /// \return Pointer to node that must replace the original binary operator, or
1239 /// null pointer if no transformation was made.
1240 Value *InstCombiner::SimplifyVectorOp(BinaryOperator &Inst) {
1241 if (!Inst.getType()->isVectorTy()) return nullptr;
1243 // It may not be safe to reorder shuffles and things like div, urem, etc.
1244 // because we may trap when executing those ops on unknown vector elements.
1246 if (!isSafeToSpeculativelyExecute(&Inst, DL)) return nullptr;
1248 unsigned VWidth = cast<VectorType>(Inst.getType())->getNumElements();
1249 Value *LHS = Inst.getOperand(0), *RHS = Inst.getOperand(1);
1250 assert(cast<VectorType>(LHS->getType())->getNumElements() == VWidth);
1251 assert(cast<VectorType>(RHS->getType())->getNumElements() == VWidth);
1253 // If both arguments of binary operation are shuffles, which use the same
1254 // mask and shuffle within a single vector, it is worthwhile to move the
1255 // shuffle after binary operation:
1256 // Op(shuffle(v1, m), shuffle(v2, m)) -> shuffle(Op(v1, v2), m)
1257 if (isa<ShuffleVectorInst>(LHS) && isa<ShuffleVectorInst>(RHS)) {
1258 ShuffleVectorInst *LShuf = cast<ShuffleVectorInst>(LHS);
1259 ShuffleVectorInst *RShuf = cast<ShuffleVectorInst>(RHS);
1260 if (isa<UndefValue>(LShuf->getOperand(1)) &&
1261 isa<UndefValue>(RShuf->getOperand(1)) &&
1262 LShuf->getOperand(0)->getType() == RShuf->getOperand(0)->getType() &&
1263 LShuf->getMask() == RShuf->getMask()) {
1264 Value *NewBO = CreateBinOpAsGiven(Inst, LShuf->getOperand(0),
1265 RShuf->getOperand(0), Builder);
1266 Value *Res = Builder->CreateShuffleVector(NewBO,
1267 UndefValue::get(NewBO->getType()), LShuf->getMask());
1272 // If one argument is a shuffle within one vector, the other is a constant,
1273 // try moving the shuffle after the binary operation.
1274 ShuffleVectorInst *Shuffle = nullptr;
1275 Constant *C1 = nullptr;
1276 if (isa<ShuffleVectorInst>(LHS)) Shuffle = cast<ShuffleVectorInst>(LHS);
1277 if (isa<ShuffleVectorInst>(RHS)) Shuffle = cast<ShuffleVectorInst>(RHS);
1278 if (isa<Constant>(LHS)) C1 = cast<Constant>(LHS);
1279 if (isa<Constant>(RHS)) C1 = cast<Constant>(RHS);
1280 if (Shuffle && C1 &&
1281 (isa<ConstantVector>(C1) || isa<ConstantDataVector>(C1)) &&
1282 isa<UndefValue>(Shuffle->getOperand(1)) &&
1283 Shuffle->getType() == Shuffle->getOperand(0)->getType()) {
1284 SmallVector<int, 16> ShMask = Shuffle->getShuffleMask();
1285 // Find constant C2 that has property:
1286 // shuffle(C2, ShMask) = C1
1287 // If such constant does not exist (example: ShMask=<0,0> and C1=<1,2>)
1288 // reorder is not possible.
1289 SmallVector<Constant*, 16> C2M(VWidth,
1290 UndefValue::get(C1->getType()->getScalarType()));
1291 bool MayChange = true;
1292 for (unsigned I = 0; I < VWidth; ++I) {
1293 if (ShMask[I] >= 0) {
1294 assert(ShMask[I] < (int)VWidth);
1295 if (!isa<UndefValue>(C2M[ShMask[I]])) {
1299 C2M[ShMask[I]] = C1->getAggregateElement(I);
1303 Constant *C2 = ConstantVector::get(C2M);
1304 Value *NewLHS, *NewRHS;
1305 if (isa<Constant>(LHS)) {
1307 NewRHS = Shuffle->getOperand(0);
1309 NewLHS = Shuffle->getOperand(0);
1312 Value *NewBO = CreateBinOpAsGiven(Inst, NewLHS, NewRHS, Builder);
1313 Value *Res = Builder->CreateShuffleVector(NewBO,
1314 UndefValue::get(Inst.getType()), Shuffle->getMask());
1322 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
1323 SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end());
1325 if (Value *V = SimplifyGEPInst(Ops, DL, TLI, DT, AT))
1326 return ReplaceInstUsesWith(GEP, V);
1328 Value *PtrOp = GEP.getOperand(0);
1330 // Eliminate unneeded casts for indices, and replace indices which displace
1331 // by multiples of a zero size type with zero.
1333 bool MadeChange = false;
1334 Type *IntPtrTy = DL->getIntPtrType(GEP.getPointerOperandType());
1336 gep_type_iterator GTI = gep_type_begin(GEP);
1337 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end();
1338 I != E; ++I, ++GTI) {
1339 // Skip indices into struct types.
1340 SequentialType *SeqTy = dyn_cast<SequentialType>(*GTI);
1341 if (!SeqTy) continue;
1343 // If the element type has zero size then any index over it is equivalent
1344 // to an index of zero, so replace it with zero if it is not zero already.
1345 if (SeqTy->getElementType()->isSized() &&
1346 DL->getTypeAllocSize(SeqTy->getElementType()) == 0)
1347 if (!isa<Constant>(*I) || !cast<Constant>(*I)->isNullValue()) {
1348 *I = Constant::getNullValue(IntPtrTy);
1352 Type *IndexTy = (*I)->getType();
1353 if (IndexTy != IntPtrTy) {
1354 // If we are using a wider index than needed for this platform, shrink
1355 // it to what we need. If narrower, sign-extend it to what we need.
1356 // This explicit cast can make subsequent optimizations more obvious.
1357 *I = Builder->CreateIntCast(*I, IntPtrTy, true);
1361 if (MadeChange) return &GEP;
1364 // Check to see if the inputs to the PHI node are getelementptr instructions.
1365 if (PHINode *PN = dyn_cast<PHINode>(PtrOp)) {
1366 GetElementPtrInst *Op1 = dyn_cast<GetElementPtrInst>(PN->getOperand(0));
1372 for (auto I = PN->op_begin()+1, E = PN->op_end(); I !=E; ++I) {
1373 GetElementPtrInst *Op2 = dyn_cast<GetElementPtrInst>(*I);
1374 if (!Op2 || Op1->getNumOperands() != Op2->getNumOperands())
1377 // Keep track of the type as we walk the GEP.
1378 Type *CurTy = Op1->getOperand(0)->getType()->getScalarType();
1380 for (unsigned J = 0, F = Op1->getNumOperands(); J != F; ++J) {
1381 if (Op1->getOperand(J)->getType() != Op2->getOperand(J)->getType())
1384 if (Op1->getOperand(J) != Op2->getOperand(J)) {
1386 // We have not seen any differences yet in the GEPs feeding the
1387 // PHI yet, so we record this one if it is allowed to be a
1390 // The first two arguments can vary for any GEP, the rest have to be
1391 // static for struct slots
1392 if (J > 1 && CurTy->isStructTy())
1397 // The GEP is different by more than one input. While this could be
1398 // extended to support GEPs that vary by more than one variable it
1399 // doesn't make sense since it greatly increases the complexity and
1400 // would result in an R+R+R addressing mode which no backend
1401 // directly supports and would need to be broken into several
1402 // simpler instructions anyway.
1407 // Sink down a layer of the type for the next iteration.
1409 if (CompositeType *CT = dyn_cast<CompositeType>(CurTy)) {
1410 CurTy = CT->getTypeAtIndex(Op1->getOperand(J));
1418 GetElementPtrInst *NewGEP = cast<GetElementPtrInst>(Op1->clone());
1421 // All the GEPs feeding the PHI are identical. Clone one down into our
1422 // BB so that it can be merged with the current GEP.
1423 GEP.getParent()->getInstList().insert(GEP.getParent()->getFirstNonPHI(),
1426 // All the GEPs feeding the PHI differ at a single offset. Clone a GEP
1427 // into the current block so it can be merged, and create a new PHI to
1429 Instruction *InsertPt = Builder->GetInsertPoint();
1430 Builder->SetInsertPoint(PN);
1431 PHINode *NewPN = Builder->CreatePHI(Op1->getOperand(DI)->getType(),
1432 PN->getNumOperands());
1433 Builder->SetInsertPoint(InsertPt);
1435 for (auto &I : PN->operands())
1436 NewPN->addIncoming(cast<GEPOperator>(I)->getOperand(DI),
1437 PN->getIncomingBlock(I));
1439 NewGEP->setOperand(DI, NewPN);
1440 GEP.getParent()->getInstList().insert(GEP.getParent()->getFirstNonPHI(),
1442 NewGEP->setOperand(DI, NewPN);
1445 GEP.setOperand(0, NewGEP);
1449 // Combine Indices - If the source pointer to this getelementptr instruction
1450 // is a getelementptr instruction, combine the indices of the two
1451 // getelementptr instructions into a single instruction.
1453 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) {
1454 if (!shouldMergeGEPs(*cast<GEPOperator>(&GEP), *Src))
1457 // Note that if our source is a gep chain itself then we wait for that
1458 // chain to be resolved before we perform this transformation. This
1459 // avoids us creating a TON of code in some cases.
1460 if (GEPOperator *SrcGEP =
1461 dyn_cast<GEPOperator>(Src->getOperand(0)))
1462 if (SrcGEP->getNumOperands() == 2 && shouldMergeGEPs(*Src, *SrcGEP))
1463 return nullptr; // Wait until our source is folded to completion.
1465 SmallVector<Value*, 8> Indices;
1467 // Find out whether the last index in the source GEP is a sequential idx.
1468 bool EndsWithSequential = false;
1469 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src);
1471 EndsWithSequential = !(*I)->isStructTy();
1473 // Can we combine the two pointer arithmetics offsets?
1474 if (EndsWithSequential) {
1475 // Replace: gep (gep %P, long B), long A, ...
1476 // With: T = long A+B; gep %P, T, ...
1479 Value *SO1 = Src->getOperand(Src->getNumOperands()-1);
1480 Value *GO1 = GEP.getOperand(1);
1481 if (SO1 == Constant::getNullValue(SO1->getType())) {
1483 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
1486 // If they aren't the same type, then the input hasn't been processed
1487 // by the loop above yet (which canonicalizes sequential index types to
1488 // intptr_t). Just avoid transforming this until the input has been
1490 if (SO1->getType() != GO1->getType())
1492 Sum = Builder->CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
1495 // Update the GEP in place if possible.
1496 if (Src->getNumOperands() == 2) {
1497 GEP.setOperand(0, Src->getOperand(0));
1498 GEP.setOperand(1, Sum);
1501 Indices.append(Src->op_begin()+1, Src->op_end()-1);
1502 Indices.push_back(Sum);
1503 Indices.append(GEP.op_begin()+2, GEP.op_end());
1504 } else if (isa<Constant>(*GEP.idx_begin()) &&
1505 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
1506 Src->getNumOperands() != 1) {
1507 // Otherwise we can do the fold if the first index of the GEP is a zero
1508 Indices.append(Src->op_begin()+1, Src->op_end());
1509 Indices.append(GEP.idx_begin()+1, GEP.idx_end());
1512 if (!Indices.empty())
1513 return (GEP.isInBounds() && Src->isInBounds()) ?
1514 GetElementPtrInst::CreateInBounds(Src->getOperand(0), Indices,
1516 GetElementPtrInst::Create(Src->getOperand(0), Indices, GEP.getName());
1519 if (DL && GEP.getNumIndices() == 1) {
1520 unsigned AS = GEP.getPointerAddressSpace();
1521 if (GEP.getOperand(1)->getType()->getScalarSizeInBits() ==
1522 DL->getPointerSizeInBits(AS)) {
1523 Type *PtrTy = GEP.getPointerOperandType();
1524 Type *Ty = PtrTy->getPointerElementType();
1525 uint64_t TyAllocSize = DL->getTypeAllocSize(Ty);
1527 bool Matched = false;
1530 if (TyAllocSize == 1) {
1531 V = GEP.getOperand(1);
1533 } else if (match(GEP.getOperand(1),
1534 m_AShr(m_Value(V), m_ConstantInt(C)))) {
1535 if (TyAllocSize == 1ULL << C)
1537 } else if (match(GEP.getOperand(1),
1538 m_SDiv(m_Value(V), m_ConstantInt(C)))) {
1539 if (TyAllocSize == C)
1544 // Canonicalize (gep i8* X, -(ptrtoint Y))
1545 // to (inttoptr (sub (ptrtoint X), (ptrtoint Y)))
1546 // The GEP pattern is emitted by the SCEV expander for certain kinds of
1547 // pointer arithmetic.
1548 if (match(V, m_Neg(m_PtrToInt(m_Value())))) {
1549 Operator *Index = cast<Operator>(V);
1550 Value *PtrToInt = Builder->CreatePtrToInt(PtrOp, Index->getType());
1551 Value *NewSub = Builder->CreateSub(PtrToInt, Index->getOperand(1));
1552 return CastInst::Create(Instruction::IntToPtr, NewSub, GEP.getType());
1554 // Canonicalize (gep i8* X, (ptrtoint Y)-(ptrtoint X))
1557 if (match(V, m_Sub(m_PtrToInt(m_Value(Y)),
1558 m_PtrToInt(m_Specific(GEP.getOperand(0)))))) {
1559 return CastInst::CreatePointerBitCastOrAddrSpaceCast(Y,
1566 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0).
1567 Value *StrippedPtr = PtrOp->stripPointerCasts();
1568 PointerType *StrippedPtrTy = dyn_cast<PointerType>(StrippedPtr->getType());
1570 // We do not handle pointer-vector geps here.
1574 if (StrippedPtr != PtrOp) {
1575 bool HasZeroPointerIndex = false;
1576 if (ConstantInt *C = dyn_cast<ConstantInt>(GEP.getOperand(1)))
1577 HasZeroPointerIndex = C->isZero();
1579 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
1580 // into : GEP [10 x i8]* X, i32 0, ...
1582 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
1583 // into : GEP i8* X, ...
1585 // This occurs when the program declares an array extern like "int X[];"
1586 if (HasZeroPointerIndex) {
1587 PointerType *CPTy = cast<PointerType>(PtrOp->getType());
1588 if (ArrayType *CATy =
1589 dyn_cast<ArrayType>(CPTy->getElementType())) {
1590 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
1591 if (CATy->getElementType() == StrippedPtrTy->getElementType()) {
1592 // -> GEP i8* X, ...
1593 SmallVector<Value*, 8> Idx(GEP.idx_begin()+1, GEP.idx_end());
1594 GetElementPtrInst *Res =
1595 GetElementPtrInst::Create(StrippedPtr, Idx, GEP.getName());
1596 Res->setIsInBounds(GEP.isInBounds());
1597 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace())
1599 // Insert Res, and create an addrspacecast.
1601 // GEP (addrspacecast i8 addrspace(1)* X to [0 x i8]*), i32 0, ...
1603 // %0 = GEP i8 addrspace(1)* X, ...
1604 // addrspacecast i8 addrspace(1)* %0 to i8*
1605 return new AddrSpaceCastInst(Builder->Insert(Res), GEP.getType());
1608 if (ArrayType *XATy =
1609 dyn_cast<ArrayType>(StrippedPtrTy->getElementType())){
1610 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
1611 if (CATy->getElementType() == XATy->getElementType()) {
1612 // -> GEP [10 x i8]* X, i32 0, ...
1613 // At this point, we know that the cast source type is a pointer
1614 // to an array of the same type as the destination pointer
1615 // array. Because the array type is never stepped over (there
1616 // is a leading zero) we can fold the cast into this GEP.
1617 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace()) {
1618 GEP.setOperand(0, StrippedPtr);
1621 // Cannot replace the base pointer directly because StrippedPtr's
1622 // address space is different. Instead, create a new GEP followed by
1623 // an addrspacecast.
1625 // GEP (addrspacecast [10 x i8] addrspace(1)* X to [0 x i8]*),
1628 // %0 = GEP [10 x i8] addrspace(1)* X, ...
1629 // addrspacecast i8 addrspace(1)* %0 to i8*
1630 SmallVector<Value*, 8> Idx(GEP.idx_begin(), GEP.idx_end());
1631 Value *NewGEP = GEP.isInBounds() ?
1632 Builder->CreateInBoundsGEP(StrippedPtr, Idx, GEP.getName()) :
1633 Builder->CreateGEP(StrippedPtr, Idx, GEP.getName());
1634 return new AddrSpaceCastInst(NewGEP, GEP.getType());
1638 } else if (GEP.getNumOperands() == 2) {
1639 // Transform things like:
1640 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
1641 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
1642 Type *SrcElTy = StrippedPtrTy->getElementType();
1643 Type *ResElTy = PtrOp->getType()->getPointerElementType();
1644 if (DL && SrcElTy->isArrayTy() &&
1645 DL->getTypeAllocSize(SrcElTy->getArrayElementType()) ==
1646 DL->getTypeAllocSize(ResElTy)) {
1647 Type *IdxType = DL->getIntPtrType(GEP.getType());
1648 Value *Idx[2] = { Constant::getNullValue(IdxType), GEP.getOperand(1) };
1649 Value *NewGEP = GEP.isInBounds() ?
1650 Builder->CreateInBoundsGEP(StrippedPtr, Idx, GEP.getName()) :
1651 Builder->CreateGEP(StrippedPtr, Idx, GEP.getName());
1653 // V and GEP are both pointer types --> BitCast
1654 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
1658 // Transform things like:
1659 // %V = mul i64 %N, 4
1660 // %t = getelementptr i8* bitcast (i32* %arr to i8*), i32 %V
1661 // into: %t1 = getelementptr i32* %arr, i32 %N; bitcast
1662 if (DL && ResElTy->isSized() && SrcElTy->isSized()) {
1663 // Check that changing the type amounts to dividing the index by a scale
1665 uint64_t ResSize = DL->getTypeAllocSize(ResElTy);
1666 uint64_t SrcSize = DL->getTypeAllocSize(SrcElTy);
1667 if (ResSize && SrcSize % ResSize == 0) {
1668 Value *Idx = GEP.getOperand(1);
1669 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
1670 uint64_t Scale = SrcSize / ResSize;
1672 // Earlier transforms ensure that the index has type IntPtrType, which
1673 // considerably simplifies the logic by eliminating implicit casts.
1674 assert(Idx->getType() == DL->getIntPtrType(GEP.getType()) &&
1675 "Index not cast to pointer width?");
1678 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
1679 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
1680 // If the multiplication NewIdx * Scale may overflow then the new
1681 // GEP may not be "inbounds".
1682 Value *NewGEP = GEP.isInBounds() && NSW ?
1683 Builder->CreateInBoundsGEP(StrippedPtr, NewIdx, GEP.getName()) :
1684 Builder->CreateGEP(StrippedPtr, NewIdx, GEP.getName());
1686 // The NewGEP must be pointer typed, so must the old one -> BitCast
1687 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
1693 // Similarly, transform things like:
1694 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
1695 // (where tmp = 8*tmp2) into:
1696 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
1697 if (DL && ResElTy->isSized() && SrcElTy->isSized() &&
1698 SrcElTy->isArrayTy()) {
1699 // Check that changing to the array element type amounts to dividing the
1700 // index by a scale factor.
1701 uint64_t ResSize = DL->getTypeAllocSize(ResElTy);
1702 uint64_t ArrayEltSize
1703 = DL->getTypeAllocSize(SrcElTy->getArrayElementType());
1704 if (ResSize && ArrayEltSize % ResSize == 0) {
1705 Value *Idx = GEP.getOperand(1);
1706 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
1707 uint64_t Scale = ArrayEltSize / ResSize;
1709 // Earlier transforms ensure that the index has type IntPtrType, which
1710 // considerably simplifies the logic by eliminating implicit casts.
1711 assert(Idx->getType() == DL->getIntPtrType(GEP.getType()) &&
1712 "Index not cast to pointer width?");
1715 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
1716 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
1717 // If the multiplication NewIdx * Scale may overflow then the new
1718 // GEP may not be "inbounds".
1720 Constant::getNullValue(DL->getIntPtrType(GEP.getType())),
1724 Value *NewGEP = GEP.isInBounds() && NSW ?
1725 Builder->CreateInBoundsGEP(StrippedPtr, Off, GEP.getName()) :
1726 Builder->CreateGEP(StrippedPtr, Off, GEP.getName());
1727 // The NewGEP must be pointer typed, so must the old one -> BitCast
1728 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
1739 // addrspacecast between types is canonicalized as a bitcast, then an
1740 // addrspacecast. To take advantage of the below bitcast + struct GEP, look
1741 // through the addrspacecast.
1742 if (AddrSpaceCastInst *ASC = dyn_cast<AddrSpaceCastInst>(PtrOp)) {
1743 // X = bitcast A addrspace(1)* to B addrspace(1)*
1744 // Y = addrspacecast A addrspace(1)* to B addrspace(2)*
1745 // Z = gep Y, <...constant indices...>
1746 // Into an addrspacecasted GEP of the struct.
1747 if (BitCastInst *BC = dyn_cast<BitCastInst>(ASC->getOperand(0)))
1751 /// See if we can simplify:
1752 /// X = bitcast A* to B*
1753 /// Y = gep X, <...constant indices...>
1754 /// into a gep of the original struct. This is important for SROA and alias
1755 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
1756 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
1757 Value *Operand = BCI->getOperand(0);
1758 PointerType *OpType = cast<PointerType>(Operand->getType());
1759 unsigned OffsetBits = DL->getPointerTypeSizeInBits(GEP.getType());
1760 APInt Offset(OffsetBits, 0);
1761 if (!isa<BitCastInst>(Operand) &&
1762 GEP.accumulateConstantOffset(*DL, Offset)) {
1764 // If this GEP instruction doesn't move the pointer, just replace the GEP
1765 // with a bitcast of the real input to the dest type.
1767 // If the bitcast is of an allocation, and the allocation will be
1768 // converted to match the type of the cast, don't touch this.
1769 if (isa<AllocaInst>(Operand) || isAllocationFn(Operand, TLI)) {
1770 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
1771 if (Instruction *I = visitBitCast(*BCI)) {
1774 BCI->getParent()->getInstList().insert(BCI, I);
1775 ReplaceInstUsesWith(*BCI, I);
1781 if (Operand->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
1782 return new AddrSpaceCastInst(Operand, GEP.getType());
1783 return new BitCastInst(Operand, GEP.getType());
1786 // Otherwise, if the offset is non-zero, we need to find out if there is a
1787 // field at Offset in 'A's type. If so, we can pull the cast through the
1789 SmallVector<Value*, 8> NewIndices;
1790 if (FindElementAtOffset(OpType, Offset.getSExtValue(), NewIndices)) {
1791 Value *NGEP = GEP.isInBounds() ?
1792 Builder->CreateInBoundsGEP(Operand, NewIndices) :
1793 Builder->CreateGEP(Operand, NewIndices);
1795 if (NGEP->getType() == GEP.getType())
1796 return ReplaceInstUsesWith(GEP, NGEP);
1797 NGEP->takeName(&GEP);
1799 if (NGEP->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
1800 return new AddrSpaceCastInst(NGEP, GEP.getType());
1801 return new BitCastInst(NGEP, GEP.getType());
1810 isAllocSiteRemovable(Instruction *AI, SmallVectorImpl<WeakVH> &Users,
1811 const TargetLibraryInfo *TLI) {
1812 SmallVector<Instruction*, 4> Worklist;
1813 Worklist.push_back(AI);
1816 Instruction *PI = Worklist.pop_back_val();
1817 for (User *U : PI->users()) {
1818 Instruction *I = cast<Instruction>(U);
1819 switch (I->getOpcode()) {
1821 // Give up the moment we see something we can't handle.
1824 case Instruction::BitCast:
1825 case Instruction::GetElementPtr:
1827 Worklist.push_back(I);
1830 case Instruction::ICmp: {
1831 ICmpInst *ICI = cast<ICmpInst>(I);
1832 // We can fold eq/ne comparisons with null to false/true, respectively.
1833 if (!ICI->isEquality() || !isa<ConstantPointerNull>(ICI->getOperand(1)))
1839 case Instruction::Call:
1840 // Ignore no-op and store intrinsics.
1841 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1842 switch (II->getIntrinsicID()) {
1846 case Intrinsic::memmove:
1847 case Intrinsic::memcpy:
1848 case Intrinsic::memset: {
1849 MemIntrinsic *MI = cast<MemIntrinsic>(II);
1850 if (MI->isVolatile() || MI->getRawDest() != PI)
1854 case Intrinsic::dbg_declare:
1855 case Intrinsic::dbg_value:
1856 case Intrinsic::invariant_start:
1857 case Intrinsic::invariant_end:
1858 case Intrinsic::lifetime_start:
1859 case Intrinsic::lifetime_end:
1860 case Intrinsic::objectsize:
1866 if (isFreeCall(I, TLI)) {
1872 case Instruction::Store: {
1873 StoreInst *SI = cast<StoreInst>(I);
1874 if (SI->isVolatile() || SI->getPointerOperand() != PI)
1880 llvm_unreachable("missing a return?");
1882 } while (!Worklist.empty());
1886 Instruction *InstCombiner::visitAllocSite(Instruction &MI) {
1887 // If we have a malloc call which is only used in any amount of comparisons
1888 // to null and free calls, delete the calls and replace the comparisons with
1889 // true or false as appropriate.
1890 SmallVector<WeakVH, 64> Users;
1891 if (isAllocSiteRemovable(&MI, Users, TLI)) {
1892 for (unsigned i = 0, e = Users.size(); i != e; ++i) {
1893 Instruction *I = cast_or_null<Instruction>(&*Users[i]);
1896 if (ICmpInst *C = dyn_cast<ICmpInst>(I)) {
1897 ReplaceInstUsesWith(*C,
1898 ConstantInt::get(Type::getInt1Ty(C->getContext()),
1899 C->isFalseWhenEqual()));
1900 } else if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) {
1901 ReplaceInstUsesWith(*I, UndefValue::get(I->getType()));
1902 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1903 if (II->getIntrinsicID() == Intrinsic::objectsize) {
1904 ConstantInt *CI = cast<ConstantInt>(II->getArgOperand(1));
1905 uint64_t DontKnow = CI->isZero() ? -1ULL : 0;
1906 ReplaceInstUsesWith(*I, ConstantInt::get(I->getType(), DontKnow));
1909 EraseInstFromFunction(*I);
1912 if (InvokeInst *II = dyn_cast<InvokeInst>(&MI)) {
1913 // Replace invoke with a NOP intrinsic to maintain the original CFG
1914 Module *M = II->getParent()->getParent()->getParent();
1915 Function *F = Intrinsic::getDeclaration(M, Intrinsic::donothing);
1916 InvokeInst::Create(F, II->getNormalDest(), II->getUnwindDest(),
1917 None, "", II->getParent());
1919 return EraseInstFromFunction(MI);
1924 /// \brief Move the call to free before a NULL test.
1926 /// Check if this free is accessed after its argument has been test
1927 /// against NULL (property 0).
1928 /// If yes, it is legal to move this call in its predecessor block.
1930 /// The move is performed only if the block containing the call to free
1931 /// will be removed, i.e.:
1932 /// 1. it has only one predecessor P, and P has two successors
1933 /// 2. it contains the call and an unconditional branch
1934 /// 3. its successor is the same as its predecessor's successor
1936 /// The profitability is out-of concern here and this function should
1937 /// be called only if the caller knows this transformation would be
1938 /// profitable (e.g., for code size).
1939 static Instruction *
1940 tryToMoveFreeBeforeNullTest(CallInst &FI) {
1941 Value *Op = FI.getArgOperand(0);
1942 BasicBlock *FreeInstrBB = FI.getParent();
1943 BasicBlock *PredBB = FreeInstrBB->getSinglePredecessor();
1945 // Validate part of constraint #1: Only one predecessor
1946 // FIXME: We can extend the number of predecessor, but in that case, we
1947 // would duplicate the call to free in each predecessor and it may
1948 // not be profitable even for code size.
1952 // Validate constraint #2: Does this block contains only the call to
1953 // free and an unconditional branch?
1954 // FIXME: We could check if we can speculate everything in the
1955 // predecessor block
1956 if (FreeInstrBB->size() != 2)
1959 if (!match(FreeInstrBB->getTerminator(), m_UnconditionalBr(SuccBB)))
1962 // Validate the rest of constraint #1 by matching on the pred branch.
1963 TerminatorInst *TI = PredBB->getTerminator();
1964 BasicBlock *TrueBB, *FalseBB;
1965 ICmpInst::Predicate Pred;
1966 if (!match(TI, m_Br(m_ICmp(Pred, m_Specific(Op), m_Zero()), TrueBB, FalseBB)))
1968 if (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE)
1971 // Validate constraint #3: Ensure the null case just falls through.
1972 if (SuccBB != (Pred == ICmpInst::ICMP_EQ ? TrueBB : FalseBB))
1974 assert(FreeInstrBB == (Pred == ICmpInst::ICMP_EQ ? FalseBB : TrueBB) &&
1975 "Broken CFG: missing edge from predecessor to successor");
1982 Instruction *InstCombiner::visitFree(CallInst &FI) {
1983 Value *Op = FI.getArgOperand(0);
1985 // free undef -> unreachable.
1986 if (isa<UndefValue>(Op)) {
1987 // Insert a new store to null because we cannot modify the CFG here.
1988 Builder->CreateStore(ConstantInt::getTrue(FI.getContext()),
1989 UndefValue::get(Type::getInt1PtrTy(FI.getContext())));
1990 return EraseInstFromFunction(FI);
1993 // If we have 'free null' delete the instruction. This can happen in stl code
1994 // when lots of inlining happens.
1995 if (isa<ConstantPointerNull>(Op))
1996 return EraseInstFromFunction(FI);
1998 // If we optimize for code size, try to move the call to free before the null
1999 // test so that simplify cfg can remove the empty block and dead code
2000 // elimination the branch. I.e., helps to turn something like:
2001 // if (foo) free(foo);
2005 if (Instruction *I = tryToMoveFreeBeforeNullTest(FI))
2011 Instruction *InstCombiner::visitReturnInst(ReturnInst &RI) {
2012 if (RI.getNumOperands() == 0) // ret void
2015 Value *ResultOp = RI.getOperand(0);
2016 Type *VTy = ResultOp->getType();
2017 if (!VTy->isIntegerTy())
2020 // There might be assume intrinsics dominating this return that completely
2021 // determine the value. If so, constant fold it.
2022 unsigned BitWidth = VTy->getPrimitiveSizeInBits();
2023 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2024 computeKnownBits(ResultOp, KnownZero, KnownOne, 0, &RI);
2025 if ((KnownZero|KnownOne).isAllOnesValue())
2026 RI.setOperand(0, Constant::getIntegerValue(VTy, KnownOne));
2031 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
2032 // Change br (not X), label True, label False to: br X, label False, True
2034 BasicBlock *TrueDest;
2035 BasicBlock *FalseDest;
2036 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
2037 !isa<Constant>(X)) {
2038 // Swap Destinations and condition...
2040 BI.swapSuccessors();
2044 // Canonicalize fcmp_one -> fcmp_oeq
2045 FCmpInst::Predicate FPred; Value *Y;
2046 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
2047 TrueDest, FalseDest)) &&
2048 BI.getCondition()->hasOneUse())
2049 if (FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
2050 FPred == FCmpInst::FCMP_OGE) {
2051 FCmpInst *Cond = cast<FCmpInst>(BI.getCondition());
2052 Cond->setPredicate(FCmpInst::getInversePredicate(FPred));
2054 // Swap Destinations and condition.
2055 BI.swapSuccessors();
2060 // Canonicalize icmp_ne -> icmp_eq
2061 ICmpInst::Predicate IPred;
2062 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
2063 TrueDest, FalseDest)) &&
2064 BI.getCondition()->hasOneUse())
2065 if (IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
2066 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
2067 IPred == ICmpInst::ICMP_SGE) {
2068 ICmpInst *Cond = cast<ICmpInst>(BI.getCondition());
2069 Cond->setPredicate(ICmpInst::getInversePredicate(IPred));
2070 // Swap Destinations and condition.
2071 BI.swapSuccessors();
2079 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
2080 Value *Cond = SI.getCondition();
2081 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
2082 if (I->getOpcode() == Instruction::Add)
2083 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2084 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
2085 // Skip the first item since that's the default case.
2086 for (SwitchInst::CaseIt i = SI.case_begin(), e = SI.case_end();
2088 ConstantInt* CaseVal = i.getCaseValue();
2089 Constant* NewCaseVal = ConstantExpr::getSub(cast<Constant>(CaseVal),
2091 assert(isa<ConstantInt>(NewCaseVal) &&
2092 "Result of expression should be constant");
2093 i.setValue(cast<ConstantInt>(NewCaseVal));
2095 SI.setCondition(I->getOperand(0));
2103 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
2104 Value *Agg = EV.getAggregateOperand();
2106 if (!EV.hasIndices())
2107 return ReplaceInstUsesWith(EV, Agg);
2109 if (Constant *C = dyn_cast<Constant>(Agg)) {
2110 if (Constant *C2 = C->getAggregateElement(*EV.idx_begin())) {
2111 if (EV.getNumIndices() == 0)
2112 return ReplaceInstUsesWith(EV, C2);
2113 // Extract the remaining indices out of the constant indexed by the
2115 return ExtractValueInst::Create(C2, EV.getIndices().slice(1));
2117 return nullptr; // Can't handle other constants
2120 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
2121 // We're extracting from an insertvalue instruction, compare the indices
2122 const unsigned *exti, *exte, *insi, *inse;
2123 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
2124 exte = EV.idx_end(), inse = IV->idx_end();
2125 exti != exte && insi != inse;
2128 // The insert and extract both reference distinctly different elements.
2129 // This means the extract is not influenced by the insert, and we can
2130 // replace the aggregate operand of the extract with the aggregate
2131 // operand of the insert. i.e., replace
2132 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
2133 // %E = extractvalue { i32, { i32 } } %I, 0
2135 // %E = extractvalue { i32, { i32 } } %A, 0
2136 return ExtractValueInst::Create(IV->getAggregateOperand(),
2139 if (exti == exte && insi == inse)
2140 // Both iterators are at the end: Index lists are identical. Replace
2141 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
2142 // %C = extractvalue { i32, { i32 } } %B, 1, 0
2144 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
2146 // The extract list is a prefix of the insert list. i.e. replace
2147 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
2148 // %E = extractvalue { i32, { i32 } } %I, 1
2150 // %X = extractvalue { i32, { i32 } } %A, 1
2151 // %E = insertvalue { i32 } %X, i32 42, 0
2152 // by switching the order of the insert and extract (though the
2153 // insertvalue should be left in, since it may have other uses).
2154 Value *NewEV = Builder->CreateExtractValue(IV->getAggregateOperand(),
2156 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
2157 makeArrayRef(insi, inse));
2160 // The insert list is a prefix of the extract list
2161 // We can simply remove the common indices from the extract and make it
2162 // operate on the inserted value instead of the insertvalue result.
2164 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
2165 // %E = extractvalue { i32, { i32 } } %I, 1, 0
2167 // %E extractvalue { i32 } { i32 42 }, 0
2168 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
2169 makeArrayRef(exti, exte));
2171 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) {
2172 // We're extracting from an intrinsic, see if we're the only user, which
2173 // allows us to simplify multiple result intrinsics to simpler things that
2174 // just get one value.
2175 if (II->hasOneUse()) {
2176 // Check if we're grabbing the overflow bit or the result of a 'with
2177 // overflow' intrinsic. If it's the latter we can remove the intrinsic
2178 // and replace it with a traditional binary instruction.
2179 switch (II->getIntrinsicID()) {
2180 case Intrinsic::uadd_with_overflow:
2181 case Intrinsic::sadd_with_overflow:
2182 if (*EV.idx_begin() == 0) { // Normal result.
2183 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2184 ReplaceInstUsesWith(*II, UndefValue::get(II->getType()));
2185 EraseInstFromFunction(*II);
2186 return BinaryOperator::CreateAdd(LHS, RHS);
2189 // If the normal result of the add is dead, and the RHS is a constant,
2190 // we can transform this into a range comparison.
2191 // overflow = uadd a, -4 --> overflow = icmp ugt a, 3
2192 if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow)
2193 if (ConstantInt *CI = dyn_cast<ConstantInt>(II->getArgOperand(1)))
2194 return new ICmpInst(ICmpInst::ICMP_UGT, II->getArgOperand(0),
2195 ConstantExpr::getNot(CI));
2197 case Intrinsic::usub_with_overflow:
2198 case Intrinsic::ssub_with_overflow:
2199 if (*EV.idx_begin() == 0) { // Normal result.
2200 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2201 ReplaceInstUsesWith(*II, UndefValue::get(II->getType()));
2202 EraseInstFromFunction(*II);
2203 return BinaryOperator::CreateSub(LHS, RHS);
2206 case Intrinsic::umul_with_overflow:
2207 case Intrinsic::smul_with_overflow:
2208 if (*EV.idx_begin() == 0) { // Normal result.
2209 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2210 ReplaceInstUsesWith(*II, UndefValue::get(II->getType()));
2211 EraseInstFromFunction(*II);
2212 return BinaryOperator::CreateMul(LHS, RHS);
2220 if (LoadInst *L = dyn_cast<LoadInst>(Agg))
2221 // If the (non-volatile) load only has one use, we can rewrite this to a
2222 // load from a GEP. This reduces the size of the load.
2223 // FIXME: If a load is used only by extractvalue instructions then this
2224 // could be done regardless of having multiple uses.
2225 if (L->isSimple() && L->hasOneUse()) {
2226 // extractvalue has integer indices, getelementptr has Value*s. Convert.
2227 SmallVector<Value*, 4> Indices;
2228 // Prefix an i32 0 since we need the first element.
2229 Indices.push_back(Builder->getInt32(0));
2230 for (ExtractValueInst::idx_iterator I = EV.idx_begin(), E = EV.idx_end();
2232 Indices.push_back(Builder->getInt32(*I));
2234 // We need to insert these at the location of the old load, not at that of
2235 // the extractvalue.
2236 Builder->SetInsertPoint(L->getParent(), L);
2237 Value *GEP = Builder->CreateInBoundsGEP(L->getPointerOperand(), Indices);
2238 // Returning the load directly will cause the main loop to insert it in
2239 // the wrong spot, so use ReplaceInstUsesWith().
2240 return ReplaceInstUsesWith(EV, Builder->CreateLoad(GEP));
2242 // We could simplify extracts from other values. Note that nested extracts may
2243 // already be simplified implicitly by the above: extract (extract (insert) )
2244 // will be translated into extract ( insert ( extract ) ) first and then just
2245 // the value inserted, if appropriate. Similarly for extracts from single-use
2246 // loads: extract (extract (load)) will be translated to extract (load (gep))
2247 // and if again single-use then via load (gep (gep)) to load (gep).
2248 // However, double extracts from e.g. function arguments or return values
2249 // aren't handled yet.
2253 enum Personality_Type {
2254 Unknown_Personality,
2255 GNU_Ada_Personality,
2256 GNU_CXX_Personality,
2257 GNU_ObjC_Personality
2260 /// RecognizePersonality - See if the given exception handling personality
2261 /// function is one that we understand. If so, return a description of it;
2262 /// otherwise return Unknown_Personality.
2263 static Personality_Type RecognizePersonality(Value *Pers) {
2264 Function *F = dyn_cast<Function>(Pers->stripPointerCasts());
2266 return Unknown_Personality;
2267 return StringSwitch<Personality_Type>(F->getName())
2268 .Case("__gnat_eh_personality", GNU_Ada_Personality)
2269 .Case("__gxx_personality_v0", GNU_CXX_Personality)
2270 .Case("__objc_personality_v0", GNU_ObjC_Personality)
2271 .Default(Unknown_Personality);
2274 /// isCatchAll - Return 'true' if the given typeinfo will match anything.
2275 static bool isCatchAll(Personality_Type Personality, Constant *TypeInfo) {
2276 switch (Personality) {
2277 case Unknown_Personality:
2279 case GNU_Ada_Personality:
2280 // While __gnat_all_others_value will match any Ada exception, it doesn't
2281 // match foreign exceptions (or didn't, before gcc-4.7).
2283 case GNU_CXX_Personality:
2284 case GNU_ObjC_Personality:
2285 return TypeInfo->isNullValue();
2287 llvm_unreachable("Unknown personality!");
2290 static bool shorter_filter(const Value *LHS, const Value *RHS) {
2292 cast<ArrayType>(LHS->getType())->getNumElements()
2294 cast<ArrayType>(RHS->getType())->getNumElements();
2297 Instruction *InstCombiner::visitLandingPadInst(LandingPadInst &LI) {
2298 // The logic here should be correct for any real-world personality function.
2299 // However if that turns out not to be true, the offending logic can always
2300 // be conditioned on the personality function, like the catch-all logic is.
2301 Personality_Type Personality = RecognizePersonality(LI.getPersonalityFn());
2303 // Simplify the list of clauses, eg by removing repeated catch clauses
2304 // (these are often created by inlining).
2305 bool MakeNewInstruction = false; // If true, recreate using the following:
2306 SmallVector<Constant *, 16> NewClauses; // - Clauses for the new instruction;
2307 bool CleanupFlag = LI.isCleanup(); // - The new instruction is a cleanup.
2309 SmallPtrSet<Value *, 16> AlreadyCaught; // Typeinfos known caught already.
2310 for (unsigned i = 0, e = LI.getNumClauses(); i != e; ++i) {
2311 bool isLastClause = i + 1 == e;
2312 if (LI.isCatch(i)) {
2314 Constant *CatchClause = LI.getClause(i);
2315 Constant *TypeInfo = CatchClause->stripPointerCasts();
2317 // If we already saw this clause, there is no point in having a second
2319 if (AlreadyCaught.insert(TypeInfo)) {
2320 // This catch clause was not already seen.
2321 NewClauses.push_back(CatchClause);
2323 // Repeated catch clause - drop the redundant copy.
2324 MakeNewInstruction = true;
2327 // If this is a catch-all then there is no point in keeping any following
2328 // clauses or marking the landingpad as having a cleanup.
2329 if (isCatchAll(Personality, TypeInfo)) {
2331 MakeNewInstruction = true;
2332 CleanupFlag = false;
2336 // A filter clause. If any of the filter elements were already caught
2337 // then they can be dropped from the filter. It is tempting to try to
2338 // exploit the filter further by saying that any typeinfo that does not
2339 // occur in the filter can't be caught later (and thus can be dropped).
2340 // However this would be wrong, since typeinfos can match without being
2341 // equal (for example if one represents a C++ class, and the other some
2342 // class derived from it).
2343 assert(LI.isFilter(i) && "Unsupported landingpad clause!");
2344 Constant *FilterClause = LI.getClause(i);
2345 ArrayType *FilterType = cast<ArrayType>(FilterClause->getType());
2346 unsigned NumTypeInfos = FilterType->getNumElements();
2348 // An empty filter catches everything, so there is no point in keeping any
2349 // following clauses or marking the landingpad as having a cleanup. By
2350 // dealing with this case here the following code is made a bit simpler.
2351 if (!NumTypeInfos) {
2352 NewClauses.push_back(FilterClause);
2354 MakeNewInstruction = true;
2355 CleanupFlag = false;
2359 bool MakeNewFilter = false; // If true, make a new filter.
2360 SmallVector<Constant *, 16> NewFilterElts; // New elements.
2361 if (isa<ConstantAggregateZero>(FilterClause)) {
2362 // Not an empty filter - it contains at least one null typeinfo.
2363 assert(NumTypeInfos > 0 && "Should have handled empty filter already!");
2364 Constant *TypeInfo =
2365 Constant::getNullValue(FilterType->getElementType());
2366 // If this typeinfo is a catch-all then the filter can never match.
2367 if (isCatchAll(Personality, TypeInfo)) {
2368 // Throw the filter away.
2369 MakeNewInstruction = true;
2373 // There is no point in having multiple copies of this typeinfo, so
2374 // discard all but the first copy if there is more than one.
2375 NewFilterElts.push_back(TypeInfo);
2376 if (NumTypeInfos > 1)
2377 MakeNewFilter = true;
2379 ConstantArray *Filter = cast<ConstantArray>(FilterClause);
2380 SmallPtrSet<Value *, 16> SeenInFilter; // For uniquing the elements.
2381 NewFilterElts.reserve(NumTypeInfos);
2383 // Remove any filter elements that were already caught or that already
2384 // occurred in the filter. While there, see if any of the elements are
2385 // catch-alls. If so, the filter can be discarded.
2386 bool SawCatchAll = false;
2387 for (unsigned j = 0; j != NumTypeInfos; ++j) {
2388 Constant *Elt = Filter->getOperand(j);
2389 Constant *TypeInfo = Elt->stripPointerCasts();
2390 if (isCatchAll(Personality, TypeInfo)) {
2391 // This element is a catch-all. Bail out, noting this fact.
2395 if (AlreadyCaught.count(TypeInfo))
2396 // Already caught by an earlier clause, so having it in the filter
2399 // There is no point in having multiple copies of the same typeinfo in
2400 // a filter, so only add it if we didn't already.
2401 if (SeenInFilter.insert(TypeInfo))
2402 NewFilterElts.push_back(cast<Constant>(Elt));
2404 // A filter containing a catch-all cannot match anything by definition.
2406 // Throw the filter away.
2407 MakeNewInstruction = true;
2411 // If we dropped something from the filter, make a new one.
2412 if (NewFilterElts.size() < NumTypeInfos)
2413 MakeNewFilter = true;
2415 if (MakeNewFilter) {
2416 FilterType = ArrayType::get(FilterType->getElementType(),
2417 NewFilterElts.size());
2418 FilterClause = ConstantArray::get(FilterType, NewFilterElts);
2419 MakeNewInstruction = true;
2422 NewClauses.push_back(FilterClause);
2424 // If the new filter is empty then it will catch everything so there is
2425 // no point in keeping any following clauses or marking the landingpad
2426 // as having a cleanup. The case of the original filter being empty was
2427 // already handled above.
2428 if (MakeNewFilter && !NewFilterElts.size()) {
2429 assert(MakeNewInstruction && "New filter but not a new instruction!");
2430 CleanupFlag = false;
2436 // If several filters occur in a row then reorder them so that the shortest
2437 // filters come first (those with the smallest number of elements). This is
2438 // advantageous because shorter filters are more likely to match, speeding up
2439 // unwinding, but mostly because it increases the effectiveness of the other
2440 // filter optimizations below.
2441 for (unsigned i = 0, e = NewClauses.size(); i + 1 < e; ) {
2443 // Find the maximal 'j' s.t. the range [i, j) consists entirely of filters.
2444 for (j = i; j != e; ++j)
2445 if (!isa<ArrayType>(NewClauses[j]->getType()))
2448 // Check whether the filters are already sorted by length. We need to know
2449 // if sorting them is actually going to do anything so that we only make a
2450 // new landingpad instruction if it does.
2451 for (unsigned k = i; k + 1 < j; ++k)
2452 if (shorter_filter(NewClauses[k+1], NewClauses[k])) {
2453 // Not sorted, so sort the filters now. Doing an unstable sort would be
2454 // correct too but reordering filters pointlessly might confuse users.
2455 std::stable_sort(NewClauses.begin() + i, NewClauses.begin() + j,
2457 MakeNewInstruction = true;
2461 // Look for the next batch of filters.
2465 // If typeinfos matched if and only if equal, then the elements of a filter L
2466 // that occurs later than a filter F could be replaced by the intersection of
2467 // the elements of F and L. In reality two typeinfos can match without being
2468 // equal (for example if one represents a C++ class, and the other some class
2469 // derived from it) so it would be wrong to perform this transform in general.
2470 // However the transform is correct and useful if F is a subset of L. In that
2471 // case L can be replaced by F, and thus removed altogether since repeating a
2472 // filter is pointless. So here we look at all pairs of filters F and L where
2473 // L follows F in the list of clauses, and remove L if every element of F is
2474 // an element of L. This can occur when inlining C++ functions with exception
2476 for (unsigned i = 0; i + 1 < NewClauses.size(); ++i) {
2477 // Examine each filter in turn.
2478 Value *Filter = NewClauses[i];
2479 ArrayType *FTy = dyn_cast<ArrayType>(Filter->getType());
2481 // Not a filter - skip it.
2483 unsigned FElts = FTy->getNumElements();
2484 // Examine each filter following this one. Doing this backwards means that
2485 // we don't have to worry about filters disappearing under us when removed.
2486 for (unsigned j = NewClauses.size() - 1; j != i; --j) {
2487 Value *LFilter = NewClauses[j];
2488 ArrayType *LTy = dyn_cast<ArrayType>(LFilter->getType());
2490 // Not a filter - skip it.
2492 // If Filter is a subset of LFilter, i.e. every element of Filter is also
2493 // an element of LFilter, then discard LFilter.
2494 SmallVectorImpl<Constant *>::iterator J = NewClauses.begin() + j;
2495 // If Filter is empty then it is a subset of LFilter.
2498 NewClauses.erase(J);
2499 MakeNewInstruction = true;
2500 // Move on to the next filter.
2503 unsigned LElts = LTy->getNumElements();
2504 // If Filter is longer than LFilter then it cannot be a subset of it.
2506 // Move on to the next filter.
2508 // At this point we know that LFilter has at least one element.
2509 if (isa<ConstantAggregateZero>(LFilter)) { // LFilter only contains zeros.
2510 // Filter is a subset of LFilter iff Filter contains only zeros (as we
2511 // already know that Filter is not longer than LFilter).
2512 if (isa<ConstantAggregateZero>(Filter)) {
2513 assert(FElts <= LElts && "Should have handled this case earlier!");
2515 NewClauses.erase(J);
2516 MakeNewInstruction = true;
2518 // Move on to the next filter.
2521 ConstantArray *LArray = cast<ConstantArray>(LFilter);
2522 if (isa<ConstantAggregateZero>(Filter)) { // Filter only contains zeros.
2523 // Since Filter is non-empty and contains only zeros, it is a subset of
2524 // LFilter iff LFilter contains a zero.
2525 assert(FElts > 0 && "Should have eliminated the empty filter earlier!");
2526 for (unsigned l = 0; l != LElts; ++l)
2527 if (LArray->getOperand(l)->isNullValue()) {
2528 // LFilter contains a zero - discard it.
2529 NewClauses.erase(J);
2530 MakeNewInstruction = true;
2533 // Move on to the next filter.
2536 // At this point we know that both filters are ConstantArrays. Loop over
2537 // operands to see whether every element of Filter is also an element of
2538 // LFilter. Since filters tend to be short this is probably faster than
2539 // using a method that scales nicely.
2540 ConstantArray *FArray = cast<ConstantArray>(Filter);
2541 bool AllFound = true;
2542 for (unsigned f = 0; f != FElts; ++f) {
2543 Value *FTypeInfo = FArray->getOperand(f)->stripPointerCasts();
2545 for (unsigned l = 0; l != LElts; ++l) {
2546 Value *LTypeInfo = LArray->getOperand(l)->stripPointerCasts();
2547 if (LTypeInfo == FTypeInfo) {
2557 NewClauses.erase(J);
2558 MakeNewInstruction = true;
2560 // Move on to the next filter.
2564 // If we changed any of the clauses, replace the old landingpad instruction
2566 if (MakeNewInstruction) {
2567 LandingPadInst *NLI = LandingPadInst::Create(LI.getType(),
2568 LI.getPersonalityFn(),
2570 for (unsigned i = 0, e = NewClauses.size(); i != e; ++i)
2571 NLI->addClause(NewClauses[i]);
2572 // A landing pad with no clauses must have the cleanup flag set. It is
2573 // theoretically possible, though highly unlikely, that we eliminated all
2574 // clauses. If so, force the cleanup flag to true.
2575 if (NewClauses.empty())
2577 NLI->setCleanup(CleanupFlag);
2581 // Even if none of the clauses changed, we may nonetheless have understood
2582 // that the cleanup flag is pointless. Clear it if so.
2583 if (LI.isCleanup() != CleanupFlag) {
2584 assert(!CleanupFlag && "Adding a cleanup, not removing one?!");
2585 LI.setCleanup(CleanupFlag);
2595 /// TryToSinkInstruction - Try to move the specified instruction from its
2596 /// current block into the beginning of DestBlock, which can only happen if it's
2597 /// safe to move the instruction past all of the instructions between it and the
2598 /// end of its block.
2599 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
2600 assert(I->hasOneUse() && "Invariants didn't hold!");
2602 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
2603 if (isa<PHINode>(I) || isa<LandingPadInst>(I) || I->mayHaveSideEffects() ||
2604 isa<TerminatorInst>(I))
2607 // Do not sink alloca instructions out of the entry block.
2608 if (isa<AllocaInst>(I) && I->getParent() ==
2609 &DestBlock->getParent()->getEntryBlock())
2612 // We can only sink load instructions if there is nothing between the load and
2613 // the end of block that could change the value.
2614 if (I->mayReadFromMemory()) {
2615 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
2617 if (Scan->mayWriteToMemory())
2621 BasicBlock::iterator InsertPos = DestBlock->getFirstInsertionPt();
2622 I->moveBefore(InsertPos);
2628 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
2629 /// all reachable code to the worklist.
2631 /// This has a couple of tricks to make the code faster and more powerful. In
2632 /// particular, we constant fold and DCE instructions as we go, to avoid adding
2633 /// them to the worklist (this significantly speeds up instcombine on code where
2634 /// many instructions are dead or constant). Additionally, if we find a branch
2635 /// whose condition is a known constant, we only visit the reachable successors.
2637 static bool AddReachableCodeToWorklist(BasicBlock *BB,
2638 SmallPtrSetImpl<BasicBlock*> &Visited,
2640 const DataLayout *DL,
2641 const TargetLibraryInfo *TLI) {
2642 bool MadeIRChange = false;
2643 SmallVector<BasicBlock*, 256> Worklist;
2644 Worklist.push_back(BB);
2646 SmallVector<Instruction*, 128> InstrsForInstCombineWorklist;
2647 DenseMap<ConstantExpr*, Constant*> FoldedConstants;
2650 BB = Worklist.pop_back_val();
2652 // We have now visited this block! If we've already been here, ignore it.
2653 if (!Visited.insert(BB)) continue;
2655 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
2656 Instruction *Inst = BBI++;
2658 // DCE instruction if trivially dead.
2659 if (isInstructionTriviallyDead(Inst, TLI)) {
2661 DEBUG(dbgs() << "IC: DCE: " << *Inst << '\n');
2662 Inst->eraseFromParent();
2666 // ConstantProp instruction if trivially constant.
2667 if (!Inst->use_empty() && isa<Constant>(Inst->getOperand(0)))
2668 if (Constant *C = ConstantFoldInstruction(Inst, DL, TLI)) {
2669 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: "
2671 Inst->replaceAllUsesWith(C);
2673 Inst->eraseFromParent();
2678 // See if we can constant fold its operands.
2679 for (User::op_iterator i = Inst->op_begin(), e = Inst->op_end();
2681 ConstantExpr *CE = dyn_cast<ConstantExpr>(i);
2682 if (CE == nullptr) continue;
2684 Constant*& FoldRes = FoldedConstants[CE];
2686 FoldRes = ConstantFoldConstantExpression(CE, DL, TLI);
2690 if (FoldRes != CE) {
2692 MadeIRChange = true;
2697 InstrsForInstCombineWorklist.push_back(Inst);
2700 // Recursively visit successors. If this is a branch or switch on a
2701 // constant, only visit the reachable successor.
2702 TerminatorInst *TI = BB->getTerminator();
2703 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
2704 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
2705 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
2706 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
2707 Worklist.push_back(ReachableBB);
2710 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
2711 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
2712 // See if this is an explicit destination.
2713 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2715 if (i.getCaseValue() == Cond) {
2716 BasicBlock *ReachableBB = i.getCaseSuccessor();
2717 Worklist.push_back(ReachableBB);
2721 // Otherwise it is the default destination.
2722 Worklist.push_back(SI->getDefaultDest());
2727 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
2728 Worklist.push_back(TI->getSuccessor(i));
2729 } while (!Worklist.empty());
2731 // Once we've found all of the instructions to add to instcombine's worklist,
2732 // add them in reverse order. This way instcombine will visit from the top
2733 // of the function down. This jives well with the way that it adds all uses
2734 // of instructions to the worklist after doing a transformation, thus avoiding
2735 // some N^2 behavior in pathological cases.
2736 IC.Worklist.AddInitialGroup(&InstrsForInstCombineWorklist[0],
2737 InstrsForInstCombineWorklist.size());
2739 return MadeIRChange;
2742 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
2743 MadeIRChange = false;
2745 DEBUG(dbgs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
2746 << F.getName() << "\n");
2749 // Do a depth-first traversal of the function, populate the worklist with
2750 // the reachable instructions. Ignore blocks that are not reachable. Keep
2751 // track of which blocks we visit.
2752 SmallPtrSet<BasicBlock*, 64> Visited;
2753 MadeIRChange |= AddReachableCodeToWorklist(F.begin(), Visited, *this, DL,
2756 // Do a quick scan over the function. If we find any blocks that are
2757 // unreachable, remove any instructions inside of them. This prevents
2758 // the instcombine code from having to deal with some bad special cases.
2759 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
2760 if (Visited.count(BB)) continue;
2762 // Delete the instructions backwards, as it has a reduced likelihood of
2763 // having to update as many def-use and use-def chains.
2764 Instruction *EndInst = BB->getTerminator(); // Last not to be deleted.
2765 while (EndInst != BB->begin()) {
2766 // Delete the next to last instruction.
2767 BasicBlock::iterator I = EndInst;
2768 Instruction *Inst = --I;
2769 if (!Inst->use_empty())
2770 Inst->replaceAllUsesWith(UndefValue::get(Inst->getType()));
2771 if (isa<LandingPadInst>(Inst)) {
2775 if (!isa<DbgInfoIntrinsic>(Inst)) {
2777 MadeIRChange = true;
2779 Inst->eraseFromParent();
2784 while (!Worklist.isEmpty()) {
2785 Instruction *I = Worklist.RemoveOne();
2786 if (I == nullptr) continue; // skip null values.
2788 // Check to see if we can DCE the instruction.
2789 if (isInstructionTriviallyDead(I, TLI)) {
2790 DEBUG(dbgs() << "IC: DCE: " << *I << '\n');
2791 EraseInstFromFunction(*I);
2793 MadeIRChange = true;
2797 // Instruction isn't dead, see if we can constant propagate it.
2798 if (!I->use_empty() && isa<Constant>(I->getOperand(0)))
2799 if (Constant *C = ConstantFoldInstruction(I, DL, TLI)) {
2800 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n');
2802 // Add operands to the worklist.
2803 ReplaceInstUsesWith(*I, C);
2805 EraseInstFromFunction(*I);
2806 MadeIRChange = true;
2810 // See if we can trivially sink this instruction to a successor basic block.
2811 if (I->hasOneUse()) {
2812 BasicBlock *BB = I->getParent();
2813 Instruction *UserInst = cast<Instruction>(*I->user_begin());
2814 BasicBlock *UserParent;
2816 // Get the block the use occurs in.
2817 if (PHINode *PN = dyn_cast<PHINode>(UserInst))
2818 UserParent = PN->getIncomingBlock(*I->use_begin());
2820 UserParent = UserInst->getParent();
2822 if (UserParent != BB) {
2823 bool UserIsSuccessor = false;
2824 // See if the user is one of our successors.
2825 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
2826 if (*SI == UserParent) {
2827 UserIsSuccessor = true;
2831 // If the user is one of our immediate successors, and if that successor
2832 // only has us as a predecessors (we'd have to split the critical edge
2833 // otherwise), we can keep going.
2834 if (UserIsSuccessor && UserParent->getSinglePredecessor()) {
2835 // Okay, the CFG is simple enough, try to sink this instruction.
2836 if (TryToSinkInstruction(I, UserParent)) {
2837 MadeIRChange = true;
2838 // We'll add uses of the sunk instruction below, but since sinking
2839 // can expose opportunities for it's *operands* add them to the
2841 for (Use &U : I->operands())
2842 if (Instruction *OpI = dyn_cast<Instruction>(U.get()))
2849 // Now that we have an instruction, try combining it to simplify it.
2850 Builder->SetInsertPoint(I->getParent(), I);
2851 Builder->SetCurrentDebugLocation(I->getDebugLoc());
2856 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str(););
2857 DEBUG(dbgs() << "IC: Visiting: " << OrigI << '\n');
2859 if (Instruction *Result = visit(*I)) {
2861 // Should we replace the old instruction with a new one?
2863 DEBUG(dbgs() << "IC: Old = " << *I << '\n'
2864 << " New = " << *Result << '\n');
2866 if (!I->getDebugLoc().isUnknown())
2867 Result->setDebugLoc(I->getDebugLoc());
2868 // Everything uses the new instruction now.
2869 I->replaceAllUsesWith(Result);
2871 // Move the name to the new instruction first.
2872 Result->takeName(I);
2874 // Push the new instruction and any users onto the worklist.
2875 Worklist.Add(Result);
2876 Worklist.AddUsersToWorkList(*Result);
2878 // Insert the new instruction into the basic block...
2879 BasicBlock *InstParent = I->getParent();
2880 BasicBlock::iterator InsertPos = I;
2882 // If we replace a PHI with something that isn't a PHI, fix up the
2884 if (!isa<PHINode>(Result) && isa<PHINode>(InsertPos))
2885 InsertPos = InstParent->getFirstInsertionPt();
2887 InstParent->getInstList().insert(InsertPos, Result);
2889 EraseInstFromFunction(*I);
2892 DEBUG(dbgs() << "IC: Mod = " << OrigI << '\n'
2893 << " New = " << *I << '\n');
2896 // If the instruction was modified, it's possible that it is now dead.
2897 // if so, remove it.
2898 if (isInstructionTriviallyDead(I, TLI)) {
2899 EraseInstFromFunction(*I);
2902 Worklist.AddUsersToWorkList(*I);
2905 MadeIRChange = true;
2910 return MadeIRChange;
2914 class InstCombinerLibCallSimplifier final : public LibCallSimplifier {
2917 InstCombinerLibCallSimplifier(const DataLayout *DL,
2918 const TargetLibraryInfo *TLI,
2920 : LibCallSimplifier(DL, TLI, EnableUnsafeFPShrink) {
2924 /// replaceAllUsesWith - override so that instruction replacement
2925 /// can be defined in terms of the instruction combiner framework.
2926 void replaceAllUsesWith(Instruction *I, Value *With) const override {
2927 IC->ReplaceInstUsesWith(*I, With);
2932 bool InstCombiner::runOnFunction(Function &F) {
2933 if (skipOptnoneFunction(F))
2936 AT = &getAnalysis<AssumptionTracker>();
2937 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
2938 DL = DLP ? &DLP->getDataLayout() : nullptr;
2939 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2940 TLI = &getAnalysis<TargetLibraryInfo>();
2943 MinimizeSize = F.getAttributes().hasAttribute(AttributeSet::FunctionIndex,
2944 Attribute::MinSize);
2946 /// Builder - This is an IRBuilder that automatically inserts new
2947 /// instructions into the worklist when they are created.
2948 IRBuilder<true, TargetFolder, InstCombineIRInserter>
2949 TheBuilder(F.getContext(), TargetFolder(DL),
2950 InstCombineIRInserter(Worklist, AT));
2951 Builder = &TheBuilder;
2953 InstCombinerLibCallSimplifier TheSimplifier(DL, TLI, this);
2954 Simplifier = &TheSimplifier;
2956 bool EverMadeChange = false;
2958 // Lower dbg.declare intrinsics otherwise their value may be clobbered
2960 EverMadeChange = LowerDbgDeclare(F);
2962 // Iterate while there is work to do.
2963 unsigned Iteration = 0;
2964 while (DoOneIteration(F, Iteration++))
2965 EverMadeChange = true;
2968 return EverMadeChange;
2971 FunctionPass *llvm::createInstructionCombiningPass() {
2972 return new InstCombiner();