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/AssumptionCache.h"
43 #include "llvm/Analysis/CFG.h"
44 #include "llvm/Analysis/ConstantFolding.h"
45 #include "llvm/Analysis/InstructionSimplify.h"
46 #include "llvm/Analysis/LoopInfo.h"
47 #include "llvm/Analysis/MemoryBuiltins.h"
48 #include "llvm/Analysis/ValueTracking.h"
49 #include "llvm/IR/CFG.h"
50 #include "llvm/IR/DataLayout.h"
51 #include "llvm/IR/Dominators.h"
52 #include "llvm/IR/GetElementPtrTypeIterator.h"
53 #include "llvm/IR/IntrinsicInst.h"
54 #include "llvm/IR/PatternMatch.h"
55 #include "llvm/IR/ValueHandle.h"
56 #include "llvm/Support/CommandLine.h"
57 #include "llvm/Support/Debug.h"
58 #include "llvm/Analysis/TargetLibraryInfo.h"
59 #include "llvm/Transforms/Utils/Local.h"
63 using namespace llvm::PatternMatch;
65 #define DEBUG_TYPE "instcombine"
67 STATISTIC(NumCombined , "Number of insts combined");
68 STATISTIC(NumConstProp, "Number of constant folds");
69 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
70 STATISTIC(NumSunkInst , "Number of instructions sunk");
71 STATISTIC(NumExpand, "Number of expansions");
72 STATISTIC(NumFactor , "Number of factorizations");
73 STATISTIC(NumReassoc , "Number of reassociations");
75 Value *InstCombiner::EmitGEPOffset(User *GEP) {
76 return llvm::EmitGEPOffset(Builder, *getDataLayout(), GEP);
79 /// ShouldChangeType - Return true if it is desirable to convert a computation
80 /// from 'From' to 'To'. We don't want to convert from a legal to an illegal
81 /// type for example, or from a smaller to a larger illegal type.
82 bool InstCombiner::ShouldChangeType(Type *From, Type *To) const {
83 assert(From->isIntegerTy() && To->isIntegerTy());
85 // If we don't have DL, we don't know if the source/dest are legal.
86 if (!DL) return false;
88 unsigned FromWidth = From->getPrimitiveSizeInBits();
89 unsigned ToWidth = To->getPrimitiveSizeInBits();
90 bool FromLegal = DL->isLegalInteger(FromWidth);
91 bool ToLegal = DL->isLegalInteger(ToWidth);
93 // If this is a legal integer from type, and the result would be an illegal
94 // type, don't do the transformation.
95 if (FromLegal && !ToLegal)
98 // Otherwise, if both are illegal, do not increase the size of the result. We
99 // do allow things like i160 -> i64, but not i64 -> i160.
100 if (!FromLegal && !ToLegal && ToWidth > FromWidth)
106 // Return true, if No Signed Wrap should be maintained for I.
107 // The No Signed Wrap flag can be kept if the operation "B (I.getOpcode) C",
108 // where both B and C should be ConstantInts, results in a constant that does
109 // not overflow. This function only handles the Add and Sub opcodes. For
110 // all other opcodes, the function conservatively returns false.
111 static bool MaintainNoSignedWrap(BinaryOperator &I, Value *B, Value *C) {
112 OverflowingBinaryOperator *OBO = dyn_cast<OverflowingBinaryOperator>(&I);
113 if (!OBO || !OBO->hasNoSignedWrap()) {
117 // We reason about Add and Sub Only.
118 Instruction::BinaryOps Opcode = I.getOpcode();
119 if (Opcode != Instruction::Add &&
120 Opcode != Instruction::Sub) {
124 ConstantInt *CB = dyn_cast<ConstantInt>(B);
125 ConstantInt *CC = dyn_cast<ConstantInt>(C);
131 const APInt &BVal = CB->getValue();
132 const APInt &CVal = CC->getValue();
133 bool Overflow = false;
135 if (Opcode == Instruction::Add) {
136 BVal.sadd_ov(CVal, Overflow);
138 BVal.ssub_ov(CVal, Overflow);
144 /// Conservatively clears subclassOptionalData after a reassociation or
145 /// commutation. We preserve fast-math flags when applicable as they can be
147 static void ClearSubclassDataAfterReassociation(BinaryOperator &I) {
148 FPMathOperator *FPMO = dyn_cast<FPMathOperator>(&I);
150 I.clearSubclassOptionalData();
154 FastMathFlags FMF = I.getFastMathFlags();
155 I.clearSubclassOptionalData();
156 I.setFastMathFlags(FMF);
159 /// SimplifyAssociativeOrCommutative - This performs a few simplifications for
160 /// operators which are associative or commutative:
162 // Commutative operators:
164 // 1. Order operands such that they are listed from right (least complex) to
165 // left (most complex). This puts constants before unary operators before
168 // Associative operators:
170 // 2. Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
171 // 3. Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
173 // Associative and commutative operators:
175 // 4. Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
176 // 5. Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
177 // 6. Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
178 // if C1 and C2 are constants.
180 bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) {
181 Instruction::BinaryOps Opcode = I.getOpcode();
182 bool Changed = false;
185 // Order operands such that they are listed from right (least complex) to
186 // left (most complex). This puts constants before unary operators before
188 if (I.isCommutative() && getComplexity(I.getOperand(0)) <
189 getComplexity(I.getOperand(1)))
190 Changed = !I.swapOperands();
192 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
193 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
195 if (I.isAssociative()) {
196 // Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
197 if (Op0 && Op0->getOpcode() == Opcode) {
198 Value *A = Op0->getOperand(0);
199 Value *B = Op0->getOperand(1);
200 Value *C = I.getOperand(1);
202 // Does "B op C" simplify?
203 if (Value *V = SimplifyBinOp(Opcode, B, C, DL)) {
204 // It simplifies to V. Form "A op V".
207 // Conservatively clear the optional flags, since they may not be
208 // preserved by the reassociation.
209 if (MaintainNoSignedWrap(I, B, C) &&
210 (!Op0 || (isa<BinaryOperator>(Op0) && Op0->hasNoSignedWrap()))) {
211 // Note: this is only valid because SimplifyBinOp doesn't look at
212 // the operands to Op0.
213 I.clearSubclassOptionalData();
214 I.setHasNoSignedWrap(true);
216 ClearSubclassDataAfterReassociation(I);
225 // Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
226 if (Op1 && Op1->getOpcode() == Opcode) {
227 Value *A = I.getOperand(0);
228 Value *B = Op1->getOperand(0);
229 Value *C = Op1->getOperand(1);
231 // Does "A op B" simplify?
232 if (Value *V = SimplifyBinOp(Opcode, A, B, DL)) {
233 // It simplifies to V. Form "V op C".
236 // Conservatively clear the optional flags, since they may not be
237 // preserved by the reassociation.
238 ClearSubclassDataAfterReassociation(I);
246 if (I.isAssociative() && I.isCommutative()) {
247 // Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
248 if (Op0 && Op0->getOpcode() == Opcode) {
249 Value *A = Op0->getOperand(0);
250 Value *B = Op0->getOperand(1);
251 Value *C = I.getOperand(1);
253 // Does "C op A" simplify?
254 if (Value *V = SimplifyBinOp(Opcode, C, A, DL)) {
255 // It simplifies to V. Form "V op B".
258 // Conservatively clear the optional flags, since they may not be
259 // preserved by the reassociation.
260 ClearSubclassDataAfterReassociation(I);
267 // Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
268 if (Op1 && Op1->getOpcode() == Opcode) {
269 Value *A = I.getOperand(0);
270 Value *B = Op1->getOperand(0);
271 Value *C = Op1->getOperand(1);
273 // Does "C op A" simplify?
274 if (Value *V = SimplifyBinOp(Opcode, C, A, DL)) {
275 // It simplifies to V. Form "B op V".
278 // Conservatively clear the optional flags, since they may not be
279 // preserved by the reassociation.
280 ClearSubclassDataAfterReassociation(I);
287 // Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
288 // if C1 and C2 are constants.
290 Op0->getOpcode() == Opcode && Op1->getOpcode() == Opcode &&
291 isa<Constant>(Op0->getOperand(1)) &&
292 isa<Constant>(Op1->getOperand(1)) &&
293 Op0->hasOneUse() && Op1->hasOneUse()) {
294 Value *A = Op0->getOperand(0);
295 Constant *C1 = cast<Constant>(Op0->getOperand(1));
296 Value *B = Op1->getOperand(0);
297 Constant *C2 = cast<Constant>(Op1->getOperand(1));
299 Constant *Folded = ConstantExpr::get(Opcode, C1, C2);
300 BinaryOperator *New = BinaryOperator::Create(Opcode, A, B);
301 if (isa<FPMathOperator>(New)) {
302 FastMathFlags Flags = I.getFastMathFlags();
303 Flags &= Op0->getFastMathFlags();
304 Flags &= Op1->getFastMathFlags();
305 New->setFastMathFlags(Flags);
307 InsertNewInstWith(New, I);
309 I.setOperand(0, New);
310 I.setOperand(1, Folded);
311 // Conservatively clear the optional flags, since they may not be
312 // preserved by the reassociation.
313 ClearSubclassDataAfterReassociation(I);
320 // No further simplifications.
325 /// LeftDistributesOverRight - Whether "X LOp (Y ROp Z)" is always equal to
326 /// "(X LOp Y) ROp (X LOp Z)".
327 static bool LeftDistributesOverRight(Instruction::BinaryOps LOp,
328 Instruction::BinaryOps ROp) {
333 case Instruction::And:
334 // And distributes over Or and Xor.
338 case Instruction::Or:
339 case Instruction::Xor:
343 case Instruction::Mul:
344 // Multiplication distributes over addition and subtraction.
348 case Instruction::Add:
349 case Instruction::Sub:
353 case Instruction::Or:
354 // Or distributes over And.
358 case Instruction::And:
364 /// RightDistributesOverLeft - Whether "(X LOp Y) ROp Z" is always equal to
365 /// "(X ROp Z) LOp (Y ROp Z)".
366 static bool RightDistributesOverLeft(Instruction::BinaryOps LOp,
367 Instruction::BinaryOps ROp) {
368 if (Instruction::isCommutative(ROp))
369 return LeftDistributesOverRight(ROp, LOp);
374 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
375 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
376 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
377 case Instruction::And:
378 case Instruction::Or:
379 case Instruction::Xor:
383 case Instruction::Shl:
384 case Instruction::LShr:
385 case Instruction::AShr:
389 // TODO: It would be nice to handle division, aka "(X + Y)/Z = X/Z + Y/Z",
390 // but this requires knowing that the addition does not overflow and other
395 /// This function returns identity value for given opcode, which can be used to
396 /// factor patterns like (X * 2) + X ==> (X * 2) + (X * 1) ==> X * (2 + 1).
397 static Value *getIdentityValue(Instruction::BinaryOps OpCode, Value *V) {
398 if (isa<Constant>(V))
401 if (OpCode == Instruction::Mul)
402 return ConstantInt::get(V->getType(), 1);
404 // TODO: We can handle other cases e.g. Instruction::And, Instruction::Or etc.
409 /// This function factors binary ops which can be combined using distributive
410 /// laws. This function tries to transform 'Op' based TopLevelOpcode to enable
411 /// factorization e.g for ADD(SHL(X , 2), MUL(X, 5)), When this function called
412 /// with TopLevelOpcode == Instruction::Add and Op = SHL(X, 2), transforms
413 /// SHL(X, 2) to MUL(X, 4) i.e. returns Instruction::Mul with LHS set to 'X' and
415 static Instruction::BinaryOps
416 getBinOpsForFactorization(Instruction::BinaryOps TopLevelOpcode,
417 BinaryOperator *Op, Value *&LHS, Value *&RHS) {
419 return Instruction::BinaryOpsEnd;
421 LHS = Op->getOperand(0);
422 RHS = Op->getOperand(1);
424 switch (TopLevelOpcode) {
426 return Op->getOpcode();
428 case Instruction::Add:
429 case Instruction::Sub:
430 if (Op->getOpcode() == Instruction::Shl) {
431 if (Constant *CST = dyn_cast<Constant>(Op->getOperand(1))) {
432 // The multiplier is really 1 << CST.
433 RHS = ConstantExpr::getShl(ConstantInt::get(Op->getType(), 1), CST);
434 return Instruction::Mul;
437 return Op->getOpcode();
440 // TODO: We can add other conversions e.g. shr => div etc.
443 /// This tries to simplify binary operations by factorizing out common terms
444 /// (e. g. "(A*B)+(A*C)" -> "A*(B+C)").
445 static Value *tryFactorization(InstCombiner::BuilderTy *Builder,
446 const DataLayout *DL, BinaryOperator &I,
447 Instruction::BinaryOps InnerOpcode, Value *A,
448 Value *B, Value *C, Value *D) {
450 // If any of A, B, C, D are null, we can not factor I, return early.
451 // Checking A and C should be enough.
452 if (!A || !C || !B || !D)
455 Value *SimplifiedInst = nullptr;
456 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
457 Instruction::BinaryOps TopLevelOpcode = I.getOpcode();
459 // Does "X op' Y" always equal "Y op' X"?
460 bool InnerCommutative = Instruction::isCommutative(InnerOpcode);
462 // Does "X op' (Y op Z)" always equal "(X op' Y) op (X op' Z)"?
463 if (LeftDistributesOverRight(InnerOpcode, TopLevelOpcode))
464 // Does the instruction have the form "(A op' B) op (A op' D)" or, in the
465 // commutative case, "(A op' B) op (C op' A)"?
466 if (A == C || (InnerCommutative && A == D)) {
469 // Consider forming "A op' (B op D)".
470 // If "B op D" simplifies then it can be formed with no cost.
471 Value *V = SimplifyBinOp(TopLevelOpcode, B, D, DL);
472 // If "B op D" doesn't simplify then only go on if both of the existing
473 // operations "A op' B" and "C op' D" will be zapped as no longer used.
474 if (!V && LHS->hasOneUse() && RHS->hasOneUse())
475 V = Builder->CreateBinOp(TopLevelOpcode, B, D, RHS->getName());
477 SimplifiedInst = Builder->CreateBinOp(InnerOpcode, A, V);
481 // Does "(X op Y) op' Z" always equal "(X op' Z) op (Y op' Z)"?
482 if (!SimplifiedInst && RightDistributesOverLeft(TopLevelOpcode, InnerOpcode))
483 // Does the instruction have the form "(A op' B) op (C op' B)" or, in the
484 // commutative case, "(A op' B) op (B op' D)"?
485 if (B == D || (InnerCommutative && B == C)) {
488 // Consider forming "(A op C) op' B".
489 // If "A op C" simplifies then it can be formed with no cost.
490 Value *V = SimplifyBinOp(TopLevelOpcode, A, C, DL);
492 // If "A op C" doesn't simplify then only go on if both of the existing
493 // operations "A op' B" and "C op' D" will be zapped as no longer used.
494 if (!V && LHS->hasOneUse() && RHS->hasOneUse())
495 V = Builder->CreateBinOp(TopLevelOpcode, A, C, LHS->getName());
497 SimplifiedInst = Builder->CreateBinOp(InnerOpcode, V, B);
501 if (SimplifiedInst) {
503 SimplifiedInst->takeName(&I);
505 // Check if we can add NSW flag to SimplifiedInst. If so, set NSW flag.
506 // TODO: Check for NUW.
507 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(SimplifiedInst)) {
508 if (isa<OverflowingBinaryOperator>(SimplifiedInst)) {
510 if (isa<OverflowingBinaryOperator>(&I))
511 HasNSW = I.hasNoSignedWrap();
513 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS))
514 if (isa<OverflowingBinaryOperator>(Op0))
515 HasNSW &= Op0->hasNoSignedWrap();
517 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS))
518 if (isa<OverflowingBinaryOperator>(Op1))
519 HasNSW &= Op1->hasNoSignedWrap();
520 BO->setHasNoSignedWrap(HasNSW);
524 return SimplifiedInst;
527 /// SimplifyUsingDistributiveLaws - This tries to simplify binary operations
528 /// which some other binary operation distributes over either by factorizing
529 /// out common terms (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this
530 /// results in simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is
531 /// a win). Returns the simplified value, or null if it didn't simplify.
532 Value *InstCombiner::SimplifyUsingDistributiveLaws(BinaryOperator &I) {
533 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
534 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
535 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
538 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
539 auto TopLevelOpcode = I.getOpcode();
540 auto LHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op0, A, B);
541 auto RHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op1, C, D);
543 // The instruction has the form "(A op' B) op (C op' D)". Try to factorize
545 if (LHSOpcode == RHSOpcode) {
546 if (Value *V = tryFactorization(Builder, DL, I, LHSOpcode, A, B, C, D))
550 // The instruction has the form "(A op' B) op (C)". Try to factorize common
552 if (Value *V = tryFactorization(Builder, DL, I, LHSOpcode, A, B, RHS,
553 getIdentityValue(LHSOpcode, RHS)))
556 // The instruction has the form "(B) op (C op' D)". Try to factorize common
558 if (Value *V = tryFactorization(Builder, DL, I, RHSOpcode, LHS,
559 getIdentityValue(RHSOpcode, LHS), C, D))
563 if (Op0 && RightDistributesOverLeft(Op0->getOpcode(), TopLevelOpcode)) {
564 // The instruction has the form "(A op' B) op C". See if expanding it out
565 // to "(A op C) op' (B op C)" results in simplifications.
566 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
567 Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op'
569 // Do "A op C" and "B op C" both simplify?
570 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, C, DL))
571 if (Value *R = SimplifyBinOp(TopLevelOpcode, B, C, DL)) {
572 // They do! Return "L op' R".
574 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
575 if ((L == A && R == B) ||
576 (Instruction::isCommutative(InnerOpcode) && L == B && R == A))
578 // Otherwise return "L op' R" if it simplifies.
579 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, DL))
581 // Otherwise, create a new instruction.
582 C = Builder->CreateBinOp(InnerOpcode, L, R);
588 if (Op1 && LeftDistributesOverRight(TopLevelOpcode, Op1->getOpcode())) {
589 // The instruction has the form "A op (B op' C)". See if expanding it out
590 // to "(A op B) op' (A op C)" results in simplifications.
591 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
592 Instruction::BinaryOps InnerOpcode = Op1->getOpcode(); // op'
594 // Do "A op B" and "A op C" both simplify?
595 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, B, DL))
596 if (Value *R = SimplifyBinOp(TopLevelOpcode, A, C, DL)) {
597 // They do! Return "L op' R".
599 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
600 if ((L == B && R == C) ||
601 (Instruction::isCommutative(InnerOpcode) && L == C && R == B))
603 // Otherwise return "L op' R" if it simplifies.
604 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, DL))
606 // Otherwise, create a new instruction.
607 A = Builder->CreateBinOp(InnerOpcode, L, R);
616 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
617 // if the LHS is a constant zero (which is the 'negate' form).
619 Value *InstCombiner::dyn_castNegVal(Value *V) const {
620 if (BinaryOperator::isNeg(V))
621 return BinaryOperator::getNegArgument(V);
623 // Constants can be considered to be negated values if they can be folded.
624 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
625 return ConstantExpr::getNeg(C);
627 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
628 if (C->getType()->getElementType()->isIntegerTy())
629 return ConstantExpr::getNeg(C);
634 // dyn_castFNegVal - Given a 'fsub' instruction, return the RHS of the
635 // instruction if the LHS is a constant negative zero (which is the 'negate'
638 Value *InstCombiner::dyn_castFNegVal(Value *V, bool IgnoreZeroSign) const {
639 if (BinaryOperator::isFNeg(V, IgnoreZeroSign))
640 return BinaryOperator::getFNegArgument(V);
642 // Constants can be considered to be negated values if they can be folded.
643 if (ConstantFP *C = dyn_cast<ConstantFP>(V))
644 return ConstantExpr::getFNeg(C);
646 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
647 if (C->getType()->getElementType()->isFloatingPointTy())
648 return ConstantExpr::getFNeg(C);
653 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
655 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
656 return IC->Builder->CreateCast(CI->getOpcode(), SO, I.getType());
659 // Figure out if the constant is the left or the right argument.
660 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
661 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
663 if (Constant *SOC = dyn_cast<Constant>(SO)) {
665 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
666 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
669 Value *Op0 = SO, *Op1 = ConstOperand;
673 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I)) {
674 Value *RI = IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1,
675 SO->getName()+".op");
676 Instruction *FPInst = dyn_cast<Instruction>(RI);
677 if (FPInst && isa<FPMathOperator>(FPInst))
678 FPInst->copyFastMathFlags(BO);
681 if (ICmpInst *CI = dyn_cast<ICmpInst>(&I))
682 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
683 SO->getName()+".cmp");
684 if (FCmpInst *CI = dyn_cast<FCmpInst>(&I))
685 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
686 SO->getName()+".cmp");
687 llvm_unreachable("Unknown binary instruction type!");
690 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
691 // constant as the other operand, try to fold the binary operator into the
692 // select arguments. This also works for Cast instructions, which obviously do
693 // not have a second operand.
694 Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) {
695 // Don't modify shared select instructions
696 if (!SI->hasOneUse()) return nullptr;
697 Value *TV = SI->getOperand(1);
698 Value *FV = SI->getOperand(2);
700 if (isa<Constant>(TV) || isa<Constant>(FV)) {
701 // Bool selects with constant operands can be folded to logical ops.
702 if (SI->getType()->isIntegerTy(1)) return nullptr;
704 // If it's a bitcast involving vectors, make sure it has the same number of
705 // elements on both sides.
706 if (BitCastInst *BC = dyn_cast<BitCastInst>(&Op)) {
707 VectorType *DestTy = dyn_cast<VectorType>(BC->getDestTy());
708 VectorType *SrcTy = dyn_cast<VectorType>(BC->getSrcTy());
710 // Verify that either both or neither are vectors.
711 if ((SrcTy == nullptr) != (DestTy == nullptr)) return nullptr;
712 // If vectors, verify that they have the same number of elements.
713 if (SrcTy && SrcTy->getNumElements() != DestTy->getNumElements())
717 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, this);
718 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, this);
720 return SelectInst::Create(SI->getCondition(),
721 SelectTrueVal, SelectFalseVal);
727 /// FoldOpIntoPhi - Given a binary operator, cast instruction, or select which
728 /// has a PHI node as operand #0, see if we can fold the instruction into the
729 /// PHI (which is only possible if all operands to the PHI are constants).
731 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
732 PHINode *PN = cast<PHINode>(I.getOperand(0));
733 unsigned NumPHIValues = PN->getNumIncomingValues();
734 if (NumPHIValues == 0)
737 // We normally only transform phis with a single use. However, if a PHI has
738 // multiple uses and they are all the same operation, we can fold *all* of the
739 // uses into the PHI.
740 if (!PN->hasOneUse()) {
741 // Walk the use list for the instruction, comparing them to I.
742 for (User *U : PN->users()) {
743 Instruction *UI = cast<Instruction>(U);
744 if (UI != &I && !I.isIdenticalTo(UI))
747 // Otherwise, we can replace *all* users with the new PHI we form.
750 // Check to see if all of the operands of the PHI are simple constants
751 // (constantint/constantfp/undef). If there is one non-constant value,
752 // remember the BB it is in. If there is more than one or if *it* is a PHI,
753 // bail out. We don't do arbitrary constant expressions here because moving
754 // their computation can be expensive without a cost model.
755 BasicBlock *NonConstBB = nullptr;
756 for (unsigned i = 0; i != NumPHIValues; ++i) {
757 Value *InVal = PN->getIncomingValue(i);
758 if (isa<Constant>(InVal) && !isa<ConstantExpr>(InVal))
761 if (isa<PHINode>(InVal)) return nullptr; // Itself a phi.
762 if (NonConstBB) return nullptr; // More than one non-const value.
764 NonConstBB = PN->getIncomingBlock(i);
766 // If the InVal is an invoke at the end of the pred block, then we can't
767 // insert a computation after it without breaking the edge.
768 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal))
769 if (II->getParent() == NonConstBB)
772 // If the incoming non-constant value is in I's block, we will remove one
773 // instruction, but insert another equivalent one, leading to infinite
775 if (isPotentiallyReachable(I.getParent(), NonConstBB, DT, LI))
779 // If there is exactly one non-constant value, we can insert a copy of the
780 // operation in that block. However, if this is a critical edge, we would be
781 // inserting the computation on some other paths (e.g. inside a loop). Only
782 // do this if the pred block is unconditionally branching into the phi block.
783 if (NonConstBB != nullptr) {
784 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
785 if (!BI || !BI->isUnconditional()) return nullptr;
788 // Okay, we can do the transformation: create the new PHI node.
789 PHINode *NewPN = PHINode::Create(I.getType(), PN->getNumIncomingValues());
790 InsertNewInstBefore(NewPN, *PN);
793 // If we are going to have to insert a new computation, do so right before the
794 // predecessors terminator.
796 Builder->SetInsertPoint(NonConstBB->getTerminator());
798 // Next, add all of the operands to the PHI.
799 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) {
800 // We only currently try to fold the condition of a select when it is a phi,
801 // not the true/false values.
802 Value *TrueV = SI->getTrueValue();
803 Value *FalseV = SI->getFalseValue();
804 BasicBlock *PhiTransBB = PN->getParent();
805 for (unsigned i = 0; i != NumPHIValues; ++i) {
806 BasicBlock *ThisBB = PN->getIncomingBlock(i);
807 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB);
808 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB);
809 Value *InV = nullptr;
810 // Beware of ConstantExpr: it may eventually evaluate to getNullValue,
811 // even if currently isNullValue gives false.
812 Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i));
813 if (InC && !isa<ConstantExpr>(InC))
814 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred;
816 InV = Builder->CreateSelect(PN->getIncomingValue(i),
817 TrueVInPred, FalseVInPred, "phitmp");
818 NewPN->addIncoming(InV, ThisBB);
820 } else if (CmpInst *CI = dyn_cast<CmpInst>(&I)) {
821 Constant *C = cast<Constant>(I.getOperand(1));
822 for (unsigned i = 0; i != NumPHIValues; ++i) {
823 Value *InV = nullptr;
824 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
825 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
826 else if (isa<ICmpInst>(CI))
827 InV = Builder->CreateICmp(CI->getPredicate(), PN->getIncomingValue(i),
830 InV = Builder->CreateFCmp(CI->getPredicate(), PN->getIncomingValue(i),
832 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
834 } else if (I.getNumOperands() == 2) {
835 Constant *C = cast<Constant>(I.getOperand(1));
836 for (unsigned i = 0; i != NumPHIValues; ++i) {
837 Value *InV = nullptr;
838 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
839 InV = ConstantExpr::get(I.getOpcode(), InC, C);
841 InV = Builder->CreateBinOp(cast<BinaryOperator>(I).getOpcode(),
842 PN->getIncomingValue(i), C, "phitmp");
843 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
846 CastInst *CI = cast<CastInst>(&I);
847 Type *RetTy = CI->getType();
848 for (unsigned i = 0; i != NumPHIValues; ++i) {
850 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
851 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
853 InV = Builder->CreateCast(CI->getOpcode(),
854 PN->getIncomingValue(i), I.getType(), "phitmp");
855 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
859 for (auto UI = PN->user_begin(), E = PN->user_end(); UI != E;) {
860 Instruction *User = cast<Instruction>(*UI++);
861 if (User == &I) continue;
862 ReplaceInstUsesWith(*User, NewPN);
863 EraseInstFromFunction(*User);
865 return ReplaceInstUsesWith(I, NewPN);
868 /// FindElementAtOffset - Given a pointer type and a constant offset, determine
869 /// whether or not there is a sequence of GEP indices into the pointed type that
870 /// will land us at the specified offset. If so, fill them into NewIndices and
871 /// return the resultant element type, otherwise return null.
872 Type *InstCombiner::FindElementAtOffset(Type *PtrTy, int64_t Offset,
873 SmallVectorImpl<Value*> &NewIndices) {
874 assert(PtrTy->isPtrOrPtrVectorTy());
879 Type *Ty = PtrTy->getPointerElementType();
883 // Start with the index over the outer type. Note that the type size
884 // might be zero (even if the offset isn't zero) if the indexed type
885 // is something like [0 x {int, int}]
886 Type *IntPtrTy = DL->getIntPtrType(PtrTy);
887 int64_t FirstIdx = 0;
888 if (int64_t TySize = DL->getTypeAllocSize(Ty)) {
889 FirstIdx = Offset/TySize;
890 Offset -= FirstIdx*TySize;
892 // Handle hosts where % returns negative instead of values [0..TySize).
898 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
901 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
903 // Index into the types. If we fail, set OrigBase to null.
905 // Indexing into tail padding between struct/array elements.
906 if (uint64_t(Offset*8) >= DL->getTypeSizeInBits(Ty))
909 if (StructType *STy = dyn_cast<StructType>(Ty)) {
910 const StructLayout *SL = DL->getStructLayout(STy);
911 assert(Offset < (int64_t)SL->getSizeInBytes() &&
912 "Offset must stay within the indexed type");
914 unsigned Elt = SL->getElementContainingOffset(Offset);
915 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
918 Offset -= SL->getElementOffset(Elt);
919 Ty = STy->getElementType(Elt);
920 } else if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
921 uint64_t EltSize = DL->getTypeAllocSize(AT->getElementType());
922 assert(EltSize && "Cannot index into a zero-sized array");
923 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
925 Ty = AT->getElementType();
927 // Otherwise, we can't index into the middle of this atomic type, bail.
935 static bool shouldMergeGEPs(GEPOperator &GEP, GEPOperator &Src) {
936 // If this GEP has only 0 indices, it is the same pointer as
937 // Src. If Src is not a trivial GEP too, don't combine
939 if (GEP.hasAllZeroIndices() && !Src.hasAllZeroIndices() &&
945 /// Descale - Return a value X such that Val = X * Scale, or null if none. If
946 /// the multiplication is known not to overflow then NoSignedWrap is set.
947 Value *InstCombiner::Descale(Value *Val, APInt Scale, bool &NoSignedWrap) {
948 assert(isa<IntegerType>(Val->getType()) && "Can only descale integers!");
949 assert(cast<IntegerType>(Val->getType())->getBitWidth() ==
950 Scale.getBitWidth() && "Scale not compatible with value!");
952 // If Val is zero or Scale is one then Val = Val * Scale.
953 if (match(Val, m_Zero()) || Scale == 1) {
958 // If Scale is zero then it does not divide Val.
959 if (Scale.isMinValue())
962 // Look through chains of multiplications, searching for a constant that is
963 // divisible by Scale. For example, descaling X*(Y*(Z*4)) by a factor of 4
964 // will find the constant factor 4 and produce X*(Y*Z). Descaling X*(Y*8) by
965 // a factor of 4 will produce X*(Y*2). The principle of operation is to bore
968 // Val = M1 * X || Analysis starts here and works down
969 // M1 = M2 * Y || Doesn't descend into terms with more
970 // M2 = Z * 4 \/ than one use
972 // Then to modify a term at the bottom:
975 // M1 = Z * Y || Replaced M2 with Z
977 // Then to work back up correcting nsw flags.
979 // Op - the term we are currently analyzing. Starts at Val then drills down.
980 // Replaced with its descaled value before exiting from the drill down loop.
983 // Parent - initially null, but after drilling down notes where Op came from.
984 // In the example above, Parent is (Val, 0) when Op is M1, because M1 is the
985 // 0'th operand of Val.
986 std::pair<Instruction*, unsigned> Parent;
988 // RequireNoSignedWrap - Set if the transform requires a descaling at deeper
989 // levels that doesn't overflow.
990 bool RequireNoSignedWrap = false;
992 // logScale - log base 2 of the scale. Negative if not a power of 2.
993 int32_t logScale = Scale.exactLogBase2();
995 for (;; Op = Parent.first->getOperand(Parent.second)) { // Drill down
997 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
998 // If Op is a constant divisible by Scale then descale to the quotient.
999 APInt Quotient(Scale), Remainder(Scale); // Init ensures right bitwidth.
1000 APInt::sdivrem(CI->getValue(), Scale, Quotient, Remainder);
1001 if (!Remainder.isMinValue())
1002 // Not divisible by Scale.
1004 // Replace with the quotient in the parent.
1005 Op = ConstantInt::get(CI->getType(), Quotient);
1006 NoSignedWrap = true;
1010 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op)) {
1012 if (BO->getOpcode() == Instruction::Mul) {
1014 NoSignedWrap = BO->hasNoSignedWrap();
1015 if (RequireNoSignedWrap && !NoSignedWrap)
1018 // There are three cases for multiplication: multiplication by exactly
1019 // the scale, multiplication by a constant different to the scale, and
1020 // multiplication by something else.
1021 Value *LHS = BO->getOperand(0);
1022 Value *RHS = BO->getOperand(1);
1024 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
1025 // Multiplication by a constant.
1026 if (CI->getValue() == Scale) {
1027 // Multiplication by exactly the scale, replace the multiplication
1028 // by its left-hand side in the parent.
1033 // Otherwise drill down into the constant.
1034 if (!Op->hasOneUse())
1037 Parent = std::make_pair(BO, 1);
1041 // Multiplication by something else. Drill down into the left-hand side
1042 // since that's where the reassociate pass puts the good stuff.
1043 if (!Op->hasOneUse())
1046 Parent = std::make_pair(BO, 0);
1050 if (logScale > 0 && BO->getOpcode() == Instruction::Shl &&
1051 isa<ConstantInt>(BO->getOperand(1))) {
1052 // Multiplication by a power of 2.
1053 NoSignedWrap = BO->hasNoSignedWrap();
1054 if (RequireNoSignedWrap && !NoSignedWrap)
1057 Value *LHS = BO->getOperand(0);
1058 int32_t Amt = cast<ConstantInt>(BO->getOperand(1))->
1059 getLimitedValue(Scale.getBitWidth());
1062 if (Amt == logScale) {
1063 // Multiplication by exactly the scale, replace the multiplication
1064 // by its left-hand side in the parent.
1068 if (Amt < logScale || !Op->hasOneUse())
1071 // Multiplication by more than the scale. Reduce the multiplying amount
1072 // by the scale in the parent.
1073 Parent = std::make_pair(BO, 1);
1074 Op = ConstantInt::get(BO->getType(), Amt - logScale);
1079 if (!Op->hasOneUse())
1082 if (CastInst *Cast = dyn_cast<CastInst>(Op)) {
1083 if (Cast->getOpcode() == Instruction::SExt) {
1084 // Op is sign-extended from a smaller type, descale in the smaller type.
1085 unsigned SmallSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1086 APInt SmallScale = Scale.trunc(SmallSize);
1087 // Suppose Op = sext X, and we descale X as Y * SmallScale. We want to
1088 // descale Op as (sext Y) * Scale. In order to have
1089 // sext (Y * SmallScale) = (sext Y) * Scale
1090 // some conditions need to hold however: SmallScale must sign-extend to
1091 // Scale and the multiplication Y * SmallScale should not overflow.
1092 if (SmallScale.sext(Scale.getBitWidth()) != Scale)
1093 // SmallScale does not sign-extend to Scale.
1095 assert(SmallScale.exactLogBase2() == logScale);
1096 // Require that Y * SmallScale must not overflow.
1097 RequireNoSignedWrap = true;
1099 // Drill down through the cast.
1100 Parent = std::make_pair(Cast, 0);
1105 if (Cast->getOpcode() == Instruction::Trunc) {
1106 // Op is truncated from a larger type, descale in the larger type.
1107 // Suppose Op = trunc X, and we descale X as Y * sext Scale. Then
1108 // trunc (Y * sext Scale) = (trunc Y) * Scale
1109 // always holds. However (trunc Y) * Scale may overflow even if
1110 // trunc (Y * sext Scale) does not, so nsw flags need to be cleared
1111 // from this point up in the expression (see later).
1112 if (RequireNoSignedWrap)
1115 // Drill down through the cast.
1116 unsigned LargeSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1117 Parent = std::make_pair(Cast, 0);
1118 Scale = Scale.sext(LargeSize);
1119 if (logScale + 1 == (int32_t)Cast->getType()->getPrimitiveSizeInBits())
1121 assert(Scale.exactLogBase2() == logScale);
1126 // Unsupported expression, bail out.
1130 // If Op is zero then Val = Op * Scale.
1131 if (match(Op, m_Zero())) {
1132 NoSignedWrap = true;
1136 // We know that we can successfully descale, so from here on we can safely
1137 // modify the IR. Op holds the descaled version of the deepest term in the
1138 // expression. NoSignedWrap is 'true' if multiplying Op by Scale is known
1142 // The expression only had one term.
1145 // Rewrite the parent using the descaled version of its operand.
1146 assert(Parent.first->hasOneUse() && "Drilled down when more than one use!");
1147 assert(Op != Parent.first->getOperand(Parent.second) &&
1148 "Descaling was a no-op?");
1149 Parent.first->setOperand(Parent.second, Op);
1150 Worklist.Add(Parent.first);
1152 // Now work back up the expression correcting nsw flags. The logic is based
1153 // on the following observation: if X * Y is known not to overflow as a signed
1154 // multiplication, and Y is replaced by a value Z with smaller absolute value,
1155 // then X * Z will not overflow as a signed multiplication either. As we work
1156 // our way up, having NoSignedWrap 'true' means that the descaled value at the
1157 // current level has strictly smaller absolute value than the original.
1158 Instruction *Ancestor = Parent.first;
1160 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Ancestor)) {
1161 // If the multiplication wasn't nsw then we can't say anything about the
1162 // value of the descaled multiplication, and we have to clear nsw flags
1163 // from this point on up.
1164 bool OpNoSignedWrap = BO->hasNoSignedWrap();
1165 NoSignedWrap &= OpNoSignedWrap;
1166 if (NoSignedWrap != OpNoSignedWrap) {
1167 BO->setHasNoSignedWrap(NoSignedWrap);
1168 Worklist.Add(Ancestor);
1170 } else if (Ancestor->getOpcode() == Instruction::Trunc) {
1171 // The fact that the descaled input to the trunc has smaller absolute
1172 // value than the original input doesn't tell us anything useful about
1173 // the absolute values of the truncations.
1174 NoSignedWrap = false;
1176 assert((Ancestor->getOpcode() != Instruction::SExt || NoSignedWrap) &&
1177 "Failed to keep proper track of nsw flags while drilling down?");
1179 if (Ancestor == Val)
1180 // Got to the top, all done!
1183 // Move up one level in the expression.
1184 assert(Ancestor->hasOneUse() && "Drilled down when more than one use!");
1185 Ancestor = Ancestor->user_back();
1189 /// \brief Creates node of binary operation with the same attributes as the
1190 /// specified one but with other operands.
1191 static Value *CreateBinOpAsGiven(BinaryOperator &Inst, Value *LHS, Value *RHS,
1192 InstCombiner::BuilderTy *B) {
1193 Value *BORes = B->CreateBinOp(Inst.getOpcode(), LHS, RHS);
1194 if (BinaryOperator *NewBO = dyn_cast<BinaryOperator>(BORes)) {
1195 if (isa<OverflowingBinaryOperator>(NewBO)) {
1196 NewBO->setHasNoSignedWrap(Inst.hasNoSignedWrap());
1197 NewBO->setHasNoUnsignedWrap(Inst.hasNoUnsignedWrap());
1199 if (isa<PossiblyExactOperator>(NewBO))
1200 NewBO->setIsExact(Inst.isExact());
1205 /// \brief Makes transformation of binary operation specific for vector types.
1206 /// \param Inst Binary operator to transform.
1207 /// \return Pointer to node that must replace the original binary operator, or
1208 /// null pointer if no transformation was made.
1209 Value *InstCombiner::SimplifyVectorOp(BinaryOperator &Inst) {
1210 if (!Inst.getType()->isVectorTy()) return nullptr;
1212 // It may not be safe to reorder shuffles and things like div, urem, etc.
1213 // because we may trap when executing those ops on unknown vector elements.
1215 if (!isSafeToSpeculativelyExecute(&Inst, DL)) return nullptr;
1217 unsigned VWidth = cast<VectorType>(Inst.getType())->getNumElements();
1218 Value *LHS = Inst.getOperand(0), *RHS = Inst.getOperand(1);
1219 assert(cast<VectorType>(LHS->getType())->getNumElements() == VWidth);
1220 assert(cast<VectorType>(RHS->getType())->getNumElements() == VWidth);
1222 // If both arguments of binary operation are shuffles, which use the same
1223 // mask and shuffle within a single vector, it is worthwhile to move the
1224 // shuffle after binary operation:
1225 // Op(shuffle(v1, m), shuffle(v2, m)) -> shuffle(Op(v1, v2), m)
1226 if (isa<ShuffleVectorInst>(LHS) && isa<ShuffleVectorInst>(RHS)) {
1227 ShuffleVectorInst *LShuf = cast<ShuffleVectorInst>(LHS);
1228 ShuffleVectorInst *RShuf = cast<ShuffleVectorInst>(RHS);
1229 if (isa<UndefValue>(LShuf->getOperand(1)) &&
1230 isa<UndefValue>(RShuf->getOperand(1)) &&
1231 LShuf->getOperand(0)->getType() == RShuf->getOperand(0)->getType() &&
1232 LShuf->getMask() == RShuf->getMask()) {
1233 Value *NewBO = CreateBinOpAsGiven(Inst, LShuf->getOperand(0),
1234 RShuf->getOperand(0), Builder);
1235 Value *Res = Builder->CreateShuffleVector(NewBO,
1236 UndefValue::get(NewBO->getType()), LShuf->getMask());
1241 // If one argument is a shuffle within one vector, the other is a constant,
1242 // try moving the shuffle after the binary operation.
1243 ShuffleVectorInst *Shuffle = nullptr;
1244 Constant *C1 = nullptr;
1245 if (isa<ShuffleVectorInst>(LHS)) Shuffle = cast<ShuffleVectorInst>(LHS);
1246 if (isa<ShuffleVectorInst>(RHS)) Shuffle = cast<ShuffleVectorInst>(RHS);
1247 if (isa<Constant>(LHS)) C1 = cast<Constant>(LHS);
1248 if (isa<Constant>(RHS)) C1 = cast<Constant>(RHS);
1249 if (Shuffle && C1 &&
1250 (isa<ConstantVector>(C1) || isa<ConstantDataVector>(C1)) &&
1251 isa<UndefValue>(Shuffle->getOperand(1)) &&
1252 Shuffle->getType() == Shuffle->getOperand(0)->getType()) {
1253 SmallVector<int, 16> ShMask = Shuffle->getShuffleMask();
1254 // Find constant C2 that has property:
1255 // shuffle(C2, ShMask) = C1
1256 // If such constant does not exist (example: ShMask=<0,0> and C1=<1,2>)
1257 // reorder is not possible.
1258 SmallVector<Constant*, 16> C2M(VWidth,
1259 UndefValue::get(C1->getType()->getScalarType()));
1260 bool MayChange = true;
1261 for (unsigned I = 0; I < VWidth; ++I) {
1262 if (ShMask[I] >= 0) {
1263 assert(ShMask[I] < (int)VWidth);
1264 if (!isa<UndefValue>(C2M[ShMask[I]])) {
1268 C2M[ShMask[I]] = C1->getAggregateElement(I);
1272 Constant *C2 = ConstantVector::get(C2M);
1273 Value *NewLHS, *NewRHS;
1274 if (isa<Constant>(LHS)) {
1276 NewRHS = Shuffle->getOperand(0);
1278 NewLHS = Shuffle->getOperand(0);
1281 Value *NewBO = CreateBinOpAsGiven(Inst, NewLHS, NewRHS, Builder);
1282 Value *Res = Builder->CreateShuffleVector(NewBO,
1283 UndefValue::get(Inst.getType()), Shuffle->getMask());
1291 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
1292 SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end());
1294 if (Value *V = SimplifyGEPInst(Ops, DL, TLI, DT, AC))
1295 return ReplaceInstUsesWith(GEP, V);
1297 Value *PtrOp = GEP.getOperand(0);
1299 // Eliminate unneeded casts for indices, and replace indices which displace
1300 // by multiples of a zero size type with zero.
1302 bool MadeChange = false;
1303 Type *IntPtrTy = DL->getIntPtrType(GEP.getPointerOperandType());
1305 gep_type_iterator GTI = gep_type_begin(GEP);
1306 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end();
1307 I != E; ++I, ++GTI) {
1308 // Skip indices into struct types.
1309 SequentialType *SeqTy = dyn_cast<SequentialType>(*GTI);
1310 if (!SeqTy) continue;
1312 // If the element type has zero size then any index over it is equivalent
1313 // to an index of zero, so replace it with zero if it is not zero already.
1314 if (SeqTy->getElementType()->isSized() &&
1315 DL->getTypeAllocSize(SeqTy->getElementType()) == 0)
1316 if (!isa<Constant>(*I) || !cast<Constant>(*I)->isNullValue()) {
1317 *I = Constant::getNullValue(IntPtrTy);
1321 Type *IndexTy = (*I)->getType();
1322 if (IndexTy != IntPtrTy) {
1323 // If we are using a wider index than needed for this platform, shrink
1324 // it to what we need. If narrower, sign-extend it to what we need.
1325 // This explicit cast can make subsequent optimizations more obvious.
1326 *I = Builder->CreateIntCast(*I, IntPtrTy, true);
1330 if (MadeChange) return &GEP;
1333 // Check to see if the inputs to the PHI node are getelementptr instructions.
1334 if (PHINode *PN = dyn_cast<PHINode>(PtrOp)) {
1335 GetElementPtrInst *Op1 = dyn_cast<GetElementPtrInst>(PN->getOperand(0));
1341 for (auto I = PN->op_begin()+1, E = PN->op_end(); I !=E; ++I) {
1342 GetElementPtrInst *Op2 = dyn_cast<GetElementPtrInst>(*I);
1343 if (!Op2 || Op1->getNumOperands() != Op2->getNumOperands())
1346 // Keep track of the type as we walk the GEP.
1347 Type *CurTy = Op1->getOperand(0)->getType()->getScalarType();
1349 for (unsigned J = 0, F = Op1->getNumOperands(); J != F; ++J) {
1350 if (Op1->getOperand(J)->getType() != Op2->getOperand(J)->getType())
1353 if (Op1->getOperand(J) != Op2->getOperand(J)) {
1355 // We have not seen any differences yet in the GEPs feeding the
1356 // PHI yet, so we record this one if it is allowed to be a
1359 // The first two arguments can vary for any GEP, the rest have to be
1360 // static for struct slots
1361 if (J > 1 && CurTy->isStructTy())
1366 // The GEP is different by more than one input. While this could be
1367 // extended to support GEPs that vary by more than one variable it
1368 // doesn't make sense since it greatly increases the complexity and
1369 // would result in an R+R+R addressing mode which no backend
1370 // directly supports and would need to be broken into several
1371 // simpler instructions anyway.
1376 // Sink down a layer of the type for the next iteration.
1378 if (CompositeType *CT = dyn_cast<CompositeType>(CurTy)) {
1379 CurTy = CT->getTypeAtIndex(Op1->getOperand(J));
1387 GetElementPtrInst *NewGEP = cast<GetElementPtrInst>(Op1->clone());
1390 // All the GEPs feeding the PHI are identical. Clone one down into our
1391 // BB so that it can be merged with the current GEP.
1392 GEP.getParent()->getInstList().insert(GEP.getParent()->getFirstNonPHI(),
1395 // All the GEPs feeding the PHI differ at a single offset. Clone a GEP
1396 // into the current block so it can be merged, and create a new PHI to
1398 Instruction *InsertPt = Builder->GetInsertPoint();
1399 Builder->SetInsertPoint(PN);
1400 PHINode *NewPN = Builder->CreatePHI(Op1->getOperand(DI)->getType(),
1401 PN->getNumOperands());
1402 Builder->SetInsertPoint(InsertPt);
1404 for (auto &I : PN->operands())
1405 NewPN->addIncoming(cast<GEPOperator>(I)->getOperand(DI),
1406 PN->getIncomingBlock(I));
1408 NewGEP->setOperand(DI, NewPN);
1409 GEP.getParent()->getInstList().insert(GEP.getParent()->getFirstNonPHI(),
1411 NewGEP->setOperand(DI, NewPN);
1414 GEP.setOperand(0, NewGEP);
1418 // Combine Indices - If the source pointer to this getelementptr instruction
1419 // is a getelementptr instruction, combine the indices of the two
1420 // getelementptr instructions into a single instruction.
1422 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) {
1423 if (!shouldMergeGEPs(*cast<GEPOperator>(&GEP), *Src))
1426 // Note that if our source is a gep chain itself then we wait for that
1427 // chain to be resolved before we perform this transformation. This
1428 // avoids us creating a TON of code in some cases.
1429 if (GEPOperator *SrcGEP =
1430 dyn_cast<GEPOperator>(Src->getOperand(0)))
1431 if (SrcGEP->getNumOperands() == 2 && shouldMergeGEPs(*Src, *SrcGEP))
1432 return nullptr; // Wait until our source is folded to completion.
1434 SmallVector<Value*, 8> Indices;
1436 // Find out whether the last index in the source GEP is a sequential idx.
1437 bool EndsWithSequential = false;
1438 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src);
1440 EndsWithSequential = !(*I)->isStructTy();
1442 // Can we combine the two pointer arithmetics offsets?
1443 if (EndsWithSequential) {
1444 // Replace: gep (gep %P, long B), long A, ...
1445 // With: T = long A+B; gep %P, T, ...
1448 Value *SO1 = Src->getOperand(Src->getNumOperands()-1);
1449 Value *GO1 = GEP.getOperand(1);
1450 if (SO1 == Constant::getNullValue(SO1->getType())) {
1452 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
1455 // If they aren't the same type, then the input hasn't been processed
1456 // by the loop above yet (which canonicalizes sequential index types to
1457 // intptr_t). Just avoid transforming this until the input has been
1459 if (SO1->getType() != GO1->getType())
1461 Sum = Builder->CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
1464 // Update the GEP in place if possible.
1465 if (Src->getNumOperands() == 2) {
1466 GEP.setOperand(0, Src->getOperand(0));
1467 GEP.setOperand(1, Sum);
1470 Indices.append(Src->op_begin()+1, Src->op_end()-1);
1471 Indices.push_back(Sum);
1472 Indices.append(GEP.op_begin()+2, GEP.op_end());
1473 } else if (isa<Constant>(*GEP.idx_begin()) &&
1474 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
1475 Src->getNumOperands() != 1) {
1476 // Otherwise we can do the fold if the first index of the GEP is a zero
1477 Indices.append(Src->op_begin()+1, Src->op_end());
1478 Indices.append(GEP.idx_begin()+1, GEP.idx_end());
1481 if (!Indices.empty())
1482 return (GEP.isInBounds() && Src->isInBounds()) ?
1483 GetElementPtrInst::CreateInBounds(Src->getOperand(0), Indices,
1485 GetElementPtrInst::Create(Src->getOperand(0), Indices, GEP.getName());
1488 if (DL && GEP.getNumIndices() == 1) {
1489 unsigned AS = GEP.getPointerAddressSpace();
1490 if (GEP.getOperand(1)->getType()->getScalarSizeInBits() ==
1491 DL->getPointerSizeInBits(AS)) {
1492 Type *PtrTy = GEP.getPointerOperandType();
1493 Type *Ty = PtrTy->getPointerElementType();
1494 uint64_t TyAllocSize = DL->getTypeAllocSize(Ty);
1496 bool Matched = false;
1499 if (TyAllocSize == 1) {
1500 V = GEP.getOperand(1);
1502 } else if (match(GEP.getOperand(1),
1503 m_AShr(m_Value(V), m_ConstantInt(C)))) {
1504 if (TyAllocSize == 1ULL << C)
1506 } else if (match(GEP.getOperand(1),
1507 m_SDiv(m_Value(V), m_ConstantInt(C)))) {
1508 if (TyAllocSize == C)
1513 // Canonicalize (gep i8* X, -(ptrtoint Y))
1514 // to (inttoptr (sub (ptrtoint X), (ptrtoint Y)))
1515 // The GEP pattern is emitted by the SCEV expander for certain kinds of
1516 // pointer arithmetic.
1517 if (match(V, m_Neg(m_PtrToInt(m_Value())))) {
1518 Operator *Index = cast<Operator>(V);
1519 Value *PtrToInt = Builder->CreatePtrToInt(PtrOp, Index->getType());
1520 Value *NewSub = Builder->CreateSub(PtrToInt, Index->getOperand(1));
1521 return CastInst::Create(Instruction::IntToPtr, NewSub, GEP.getType());
1523 // Canonicalize (gep i8* X, (ptrtoint Y)-(ptrtoint X))
1526 if (match(V, m_Sub(m_PtrToInt(m_Value(Y)),
1527 m_PtrToInt(m_Specific(GEP.getOperand(0)))))) {
1528 return CastInst::CreatePointerBitCastOrAddrSpaceCast(Y,
1535 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0).
1536 Value *StrippedPtr = PtrOp->stripPointerCasts();
1537 PointerType *StrippedPtrTy = dyn_cast<PointerType>(StrippedPtr->getType());
1539 // We do not handle pointer-vector geps here.
1543 if (StrippedPtr != PtrOp) {
1544 bool HasZeroPointerIndex = false;
1545 if (ConstantInt *C = dyn_cast<ConstantInt>(GEP.getOperand(1)))
1546 HasZeroPointerIndex = C->isZero();
1548 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
1549 // into : GEP [10 x i8]* X, i32 0, ...
1551 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
1552 // into : GEP i8* X, ...
1554 // This occurs when the program declares an array extern like "int X[];"
1555 if (HasZeroPointerIndex) {
1556 PointerType *CPTy = cast<PointerType>(PtrOp->getType());
1557 if (ArrayType *CATy =
1558 dyn_cast<ArrayType>(CPTy->getElementType())) {
1559 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
1560 if (CATy->getElementType() == StrippedPtrTy->getElementType()) {
1561 // -> GEP i8* X, ...
1562 SmallVector<Value*, 8> Idx(GEP.idx_begin()+1, GEP.idx_end());
1563 GetElementPtrInst *Res =
1564 GetElementPtrInst::Create(StrippedPtr, Idx, GEP.getName());
1565 Res->setIsInBounds(GEP.isInBounds());
1566 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace())
1568 // Insert Res, and create an addrspacecast.
1570 // GEP (addrspacecast i8 addrspace(1)* X to [0 x i8]*), i32 0, ...
1572 // %0 = GEP i8 addrspace(1)* X, ...
1573 // addrspacecast i8 addrspace(1)* %0 to i8*
1574 return new AddrSpaceCastInst(Builder->Insert(Res), GEP.getType());
1577 if (ArrayType *XATy =
1578 dyn_cast<ArrayType>(StrippedPtrTy->getElementType())){
1579 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
1580 if (CATy->getElementType() == XATy->getElementType()) {
1581 // -> GEP [10 x i8]* X, i32 0, ...
1582 // At this point, we know that the cast source type is a pointer
1583 // to an array of the same type as the destination pointer
1584 // array. Because the array type is never stepped over (there
1585 // is a leading zero) we can fold the cast into this GEP.
1586 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace()) {
1587 GEP.setOperand(0, StrippedPtr);
1590 // Cannot replace the base pointer directly because StrippedPtr's
1591 // address space is different. Instead, create a new GEP followed by
1592 // an addrspacecast.
1594 // GEP (addrspacecast [10 x i8] addrspace(1)* X to [0 x i8]*),
1597 // %0 = GEP [10 x i8] addrspace(1)* X, ...
1598 // addrspacecast i8 addrspace(1)* %0 to i8*
1599 SmallVector<Value*, 8> Idx(GEP.idx_begin(), GEP.idx_end());
1600 Value *NewGEP = GEP.isInBounds() ?
1601 Builder->CreateInBoundsGEP(StrippedPtr, Idx, GEP.getName()) :
1602 Builder->CreateGEP(StrippedPtr, Idx, GEP.getName());
1603 return new AddrSpaceCastInst(NewGEP, GEP.getType());
1607 } else if (GEP.getNumOperands() == 2) {
1608 // Transform things like:
1609 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
1610 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
1611 Type *SrcElTy = StrippedPtrTy->getElementType();
1612 Type *ResElTy = PtrOp->getType()->getPointerElementType();
1613 if (DL && SrcElTy->isArrayTy() &&
1614 DL->getTypeAllocSize(SrcElTy->getArrayElementType()) ==
1615 DL->getTypeAllocSize(ResElTy)) {
1616 Type *IdxType = DL->getIntPtrType(GEP.getType());
1617 Value *Idx[2] = { Constant::getNullValue(IdxType), GEP.getOperand(1) };
1618 Value *NewGEP = GEP.isInBounds() ?
1619 Builder->CreateInBoundsGEP(StrippedPtr, Idx, GEP.getName()) :
1620 Builder->CreateGEP(StrippedPtr, Idx, GEP.getName());
1622 // V and GEP are both pointer types --> BitCast
1623 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
1627 // Transform things like:
1628 // %V = mul i64 %N, 4
1629 // %t = getelementptr i8* bitcast (i32* %arr to i8*), i32 %V
1630 // into: %t1 = getelementptr i32* %arr, i32 %N; bitcast
1631 if (DL && ResElTy->isSized() && SrcElTy->isSized()) {
1632 // Check that changing the type amounts to dividing the index by a scale
1634 uint64_t ResSize = DL->getTypeAllocSize(ResElTy);
1635 uint64_t SrcSize = DL->getTypeAllocSize(SrcElTy);
1636 if (ResSize && SrcSize % ResSize == 0) {
1637 Value *Idx = GEP.getOperand(1);
1638 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
1639 uint64_t Scale = SrcSize / ResSize;
1641 // Earlier transforms ensure that the index has type IntPtrType, which
1642 // considerably simplifies the logic by eliminating implicit casts.
1643 assert(Idx->getType() == DL->getIntPtrType(GEP.getType()) &&
1644 "Index not cast to pointer width?");
1647 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
1648 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
1649 // If the multiplication NewIdx * Scale may overflow then the new
1650 // GEP may not be "inbounds".
1651 Value *NewGEP = GEP.isInBounds() && NSW ?
1652 Builder->CreateInBoundsGEP(StrippedPtr, NewIdx, GEP.getName()) :
1653 Builder->CreateGEP(StrippedPtr, NewIdx, GEP.getName());
1655 // The NewGEP must be pointer typed, so must the old one -> BitCast
1656 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
1662 // Similarly, transform things like:
1663 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
1664 // (where tmp = 8*tmp2) into:
1665 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
1666 if (DL && ResElTy->isSized() && SrcElTy->isSized() &&
1667 SrcElTy->isArrayTy()) {
1668 // Check that changing to the array element type amounts to dividing the
1669 // index by a scale factor.
1670 uint64_t ResSize = DL->getTypeAllocSize(ResElTy);
1671 uint64_t ArrayEltSize
1672 = DL->getTypeAllocSize(SrcElTy->getArrayElementType());
1673 if (ResSize && ArrayEltSize % ResSize == 0) {
1674 Value *Idx = GEP.getOperand(1);
1675 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
1676 uint64_t Scale = ArrayEltSize / ResSize;
1678 // Earlier transforms ensure that the index has type IntPtrType, which
1679 // considerably simplifies the logic by eliminating implicit casts.
1680 assert(Idx->getType() == DL->getIntPtrType(GEP.getType()) &&
1681 "Index not cast to pointer width?");
1684 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
1685 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
1686 // If the multiplication NewIdx * Scale may overflow then the new
1687 // GEP may not be "inbounds".
1689 Constant::getNullValue(DL->getIntPtrType(GEP.getType())),
1693 Value *NewGEP = GEP.isInBounds() && NSW ?
1694 Builder->CreateInBoundsGEP(StrippedPtr, Off, GEP.getName()) :
1695 Builder->CreateGEP(StrippedPtr, Off, GEP.getName());
1696 // The NewGEP must be pointer typed, so must the old one -> BitCast
1697 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
1708 // addrspacecast between types is canonicalized as a bitcast, then an
1709 // addrspacecast. To take advantage of the below bitcast + struct GEP, look
1710 // through the addrspacecast.
1711 if (AddrSpaceCastInst *ASC = dyn_cast<AddrSpaceCastInst>(PtrOp)) {
1712 // X = bitcast A addrspace(1)* to B addrspace(1)*
1713 // Y = addrspacecast A addrspace(1)* to B addrspace(2)*
1714 // Z = gep Y, <...constant indices...>
1715 // Into an addrspacecasted GEP of the struct.
1716 if (BitCastInst *BC = dyn_cast<BitCastInst>(ASC->getOperand(0)))
1720 /// See if we can simplify:
1721 /// X = bitcast A* to B*
1722 /// Y = gep X, <...constant indices...>
1723 /// into a gep of the original struct. This is important for SROA and alias
1724 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
1725 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
1726 Value *Operand = BCI->getOperand(0);
1727 PointerType *OpType = cast<PointerType>(Operand->getType());
1728 unsigned OffsetBits = DL->getPointerTypeSizeInBits(GEP.getType());
1729 APInt Offset(OffsetBits, 0);
1730 if (!isa<BitCastInst>(Operand) &&
1731 GEP.accumulateConstantOffset(*DL, Offset)) {
1733 // If this GEP instruction doesn't move the pointer, just replace the GEP
1734 // with a bitcast of the real input to the dest type.
1736 // If the bitcast is of an allocation, and the allocation will be
1737 // converted to match the type of the cast, don't touch this.
1738 if (isa<AllocaInst>(Operand) || isAllocationFn(Operand, TLI)) {
1739 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
1740 if (Instruction *I = visitBitCast(*BCI)) {
1743 BCI->getParent()->getInstList().insert(BCI, I);
1744 ReplaceInstUsesWith(*BCI, I);
1750 if (Operand->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
1751 return new AddrSpaceCastInst(Operand, GEP.getType());
1752 return new BitCastInst(Operand, GEP.getType());
1755 // Otherwise, if the offset is non-zero, we need to find out if there is a
1756 // field at Offset in 'A's type. If so, we can pull the cast through the
1758 SmallVector<Value*, 8> NewIndices;
1759 if (FindElementAtOffset(OpType, Offset.getSExtValue(), NewIndices)) {
1760 Value *NGEP = GEP.isInBounds() ?
1761 Builder->CreateInBoundsGEP(Operand, NewIndices) :
1762 Builder->CreateGEP(Operand, NewIndices);
1764 if (NGEP->getType() == GEP.getType())
1765 return ReplaceInstUsesWith(GEP, NGEP);
1766 NGEP->takeName(&GEP);
1768 if (NGEP->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
1769 return new AddrSpaceCastInst(NGEP, GEP.getType());
1770 return new BitCastInst(NGEP, GEP.getType());
1779 isAllocSiteRemovable(Instruction *AI, SmallVectorImpl<WeakVH> &Users,
1780 const TargetLibraryInfo *TLI) {
1781 SmallVector<Instruction*, 4> Worklist;
1782 Worklist.push_back(AI);
1785 Instruction *PI = Worklist.pop_back_val();
1786 for (User *U : PI->users()) {
1787 Instruction *I = cast<Instruction>(U);
1788 switch (I->getOpcode()) {
1790 // Give up the moment we see something we can't handle.
1793 case Instruction::BitCast:
1794 case Instruction::GetElementPtr:
1796 Worklist.push_back(I);
1799 case Instruction::ICmp: {
1800 ICmpInst *ICI = cast<ICmpInst>(I);
1801 // We can fold eq/ne comparisons with null to false/true, respectively.
1802 if (!ICI->isEquality() || !isa<ConstantPointerNull>(ICI->getOperand(1)))
1808 case Instruction::Call:
1809 // Ignore no-op and store intrinsics.
1810 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1811 switch (II->getIntrinsicID()) {
1815 case Intrinsic::memmove:
1816 case Intrinsic::memcpy:
1817 case Intrinsic::memset: {
1818 MemIntrinsic *MI = cast<MemIntrinsic>(II);
1819 if (MI->isVolatile() || MI->getRawDest() != PI)
1823 case Intrinsic::dbg_declare:
1824 case Intrinsic::dbg_value:
1825 case Intrinsic::invariant_start:
1826 case Intrinsic::invariant_end:
1827 case Intrinsic::lifetime_start:
1828 case Intrinsic::lifetime_end:
1829 case Intrinsic::objectsize:
1835 if (isFreeCall(I, TLI)) {
1841 case Instruction::Store: {
1842 StoreInst *SI = cast<StoreInst>(I);
1843 if (SI->isVolatile() || SI->getPointerOperand() != PI)
1849 llvm_unreachable("missing a return?");
1851 } while (!Worklist.empty());
1855 Instruction *InstCombiner::visitAllocSite(Instruction &MI) {
1856 // If we have a malloc call which is only used in any amount of comparisons
1857 // to null and free calls, delete the calls and replace the comparisons with
1858 // true or false as appropriate.
1859 SmallVector<WeakVH, 64> Users;
1860 if (isAllocSiteRemovable(&MI, Users, TLI)) {
1861 for (unsigned i = 0, e = Users.size(); i != e; ++i) {
1862 Instruction *I = cast_or_null<Instruction>(&*Users[i]);
1865 if (ICmpInst *C = dyn_cast<ICmpInst>(I)) {
1866 ReplaceInstUsesWith(*C,
1867 ConstantInt::get(Type::getInt1Ty(C->getContext()),
1868 C->isFalseWhenEqual()));
1869 } else if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) {
1870 ReplaceInstUsesWith(*I, UndefValue::get(I->getType()));
1871 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1872 if (II->getIntrinsicID() == Intrinsic::objectsize) {
1873 ConstantInt *CI = cast<ConstantInt>(II->getArgOperand(1));
1874 uint64_t DontKnow = CI->isZero() ? -1ULL : 0;
1875 ReplaceInstUsesWith(*I, ConstantInt::get(I->getType(), DontKnow));
1878 EraseInstFromFunction(*I);
1881 if (InvokeInst *II = dyn_cast<InvokeInst>(&MI)) {
1882 // Replace invoke with a NOP intrinsic to maintain the original CFG
1883 Module *M = II->getParent()->getParent()->getParent();
1884 Function *F = Intrinsic::getDeclaration(M, Intrinsic::donothing);
1885 InvokeInst::Create(F, II->getNormalDest(), II->getUnwindDest(),
1886 None, "", II->getParent());
1888 return EraseInstFromFunction(MI);
1893 /// \brief Move the call to free before a NULL test.
1895 /// Check if this free is accessed after its argument has been test
1896 /// against NULL (property 0).
1897 /// If yes, it is legal to move this call in its predecessor block.
1899 /// The move is performed only if the block containing the call to free
1900 /// will be removed, i.e.:
1901 /// 1. it has only one predecessor P, and P has two successors
1902 /// 2. it contains the call and an unconditional branch
1903 /// 3. its successor is the same as its predecessor's successor
1905 /// The profitability is out-of concern here and this function should
1906 /// be called only if the caller knows this transformation would be
1907 /// profitable (e.g., for code size).
1908 static Instruction *
1909 tryToMoveFreeBeforeNullTest(CallInst &FI) {
1910 Value *Op = FI.getArgOperand(0);
1911 BasicBlock *FreeInstrBB = FI.getParent();
1912 BasicBlock *PredBB = FreeInstrBB->getSinglePredecessor();
1914 // Validate part of constraint #1: Only one predecessor
1915 // FIXME: We can extend the number of predecessor, but in that case, we
1916 // would duplicate the call to free in each predecessor and it may
1917 // not be profitable even for code size.
1921 // Validate constraint #2: Does this block contains only the call to
1922 // free and an unconditional branch?
1923 // FIXME: We could check if we can speculate everything in the
1924 // predecessor block
1925 if (FreeInstrBB->size() != 2)
1928 if (!match(FreeInstrBB->getTerminator(), m_UnconditionalBr(SuccBB)))
1931 // Validate the rest of constraint #1 by matching on the pred branch.
1932 TerminatorInst *TI = PredBB->getTerminator();
1933 BasicBlock *TrueBB, *FalseBB;
1934 ICmpInst::Predicate Pred;
1935 if (!match(TI, m_Br(m_ICmp(Pred, m_Specific(Op), m_Zero()), TrueBB, FalseBB)))
1937 if (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE)
1940 // Validate constraint #3: Ensure the null case just falls through.
1941 if (SuccBB != (Pred == ICmpInst::ICMP_EQ ? TrueBB : FalseBB))
1943 assert(FreeInstrBB == (Pred == ICmpInst::ICMP_EQ ? FalseBB : TrueBB) &&
1944 "Broken CFG: missing edge from predecessor to successor");
1951 Instruction *InstCombiner::visitFree(CallInst &FI) {
1952 Value *Op = FI.getArgOperand(0);
1954 // free undef -> unreachable.
1955 if (isa<UndefValue>(Op)) {
1956 // Insert a new store to null because we cannot modify the CFG here.
1957 Builder->CreateStore(ConstantInt::getTrue(FI.getContext()),
1958 UndefValue::get(Type::getInt1PtrTy(FI.getContext())));
1959 return EraseInstFromFunction(FI);
1962 // If we have 'free null' delete the instruction. This can happen in stl code
1963 // when lots of inlining happens.
1964 if (isa<ConstantPointerNull>(Op))
1965 return EraseInstFromFunction(FI);
1967 // If we optimize for code size, try to move the call to free before the null
1968 // test so that simplify cfg can remove the empty block and dead code
1969 // elimination the branch. I.e., helps to turn something like:
1970 // if (foo) free(foo);
1974 if (Instruction *I = tryToMoveFreeBeforeNullTest(FI))
1980 Instruction *InstCombiner::visitReturnInst(ReturnInst &RI) {
1981 if (RI.getNumOperands() == 0) // ret void
1984 Value *ResultOp = RI.getOperand(0);
1985 Type *VTy = ResultOp->getType();
1986 if (!VTy->isIntegerTy())
1989 // There might be assume intrinsics dominating this return that completely
1990 // determine the value. If so, constant fold it.
1991 unsigned BitWidth = VTy->getPrimitiveSizeInBits();
1992 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
1993 computeKnownBits(ResultOp, KnownZero, KnownOne, 0, &RI);
1994 if ((KnownZero|KnownOne).isAllOnesValue())
1995 RI.setOperand(0, Constant::getIntegerValue(VTy, KnownOne));
2000 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
2001 // Change br (not X), label True, label False to: br X, label False, True
2003 BasicBlock *TrueDest;
2004 BasicBlock *FalseDest;
2005 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
2006 !isa<Constant>(X)) {
2007 // Swap Destinations and condition...
2009 BI.swapSuccessors();
2013 // Canonicalize fcmp_one -> fcmp_oeq
2014 FCmpInst::Predicate FPred; Value *Y;
2015 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
2016 TrueDest, FalseDest)) &&
2017 BI.getCondition()->hasOneUse())
2018 if (FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
2019 FPred == FCmpInst::FCMP_OGE) {
2020 FCmpInst *Cond = cast<FCmpInst>(BI.getCondition());
2021 Cond->setPredicate(FCmpInst::getInversePredicate(FPred));
2023 // Swap Destinations and condition.
2024 BI.swapSuccessors();
2029 // Canonicalize icmp_ne -> icmp_eq
2030 ICmpInst::Predicate IPred;
2031 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
2032 TrueDest, FalseDest)) &&
2033 BI.getCondition()->hasOneUse())
2034 if (IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
2035 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
2036 IPred == ICmpInst::ICMP_SGE) {
2037 ICmpInst *Cond = cast<ICmpInst>(BI.getCondition());
2038 Cond->setPredicate(ICmpInst::getInversePredicate(IPred));
2039 // Swap Destinations and condition.
2040 BI.swapSuccessors();
2048 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
2049 Value *Cond = SI.getCondition();
2050 unsigned BitWidth = cast<IntegerType>(Cond->getType())->getBitWidth();
2051 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2052 computeKnownBits(Cond, KnownZero, KnownOne);
2053 unsigned LeadingKnownZeros = KnownZero.countLeadingOnes();
2054 unsigned LeadingKnownOnes = KnownOne.countLeadingOnes();
2056 // Compute the number of leading bits we can ignore.
2057 for (auto &C : SI.cases()) {
2058 LeadingKnownZeros = std::min(
2059 LeadingKnownZeros, C.getCaseValue()->getValue().countLeadingZeros());
2060 LeadingKnownOnes = std::min(
2061 LeadingKnownOnes, C.getCaseValue()->getValue().countLeadingOnes());
2064 unsigned NewWidth = BitWidth - std::max(LeadingKnownZeros, LeadingKnownOnes);
2066 // Truncate the condition operand if the new type is equal to or larger than
2067 // the largest legal integer type. We need to be conservative here since
2068 // x86 generates redundant zero-extenstion instructions if the operand is
2069 // truncated to i8 or i16.
2070 bool TruncCond = false;
2071 if (DL && BitWidth > NewWidth &&
2072 NewWidth >= DL->getLargestLegalIntTypeSize()) {
2074 IntegerType *Ty = IntegerType::get(SI.getContext(), NewWidth);
2075 Builder->SetInsertPoint(&SI);
2076 Value *NewCond = Builder->CreateTrunc(SI.getCondition(), Ty, "trunc");
2077 SI.setCondition(NewCond);
2079 for (auto &C : SI.cases())
2080 static_cast<SwitchInst::CaseIt *>(&C)->setValue(ConstantInt::get(
2081 SI.getContext(), C.getCaseValue()->getValue().trunc(NewWidth)));
2084 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
2085 if (I->getOpcode() == Instruction::Add)
2086 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2087 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
2088 // Skip the first item since that's the default case.
2089 for (SwitchInst::CaseIt i = SI.case_begin(), e = SI.case_end();
2091 ConstantInt* CaseVal = i.getCaseValue();
2092 Constant *LHS = CaseVal;
2094 LHS = LeadingKnownZeros
2095 ? ConstantExpr::getZExt(CaseVal, Cond->getType())
2096 : ConstantExpr::getSExt(CaseVal, Cond->getType());
2097 Constant* NewCaseVal = ConstantExpr::getSub(LHS, AddRHS);
2098 assert(isa<ConstantInt>(NewCaseVal) &&
2099 "Result of expression should be constant");
2100 i.setValue(cast<ConstantInt>(NewCaseVal));
2102 SI.setCondition(I->getOperand(0));
2108 return TruncCond ? &SI : nullptr;
2111 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
2112 Value *Agg = EV.getAggregateOperand();
2114 if (!EV.hasIndices())
2115 return ReplaceInstUsesWith(EV, Agg);
2117 if (Constant *C = dyn_cast<Constant>(Agg)) {
2118 if (Constant *C2 = C->getAggregateElement(*EV.idx_begin())) {
2119 if (EV.getNumIndices() == 0)
2120 return ReplaceInstUsesWith(EV, C2);
2121 // Extract the remaining indices out of the constant indexed by the
2123 return ExtractValueInst::Create(C2, EV.getIndices().slice(1));
2125 return nullptr; // Can't handle other constants
2128 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
2129 // We're extracting from an insertvalue instruction, compare the indices
2130 const unsigned *exti, *exte, *insi, *inse;
2131 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
2132 exte = EV.idx_end(), inse = IV->idx_end();
2133 exti != exte && insi != inse;
2136 // The insert and extract both reference distinctly different elements.
2137 // This means the extract is not influenced by the insert, and we can
2138 // replace the aggregate operand of the extract with the aggregate
2139 // operand of the insert. i.e., replace
2140 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
2141 // %E = extractvalue { i32, { i32 } } %I, 0
2143 // %E = extractvalue { i32, { i32 } } %A, 0
2144 return ExtractValueInst::Create(IV->getAggregateOperand(),
2147 if (exti == exte && insi == inse)
2148 // Both iterators are at the end: Index lists are identical. Replace
2149 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
2150 // %C = extractvalue { i32, { i32 } } %B, 1, 0
2152 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
2154 // The extract list is a prefix of the insert list. i.e. replace
2155 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
2156 // %E = extractvalue { i32, { i32 } } %I, 1
2158 // %X = extractvalue { i32, { i32 } } %A, 1
2159 // %E = insertvalue { i32 } %X, i32 42, 0
2160 // by switching the order of the insert and extract (though the
2161 // insertvalue should be left in, since it may have other uses).
2162 Value *NewEV = Builder->CreateExtractValue(IV->getAggregateOperand(),
2164 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
2165 makeArrayRef(insi, inse));
2168 // The insert list is a prefix of the extract list
2169 // We can simply remove the common indices from the extract and make it
2170 // operate on the inserted value instead of the insertvalue result.
2172 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
2173 // %E = extractvalue { i32, { i32 } } %I, 1, 0
2175 // %E extractvalue { i32 } { i32 42 }, 0
2176 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
2177 makeArrayRef(exti, exte));
2179 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) {
2180 // We're extracting from an intrinsic, see if we're the only user, which
2181 // allows us to simplify multiple result intrinsics to simpler things that
2182 // just get one value.
2183 if (II->hasOneUse()) {
2184 // Check if we're grabbing the overflow bit or the result of a 'with
2185 // overflow' intrinsic. If it's the latter we can remove the intrinsic
2186 // and replace it with a traditional binary instruction.
2187 switch (II->getIntrinsicID()) {
2188 case Intrinsic::uadd_with_overflow:
2189 case Intrinsic::sadd_with_overflow:
2190 if (*EV.idx_begin() == 0) { // Normal result.
2191 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2192 ReplaceInstUsesWith(*II, UndefValue::get(II->getType()));
2193 EraseInstFromFunction(*II);
2194 return BinaryOperator::CreateAdd(LHS, RHS);
2197 // If the normal result of the add is dead, and the RHS is a constant,
2198 // we can transform this into a range comparison.
2199 // overflow = uadd a, -4 --> overflow = icmp ugt a, 3
2200 if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow)
2201 if (ConstantInt *CI = dyn_cast<ConstantInt>(II->getArgOperand(1)))
2202 return new ICmpInst(ICmpInst::ICMP_UGT, II->getArgOperand(0),
2203 ConstantExpr::getNot(CI));
2205 case Intrinsic::usub_with_overflow:
2206 case Intrinsic::ssub_with_overflow:
2207 if (*EV.idx_begin() == 0) { // Normal result.
2208 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2209 ReplaceInstUsesWith(*II, UndefValue::get(II->getType()));
2210 EraseInstFromFunction(*II);
2211 return BinaryOperator::CreateSub(LHS, RHS);
2214 case Intrinsic::umul_with_overflow:
2215 case Intrinsic::smul_with_overflow:
2216 if (*EV.idx_begin() == 0) { // Normal result.
2217 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2218 ReplaceInstUsesWith(*II, UndefValue::get(II->getType()));
2219 EraseInstFromFunction(*II);
2220 return BinaryOperator::CreateMul(LHS, RHS);
2228 if (LoadInst *L = dyn_cast<LoadInst>(Agg))
2229 // If the (non-volatile) load only has one use, we can rewrite this to a
2230 // load from a GEP. This reduces the size of the load.
2231 // FIXME: If a load is used only by extractvalue instructions then this
2232 // could be done regardless of having multiple uses.
2233 if (L->isSimple() && L->hasOneUse()) {
2234 // extractvalue has integer indices, getelementptr has Value*s. Convert.
2235 SmallVector<Value*, 4> Indices;
2236 // Prefix an i32 0 since we need the first element.
2237 Indices.push_back(Builder->getInt32(0));
2238 for (ExtractValueInst::idx_iterator I = EV.idx_begin(), E = EV.idx_end();
2240 Indices.push_back(Builder->getInt32(*I));
2242 // We need to insert these at the location of the old load, not at that of
2243 // the extractvalue.
2244 Builder->SetInsertPoint(L->getParent(), L);
2245 Value *GEP = Builder->CreateInBoundsGEP(L->getPointerOperand(), Indices);
2246 // Returning the load directly will cause the main loop to insert it in
2247 // the wrong spot, so use ReplaceInstUsesWith().
2248 return ReplaceInstUsesWith(EV, Builder->CreateLoad(GEP));
2250 // We could simplify extracts from other values. Note that nested extracts may
2251 // already be simplified implicitly by the above: extract (extract (insert) )
2252 // will be translated into extract ( insert ( extract ) ) first and then just
2253 // the value inserted, if appropriate. Similarly for extracts from single-use
2254 // loads: extract (extract (load)) will be translated to extract (load (gep))
2255 // and if again single-use then via load (gep (gep)) to load (gep).
2256 // However, double extracts from e.g. function arguments or return values
2257 // aren't handled yet.
2261 enum Personality_Type {
2262 Unknown_Personality,
2263 GNU_Ada_Personality,
2264 GNU_CXX_Personality,
2265 GNU_ObjC_Personality
2268 /// RecognizePersonality - See if the given exception handling personality
2269 /// function is one that we understand. If so, return a description of it;
2270 /// otherwise return Unknown_Personality.
2271 static Personality_Type RecognizePersonality(Value *Pers) {
2272 Function *F = dyn_cast<Function>(Pers->stripPointerCasts());
2274 return Unknown_Personality;
2275 return StringSwitch<Personality_Type>(F->getName())
2276 .Case("__gnat_eh_personality", GNU_Ada_Personality)
2277 .Case("__gxx_personality_v0", GNU_CXX_Personality)
2278 .Case("__objc_personality_v0", GNU_ObjC_Personality)
2279 .Default(Unknown_Personality);
2282 /// isCatchAll - Return 'true' if the given typeinfo will match anything.
2283 static bool isCatchAll(Personality_Type Personality, Constant *TypeInfo) {
2284 switch (Personality) {
2285 case Unknown_Personality:
2287 case GNU_Ada_Personality:
2288 // While __gnat_all_others_value will match any Ada exception, it doesn't
2289 // match foreign exceptions (or didn't, before gcc-4.7).
2291 case GNU_CXX_Personality:
2292 case GNU_ObjC_Personality:
2293 return TypeInfo->isNullValue();
2295 llvm_unreachable("Unknown personality!");
2298 static bool shorter_filter(const Value *LHS, const Value *RHS) {
2300 cast<ArrayType>(LHS->getType())->getNumElements()
2302 cast<ArrayType>(RHS->getType())->getNumElements();
2305 Instruction *InstCombiner::visitLandingPadInst(LandingPadInst &LI) {
2306 // The logic here should be correct for any real-world personality function.
2307 // However if that turns out not to be true, the offending logic can always
2308 // be conditioned on the personality function, like the catch-all logic is.
2309 Personality_Type Personality = RecognizePersonality(LI.getPersonalityFn());
2311 // Simplify the list of clauses, eg by removing repeated catch clauses
2312 // (these are often created by inlining).
2313 bool MakeNewInstruction = false; // If true, recreate using the following:
2314 SmallVector<Constant *, 16> NewClauses; // - Clauses for the new instruction;
2315 bool CleanupFlag = LI.isCleanup(); // - The new instruction is a cleanup.
2317 SmallPtrSet<Value *, 16> AlreadyCaught; // Typeinfos known caught already.
2318 for (unsigned i = 0, e = LI.getNumClauses(); i != e; ++i) {
2319 bool isLastClause = i + 1 == e;
2320 if (LI.isCatch(i)) {
2322 Constant *CatchClause = LI.getClause(i);
2323 Constant *TypeInfo = CatchClause->stripPointerCasts();
2325 // If we already saw this clause, there is no point in having a second
2327 if (AlreadyCaught.insert(TypeInfo).second) {
2328 // This catch clause was not already seen.
2329 NewClauses.push_back(CatchClause);
2331 // Repeated catch clause - drop the redundant copy.
2332 MakeNewInstruction = true;
2335 // If this is a catch-all then there is no point in keeping any following
2336 // clauses or marking the landingpad as having a cleanup.
2337 if (isCatchAll(Personality, TypeInfo)) {
2339 MakeNewInstruction = true;
2340 CleanupFlag = false;
2344 // A filter clause. If any of the filter elements were already caught
2345 // then they can be dropped from the filter. It is tempting to try to
2346 // exploit the filter further by saying that any typeinfo that does not
2347 // occur in the filter can't be caught later (and thus can be dropped).
2348 // However this would be wrong, since typeinfos can match without being
2349 // equal (for example if one represents a C++ class, and the other some
2350 // class derived from it).
2351 assert(LI.isFilter(i) && "Unsupported landingpad clause!");
2352 Constant *FilterClause = LI.getClause(i);
2353 ArrayType *FilterType = cast<ArrayType>(FilterClause->getType());
2354 unsigned NumTypeInfos = FilterType->getNumElements();
2356 // An empty filter catches everything, so there is no point in keeping any
2357 // following clauses or marking the landingpad as having a cleanup. By
2358 // dealing with this case here the following code is made a bit simpler.
2359 if (!NumTypeInfos) {
2360 NewClauses.push_back(FilterClause);
2362 MakeNewInstruction = true;
2363 CleanupFlag = false;
2367 bool MakeNewFilter = false; // If true, make a new filter.
2368 SmallVector<Constant *, 16> NewFilterElts; // New elements.
2369 if (isa<ConstantAggregateZero>(FilterClause)) {
2370 // Not an empty filter - it contains at least one null typeinfo.
2371 assert(NumTypeInfos > 0 && "Should have handled empty filter already!");
2372 Constant *TypeInfo =
2373 Constant::getNullValue(FilterType->getElementType());
2374 // If this typeinfo is a catch-all then the filter can never match.
2375 if (isCatchAll(Personality, TypeInfo)) {
2376 // Throw the filter away.
2377 MakeNewInstruction = true;
2381 // There is no point in having multiple copies of this typeinfo, so
2382 // discard all but the first copy if there is more than one.
2383 NewFilterElts.push_back(TypeInfo);
2384 if (NumTypeInfos > 1)
2385 MakeNewFilter = true;
2387 ConstantArray *Filter = cast<ConstantArray>(FilterClause);
2388 SmallPtrSet<Value *, 16> SeenInFilter; // For uniquing the elements.
2389 NewFilterElts.reserve(NumTypeInfos);
2391 // Remove any filter elements that were already caught or that already
2392 // occurred in the filter. While there, see if any of the elements are
2393 // catch-alls. If so, the filter can be discarded.
2394 bool SawCatchAll = false;
2395 for (unsigned j = 0; j != NumTypeInfos; ++j) {
2396 Constant *Elt = Filter->getOperand(j);
2397 Constant *TypeInfo = Elt->stripPointerCasts();
2398 if (isCatchAll(Personality, TypeInfo)) {
2399 // This element is a catch-all. Bail out, noting this fact.
2403 if (AlreadyCaught.count(TypeInfo))
2404 // Already caught by an earlier clause, so having it in the filter
2407 // There is no point in having multiple copies of the same typeinfo in
2408 // a filter, so only add it if we didn't already.
2409 if (SeenInFilter.insert(TypeInfo).second)
2410 NewFilterElts.push_back(cast<Constant>(Elt));
2412 // A filter containing a catch-all cannot match anything by definition.
2414 // Throw the filter away.
2415 MakeNewInstruction = true;
2419 // If we dropped something from the filter, make a new one.
2420 if (NewFilterElts.size() < NumTypeInfos)
2421 MakeNewFilter = true;
2423 if (MakeNewFilter) {
2424 FilterType = ArrayType::get(FilterType->getElementType(),
2425 NewFilterElts.size());
2426 FilterClause = ConstantArray::get(FilterType, NewFilterElts);
2427 MakeNewInstruction = true;
2430 NewClauses.push_back(FilterClause);
2432 // If the new filter is empty then it will catch everything so there is
2433 // no point in keeping any following clauses or marking the landingpad
2434 // as having a cleanup. The case of the original filter being empty was
2435 // already handled above.
2436 if (MakeNewFilter && !NewFilterElts.size()) {
2437 assert(MakeNewInstruction && "New filter but not a new instruction!");
2438 CleanupFlag = false;
2444 // If several filters occur in a row then reorder them so that the shortest
2445 // filters come first (those with the smallest number of elements). This is
2446 // advantageous because shorter filters are more likely to match, speeding up
2447 // unwinding, but mostly because it increases the effectiveness of the other
2448 // filter optimizations below.
2449 for (unsigned i = 0, e = NewClauses.size(); i + 1 < e; ) {
2451 // Find the maximal 'j' s.t. the range [i, j) consists entirely of filters.
2452 for (j = i; j != e; ++j)
2453 if (!isa<ArrayType>(NewClauses[j]->getType()))
2456 // Check whether the filters are already sorted by length. We need to know
2457 // if sorting them is actually going to do anything so that we only make a
2458 // new landingpad instruction if it does.
2459 for (unsigned k = i; k + 1 < j; ++k)
2460 if (shorter_filter(NewClauses[k+1], NewClauses[k])) {
2461 // Not sorted, so sort the filters now. Doing an unstable sort would be
2462 // correct too but reordering filters pointlessly might confuse users.
2463 std::stable_sort(NewClauses.begin() + i, NewClauses.begin() + j,
2465 MakeNewInstruction = true;
2469 // Look for the next batch of filters.
2473 // If typeinfos matched if and only if equal, then the elements of a filter L
2474 // that occurs later than a filter F could be replaced by the intersection of
2475 // the elements of F and L. In reality two typeinfos can match without being
2476 // equal (for example if one represents a C++ class, and the other some class
2477 // derived from it) so it would be wrong to perform this transform in general.
2478 // However the transform is correct and useful if F is a subset of L. In that
2479 // case L can be replaced by F, and thus removed altogether since repeating a
2480 // filter is pointless. So here we look at all pairs of filters F and L where
2481 // L follows F in the list of clauses, and remove L if every element of F is
2482 // an element of L. This can occur when inlining C++ functions with exception
2484 for (unsigned i = 0; i + 1 < NewClauses.size(); ++i) {
2485 // Examine each filter in turn.
2486 Value *Filter = NewClauses[i];
2487 ArrayType *FTy = dyn_cast<ArrayType>(Filter->getType());
2489 // Not a filter - skip it.
2491 unsigned FElts = FTy->getNumElements();
2492 // Examine each filter following this one. Doing this backwards means that
2493 // we don't have to worry about filters disappearing under us when removed.
2494 for (unsigned j = NewClauses.size() - 1; j != i; --j) {
2495 Value *LFilter = NewClauses[j];
2496 ArrayType *LTy = dyn_cast<ArrayType>(LFilter->getType());
2498 // Not a filter - skip it.
2500 // If Filter is a subset of LFilter, i.e. every element of Filter is also
2501 // an element of LFilter, then discard LFilter.
2502 SmallVectorImpl<Constant *>::iterator J = NewClauses.begin() + j;
2503 // If Filter is empty then it is a subset of LFilter.
2506 NewClauses.erase(J);
2507 MakeNewInstruction = true;
2508 // Move on to the next filter.
2511 unsigned LElts = LTy->getNumElements();
2512 // If Filter is longer than LFilter then it cannot be a subset of it.
2514 // Move on to the next filter.
2516 // At this point we know that LFilter has at least one element.
2517 if (isa<ConstantAggregateZero>(LFilter)) { // LFilter only contains zeros.
2518 // Filter is a subset of LFilter iff Filter contains only zeros (as we
2519 // already know that Filter is not longer than LFilter).
2520 if (isa<ConstantAggregateZero>(Filter)) {
2521 assert(FElts <= LElts && "Should have handled this case earlier!");
2523 NewClauses.erase(J);
2524 MakeNewInstruction = true;
2526 // Move on to the next filter.
2529 ConstantArray *LArray = cast<ConstantArray>(LFilter);
2530 if (isa<ConstantAggregateZero>(Filter)) { // Filter only contains zeros.
2531 // Since Filter is non-empty and contains only zeros, it is a subset of
2532 // LFilter iff LFilter contains a zero.
2533 assert(FElts > 0 && "Should have eliminated the empty filter earlier!");
2534 for (unsigned l = 0; l != LElts; ++l)
2535 if (LArray->getOperand(l)->isNullValue()) {
2536 // LFilter contains a zero - discard it.
2537 NewClauses.erase(J);
2538 MakeNewInstruction = true;
2541 // Move on to the next filter.
2544 // At this point we know that both filters are ConstantArrays. Loop over
2545 // operands to see whether every element of Filter is also an element of
2546 // LFilter. Since filters tend to be short this is probably faster than
2547 // using a method that scales nicely.
2548 ConstantArray *FArray = cast<ConstantArray>(Filter);
2549 bool AllFound = true;
2550 for (unsigned f = 0; f != FElts; ++f) {
2551 Value *FTypeInfo = FArray->getOperand(f)->stripPointerCasts();
2553 for (unsigned l = 0; l != LElts; ++l) {
2554 Value *LTypeInfo = LArray->getOperand(l)->stripPointerCasts();
2555 if (LTypeInfo == FTypeInfo) {
2565 NewClauses.erase(J);
2566 MakeNewInstruction = true;
2568 // Move on to the next filter.
2572 // If we changed any of the clauses, replace the old landingpad instruction
2574 if (MakeNewInstruction) {
2575 LandingPadInst *NLI = LandingPadInst::Create(LI.getType(),
2576 LI.getPersonalityFn(),
2578 for (unsigned i = 0, e = NewClauses.size(); i != e; ++i)
2579 NLI->addClause(NewClauses[i]);
2580 // A landing pad with no clauses must have the cleanup flag set. It is
2581 // theoretically possible, though highly unlikely, that we eliminated all
2582 // clauses. If so, force the cleanup flag to true.
2583 if (NewClauses.empty())
2585 NLI->setCleanup(CleanupFlag);
2589 // Even if none of the clauses changed, we may nonetheless have understood
2590 // that the cleanup flag is pointless. Clear it if so.
2591 if (LI.isCleanup() != CleanupFlag) {
2592 assert(!CleanupFlag && "Adding a cleanup, not removing one?!");
2593 LI.setCleanup(CleanupFlag);
2603 /// TryToSinkInstruction - Try to move the specified instruction from its
2604 /// current block into the beginning of DestBlock, which can only happen if it's
2605 /// safe to move the instruction past all of the instructions between it and the
2606 /// end of its block.
2607 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
2608 assert(I->hasOneUse() && "Invariants didn't hold!");
2610 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
2611 if (isa<PHINode>(I) || isa<LandingPadInst>(I) || I->mayHaveSideEffects() ||
2612 isa<TerminatorInst>(I))
2615 // Do not sink alloca instructions out of the entry block.
2616 if (isa<AllocaInst>(I) && I->getParent() ==
2617 &DestBlock->getParent()->getEntryBlock())
2620 // We can only sink load instructions if there is nothing between the load and
2621 // the end of block that could change the value.
2622 if (I->mayReadFromMemory()) {
2623 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
2625 if (Scan->mayWriteToMemory())
2629 BasicBlock::iterator InsertPos = DestBlock->getFirstInsertionPt();
2630 I->moveBefore(InsertPos);
2636 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
2637 /// all reachable code to the worklist.
2639 /// This has a couple of tricks to make the code faster and more powerful. In
2640 /// particular, we constant fold and DCE instructions as we go, to avoid adding
2641 /// them to the worklist (this significantly speeds up instcombine on code where
2642 /// many instructions are dead or constant). Additionally, if we find a branch
2643 /// whose condition is a known constant, we only visit the reachable successors.
2645 static bool AddReachableCodeToWorklist(BasicBlock *BB,
2646 SmallPtrSetImpl<BasicBlock*> &Visited,
2648 const DataLayout *DL,
2649 const TargetLibraryInfo *TLI) {
2650 bool MadeIRChange = false;
2651 SmallVector<BasicBlock*, 256> Worklist;
2652 Worklist.push_back(BB);
2654 SmallVector<Instruction*, 128> InstrsForInstCombineWorklist;
2655 DenseMap<ConstantExpr*, Constant*> FoldedConstants;
2658 BB = Worklist.pop_back_val();
2660 // We have now visited this block! If we've already been here, ignore it.
2661 if (!Visited.insert(BB).second)
2664 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
2665 Instruction *Inst = BBI++;
2667 // DCE instruction if trivially dead.
2668 if (isInstructionTriviallyDead(Inst, TLI)) {
2670 DEBUG(dbgs() << "IC: DCE: " << *Inst << '\n');
2671 Inst->eraseFromParent();
2675 // ConstantProp instruction if trivially constant.
2676 if (!Inst->use_empty() && isa<Constant>(Inst->getOperand(0)))
2677 if (Constant *C = ConstantFoldInstruction(Inst, DL, TLI)) {
2678 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: "
2680 Inst->replaceAllUsesWith(C);
2682 Inst->eraseFromParent();
2687 // See if we can constant fold its operands.
2688 for (User::op_iterator i = Inst->op_begin(), e = Inst->op_end();
2690 ConstantExpr *CE = dyn_cast<ConstantExpr>(i);
2691 if (CE == nullptr) continue;
2693 Constant*& FoldRes = FoldedConstants[CE];
2695 FoldRes = ConstantFoldConstantExpression(CE, DL, TLI);
2699 if (FoldRes != CE) {
2701 MadeIRChange = true;
2706 InstrsForInstCombineWorklist.push_back(Inst);
2709 // Recursively visit successors. If this is a branch or switch on a
2710 // constant, only visit the reachable successor.
2711 TerminatorInst *TI = BB->getTerminator();
2712 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
2713 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
2714 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
2715 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
2716 Worklist.push_back(ReachableBB);
2719 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
2720 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
2721 // See if this is an explicit destination.
2722 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2724 if (i.getCaseValue() == Cond) {
2725 BasicBlock *ReachableBB = i.getCaseSuccessor();
2726 Worklist.push_back(ReachableBB);
2730 // Otherwise it is the default destination.
2731 Worklist.push_back(SI->getDefaultDest());
2736 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
2737 Worklist.push_back(TI->getSuccessor(i));
2738 } while (!Worklist.empty());
2740 // Once we've found all of the instructions to add to instcombine's worklist,
2741 // add them in reverse order. This way instcombine will visit from the top
2742 // of the function down. This jives well with the way that it adds all uses
2743 // of instructions to the worklist after doing a transformation, thus avoiding
2744 // some N^2 behavior in pathological cases.
2745 IC.Worklist.AddInitialGroup(&InstrsForInstCombineWorklist[0],
2746 InstrsForInstCombineWorklist.size());
2748 return MadeIRChange;
2751 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
2752 MadeIRChange = false;
2754 DEBUG(dbgs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
2755 << F.getName() << "\n");
2758 // Do a depth-first traversal of the function, populate the worklist with
2759 // the reachable instructions. Ignore blocks that are not reachable. Keep
2760 // track of which blocks we visit.
2761 SmallPtrSet<BasicBlock*, 64> Visited;
2762 MadeIRChange |= AddReachableCodeToWorklist(F.begin(), Visited, *this, DL,
2765 // Do a quick scan over the function. If we find any blocks that are
2766 // unreachable, remove any instructions inside of them. This prevents
2767 // the instcombine code from having to deal with some bad special cases.
2768 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
2769 if (Visited.count(BB)) continue;
2771 // Delete the instructions backwards, as it has a reduced likelihood of
2772 // having to update as many def-use and use-def chains.
2773 Instruction *EndInst = BB->getTerminator(); // Last not to be deleted.
2774 while (EndInst != BB->begin()) {
2775 // Delete the next to last instruction.
2776 BasicBlock::iterator I = EndInst;
2777 Instruction *Inst = --I;
2778 if (!Inst->use_empty())
2779 Inst->replaceAllUsesWith(UndefValue::get(Inst->getType()));
2780 if (isa<LandingPadInst>(Inst)) {
2784 if (!isa<DbgInfoIntrinsic>(Inst)) {
2786 MadeIRChange = true;
2788 Inst->eraseFromParent();
2793 while (!Worklist.isEmpty()) {
2794 Instruction *I = Worklist.RemoveOne();
2795 if (I == nullptr) continue; // skip null values.
2797 // Check to see if we can DCE the instruction.
2798 if (isInstructionTriviallyDead(I, TLI)) {
2799 DEBUG(dbgs() << "IC: DCE: " << *I << '\n');
2800 EraseInstFromFunction(*I);
2802 MadeIRChange = true;
2806 // Instruction isn't dead, see if we can constant propagate it.
2807 if (!I->use_empty() && isa<Constant>(I->getOperand(0)))
2808 if (Constant *C = ConstantFoldInstruction(I, DL, TLI)) {
2809 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n');
2811 // Add operands to the worklist.
2812 ReplaceInstUsesWith(*I, C);
2814 EraseInstFromFunction(*I);
2815 MadeIRChange = true;
2819 // See if we can trivially sink this instruction to a successor basic block.
2820 if (I->hasOneUse()) {
2821 BasicBlock *BB = I->getParent();
2822 Instruction *UserInst = cast<Instruction>(*I->user_begin());
2823 BasicBlock *UserParent;
2825 // Get the block the use occurs in.
2826 if (PHINode *PN = dyn_cast<PHINode>(UserInst))
2827 UserParent = PN->getIncomingBlock(*I->use_begin());
2829 UserParent = UserInst->getParent();
2831 if (UserParent != BB) {
2832 bool UserIsSuccessor = false;
2833 // See if the user is one of our successors.
2834 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
2835 if (*SI == UserParent) {
2836 UserIsSuccessor = true;
2840 // If the user is one of our immediate successors, and if that successor
2841 // only has us as a predecessors (we'd have to split the critical edge
2842 // otherwise), we can keep going.
2843 if (UserIsSuccessor && UserParent->getSinglePredecessor()) {
2844 // Okay, the CFG is simple enough, try to sink this instruction.
2845 if (TryToSinkInstruction(I, UserParent)) {
2846 MadeIRChange = true;
2847 // We'll add uses of the sunk instruction below, but since sinking
2848 // can expose opportunities for it's *operands* add them to the
2850 for (Use &U : I->operands())
2851 if (Instruction *OpI = dyn_cast<Instruction>(U.get()))
2858 // Now that we have an instruction, try combining it to simplify it.
2859 Builder->SetInsertPoint(I->getParent(), I);
2860 Builder->SetCurrentDebugLocation(I->getDebugLoc());
2865 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str(););
2866 DEBUG(dbgs() << "IC: Visiting: " << OrigI << '\n');
2868 if (Instruction *Result = visit(*I)) {
2870 // Should we replace the old instruction with a new one?
2872 DEBUG(dbgs() << "IC: Old = " << *I << '\n'
2873 << " New = " << *Result << '\n');
2875 if (!I->getDebugLoc().isUnknown())
2876 Result->setDebugLoc(I->getDebugLoc());
2877 // Everything uses the new instruction now.
2878 I->replaceAllUsesWith(Result);
2880 // Move the name to the new instruction first.
2881 Result->takeName(I);
2883 // Push the new instruction and any users onto the worklist.
2884 Worklist.Add(Result);
2885 Worklist.AddUsersToWorkList(*Result);
2887 // Insert the new instruction into the basic block...
2888 BasicBlock *InstParent = I->getParent();
2889 BasicBlock::iterator InsertPos = I;
2891 // If we replace a PHI with something that isn't a PHI, fix up the
2893 if (!isa<PHINode>(Result) && isa<PHINode>(InsertPos))
2894 InsertPos = InstParent->getFirstInsertionPt();
2896 InstParent->getInstList().insert(InsertPos, Result);
2898 EraseInstFromFunction(*I);
2901 DEBUG(dbgs() << "IC: Mod = " << OrigI << '\n'
2902 << " New = " << *I << '\n');
2905 // If the instruction was modified, it's possible that it is now dead.
2906 // if so, remove it.
2907 if (isInstructionTriviallyDead(I, TLI)) {
2908 EraseInstFromFunction(*I);
2911 Worklist.AddUsersToWorkList(*I);
2914 MadeIRChange = true;
2919 return MadeIRChange;
2922 // FIXME: Passing all of the analyses here in the run method is ugly. We should
2923 // separate out the worklist from the combiner so that we can construct
2924 // a combiner once per function while re-using the storage of an external
2926 bool InstCombiner::run(Function &F, AssumptionCache *AC, const DataLayout *DL,
2927 TargetLibraryInfo *TLI, DominatorTree *DT,
2929 // Set up our analysis pointers.
2937 MinimizeSize = F.getAttributes().hasAttribute(AttributeSet::FunctionIndex,
2938 Attribute::MinSize);
2940 /// Builder - This is an IRBuilder that automatically inserts new
2941 /// instructions into the worklist when they are created.
2942 IRBuilder<true, TargetFolder, InstCombineIRInserter> TheBuilder(
2943 F.getContext(), TargetFolder(DL), InstCombineIRInserter(Worklist, AC));
2944 Builder = &TheBuilder;
2946 auto InstCombineRAUW = [this](Instruction *From, Value *With) {
2947 ReplaceInstUsesWith(*From, With);
2949 LibCallSimplifier TheSimplifier(DL, TLI, InstCombineRAUW);
2950 Simplifier = &TheSimplifier;
2952 bool EverMadeChange = false;
2954 // Lower dbg.declare intrinsics otherwise their value may be clobbered
2956 EverMadeChange = LowerDbgDeclare(F);
2958 // Iterate while there is work to do.
2959 unsigned Iteration = 0;
2960 while (DoOneIteration(F, Iteration++))
2961 EverMadeChange = true;
2964 return EverMadeChange;
2968 /// \brief The legacy pass manager's instcombine pass.
2970 /// This is a basic whole-function wrapper around the instcombine utility. It
2971 /// will try to combine all instructions in the function.
2972 class InstructionCombiningPass : public FunctionPass {
2976 static char ID; // Pass identification, replacement for typeid
2978 InstructionCombiningPass() : FunctionPass(ID) {
2979 initializeInstructionCombiningPassPass(*PassRegistry::getPassRegistry());
2982 void getAnalysisUsage(AnalysisUsage &AU) const override;
2983 bool runOnFunction(Function &F) override;
2987 void InstructionCombiningPass::getAnalysisUsage(AnalysisUsage &AU) const {
2988 AU.setPreservesCFG();
2989 AU.addRequired<AssumptionCacheTracker>();
2990 AU.addRequired<TargetLibraryInfoWrapperPass>();
2991 AU.addRequired<DominatorTreeWrapperPass>();
2992 AU.addPreserved<DominatorTreeWrapperPass>();
2995 bool InstructionCombiningPass::runOnFunction(Function &F) {
2996 if (skipOptnoneFunction(F))
2999 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
3000 auto *DLP = getAnalysisIfAvailable<DataLayoutPass>();
3001 auto *DL = DLP ? &DLP->getDataLayout() : nullptr;
3002 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
3003 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
3004 auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>();
3005 auto *LI = LIWP ? &LIWP->getLoopInfo() : nullptr;
3007 return IC.run(F, &AC, DL, &TLI, &DT, LI);
3010 char InstructionCombiningPass::ID = 0;
3011 INITIALIZE_PASS_BEGIN(InstructionCombiningPass, "instcombine",
3012 "Combine redundant instructions", false, false)
3013 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
3014 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
3015 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
3016 INITIALIZE_PASS_END(InstructionCombiningPass, "instcombine",
3017 "Combine redundant instructions", false, false)
3019 // Initialization Routines
3020 void llvm::initializeInstCombine(PassRegistry &Registry) {
3021 initializeInstructionCombiningPassPass(Registry);
3024 void LLVMInitializeInstCombine(LLVMPassRegistryRef R) {
3025 initializeInstructionCombiningPassPass(*unwrap(R));
3028 FunctionPass *llvm::createInstructionCombiningPass() {
3029 return new InstructionCombiningPass();