1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // InstructionCombining - Combine instructions to form fewer, simple
11 // instructions. This pass does not modify the CFG. This pass is where
12 // algebraic simplification happens.
14 // This pass combines things like:
20 // This is a simple worklist driven algorithm.
22 // This pass guarantees that the following canonicalizations are performed on
24 // 1. If a binary operator has a constant operand, it is moved to the RHS
25 // 2. Bitwise operators with constant operands are always grouped so that
26 // shifts are performed first, then or's, then and's, then xor's.
27 // 3. Compare instructions are converted from <,>,<=,>= to ==,!= if possible
28 // 4. All cmp instructions on boolean values are replaced with logical ops
29 // 5. add X, X is represented as (X*2) => (X << 1)
30 // 6. Multiplies with a power-of-two constant argument are transformed into
34 //===----------------------------------------------------------------------===//
36 #include "llvm/Transforms/Scalar.h"
37 #include "InstCombine.h"
38 #include "llvm-c/Initialization.h"
39 #include "llvm/ADT/SmallPtrSet.h"
40 #include "llvm/ADT/Statistic.h"
41 #include "llvm/ADT/StringSwitch.h"
42 #include "llvm/Analysis/AssumptionTracker.h"
43 #include "llvm/Analysis/CFG.h"
44 #include "llvm/Analysis/ConstantFolding.h"
45 #include "llvm/Analysis/InstructionSimplify.h"
46 #include "llvm/Analysis/LoopInfo.h"
47 #include "llvm/Analysis/MemoryBuiltins.h"
48 #include "llvm/Analysis/ValueTracking.h"
49 #include "llvm/IR/CFG.h"
50 #include "llvm/IR/DataLayout.h"
51 #include "llvm/IR/Dominators.h"
52 #include "llvm/IR/GetElementPtrTypeIterator.h"
53 #include "llvm/IR/IntrinsicInst.h"
54 #include "llvm/IR/PatternMatch.h"
55 #include "llvm/IR/ValueHandle.h"
56 #include "llvm/Support/CommandLine.h"
57 #include "llvm/Support/Debug.h"
58 #include "llvm/Target/TargetLibraryInfo.h"
59 #include "llvm/Transforms/Utils/Local.h"
63 using namespace llvm::PatternMatch;
65 #define DEBUG_TYPE "instcombine"
67 STATISTIC(NumCombined , "Number of insts combined");
68 STATISTIC(NumConstProp, "Number of constant folds");
69 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
70 STATISTIC(NumSunkInst , "Number of instructions sunk");
71 STATISTIC(NumExpand, "Number of expansions");
72 STATISTIC(NumFactor , "Number of factorizations");
73 STATISTIC(NumReassoc , "Number of reassociations");
75 // Initialization Routines
76 void llvm::initializeInstCombine(PassRegistry &Registry) {
77 initializeInstCombinerPass(Registry);
80 void LLVMInitializeInstCombine(LLVMPassRegistryRef R) {
81 initializeInstCombine(*unwrap(R));
84 char InstCombiner::ID = 0;
85 INITIALIZE_PASS_BEGIN(InstCombiner, "instcombine",
86 "Combine redundant instructions", false, false)
87 INITIALIZE_PASS_DEPENDENCY(AssumptionTracker)
88 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
89 INITIALIZE_PASS_END(InstCombiner, "instcombine",
90 "Combine redundant instructions", false, false)
92 void InstCombiner::getAnalysisUsage(AnalysisUsage &AU) const {
94 AU.addRequired<AssumptionTracker>();
95 AU.addRequired<TargetLibraryInfo>();
99 Value *InstCombiner::EmitGEPOffset(User *GEP) {
100 return llvm::EmitGEPOffset(Builder, *getDataLayout(), GEP);
103 /// ShouldChangeType - Return true if it is desirable to convert a computation
104 /// from 'From' to 'To'. We don't want to convert from a legal to an illegal
105 /// type for example, or from a smaller to a larger illegal type.
106 bool InstCombiner::ShouldChangeType(Type *From, Type *To) const {
107 assert(From->isIntegerTy() && To->isIntegerTy());
109 // If we don't have DL, we don't know if the source/dest are legal.
110 if (!DL) return false;
112 unsigned FromWidth = From->getPrimitiveSizeInBits();
113 unsigned ToWidth = To->getPrimitiveSizeInBits();
114 bool FromLegal = DL->isLegalInteger(FromWidth);
115 bool ToLegal = DL->isLegalInteger(ToWidth);
117 // If this is a legal integer from type, and the result would be an illegal
118 // type, don't do the transformation.
119 if (FromLegal && !ToLegal)
122 // Otherwise, if both are illegal, do not increase the size of the result. We
123 // do allow things like i160 -> i64, but not i64 -> i160.
124 if (!FromLegal && !ToLegal && ToWidth > FromWidth)
130 // Return true, if No Signed Wrap should be maintained for I.
131 // The No Signed Wrap flag can be kept if the operation "B (I.getOpcode) C",
132 // where both B and C should be ConstantInts, results in a constant that does
133 // not overflow. This function only handles the Add and Sub opcodes. For
134 // all other opcodes, the function conservatively returns false.
135 static bool MaintainNoSignedWrap(BinaryOperator &I, Value *B, Value *C) {
136 OverflowingBinaryOperator *OBO = dyn_cast<OverflowingBinaryOperator>(&I);
137 if (!OBO || !OBO->hasNoSignedWrap()) {
141 // We reason about Add and Sub Only.
142 Instruction::BinaryOps Opcode = I.getOpcode();
143 if (Opcode != Instruction::Add &&
144 Opcode != Instruction::Sub) {
148 ConstantInt *CB = dyn_cast<ConstantInt>(B);
149 ConstantInt *CC = dyn_cast<ConstantInt>(C);
155 const APInt &BVal = CB->getValue();
156 const APInt &CVal = CC->getValue();
157 bool Overflow = false;
159 if (Opcode == Instruction::Add) {
160 BVal.sadd_ov(CVal, Overflow);
162 BVal.ssub_ov(CVal, Overflow);
168 /// Conservatively clears subclassOptionalData after a reassociation or
169 /// commutation. We preserve fast-math flags when applicable as they can be
171 static void ClearSubclassDataAfterReassociation(BinaryOperator &I) {
172 FPMathOperator *FPMO = dyn_cast<FPMathOperator>(&I);
174 I.clearSubclassOptionalData();
178 FastMathFlags FMF = I.getFastMathFlags();
179 I.clearSubclassOptionalData();
180 I.setFastMathFlags(FMF);
183 /// SimplifyAssociativeOrCommutative - This performs a few simplifications for
184 /// operators which are associative or commutative:
186 // Commutative operators:
188 // 1. Order operands such that they are listed from right (least complex) to
189 // left (most complex). This puts constants before unary operators before
192 // Associative operators:
194 // 2. Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
195 // 3. Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
197 // Associative and commutative operators:
199 // 4. Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
200 // 5. Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
201 // 6. Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
202 // if C1 and C2 are constants.
204 bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) {
205 Instruction::BinaryOps Opcode = I.getOpcode();
206 bool Changed = false;
209 // Order operands such that they are listed from right (least complex) to
210 // left (most complex). This puts constants before unary operators before
212 if (I.isCommutative() && getComplexity(I.getOperand(0)) <
213 getComplexity(I.getOperand(1)))
214 Changed = !I.swapOperands();
216 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
217 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
219 if (I.isAssociative()) {
220 // Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
221 if (Op0 && Op0->getOpcode() == Opcode) {
222 Value *A = Op0->getOperand(0);
223 Value *B = Op0->getOperand(1);
224 Value *C = I.getOperand(1);
226 // Does "B op C" simplify?
227 if (Value *V = SimplifyBinOp(Opcode, B, C, DL)) {
228 // It simplifies to V. Form "A op V".
231 // Conservatively clear the optional flags, since they may not be
232 // preserved by the reassociation.
233 if (MaintainNoSignedWrap(I, B, C) &&
234 (!Op0 || (isa<BinaryOperator>(Op0) && Op0->hasNoSignedWrap()))) {
235 // Note: this is only valid because SimplifyBinOp doesn't look at
236 // the operands to Op0.
237 I.clearSubclassOptionalData();
238 I.setHasNoSignedWrap(true);
240 ClearSubclassDataAfterReassociation(I);
249 // Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
250 if (Op1 && Op1->getOpcode() == Opcode) {
251 Value *A = I.getOperand(0);
252 Value *B = Op1->getOperand(0);
253 Value *C = Op1->getOperand(1);
255 // Does "A op B" simplify?
256 if (Value *V = SimplifyBinOp(Opcode, A, B, DL)) {
257 // It simplifies to V. Form "V op C".
260 // Conservatively clear the optional flags, since they may not be
261 // preserved by the reassociation.
262 ClearSubclassDataAfterReassociation(I);
270 if (I.isAssociative() && I.isCommutative()) {
271 // Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
272 if (Op0 && Op0->getOpcode() == Opcode) {
273 Value *A = Op0->getOperand(0);
274 Value *B = Op0->getOperand(1);
275 Value *C = I.getOperand(1);
277 // Does "C op A" simplify?
278 if (Value *V = SimplifyBinOp(Opcode, C, A, DL)) {
279 // It simplifies to V. Form "V op B".
282 // Conservatively clear the optional flags, since they may not be
283 // preserved by the reassociation.
284 ClearSubclassDataAfterReassociation(I);
291 // Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
292 if (Op1 && Op1->getOpcode() == Opcode) {
293 Value *A = I.getOperand(0);
294 Value *B = Op1->getOperand(0);
295 Value *C = Op1->getOperand(1);
297 // Does "C op A" simplify?
298 if (Value *V = SimplifyBinOp(Opcode, C, A, DL)) {
299 // It simplifies to V. Form "B op V".
302 // Conservatively clear the optional flags, since they may not be
303 // preserved by the reassociation.
304 ClearSubclassDataAfterReassociation(I);
311 // Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
312 // if C1 and C2 are constants.
314 Op0->getOpcode() == Opcode && Op1->getOpcode() == Opcode &&
315 isa<Constant>(Op0->getOperand(1)) &&
316 isa<Constant>(Op1->getOperand(1)) &&
317 Op0->hasOneUse() && Op1->hasOneUse()) {
318 Value *A = Op0->getOperand(0);
319 Constant *C1 = cast<Constant>(Op0->getOperand(1));
320 Value *B = Op1->getOperand(0);
321 Constant *C2 = cast<Constant>(Op1->getOperand(1));
323 Constant *Folded = ConstantExpr::get(Opcode, C1, C2);
324 BinaryOperator *New = BinaryOperator::Create(Opcode, A, B);
325 if (isa<FPMathOperator>(New)) {
326 FastMathFlags Flags = I.getFastMathFlags();
327 Flags &= Op0->getFastMathFlags();
328 Flags &= Op1->getFastMathFlags();
329 New->setFastMathFlags(Flags);
331 InsertNewInstWith(New, I);
333 I.setOperand(0, New);
334 I.setOperand(1, Folded);
335 // Conservatively clear the optional flags, since they may not be
336 // preserved by the reassociation.
337 ClearSubclassDataAfterReassociation(I);
344 // No further simplifications.
349 /// LeftDistributesOverRight - Whether "X LOp (Y ROp Z)" is always equal to
350 /// "(X LOp Y) ROp (X LOp Z)".
351 static bool LeftDistributesOverRight(Instruction::BinaryOps LOp,
352 Instruction::BinaryOps ROp) {
357 case Instruction::And:
358 // And distributes over Or and Xor.
362 case Instruction::Or:
363 case Instruction::Xor:
367 case Instruction::Mul:
368 // Multiplication distributes over addition and subtraction.
372 case Instruction::Add:
373 case Instruction::Sub:
377 case Instruction::Or:
378 // Or distributes over And.
382 case Instruction::And:
388 /// RightDistributesOverLeft - Whether "(X LOp Y) ROp Z" is always equal to
389 /// "(X ROp Z) LOp (Y ROp Z)".
390 static bool RightDistributesOverLeft(Instruction::BinaryOps LOp,
391 Instruction::BinaryOps ROp) {
392 if (Instruction::isCommutative(ROp))
393 return LeftDistributesOverRight(ROp, LOp);
398 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
399 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
400 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
401 case Instruction::And:
402 case Instruction::Or:
403 case Instruction::Xor:
407 case Instruction::Shl:
408 case Instruction::LShr:
409 case Instruction::AShr:
413 // TODO: It would be nice to handle division, aka "(X + Y)/Z = X/Z + Y/Z",
414 // but this requires knowing that the addition does not overflow and other
419 /// This function returns identity value for given opcode, which can be used to
420 /// factor patterns like (X * 2) + X ==> (X * 2) + (X * 1) ==> X * (2 + 1).
421 static Value *getIdentityValue(Instruction::BinaryOps OpCode, Value *V) {
422 if (isa<Constant>(V))
425 if (OpCode == Instruction::Mul)
426 return ConstantInt::get(V->getType(), 1);
428 // TODO: We can handle other cases e.g. Instruction::And, Instruction::Or etc.
433 /// This function factors binary ops which can be combined using distributive
434 /// laws. This function tries to transform 'Op' based TopLevelOpcode to enable
435 /// factorization e.g for ADD(SHL(X , 2), MUL(X, 5)), When this function called
436 /// with TopLevelOpcode == Instruction::Add and Op = SHL(X, 2), transforms
437 /// SHL(X, 2) to MUL(X, 4) i.e. returns Instruction::Mul with LHS set to 'X' and
439 static Instruction::BinaryOps
440 getBinOpsForFactorization(Instruction::BinaryOps TopLevelOpcode,
441 BinaryOperator *Op, Value *&LHS, Value *&RHS) {
443 return Instruction::BinaryOpsEnd;
445 LHS = Op->getOperand(0);
446 RHS = Op->getOperand(1);
448 switch (TopLevelOpcode) {
450 return Op->getOpcode();
452 case Instruction::Add:
453 case Instruction::Sub:
454 if (Op->getOpcode() == Instruction::Shl) {
455 if (Constant *CST = dyn_cast<Constant>(Op->getOperand(1))) {
456 // The multiplier is really 1 << CST.
457 RHS = ConstantExpr::getShl(ConstantInt::get(Op->getType(), 1), CST);
458 return Instruction::Mul;
461 return Op->getOpcode();
464 // TODO: We can add other conversions e.g. shr => div etc.
467 /// This tries to simplify binary operations by factorizing out common terms
468 /// (e. g. "(A*B)+(A*C)" -> "A*(B+C)").
469 static Value *tryFactorization(InstCombiner::BuilderTy *Builder,
470 const DataLayout *DL, BinaryOperator &I,
471 Instruction::BinaryOps InnerOpcode, Value *A,
472 Value *B, Value *C, Value *D) {
474 // If any of A, B, C, D are null, we can not factor I, return early.
475 // Checking A and C should be enough.
476 if (!A || !C || !B || !D)
479 Value *SimplifiedInst = nullptr;
480 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
481 Instruction::BinaryOps TopLevelOpcode = I.getOpcode();
483 // Does "X op' Y" always equal "Y op' X"?
484 bool InnerCommutative = Instruction::isCommutative(InnerOpcode);
486 // Does "X op' (Y op Z)" always equal "(X op' Y) op (X op' Z)"?
487 if (LeftDistributesOverRight(InnerOpcode, TopLevelOpcode))
488 // Does the instruction have the form "(A op' B) op (A op' D)" or, in the
489 // commutative case, "(A op' B) op (C op' A)"?
490 if (A == C || (InnerCommutative && A == D)) {
493 // Consider forming "A op' (B op D)".
494 // If "B op D" simplifies then it can be formed with no cost.
495 Value *V = SimplifyBinOp(TopLevelOpcode, B, D, DL);
496 // If "B op D" doesn't simplify then only go on if both of the existing
497 // operations "A op' B" and "C op' D" will be zapped as no longer used.
498 if (!V && LHS->hasOneUse() && RHS->hasOneUse())
499 V = Builder->CreateBinOp(TopLevelOpcode, B, D, RHS->getName());
501 SimplifiedInst = Builder->CreateBinOp(InnerOpcode, A, V);
505 // Does "(X op Y) op' Z" always equal "(X op' Z) op (Y op' Z)"?
506 if (!SimplifiedInst && RightDistributesOverLeft(TopLevelOpcode, InnerOpcode))
507 // Does the instruction have the form "(A op' B) op (C op' B)" or, in the
508 // commutative case, "(A op' B) op (B op' D)"?
509 if (B == D || (InnerCommutative && B == C)) {
512 // Consider forming "(A op C) op' B".
513 // If "A op C" simplifies then it can be formed with no cost.
514 Value *V = SimplifyBinOp(TopLevelOpcode, A, C, DL);
516 // If "A op C" doesn't simplify then only go on if both of the existing
517 // operations "A op' B" and "C op' D" will be zapped as no longer used.
518 if (!V && LHS->hasOneUse() && RHS->hasOneUse())
519 V = Builder->CreateBinOp(TopLevelOpcode, A, C, LHS->getName());
521 SimplifiedInst = Builder->CreateBinOp(InnerOpcode, V, B);
525 if (SimplifiedInst) {
527 SimplifiedInst->takeName(&I);
529 // Check if we can add NSW flag to SimplifiedInst. If so, set NSW flag.
530 // TODO: Check for NUW.
531 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(SimplifiedInst)) {
532 if (isa<OverflowingBinaryOperator>(SimplifiedInst)) {
534 if (isa<OverflowingBinaryOperator>(&I))
535 HasNSW = I.hasNoSignedWrap();
537 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS))
538 if (isa<OverflowingBinaryOperator>(Op0))
539 HasNSW &= Op0->hasNoSignedWrap();
541 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS))
542 if (isa<OverflowingBinaryOperator>(Op1))
543 HasNSW &= Op1->hasNoSignedWrap();
544 BO->setHasNoSignedWrap(HasNSW);
548 return SimplifiedInst;
551 /// SimplifyUsingDistributiveLaws - This tries to simplify binary operations
552 /// which some other binary operation distributes over either by factorizing
553 /// out common terms (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this
554 /// results in simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is
555 /// a win). Returns the simplified value, or null if it didn't simplify.
556 Value *InstCombiner::SimplifyUsingDistributiveLaws(BinaryOperator &I) {
557 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
558 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
559 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
562 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
563 auto TopLevelOpcode = I.getOpcode();
564 auto LHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op0, A, B);
565 auto RHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op1, C, D);
567 // The instruction has the form "(A op' B) op (C op' D)". Try to factorize
569 if (LHSOpcode == RHSOpcode) {
570 if (Value *V = tryFactorization(Builder, DL, I, LHSOpcode, A, B, C, D))
574 // The instruction has the form "(A op' B) op (C)". Try to factorize common
576 if (Value *V = tryFactorization(Builder, DL, I, LHSOpcode, A, B, RHS,
577 getIdentityValue(LHSOpcode, RHS)))
580 // The instruction has the form "(B) op (C op' D)". Try to factorize common
582 if (Value *V = tryFactorization(Builder, DL, I, RHSOpcode, LHS,
583 getIdentityValue(RHSOpcode, LHS), C, D))
587 if (Op0 && RightDistributesOverLeft(Op0->getOpcode(), TopLevelOpcode)) {
588 // The instruction has the form "(A op' B) op C". See if expanding it out
589 // to "(A op C) op' (B op C)" results in simplifications.
590 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
591 Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op'
593 // Do "A op C" and "B op C" both simplify?
594 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, C, DL))
595 if (Value *R = SimplifyBinOp(TopLevelOpcode, B, C, DL)) {
596 // They do! Return "L op' R".
598 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
599 if ((L == A && R == B) ||
600 (Instruction::isCommutative(InnerOpcode) && L == B && R == A))
602 // Otherwise return "L op' R" if it simplifies.
603 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, DL))
605 // Otherwise, create a new instruction.
606 C = Builder->CreateBinOp(InnerOpcode, L, R);
612 if (Op1 && LeftDistributesOverRight(TopLevelOpcode, Op1->getOpcode())) {
613 // The instruction has the form "A op (B op' C)". See if expanding it out
614 // to "(A op B) op' (A op C)" results in simplifications.
615 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
616 Instruction::BinaryOps InnerOpcode = Op1->getOpcode(); // op'
618 // Do "A op B" and "A op C" both simplify?
619 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, B, DL))
620 if (Value *R = SimplifyBinOp(TopLevelOpcode, A, C, DL)) {
621 // They do! Return "L op' R".
623 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
624 if ((L == B && R == C) ||
625 (Instruction::isCommutative(InnerOpcode) && L == C && R == B))
627 // Otherwise return "L op' R" if it simplifies.
628 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, DL))
630 // Otherwise, create a new instruction.
631 A = Builder->CreateBinOp(InnerOpcode, L, R);
640 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
641 // if the LHS is a constant zero (which is the 'negate' form).
643 Value *InstCombiner::dyn_castNegVal(Value *V) const {
644 if (BinaryOperator::isNeg(V))
645 return BinaryOperator::getNegArgument(V);
647 // Constants can be considered to be negated values if they can be folded.
648 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
649 return ConstantExpr::getNeg(C);
651 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
652 if (C->getType()->getElementType()->isIntegerTy())
653 return ConstantExpr::getNeg(C);
658 // dyn_castFNegVal - Given a 'fsub' instruction, return the RHS of the
659 // instruction if the LHS is a constant negative zero (which is the 'negate'
662 Value *InstCombiner::dyn_castFNegVal(Value *V, bool IgnoreZeroSign) const {
663 if (BinaryOperator::isFNeg(V, IgnoreZeroSign))
664 return BinaryOperator::getFNegArgument(V);
666 // Constants can be considered to be negated values if they can be folded.
667 if (ConstantFP *C = dyn_cast<ConstantFP>(V))
668 return ConstantExpr::getFNeg(C);
670 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
671 if (C->getType()->getElementType()->isFloatingPointTy())
672 return ConstantExpr::getFNeg(C);
677 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
679 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
680 return IC->Builder->CreateCast(CI->getOpcode(), SO, I.getType());
683 // Figure out if the constant is the left or the right argument.
684 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
685 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
687 if (Constant *SOC = dyn_cast<Constant>(SO)) {
689 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
690 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
693 Value *Op0 = SO, *Op1 = ConstOperand;
697 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I)) {
698 Value *RI = IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1,
699 SO->getName()+".op");
700 Instruction *FPInst = dyn_cast<Instruction>(RI);
701 if (FPInst && isa<FPMathOperator>(FPInst))
702 FPInst->copyFastMathFlags(BO);
705 if (ICmpInst *CI = dyn_cast<ICmpInst>(&I))
706 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
707 SO->getName()+".cmp");
708 if (FCmpInst *CI = dyn_cast<FCmpInst>(&I))
709 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
710 SO->getName()+".cmp");
711 llvm_unreachable("Unknown binary instruction type!");
714 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
715 // constant as the other operand, try to fold the binary operator into the
716 // select arguments. This also works for Cast instructions, which obviously do
717 // not have a second operand.
718 Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) {
719 // Don't modify shared select instructions
720 if (!SI->hasOneUse()) return nullptr;
721 Value *TV = SI->getOperand(1);
722 Value *FV = SI->getOperand(2);
724 if (isa<Constant>(TV) || isa<Constant>(FV)) {
725 // Bool selects with constant operands can be folded to logical ops.
726 if (SI->getType()->isIntegerTy(1)) return nullptr;
728 // If it's a bitcast involving vectors, make sure it has the same number of
729 // elements on both sides.
730 if (BitCastInst *BC = dyn_cast<BitCastInst>(&Op)) {
731 VectorType *DestTy = dyn_cast<VectorType>(BC->getDestTy());
732 VectorType *SrcTy = dyn_cast<VectorType>(BC->getSrcTy());
734 // Verify that either both or neither are vectors.
735 if ((SrcTy == nullptr) != (DestTy == nullptr)) return nullptr;
736 // If vectors, verify that they have the same number of elements.
737 if (SrcTy && SrcTy->getNumElements() != DestTy->getNumElements())
741 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, this);
742 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, this);
744 return SelectInst::Create(SI->getCondition(),
745 SelectTrueVal, SelectFalseVal);
751 /// FoldOpIntoPhi - Given a binary operator, cast instruction, or select which
752 /// has a PHI node as operand #0, see if we can fold the instruction into the
753 /// PHI (which is only possible if all operands to the PHI are constants).
755 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
756 PHINode *PN = cast<PHINode>(I.getOperand(0));
757 unsigned NumPHIValues = PN->getNumIncomingValues();
758 if (NumPHIValues == 0)
761 // We normally only transform phis with a single use. However, if a PHI has
762 // multiple uses and they are all the same operation, we can fold *all* of the
763 // uses into the PHI.
764 if (!PN->hasOneUse()) {
765 // Walk the use list for the instruction, comparing them to I.
766 for (User *U : PN->users()) {
767 Instruction *UI = cast<Instruction>(U);
768 if (UI != &I && !I.isIdenticalTo(UI))
771 // Otherwise, we can replace *all* users with the new PHI we form.
774 // Check to see if all of the operands of the PHI are simple constants
775 // (constantint/constantfp/undef). If there is one non-constant value,
776 // remember the BB it is in. If there is more than one or if *it* is a PHI,
777 // bail out. We don't do arbitrary constant expressions here because moving
778 // their computation can be expensive without a cost model.
779 BasicBlock *NonConstBB = nullptr;
780 for (unsigned i = 0; i != NumPHIValues; ++i) {
781 Value *InVal = PN->getIncomingValue(i);
782 if (isa<Constant>(InVal) && !isa<ConstantExpr>(InVal))
785 if (isa<PHINode>(InVal)) return nullptr; // Itself a phi.
786 if (NonConstBB) return nullptr; // More than one non-const value.
788 NonConstBB = PN->getIncomingBlock(i);
790 // If the InVal is an invoke at the end of the pred block, then we can't
791 // insert a computation after it without breaking the edge.
792 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal))
793 if (II->getParent() == NonConstBB)
796 // If the incoming non-constant value is in I's block, we will remove one
797 // instruction, but insert another equivalent one, leading to infinite
799 if (isPotentiallyReachable(I.getParent(), NonConstBB, DT,
800 getAnalysisIfAvailable<LoopInfo>()))
804 // If there is exactly one non-constant value, we can insert a copy of the
805 // operation in that block. However, if this is a critical edge, we would be
806 // inserting the computation on some other paths (e.g. inside a loop). Only
807 // do this if the pred block is unconditionally branching into the phi block.
808 if (NonConstBB != nullptr) {
809 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
810 if (!BI || !BI->isUnconditional()) return nullptr;
813 // Okay, we can do the transformation: create the new PHI node.
814 PHINode *NewPN = PHINode::Create(I.getType(), PN->getNumIncomingValues());
815 InsertNewInstBefore(NewPN, *PN);
818 // If we are going to have to insert a new computation, do so right before the
819 // predecessors terminator.
821 Builder->SetInsertPoint(NonConstBB->getTerminator());
823 // Next, add all of the operands to the PHI.
824 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) {
825 // We only currently try to fold the condition of a select when it is a phi,
826 // not the true/false values.
827 Value *TrueV = SI->getTrueValue();
828 Value *FalseV = SI->getFalseValue();
829 BasicBlock *PhiTransBB = PN->getParent();
830 for (unsigned i = 0; i != NumPHIValues; ++i) {
831 BasicBlock *ThisBB = PN->getIncomingBlock(i);
832 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB);
833 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB);
834 Value *InV = nullptr;
835 // Beware of ConstantExpr: it may eventually evaluate to getNullValue,
836 // even if currently isNullValue gives false.
837 Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i));
838 if (InC && !isa<ConstantExpr>(InC))
839 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred;
841 InV = Builder->CreateSelect(PN->getIncomingValue(i),
842 TrueVInPred, FalseVInPred, "phitmp");
843 NewPN->addIncoming(InV, ThisBB);
845 } else if (CmpInst *CI = dyn_cast<CmpInst>(&I)) {
846 Constant *C = cast<Constant>(I.getOperand(1));
847 for (unsigned i = 0; i != NumPHIValues; ++i) {
848 Value *InV = nullptr;
849 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
850 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
851 else if (isa<ICmpInst>(CI))
852 InV = Builder->CreateICmp(CI->getPredicate(), PN->getIncomingValue(i),
855 InV = Builder->CreateFCmp(CI->getPredicate(), PN->getIncomingValue(i),
857 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
859 } else if (I.getNumOperands() == 2) {
860 Constant *C = cast<Constant>(I.getOperand(1));
861 for (unsigned i = 0; i != NumPHIValues; ++i) {
862 Value *InV = nullptr;
863 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
864 InV = ConstantExpr::get(I.getOpcode(), InC, C);
866 InV = Builder->CreateBinOp(cast<BinaryOperator>(I).getOpcode(),
867 PN->getIncomingValue(i), C, "phitmp");
868 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
871 CastInst *CI = cast<CastInst>(&I);
872 Type *RetTy = CI->getType();
873 for (unsigned i = 0; i != NumPHIValues; ++i) {
875 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
876 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
878 InV = Builder->CreateCast(CI->getOpcode(),
879 PN->getIncomingValue(i), I.getType(), "phitmp");
880 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
884 for (auto UI = PN->user_begin(), E = PN->user_end(); UI != E;) {
885 Instruction *User = cast<Instruction>(*UI++);
886 if (User == &I) continue;
887 ReplaceInstUsesWith(*User, NewPN);
888 EraseInstFromFunction(*User);
890 return ReplaceInstUsesWith(I, NewPN);
893 /// FindElementAtOffset - Given a pointer type and a constant offset, determine
894 /// whether or not there is a sequence of GEP indices into the pointed type that
895 /// will land us at the specified offset. If so, fill them into NewIndices and
896 /// return the resultant element type, otherwise return null.
897 Type *InstCombiner::FindElementAtOffset(Type *PtrTy, int64_t Offset,
898 SmallVectorImpl<Value*> &NewIndices) {
899 assert(PtrTy->isPtrOrPtrVectorTy());
904 Type *Ty = PtrTy->getPointerElementType();
908 // Start with the index over the outer type. Note that the type size
909 // might be zero (even if the offset isn't zero) if the indexed type
910 // is something like [0 x {int, int}]
911 Type *IntPtrTy = DL->getIntPtrType(PtrTy);
912 int64_t FirstIdx = 0;
913 if (int64_t TySize = DL->getTypeAllocSize(Ty)) {
914 FirstIdx = Offset/TySize;
915 Offset -= FirstIdx*TySize;
917 // Handle hosts where % returns negative instead of values [0..TySize).
923 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
926 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
928 // Index into the types. If we fail, set OrigBase to null.
930 // Indexing into tail padding between struct/array elements.
931 if (uint64_t(Offset*8) >= DL->getTypeSizeInBits(Ty))
934 if (StructType *STy = dyn_cast<StructType>(Ty)) {
935 const StructLayout *SL = DL->getStructLayout(STy);
936 assert(Offset < (int64_t)SL->getSizeInBytes() &&
937 "Offset must stay within the indexed type");
939 unsigned Elt = SL->getElementContainingOffset(Offset);
940 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
943 Offset -= SL->getElementOffset(Elt);
944 Ty = STy->getElementType(Elt);
945 } else if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
946 uint64_t EltSize = DL->getTypeAllocSize(AT->getElementType());
947 assert(EltSize && "Cannot index into a zero-sized array");
948 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
950 Ty = AT->getElementType();
952 // Otherwise, we can't index into the middle of this atomic type, bail.
960 static bool shouldMergeGEPs(GEPOperator &GEP, GEPOperator &Src) {
961 // If this GEP has only 0 indices, it is the same pointer as
962 // Src. If Src is not a trivial GEP too, don't combine
964 if (GEP.hasAllZeroIndices() && !Src.hasAllZeroIndices() &&
970 /// Descale - Return a value X such that Val = X * Scale, or null if none. If
971 /// the multiplication is known not to overflow then NoSignedWrap is set.
972 Value *InstCombiner::Descale(Value *Val, APInt Scale, bool &NoSignedWrap) {
973 assert(isa<IntegerType>(Val->getType()) && "Can only descale integers!");
974 assert(cast<IntegerType>(Val->getType())->getBitWidth() ==
975 Scale.getBitWidth() && "Scale not compatible with value!");
977 // If Val is zero or Scale is one then Val = Val * Scale.
978 if (match(Val, m_Zero()) || Scale == 1) {
983 // If Scale is zero then it does not divide Val.
984 if (Scale.isMinValue())
987 // Look through chains of multiplications, searching for a constant that is
988 // divisible by Scale. For example, descaling X*(Y*(Z*4)) by a factor of 4
989 // will find the constant factor 4 and produce X*(Y*Z). Descaling X*(Y*8) by
990 // a factor of 4 will produce X*(Y*2). The principle of operation is to bore
993 // Val = M1 * X || Analysis starts here and works down
994 // M1 = M2 * Y || Doesn't descend into terms with more
995 // M2 = Z * 4 \/ than one use
997 // Then to modify a term at the bottom:
1000 // M1 = Z * Y || Replaced M2 with Z
1002 // Then to work back up correcting nsw flags.
1004 // Op - the term we are currently analyzing. Starts at Val then drills down.
1005 // Replaced with its descaled value before exiting from the drill down loop.
1008 // Parent - initially null, but after drilling down notes where Op came from.
1009 // In the example above, Parent is (Val, 0) when Op is M1, because M1 is the
1010 // 0'th operand of Val.
1011 std::pair<Instruction*, unsigned> Parent;
1013 // RequireNoSignedWrap - Set if the transform requires a descaling at deeper
1014 // levels that doesn't overflow.
1015 bool RequireNoSignedWrap = false;
1017 // logScale - log base 2 of the scale. Negative if not a power of 2.
1018 int32_t logScale = Scale.exactLogBase2();
1020 for (;; Op = Parent.first->getOperand(Parent.second)) { // Drill down
1022 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
1023 // If Op is a constant divisible by Scale then descale to the quotient.
1024 APInt Quotient(Scale), Remainder(Scale); // Init ensures right bitwidth.
1025 APInt::sdivrem(CI->getValue(), Scale, Quotient, Remainder);
1026 if (!Remainder.isMinValue())
1027 // Not divisible by Scale.
1029 // Replace with the quotient in the parent.
1030 Op = ConstantInt::get(CI->getType(), Quotient);
1031 NoSignedWrap = true;
1035 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op)) {
1037 if (BO->getOpcode() == Instruction::Mul) {
1039 NoSignedWrap = BO->hasNoSignedWrap();
1040 if (RequireNoSignedWrap && !NoSignedWrap)
1043 // There are three cases for multiplication: multiplication by exactly
1044 // the scale, multiplication by a constant different to the scale, and
1045 // multiplication by something else.
1046 Value *LHS = BO->getOperand(0);
1047 Value *RHS = BO->getOperand(1);
1049 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
1050 // Multiplication by a constant.
1051 if (CI->getValue() == Scale) {
1052 // Multiplication by exactly the scale, replace the multiplication
1053 // by its left-hand side in the parent.
1058 // Otherwise drill down into the constant.
1059 if (!Op->hasOneUse())
1062 Parent = std::make_pair(BO, 1);
1066 // Multiplication by something else. Drill down into the left-hand side
1067 // since that's where the reassociate pass puts the good stuff.
1068 if (!Op->hasOneUse())
1071 Parent = std::make_pair(BO, 0);
1075 if (logScale > 0 && BO->getOpcode() == Instruction::Shl &&
1076 isa<ConstantInt>(BO->getOperand(1))) {
1077 // Multiplication by a power of 2.
1078 NoSignedWrap = BO->hasNoSignedWrap();
1079 if (RequireNoSignedWrap && !NoSignedWrap)
1082 Value *LHS = BO->getOperand(0);
1083 int32_t Amt = cast<ConstantInt>(BO->getOperand(1))->
1084 getLimitedValue(Scale.getBitWidth());
1087 if (Amt == logScale) {
1088 // Multiplication by exactly the scale, replace the multiplication
1089 // by its left-hand side in the parent.
1093 if (Amt < logScale || !Op->hasOneUse())
1096 // Multiplication by more than the scale. Reduce the multiplying amount
1097 // by the scale in the parent.
1098 Parent = std::make_pair(BO, 1);
1099 Op = ConstantInt::get(BO->getType(), Amt - logScale);
1104 if (!Op->hasOneUse())
1107 if (CastInst *Cast = dyn_cast<CastInst>(Op)) {
1108 if (Cast->getOpcode() == Instruction::SExt) {
1109 // Op is sign-extended from a smaller type, descale in the smaller type.
1110 unsigned SmallSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1111 APInt SmallScale = Scale.trunc(SmallSize);
1112 // Suppose Op = sext X, and we descale X as Y * SmallScale. We want to
1113 // descale Op as (sext Y) * Scale. In order to have
1114 // sext (Y * SmallScale) = (sext Y) * Scale
1115 // some conditions need to hold however: SmallScale must sign-extend to
1116 // Scale and the multiplication Y * SmallScale should not overflow.
1117 if (SmallScale.sext(Scale.getBitWidth()) != Scale)
1118 // SmallScale does not sign-extend to Scale.
1120 assert(SmallScale.exactLogBase2() == logScale);
1121 // Require that Y * SmallScale must not overflow.
1122 RequireNoSignedWrap = true;
1124 // Drill down through the cast.
1125 Parent = std::make_pair(Cast, 0);
1130 if (Cast->getOpcode() == Instruction::Trunc) {
1131 // Op is truncated from a larger type, descale in the larger type.
1132 // Suppose Op = trunc X, and we descale X as Y * sext Scale. Then
1133 // trunc (Y * sext Scale) = (trunc Y) * Scale
1134 // always holds. However (trunc Y) * Scale may overflow even if
1135 // trunc (Y * sext Scale) does not, so nsw flags need to be cleared
1136 // from this point up in the expression (see later).
1137 if (RequireNoSignedWrap)
1140 // Drill down through the cast.
1141 unsigned LargeSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1142 Parent = std::make_pair(Cast, 0);
1143 Scale = Scale.sext(LargeSize);
1144 if (logScale + 1 == (int32_t)Cast->getType()->getPrimitiveSizeInBits())
1146 assert(Scale.exactLogBase2() == logScale);
1151 // Unsupported expression, bail out.
1155 // If Op is zero then Val = Op * Scale.
1156 if (match(Op, m_Zero())) {
1157 NoSignedWrap = true;
1161 // We know that we can successfully descale, so from here on we can safely
1162 // modify the IR. Op holds the descaled version of the deepest term in the
1163 // expression. NoSignedWrap is 'true' if multiplying Op by Scale is known
1167 // The expression only had one term.
1170 // Rewrite the parent using the descaled version of its operand.
1171 assert(Parent.first->hasOneUse() && "Drilled down when more than one use!");
1172 assert(Op != Parent.first->getOperand(Parent.second) &&
1173 "Descaling was a no-op?");
1174 Parent.first->setOperand(Parent.second, Op);
1175 Worklist.Add(Parent.first);
1177 // Now work back up the expression correcting nsw flags. The logic is based
1178 // on the following observation: if X * Y is known not to overflow as a signed
1179 // multiplication, and Y is replaced by a value Z with smaller absolute value,
1180 // then X * Z will not overflow as a signed multiplication either. As we work
1181 // our way up, having NoSignedWrap 'true' means that the descaled value at the
1182 // current level has strictly smaller absolute value than the original.
1183 Instruction *Ancestor = Parent.first;
1185 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Ancestor)) {
1186 // If the multiplication wasn't nsw then we can't say anything about the
1187 // value of the descaled multiplication, and we have to clear nsw flags
1188 // from this point on up.
1189 bool OpNoSignedWrap = BO->hasNoSignedWrap();
1190 NoSignedWrap &= OpNoSignedWrap;
1191 if (NoSignedWrap != OpNoSignedWrap) {
1192 BO->setHasNoSignedWrap(NoSignedWrap);
1193 Worklist.Add(Ancestor);
1195 } else if (Ancestor->getOpcode() == Instruction::Trunc) {
1196 // The fact that the descaled input to the trunc has smaller absolute
1197 // value than the original input doesn't tell us anything useful about
1198 // the absolute values of the truncations.
1199 NoSignedWrap = false;
1201 assert((Ancestor->getOpcode() != Instruction::SExt || NoSignedWrap) &&
1202 "Failed to keep proper track of nsw flags while drilling down?");
1204 if (Ancestor == Val)
1205 // Got to the top, all done!
1208 // Move up one level in the expression.
1209 assert(Ancestor->hasOneUse() && "Drilled down when more than one use!");
1210 Ancestor = Ancestor->user_back();
1214 /// \brief Creates node of binary operation with the same attributes as the
1215 /// specified one but with other operands.
1216 static Value *CreateBinOpAsGiven(BinaryOperator &Inst, Value *LHS, Value *RHS,
1217 InstCombiner::BuilderTy *B) {
1218 Value *BORes = B->CreateBinOp(Inst.getOpcode(), LHS, RHS);
1219 if (BinaryOperator *NewBO = dyn_cast<BinaryOperator>(BORes)) {
1220 if (isa<OverflowingBinaryOperator>(NewBO)) {
1221 NewBO->setHasNoSignedWrap(Inst.hasNoSignedWrap());
1222 NewBO->setHasNoUnsignedWrap(Inst.hasNoUnsignedWrap());
1224 if (isa<PossiblyExactOperator>(NewBO))
1225 NewBO->setIsExact(Inst.isExact());
1230 /// \brief Makes transformation of binary operation specific for vector types.
1231 /// \param Inst Binary operator to transform.
1232 /// \return Pointer to node that must replace the original binary operator, or
1233 /// null pointer if no transformation was made.
1234 Value *InstCombiner::SimplifyVectorOp(BinaryOperator &Inst) {
1235 if (!Inst.getType()->isVectorTy()) return nullptr;
1237 // It may not be safe to reorder shuffles and things like div, urem, etc.
1238 // because we may trap when executing those ops on unknown vector elements.
1240 if (!isSafeToSpeculativelyExecute(&Inst, DL)) return nullptr;
1242 unsigned VWidth = cast<VectorType>(Inst.getType())->getNumElements();
1243 Value *LHS = Inst.getOperand(0), *RHS = Inst.getOperand(1);
1244 assert(cast<VectorType>(LHS->getType())->getNumElements() == VWidth);
1245 assert(cast<VectorType>(RHS->getType())->getNumElements() == VWidth);
1247 // If both arguments of binary operation are shuffles, which use the same
1248 // mask and shuffle within a single vector, it is worthwhile to move the
1249 // shuffle after binary operation:
1250 // Op(shuffle(v1, m), shuffle(v2, m)) -> shuffle(Op(v1, v2), m)
1251 if (isa<ShuffleVectorInst>(LHS) && isa<ShuffleVectorInst>(RHS)) {
1252 ShuffleVectorInst *LShuf = cast<ShuffleVectorInst>(LHS);
1253 ShuffleVectorInst *RShuf = cast<ShuffleVectorInst>(RHS);
1254 if (isa<UndefValue>(LShuf->getOperand(1)) &&
1255 isa<UndefValue>(RShuf->getOperand(1)) &&
1256 LShuf->getOperand(0)->getType() == RShuf->getOperand(0)->getType() &&
1257 LShuf->getMask() == RShuf->getMask()) {
1258 Value *NewBO = CreateBinOpAsGiven(Inst, LShuf->getOperand(0),
1259 RShuf->getOperand(0), Builder);
1260 Value *Res = Builder->CreateShuffleVector(NewBO,
1261 UndefValue::get(NewBO->getType()), LShuf->getMask());
1266 // If one argument is a shuffle within one vector, the other is a constant,
1267 // try moving the shuffle after the binary operation.
1268 ShuffleVectorInst *Shuffle = nullptr;
1269 Constant *C1 = nullptr;
1270 if (isa<ShuffleVectorInst>(LHS)) Shuffle = cast<ShuffleVectorInst>(LHS);
1271 if (isa<ShuffleVectorInst>(RHS)) Shuffle = cast<ShuffleVectorInst>(RHS);
1272 if (isa<Constant>(LHS)) C1 = cast<Constant>(LHS);
1273 if (isa<Constant>(RHS)) C1 = cast<Constant>(RHS);
1274 if (Shuffle && C1 &&
1275 (isa<ConstantVector>(C1) || isa<ConstantDataVector>(C1)) &&
1276 isa<UndefValue>(Shuffle->getOperand(1)) &&
1277 Shuffle->getType() == Shuffle->getOperand(0)->getType()) {
1278 SmallVector<int, 16> ShMask = Shuffle->getShuffleMask();
1279 // Find constant C2 that has property:
1280 // shuffle(C2, ShMask) = C1
1281 // If such constant does not exist (example: ShMask=<0,0> and C1=<1,2>)
1282 // reorder is not possible.
1283 SmallVector<Constant*, 16> C2M(VWidth,
1284 UndefValue::get(C1->getType()->getScalarType()));
1285 bool MayChange = true;
1286 for (unsigned I = 0; I < VWidth; ++I) {
1287 if (ShMask[I] >= 0) {
1288 assert(ShMask[I] < (int)VWidth);
1289 if (!isa<UndefValue>(C2M[ShMask[I]])) {
1293 C2M[ShMask[I]] = C1->getAggregateElement(I);
1297 Constant *C2 = ConstantVector::get(C2M);
1298 Value *NewLHS, *NewRHS;
1299 if (isa<Constant>(LHS)) {
1301 NewRHS = Shuffle->getOperand(0);
1303 NewLHS = Shuffle->getOperand(0);
1306 Value *NewBO = CreateBinOpAsGiven(Inst, NewLHS, NewRHS, Builder);
1307 Value *Res = Builder->CreateShuffleVector(NewBO,
1308 UndefValue::get(Inst.getType()), Shuffle->getMask());
1316 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
1317 SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end());
1319 if (Value *V = SimplifyGEPInst(Ops, DL, TLI, DT, AT))
1320 return ReplaceInstUsesWith(GEP, V);
1322 Value *PtrOp = GEP.getOperand(0);
1324 // Eliminate unneeded casts for indices, and replace indices which displace
1325 // by multiples of a zero size type with zero.
1327 bool MadeChange = false;
1328 Type *IntPtrTy = DL->getIntPtrType(GEP.getPointerOperandType());
1330 gep_type_iterator GTI = gep_type_begin(GEP);
1331 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end();
1332 I != E; ++I, ++GTI) {
1333 // Skip indices into struct types.
1334 SequentialType *SeqTy = dyn_cast<SequentialType>(*GTI);
1335 if (!SeqTy) continue;
1337 // If the element type has zero size then any index over it is equivalent
1338 // to an index of zero, so replace it with zero if it is not zero already.
1339 if (SeqTy->getElementType()->isSized() &&
1340 DL->getTypeAllocSize(SeqTy->getElementType()) == 0)
1341 if (!isa<Constant>(*I) || !cast<Constant>(*I)->isNullValue()) {
1342 *I = Constant::getNullValue(IntPtrTy);
1346 Type *IndexTy = (*I)->getType();
1347 if (IndexTy != IntPtrTy) {
1348 // If we are using a wider index than needed for this platform, shrink
1349 // it to what we need. If narrower, sign-extend it to what we need.
1350 // This explicit cast can make subsequent optimizations more obvious.
1351 *I = Builder->CreateIntCast(*I, IntPtrTy, true);
1355 if (MadeChange) return &GEP;
1358 // Check to see if the inputs to the PHI node are getelementptr instructions.
1359 if (PHINode *PN = dyn_cast<PHINode>(PtrOp)) {
1360 GetElementPtrInst *Op1 = dyn_cast<GetElementPtrInst>(PN->getOperand(0));
1366 for (auto I = PN->op_begin()+1, E = PN->op_end(); I !=E; ++I) {
1367 GetElementPtrInst *Op2 = dyn_cast<GetElementPtrInst>(*I);
1368 if (!Op2 || Op1->getNumOperands() != Op2->getNumOperands())
1371 // Keep track of the type as we walk the GEP.
1372 Type *CurTy = Op1->getOperand(0)->getType()->getScalarType();
1374 for (unsigned J = 0, F = Op1->getNumOperands(); J != F; ++J) {
1375 if (Op1->getOperand(J)->getType() != Op2->getOperand(J)->getType())
1378 if (Op1->getOperand(J) != Op2->getOperand(J)) {
1380 // We have not seen any differences yet in the GEPs feeding the
1381 // PHI yet, so we record this one if it is allowed to be a
1384 // The first two arguments can vary for any GEP, the rest have to be
1385 // static for struct slots
1386 if (J > 1 && CurTy->isStructTy())
1391 // The GEP is different by more than one input. While this could be
1392 // extended to support GEPs that vary by more than one variable it
1393 // doesn't make sense since it greatly increases the complexity and
1394 // would result in an R+R+R addressing mode which no backend
1395 // directly supports and would need to be broken into several
1396 // simpler instructions anyway.
1401 // Sink down a layer of the type for the next iteration.
1403 if (CompositeType *CT = dyn_cast<CompositeType>(CurTy)) {
1404 CurTy = CT->getTypeAtIndex(Op1->getOperand(J));
1412 GetElementPtrInst *NewGEP = cast<GetElementPtrInst>(Op1->clone());
1415 // All the GEPs feeding the PHI are identical. Clone one down into our
1416 // BB so that it can be merged with the current GEP.
1417 GEP.getParent()->getInstList().insert(GEP.getParent()->getFirstNonPHI(),
1420 // All the GEPs feeding the PHI differ at a single offset. Clone a GEP
1421 // into the current block so it can be merged, and create a new PHI to
1423 Instruction *InsertPt = Builder->GetInsertPoint();
1424 Builder->SetInsertPoint(PN);
1425 PHINode *NewPN = Builder->CreatePHI(Op1->getOperand(DI)->getType(),
1426 PN->getNumOperands());
1427 Builder->SetInsertPoint(InsertPt);
1429 for (auto &I : PN->operands())
1430 NewPN->addIncoming(cast<GEPOperator>(I)->getOperand(DI),
1431 PN->getIncomingBlock(I));
1433 NewGEP->setOperand(DI, NewPN);
1434 GEP.getParent()->getInstList().insert(GEP.getParent()->getFirstNonPHI(),
1436 NewGEP->setOperand(DI, NewPN);
1439 GEP.setOperand(0, NewGEP);
1443 // Combine Indices - If the source pointer to this getelementptr instruction
1444 // is a getelementptr instruction, combine the indices of the two
1445 // getelementptr instructions into a single instruction.
1447 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) {
1448 if (!shouldMergeGEPs(*cast<GEPOperator>(&GEP), *Src))
1451 // Note that if our source is a gep chain itself then we wait for that
1452 // chain to be resolved before we perform this transformation. This
1453 // avoids us creating a TON of code in some cases.
1454 if (GEPOperator *SrcGEP =
1455 dyn_cast<GEPOperator>(Src->getOperand(0)))
1456 if (SrcGEP->getNumOperands() == 2 && shouldMergeGEPs(*Src, *SrcGEP))
1457 return nullptr; // Wait until our source is folded to completion.
1459 SmallVector<Value*, 8> Indices;
1461 // Find out whether the last index in the source GEP is a sequential idx.
1462 bool EndsWithSequential = false;
1463 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src);
1465 EndsWithSequential = !(*I)->isStructTy();
1467 // Can we combine the two pointer arithmetics offsets?
1468 if (EndsWithSequential) {
1469 // Replace: gep (gep %P, long B), long A, ...
1470 // With: T = long A+B; gep %P, T, ...
1473 Value *SO1 = Src->getOperand(Src->getNumOperands()-1);
1474 Value *GO1 = GEP.getOperand(1);
1475 if (SO1 == Constant::getNullValue(SO1->getType())) {
1477 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
1480 // If they aren't the same type, then the input hasn't been processed
1481 // by the loop above yet (which canonicalizes sequential index types to
1482 // intptr_t). Just avoid transforming this until the input has been
1484 if (SO1->getType() != GO1->getType())
1486 Sum = Builder->CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
1489 // Update the GEP in place if possible.
1490 if (Src->getNumOperands() == 2) {
1491 GEP.setOperand(0, Src->getOperand(0));
1492 GEP.setOperand(1, Sum);
1495 Indices.append(Src->op_begin()+1, Src->op_end()-1);
1496 Indices.push_back(Sum);
1497 Indices.append(GEP.op_begin()+2, GEP.op_end());
1498 } else if (isa<Constant>(*GEP.idx_begin()) &&
1499 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
1500 Src->getNumOperands() != 1) {
1501 // Otherwise we can do the fold if the first index of the GEP is a zero
1502 Indices.append(Src->op_begin()+1, Src->op_end());
1503 Indices.append(GEP.idx_begin()+1, GEP.idx_end());
1506 if (!Indices.empty())
1507 return (GEP.isInBounds() && Src->isInBounds()) ?
1508 GetElementPtrInst::CreateInBounds(Src->getOperand(0), Indices,
1510 GetElementPtrInst::Create(Src->getOperand(0), Indices, GEP.getName());
1513 if (DL && GEP.getNumIndices() == 1) {
1514 unsigned AS = GEP.getPointerAddressSpace();
1515 if (GEP.getOperand(1)->getType()->getScalarSizeInBits() ==
1516 DL->getPointerSizeInBits(AS)) {
1517 Type *PtrTy = GEP.getPointerOperandType();
1518 Type *Ty = PtrTy->getPointerElementType();
1519 uint64_t TyAllocSize = DL->getTypeAllocSize(Ty);
1521 bool Matched = false;
1524 if (TyAllocSize == 1) {
1525 V = GEP.getOperand(1);
1527 } else if (match(GEP.getOperand(1),
1528 m_AShr(m_Value(V), m_ConstantInt(C)))) {
1529 if (TyAllocSize == 1ULL << C)
1531 } else if (match(GEP.getOperand(1),
1532 m_SDiv(m_Value(V), m_ConstantInt(C)))) {
1533 if (TyAllocSize == C)
1538 // Canonicalize (gep i8* X, -(ptrtoint Y))
1539 // to (inttoptr (sub (ptrtoint X), (ptrtoint Y)))
1540 // The GEP pattern is emitted by the SCEV expander for certain kinds of
1541 // pointer arithmetic.
1542 if (match(V, m_Neg(m_PtrToInt(m_Value())))) {
1543 Operator *Index = cast<Operator>(V);
1544 Value *PtrToInt = Builder->CreatePtrToInt(PtrOp, Index->getType());
1545 Value *NewSub = Builder->CreateSub(PtrToInt, Index->getOperand(1));
1546 return CastInst::Create(Instruction::IntToPtr, NewSub, GEP.getType());
1548 // Canonicalize (gep i8* X, (ptrtoint Y)-(ptrtoint X))
1551 if (match(V, m_Sub(m_PtrToInt(m_Value(Y)),
1552 m_PtrToInt(m_Specific(GEP.getOperand(0)))))) {
1553 return CastInst::CreatePointerBitCastOrAddrSpaceCast(Y,
1560 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0).
1561 Value *StrippedPtr = PtrOp->stripPointerCasts();
1562 PointerType *StrippedPtrTy = dyn_cast<PointerType>(StrippedPtr->getType());
1564 // We do not handle pointer-vector geps here.
1568 if (StrippedPtr != PtrOp) {
1569 bool HasZeroPointerIndex = false;
1570 if (ConstantInt *C = dyn_cast<ConstantInt>(GEP.getOperand(1)))
1571 HasZeroPointerIndex = C->isZero();
1573 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
1574 // into : GEP [10 x i8]* X, i32 0, ...
1576 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
1577 // into : GEP i8* X, ...
1579 // This occurs when the program declares an array extern like "int X[];"
1580 if (HasZeroPointerIndex) {
1581 PointerType *CPTy = cast<PointerType>(PtrOp->getType());
1582 if (ArrayType *CATy =
1583 dyn_cast<ArrayType>(CPTy->getElementType())) {
1584 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
1585 if (CATy->getElementType() == StrippedPtrTy->getElementType()) {
1586 // -> GEP i8* X, ...
1587 SmallVector<Value*, 8> Idx(GEP.idx_begin()+1, GEP.idx_end());
1588 GetElementPtrInst *Res =
1589 GetElementPtrInst::Create(StrippedPtr, Idx, GEP.getName());
1590 Res->setIsInBounds(GEP.isInBounds());
1591 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace())
1593 // Insert Res, and create an addrspacecast.
1595 // GEP (addrspacecast i8 addrspace(1)* X to [0 x i8]*), i32 0, ...
1597 // %0 = GEP i8 addrspace(1)* X, ...
1598 // addrspacecast i8 addrspace(1)* %0 to i8*
1599 return new AddrSpaceCastInst(Builder->Insert(Res), GEP.getType());
1602 if (ArrayType *XATy =
1603 dyn_cast<ArrayType>(StrippedPtrTy->getElementType())){
1604 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
1605 if (CATy->getElementType() == XATy->getElementType()) {
1606 // -> GEP [10 x i8]* X, i32 0, ...
1607 // At this point, we know that the cast source type is a pointer
1608 // to an array of the same type as the destination pointer
1609 // array. Because the array type is never stepped over (there
1610 // is a leading zero) we can fold the cast into this GEP.
1611 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace()) {
1612 GEP.setOperand(0, StrippedPtr);
1615 // Cannot replace the base pointer directly because StrippedPtr's
1616 // address space is different. Instead, create a new GEP followed by
1617 // an addrspacecast.
1619 // GEP (addrspacecast [10 x i8] addrspace(1)* X to [0 x i8]*),
1622 // %0 = GEP [10 x i8] addrspace(1)* X, ...
1623 // addrspacecast i8 addrspace(1)* %0 to i8*
1624 SmallVector<Value*, 8> Idx(GEP.idx_begin(), GEP.idx_end());
1625 Value *NewGEP = GEP.isInBounds() ?
1626 Builder->CreateInBoundsGEP(StrippedPtr, Idx, GEP.getName()) :
1627 Builder->CreateGEP(StrippedPtr, Idx, GEP.getName());
1628 return new AddrSpaceCastInst(NewGEP, GEP.getType());
1632 } else if (GEP.getNumOperands() == 2) {
1633 // Transform things like:
1634 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
1635 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
1636 Type *SrcElTy = StrippedPtrTy->getElementType();
1637 Type *ResElTy = PtrOp->getType()->getPointerElementType();
1638 if (DL && SrcElTy->isArrayTy() &&
1639 DL->getTypeAllocSize(SrcElTy->getArrayElementType()) ==
1640 DL->getTypeAllocSize(ResElTy)) {
1641 Type *IdxType = DL->getIntPtrType(GEP.getType());
1642 Value *Idx[2] = { Constant::getNullValue(IdxType), GEP.getOperand(1) };
1643 Value *NewGEP = GEP.isInBounds() ?
1644 Builder->CreateInBoundsGEP(StrippedPtr, Idx, GEP.getName()) :
1645 Builder->CreateGEP(StrippedPtr, Idx, GEP.getName());
1647 // V and GEP are both pointer types --> BitCast
1648 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
1652 // Transform things like:
1653 // %V = mul i64 %N, 4
1654 // %t = getelementptr i8* bitcast (i32* %arr to i8*), i32 %V
1655 // into: %t1 = getelementptr i32* %arr, i32 %N; bitcast
1656 if (DL && ResElTy->isSized() && SrcElTy->isSized()) {
1657 // Check that changing the type amounts to dividing the index by a scale
1659 uint64_t ResSize = DL->getTypeAllocSize(ResElTy);
1660 uint64_t SrcSize = DL->getTypeAllocSize(SrcElTy);
1661 if (ResSize && SrcSize % ResSize == 0) {
1662 Value *Idx = GEP.getOperand(1);
1663 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
1664 uint64_t Scale = SrcSize / ResSize;
1666 // Earlier transforms ensure that the index has type IntPtrType, which
1667 // considerably simplifies the logic by eliminating implicit casts.
1668 assert(Idx->getType() == DL->getIntPtrType(GEP.getType()) &&
1669 "Index not cast to pointer width?");
1672 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
1673 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
1674 // If the multiplication NewIdx * Scale may overflow then the new
1675 // GEP may not be "inbounds".
1676 Value *NewGEP = GEP.isInBounds() && NSW ?
1677 Builder->CreateInBoundsGEP(StrippedPtr, NewIdx, GEP.getName()) :
1678 Builder->CreateGEP(StrippedPtr, NewIdx, GEP.getName());
1680 // The NewGEP must be pointer typed, so must the old one -> BitCast
1681 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
1687 // Similarly, transform things like:
1688 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
1689 // (where tmp = 8*tmp2) into:
1690 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
1691 if (DL && ResElTy->isSized() && SrcElTy->isSized() &&
1692 SrcElTy->isArrayTy()) {
1693 // Check that changing to the array element type amounts to dividing the
1694 // index by a scale factor.
1695 uint64_t ResSize = DL->getTypeAllocSize(ResElTy);
1696 uint64_t ArrayEltSize
1697 = DL->getTypeAllocSize(SrcElTy->getArrayElementType());
1698 if (ResSize && ArrayEltSize % ResSize == 0) {
1699 Value *Idx = GEP.getOperand(1);
1700 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
1701 uint64_t Scale = ArrayEltSize / ResSize;
1703 // Earlier transforms ensure that the index has type IntPtrType, which
1704 // considerably simplifies the logic by eliminating implicit casts.
1705 assert(Idx->getType() == DL->getIntPtrType(GEP.getType()) &&
1706 "Index not cast to pointer width?");
1709 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
1710 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
1711 // If the multiplication NewIdx * Scale may overflow then the new
1712 // GEP may not be "inbounds".
1714 Constant::getNullValue(DL->getIntPtrType(GEP.getType())),
1718 Value *NewGEP = GEP.isInBounds() && NSW ?
1719 Builder->CreateInBoundsGEP(StrippedPtr, Off, GEP.getName()) :
1720 Builder->CreateGEP(StrippedPtr, Off, GEP.getName());
1721 // The NewGEP must be pointer typed, so must the old one -> BitCast
1722 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
1733 // addrspacecast between types is canonicalized as a bitcast, then an
1734 // addrspacecast. To take advantage of the below bitcast + struct GEP, look
1735 // through the addrspacecast.
1736 if (AddrSpaceCastInst *ASC = dyn_cast<AddrSpaceCastInst>(PtrOp)) {
1737 // X = bitcast A addrspace(1)* to B addrspace(1)*
1738 // Y = addrspacecast A addrspace(1)* to B addrspace(2)*
1739 // Z = gep Y, <...constant indices...>
1740 // Into an addrspacecasted GEP of the struct.
1741 if (BitCastInst *BC = dyn_cast<BitCastInst>(ASC->getOperand(0)))
1745 /// See if we can simplify:
1746 /// X = bitcast A* to B*
1747 /// Y = gep X, <...constant indices...>
1748 /// into a gep of the original struct. This is important for SROA and alias
1749 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
1750 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
1751 Value *Operand = BCI->getOperand(0);
1752 PointerType *OpType = cast<PointerType>(Operand->getType());
1753 unsigned OffsetBits = DL->getPointerTypeSizeInBits(GEP.getType());
1754 APInt Offset(OffsetBits, 0);
1755 if (!isa<BitCastInst>(Operand) &&
1756 GEP.accumulateConstantOffset(*DL, Offset)) {
1758 // If this GEP instruction doesn't move the pointer, just replace the GEP
1759 // with a bitcast of the real input to the dest type.
1761 // If the bitcast is of an allocation, and the allocation will be
1762 // converted to match the type of the cast, don't touch this.
1763 if (isa<AllocaInst>(Operand) || isAllocationFn(Operand, TLI)) {
1764 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
1765 if (Instruction *I = visitBitCast(*BCI)) {
1768 BCI->getParent()->getInstList().insert(BCI, I);
1769 ReplaceInstUsesWith(*BCI, I);
1775 if (Operand->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
1776 return new AddrSpaceCastInst(Operand, GEP.getType());
1777 return new BitCastInst(Operand, GEP.getType());
1780 // Otherwise, if the offset is non-zero, we need to find out if there is a
1781 // field at Offset in 'A's type. If so, we can pull the cast through the
1783 SmallVector<Value*, 8> NewIndices;
1784 if (FindElementAtOffset(OpType, Offset.getSExtValue(), NewIndices)) {
1785 Value *NGEP = GEP.isInBounds() ?
1786 Builder->CreateInBoundsGEP(Operand, NewIndices) :
1787 Builder->CreateGEP(Operand, NewIndices);
1789 if (NGEP->getType() == GEP.getType())
1790 return ReplaceInstUsesWith(GEP, NGEP);
1791 NGEP->takeName(&GEP);
1793 if (NGEP->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
1794 return new AddrSpaceCastInst(NGEP, GEP.getType());
1795 return new BitCastInst(NGEP, GEP.getType());
1804 isAllocSiteRemovable(Instruction *AI, SmallVectorImpl<WeakVH> &Users,
1805 const TargetLibraryInfo *TLI) {
1806 SmallVector<Instruction*, 4> Worklist;
1807 Worklist.push_back(AI);
1810 Instruction *PI = Worklist.pop_back_val();
1811 for (User *U : PI->users()) {
1812 Instruction *I = cast<Instruction>(U);
1813 switch (I->getOpcode()) {
1815 // Give up the moment we see something we can't handle.
1818 case Instruction::BitCast:
1819 case Instruction::GetElementPtr:
1821 Worklist.push_back(I);
1824 case Instruction::ICmp: {
1825 ICmpInst *ICI = cast<ICmpInst>(I);
1826 // We can fold eq/ne comparisons with null to false/true, respectively.
1827 if (!ICI->isEquality() || !isa<ConstantPointerNull>(ICI->getOperand(1)))
1833 case Instruction::Call:
1834 // Ignore no-op and store intrinsics.
1835 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1836 switch (II->getIntrinsicID()) {
1840 case Intrinsic::memmove:
1841 case Intrinsic::memcpy:
1842 case Intrinsic::memset: {
1843 MemIntrinsic *MI = cast<MemIntrinsic>(II);
1844 if (MI->isVolatile() || MI->getRawDest() != PI)
1848 case Intrinsic::dbg_declare:
1849 case Intrinsic::dbg_value:
1850 case Intrinsic::invariant_start:
1851 case Intrinsic::invariant_end:
1852 case Intrinsic::lifetime_start:
1853 case Intrinsic::lifetime_end:
1854 case Intrinsic::objectsize:
1860 if (isFreeCall(I, TLI)) {
1866 case Instruction::Store: {
1867 StoreInst *SI = cast<StoreInst>(I);
1868 if (SI->isVolatile() || SI->getPointerOperand() != PI)
1874 llvm_unreachable("missing a return?");
1876 } while (!Worklist.empty());
1880 Instruction *InstCombiner::visitAllocSite(Instruction &MI) {
1881 // If we have a malloc call which is only used in any amount of comparisons
1882 // to null and free calls, delete the calls and replace the comparisons with
1883 // true or false as appropriate.
1884 SmallVector<WeakVH, 64> Users;
1885 if (isAllocSiteRemovable(&MI, Users, TLI)) {
1886 for (unsigned i = 0, e = Users.size(); i != e; ++i) {
1887 Instruction *I = cast_or_null<Instruction>(&*Users[i]);
1890 if (ICmpInst *C = dyn_cast<ICmpInst>(I)) {
1891 ReplaceInstUsesWith(*C,
1892 ConstantInt::get(Type::getInt1Ty(C->getContext()),
1893 C->isFalseWhenEqual()));
1894 } else if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) {
1895 ReplaceInstUsesWith(*I, UndefValue::get(I->getType()));
1896 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1897 if (II->getIntrinsicID() == Intrinsic::objectsize) {
1898 ConstantInt *CI = cast<ConstantInt>(II->getArgOperand(1));
1899 uint64_t DontKnow = CI->isZero() ? -1ULL : 0;
1900 ReplaceInstUsesWith(*I, ConstantInt::get(I->getType(), DontKnow));
1903 EraseInstFromFunction(*I);
1906 if (InvokeInst *II = dyn_cast<InvokeInst>(&MI)) {
1907 // Replace invoke with a NOP intrinsic to maintain the original CFG
1908 Module *M = II->getParent()->getParent()->getParent();
1909 Function *F = Intrinsic::getDeclaration(M, Intrinsic::donothing);
1910 InvokeInst::Create(F, II->getNormalDest(), II->getUnwindDest(),
1911 None, "", II->getParent());
1913 return EraseInstFromFunction(MI);
1918 /// \brief Move the call to free before a NULL test.
1920 /// Check if this free is accessed after its argument has been test
1921 /// against NULL (property 0).
1922 /// If yes, it is legal to move this call in its predecessor block.
1924 /// The move is performed only if the block containing the call to free
1925 /// will be removed, i.e.:
1926 /// 1. it has only one predecessor P, and P has two successors
1927 /// 2. it contains the call and an unconditional branch
1928 /// 3. its successor is the same as its predecessor's successor
1930 /// The profitability is out-of concern here and this function should
1931 /// be called only if the caller knows this transformation would be
1932 /// profitable (e.g., for code size).
1933 static Instruction *
1934 tryToMoveFreeBeforeNullTest(CallInst &FI) {
1935 Value *Op = FI.getArgOperand(0);
1936 BasicBlock *FreeInstrBB = FI.getParent();
1937 BasicBlock *PredBB = FreeInstrBB->getSinglePredecessor();
1939 // Validate part of constraint #1: Only one predecessor
1940 // FIXME: We can extend the number of predecessor, but in that case, we
1941 // would duplicate the call to free in each predecessor and it may
1942 // not be profitable even for code size.
1946 // Validate constraint #2: Does this block contains only the call to
1947 // free and an unconditional branch?
1948 // FIXME: We could check if we can speculate everything in the
1949 // predecessor block
1950 if (FreeInstrBB->size() != 2)
1953 if (!match(FreeInstrBB->getTerminator(), m_UnconditionalBr(SuccBB)))
1956 // Validate the rest of constraint #1 by matching on the pred branch.
1957 TerminatorInst *TI = PredBB->getTerminator();
1958 BasicBlock *TrueBB, *FalseBB;
1959 ICmpInst::Predicate Pred;
1960 if (!match(TI, m_Br(m_ICmp(Pred, m_Specific(Op), m_Zero()), TrueBB, FalseBB)))
1962 if (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE)
1965 // Validate constraint #3: Ensure the null case just falls through.
1966 if (SuccBB != (Pred == ICmpInst::ICMP_EQ ? TrueBB : FalseBB))
1968 assert(FreeInstrBB == (Pred == ICmpInst::ICMP_EQ ? FalseBB : TrueBB) &&
1969 "Broken CFG: missing edge from predecessor to successor");
1976 Instruction *InstCombiner::visitFree(CallInst &FI) {
1977 Value *Op = FI.getArgOperand(0);
1979 // free undef -> unreachable.
1980 if (isa<UndefValue>(Op)) {
1981 // Insert a new store to null because we cannot modify the CFG here.
1982 Builder->CreateStore(ConstantInt::getTrue(FI.getContext()),
1983 UndefValue::get(Type::getInt1PtrTy(FI.getContext())));
1984 return EraseInstFromFunction(FI);
1987 // If we have 'free null' delete the instruction. This can happen in stl code
1988 // when lots of inlining happens.
1989 if (isa<ConstantPointerNull>(Op))
1990 return EraseInstFromFunction(FI);
1992 // If we optimize for code size, try to move the call to free before the null
1993 // test so that simplify cfg can remove the empty block and dead code
1994 // elimination the branch. I.e., helps to turn something like:
1995 // if (foo) free(foo);
1999 if (Instruction *I = tryToMoveFreeBeforeNullTest(FI))
2005 Instruction *InstCombiner::visitReturnInst(ReturnInst &RI) {
2006 if (RI.getNumOperands() == 0) // ret void
2009 Value *ResultOp = RI.getOperand(0);
2010 Type *VTy = ResultOp->getType();
2011 if (!VTy->isIntegerTy())
2014 // There might be assume intrinsics dominating this return that completely
2015 // determine the value. If so, constant fold it.
2016 unsigned BitWidth = VTy->getPrimitiveSizeInBits();
2017 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2018 computeKnownBits(ResultOp, KnownZero, KnownOne, 0, &RI);
2019 if ((KnownZero|KnownOne).isAllOnesValue())
2020 RI.setOperand(0, Constant::getIntegerValue(VTy, KnownOne));
2025 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
2026 // Change br (not X), label True, label False to: br X, label False, True
2028 BasicBlock *TrueDest;
2029 BasicBlock *FalseDest;
2030 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
2031 !isa<Constant>(X)) {
2032 // Swap Destinations and condition...
2034 BI.swapSuccessors();
2038 // Canonicalize fcmp_one -> fcmp_oeq
2039 FCmpInst::Predicate FPred; Value *Y;
2040 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
2041 TrueDest, FalseDest)) &&
2042 BI.getCondition()->hasOneUse())
2043 if (FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
2044 FPred == FCmpInst::FCMP_OGE) {
2045 FCmpInst *Cond = cast<FCmpInst>(BI.getCondition());
2046 Cond->setPredicate(FCmpInst::getInversePredicate(FPred));
2048 // Swap Destinations and condition.
2049 BI.swapSuccessors();
2054 // Canonicalize icmp_ne -> icmp_eq
2055 ICmpInst::Predicate IPred;
2056 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
2057 TrueDest, FalseDest)) &&
2058 BI.getCondition()->hasOneUse())
2059 if (IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
2060 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
2061 IPred == ICmpInst::ICMP_SGE) {
2062 ICmpInst *Cond = cast<ICmpInst>(BI.getCondition());
2063 Cond->setPredicate(ICmpInst::getInversePredicate(IPred));
2064 // Swap Destinations and condition.
2065 BI.swapSuccessors();
2073 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
2074 Value *Cond = SI.getCondition();
2075 unsigned BitWidth = cast<IntegerType>(Cond->getType())->getBitWidth();
2076 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2077 computeKnownBits(Cond, KnownZero, KnownOne);
2078 unsigned LeadingKnownZeros = KnownZero.countLeadingOnes();
2079 unsigned LeadingKnownOnes = KnownOne.countLeadingOnes();
2081 // Compute the number of leading bits we can ignore.
2082 for (auto &C : SI.cases()) {
2083 LeadingKnownZeros = std::min(
2084 LeadingKnownZeros, C.getCaseValue()->getValue().countLeadingZeros());
2085 LeadingKnownOnes = std::min(
2086 LeadingKnownOnes, C.getCaseValue()->getValue().countLeadingOnes());
2089 unsigned NewWidth = BitWidth - std::max(LeadingKnownZeros, LeadingKnownOnes);
2091 // Truncate the condition operand if the new type is equal to or larger than
2092 // the largest legal integer type. We need to be conservative here since
2093 // x86 generates redundant zero-extenstion instructions if the operand is
2094 // truncated to i8 or i16.
2095 if (BitWidth > NewWidth && NewWidth >= DL->getLargestLegalIntTypeSize()) {
2096 IntegerType *Ty = IntegerType::get(SI.getContext(), NewWidth);
2097 Builder->SetInsertPoint(&SI);
2098 Value *NewCond = Builder->CreateTrunc(SI.getCondition(), Ty, "trunc");
2099 SI.setCondition(NewCond);
2101 for (auto &C : SI.cases())
2102 static_cast<SwitchInst::CaseIt *>(&C)->setValue(ConstantInt::get(
2103 SI.getContext(), C.getCaseValue()->getValue().trunc(NewWidth)));
2106 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
2107 if (I->getOpcode() == Instruction::Add)
2108 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2109 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
2110 // Skip the first item since that's the default case.
2111 for (SwitchInst::CaseIt i = SI.case_begin(), e = SI.case_end();
2113 ConstantInt* CaseVal = i.getCaseValue();
2114 Constant* NewCaseVal = ConstantExpr::getSub(cast<Constant>(CaseVal),
2116 assert(isa<ConstantInt>(NewCaseVal) &&
2117 "Result of expression should be constant");
2118 i.setValue(cast<ConstantInt>(NewCaseVal));
2120 SI.setCondition(I->getOperand(0));
2128 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
2129 Value *Agg = EV.getAggregateOperand();
2131 if (!EV.hasIndices())
2132 return ReplaceInstUsesWith(EV, Agg);
2134 if (Constant *C = dyn_cast<Constant>(Agg)) {
2135 if (Constant *C2 = C->getAggregateElement(*EV.idx_begin())) {
2136 if (EV.getNumIndices() == 0)
2137 return ReplaceInstUsesWith(EV, C2);
2138 // Extract the remaining indices out of the constant indexed by the
2140 return ExtractValueInst::Create(C2, EV.getIndices().slice(1));
2142 return nullptr; // Can't handle other constants
2145 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
2146 // We're extracting from an insertvalue instruction, compare the indices
2147 const unsigned *exti, *exte, *insi, *inse;
2148 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
2149 exte = EV.idx_end(), inse = IV->idx_end();
2150 exti != exte && insi != inse;
2153 // The insert and extract both reference distinctly different elements.
2154 // This means the extract is not influenced by the insert, and we can
2155 // replace the aggregate operand of the extract with the aggregate
2156 // operand of the insert. i.e., replace
2157 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
2158 // %E = extractvalue { i32, { i32 } } %I, 0
2160 // %E = extractvalue { i32, { i32 } } %A, 0
2161 return ExtractValueInst::Create(IV->getAggregateOperand(),
2164 if (exti == exte && insi == inse)
2165 // Both iterators are at the end: Index lists are identical. Replace
2166 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
2167 // %C = extractvalue { i32, { i32 } } %B, 1, 0
2169 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
2171 // The extract list is a prefix of the insert list. i.e. replace
2172 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
2173 // %E = extractvalue { i32, { i32 } } %I, 1
2175 // %X = extractvalue { i32, { i32 } } %A, 1
2176 // %E = insertvalue { i32 } %X, i32 42, 0
2177 // by switching the order of the insert and extract (though the
2178 // insertvalue should be left in, since it may have other uses).
2179 Value *NewEV = Builder->CreateExtractValue(IV->getAggregateOperand(),
2181 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
2182 makeArrayRef(insi, inse));
2185 // The insert list is a prefix of the extract list
2186 // We can simply remove the common indices from the extract and make it
2187 // operate on the inserted value instead of the insertvalue result.
2189 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
2190 // %E = extractvalue { i32, { i32 } } %I, 1, 0
2192 // %E extractvalue { i32 } { i32 42 }, 0
2193 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
2194 makeArrayRef(exti, exte));
2196 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) {
2197 // We're extracting from an intrinsic, see if we're the only user, which
2198 // allows us to simplify multiple result intrinsics to simpler things that
2199 // just get one value.
2200 if (II->hasOneUse()) {
2201 // Check if we're grabbing the overflow bit or the result of a 'with
2202 // overflow' intrinsic. If it's the latter we can remove the intrinsic
2203 // and replace it with a traditional binary instruction.
2204 switch (II->getIntrinsicID()) {
2205 case Intrinsic::uadd_with_overflow:
2206 case Intrinsic::sadd_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::CreateAdd(LHS, RHS);
2214 // If the normal result of the add is dead, and the RHS is a constant,
2215 // we can transform this into a range comparison.
2216 // overflow = uadd a, -4 --> overflow = icmp ugt a, 3
2217 if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow)
2218 if (ConstantInt *CI = dyn_cast<ConstantInt>(II->getArgOperand(1)))
2219 return new ICmpInst(ICmpInst::ICMP_UGT, II->getArgOperand(0),
2220 ConstantExpr::getNot(CI));
2222 case Intrinsic::usub_with_overflow:
2223 case Intrinsic::ssub_with_overflow:
2224 if (*EV.idx_begin() == 0) { // Normal result.
2225 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2226 ReplaceInstUsesWith(*II, UndefValue::get(II->getType()));
2227 EraseInstFromFunction(*II);
2228 return BinaryOperator::CreateSub(LHS, RHS);
2231 case Intrinsic::umul_with_overflow:
2232 case Intrinsic::smul_with_overflow:
2233 if (*EV.idx_begin() == 0) { // Normal result.
2234 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2235 ReplaceInstUsesWith(*II, UndefValue::get(II->getType()));
2236 EraseInstFromFunction(*II);
2237 return BinaryOperator::CreateMul(LHS, RHS);
2245 if (LoadInst *L = dyn_cast<LoadInst>(Agg))
2246 // If the (non-volatile) load only has one use, we can rewrite this to a
2247 // load from a GEP. This reduces the size of the load.
2248 // FIXME: If a load is used only by extractvalue instructions then this
2249 // could be done regardless of having multiple uses.
2250 if (L->isSimple() && L->hasOneUse()) {
2251 // extractvalue has integer indices, getelementptr has Value*s. Convert.
2252 SmallVector<Value*, 4> Indices;
2253 // Prefix an i32 0 since we need the first element.
2254 Indices.push_back(Builder->getInt32(0));
2255 for (ExtractValueInst::idx_iterator I = EV.idx_begin(), E = EV.idx_end();
2257 Indices.push_back(Builder->getInt32(*I));
2259 // We need to insert these at the location of the old load, not at that of
2260 // the extractvalue.
2261 Builder->SetInsertPoint(L->getParent(), L);
2262 Value *GEP = Builder->CreateInBoundsGEP(L->getPointerOperand(), Indices);
2263 // Returning the load directly will cause the main loop to insert it in
2264 // the wrong spot, so use ReplaceInstUsesWith().
2265 return ReplaceInstUsesWith(EV, Builder->CreateLoad(GEP));
2267 // We could simplify extracts from other values. Note that nested extracts may
2268 // already be simplified implicitly by the above: extract (extract (insert) )
2269 // will be translated into extract ( insert ( extract ) ) first and then just
2270 // the value inserted, if appropriate. Similarly for extracts from single-use
2271 // loads: extract (extract (load)) will be translated to extract (load (gep))
2272 // and if again single-use then via load (gep (gep)) to load (gep).
2273 // However, double extracts from e.g. function arguments or return values
2274 // aren't handled yet.
2278 enum Personality_Type {
2279 Unknown_Personality,
2280 GNU_Ada_Personality,
2281 GNU_CXX_Personality,
2282 GNU_ObjC_Personality
2285 /// RecognizePersonality - See if the given exception handling personality
2286 /// function is one that we understand. If so, return a description of it;
2287 /// otherwise return Unknown_Personality.
2288 static Personality_Type RecognizePersonality(Value *Pers) {
2289 Function *F = dyn_cast<Function>(Pers->stripPointerCasts());
2291 return Unknown_Personality;
2292 return StringSwitch<Personality_Type>(F->getName())
2293 .Case("__gnat_eh_personality", GNU_Ada_Personality)
2294 .Case("__gxx_personality_v0", GNU_CXX_Personality)
2295 .Case("__objc_personality_v0", GNU_ObjC_Personality)
2296 .Default(Unknown_Personality);
2299 /// isCatchAll - Return 'true' if the given typeinfo will match anything.
2300 static bool isCatchAll(Personality_Type Personality, Constant *TypeInfo) {
2301 switch (Personality) {
2302 case Unknown_Personality:
2304 case GNU_Ada_Personality:
2305 // While __gnat_all_others_value will match any Ada exception, it doesn't
2306 // match foreign exceptions (or didn't, before gcc-4.7).
2308 case GNU_CXX_Personality:
2309 case GNU_ObjC_Personality:
2310 return TypeInfo->isNullValue();
2312 llvm_unreachable("Unknown personality!");
2315 static bool shorter_filter(const Value *LHS, const Value *RHS) {
2317 cast<ArrayType>(LHS->getType())->getNumElements()
2319 cast<ArrayType>(RHS->getType())->getNumElements();
2322 Instruction *InstCombiner::visitLandingPadInst(LandingPadInst &LI) {
2323 // The logic here should be correct for any real-world personality function.
2324 // However if that turns out not to be true, the offending logic can always
2325 // be conditioned on the personality function, like the catch-all logic is.
2326 Personality_Type Personality = RecognizePersonality(LI.getPersonalityFn());
2328 // Simplify the list of clauses, eg by removing repeated catch clauses
2329 // (these are often created by inlining).
2330 bool MakeNewInstruction = false; // If true, recreate using the following:
2331 SmallVector<Constant *, 16> NewClauses; // - Clauses for the new instruction;
2332 bool CleanupFlag = LI.isCleanup(); // - The new instruction is a cleanup.
2334 SmallPtrSet<Value *, 16> AlreadyCaught; // Typeinfos known caught already.
2335 for (unsigned i = 0, e = LI.getNumClauses(); i != e; ++i) {
2336 bool isLastClause = i + 1 == e;
2337 if (LI.isCatch(i)) {
2339 Constant *CatchClause = LI.getClause(i);
2340 Constant *TypeInfo = CatchClause->stripPointerCasts();
2342 // If we already saw this clause, there is no point in having a second
2344 if (AlreadyCaught.insert(TypeInfo)) {
2345 // This catch clause was not already seen.
2346 NewClauses.push_back(CatchClause);
2348 // Repeated catch clause - drop the redundant copy.
2349 MakeNewInstruction = true;
2352 // If this is a catch-all then there is no point in keeping any following
2353 // clauses or marking the landingpad as having a cleanup.
2354 if (isCatchAll(Personality, TypeInfo)) {
2356 MakeNewInstruction = true;
2357 CleanupFlag = false;
2361 // A filter clause. If any of the filter elements were already caught
2362 // then they can be dropped from the filter. It is tempting to try to
2363 // exploit the filter further by saying that any typeinfo that does not
2364 // occur in the filter can't be caught later (and thus can be dropped).
2365 // However this would be wrong, since typeinfos can match without being
2366 // equal (for example if one represents a C++ class, and the other some
2367 // class derived from it).
2368 assert(LI.isFilter(i) && "Unsupported landingpad clause!");
2369 Constant *FilterClause = LI.getClause(i);
2370 ArrayType *FilterType = cast<ArrayType>(FilterClause->getType());
2371 unsigned NumTypeInfos = FilterType->getNumElements();
2373 // An empty filter catches everything, so there is no point in keeping any
2374 // following clauses or marking the landingpad as having a cleanup. By
2375 // dealing with this case here the following code is made a bit simpler.
2376 if (!NumTypeInfos) {
2377 NewClauses.push_back(FilterClause);
2379 MakeNewInstruction = true;
2380 CleanupFlag = false;
2384 bool MakeNewFilter = false; // If true, make a new filter.
2385 SmallVector<Constant *, 16> NewFilterElts; // New elements.
2386 if (isa<ConstantAggregateZero>(FilterClause)) {
2387 // Not an empty filter - it contains at least one null typeinfo.
2388 assert(NumTypeInfos > 0 && "Should have handled empty filter already!");
2389 Constant *TypeInfo =
2390 Constant::getNullValue(FilterType->getElementType());
2391 // If this typeinfo is a catch-all then the filter can never match.
2392 if (isCatchAll(Personality, TypeInfo)) {
2393 // Throw the filter away.
2394 MakeNewInstruction = true;
2398 // There is no point in having multiple copies of this typeinfo, so
2399 // discard all but the first copy if there is more than one.
2400 NewFilterElts.push_back(TypeInfo);
2401 if (NumTypeInfos > 1)
2402 MakeNewFilter = true;
2404 ConstantArray *Filter = cast<ConstantArray>(FilterClause);
2405 SmallPtrSet<Value *, 16> SeenInFilter; // For uniquing the elements.
2406 NewFilterElts.reserve(NumTypeInfos);
2408 // Remove any filter elements that were already caught or that already
2409 // occurred in the filter. While there, see if any of the elements are
2410 // catch-alls. If so, the filter can be discarded.
2411 bool SawCatchAll = false;
2412 for (unsigned j = 0; j != NumTypeInfos; ++j) {
2413 Constant *Elt = Filter->getOperand(j);
2414 Constant *TypeInfo = Elt->stripPointerCasts();
2415 if (isCatchAll(Personality, TypeInfo)) {
2416 // This element is a catch-all. Bail out, noting this fact.
2420 if (AlreadyCaught.count(TypeInfo))
2421 // Already caught by an earlier clause, so having it in the filter
2424 // There is no point in having multiple copies of the same typeinfo in
2425 // a filter, so only add it if we didn't already.
2426 if (SeenInFilter.insert(TypeInfo))
2427 NewFilterElts.push_back(cast<Constant>(Elt));
2429 // A filter containing a catch-all cannot match anything by definition.
2431 // Throw the filter away.
2432 MakeNewInstruction = true;
2436 // If we dropped something from the filter, make a new one.
2437 if (NewFilterElts.size() < NumTypeInfos)
2438 MakeNewFilter = true;
2440 if (MakeNewFilter) {
2441 FilterType = ArrayType::get(FilterType->getElementType(),
2442 NewFilterElts.size());
2443 FilterClause = ConstantArray::get(FilterType, NewFilterElts);
2444 MakeNewInstruction = true;
2447 NewClauses.push_back(FilterClause);
2449 // If the new filter is empty then it will catch everything so there is
2450 // no point in keeping any following clauses or marking the landingpad
2451 // as having a cleanup. The case of the original filter being empty was
2452 // already handled above.
2453 if (MakeNewFilter && !NewFilterElts.size()) {
2454 assert(MakeNewInstruction && "New filter but not a new instruction!");
2455 CleanupFlag = false;
2461 // If several filters occur in a row then reorder them so that the shortest
2462 // filters come first (those with the smallest number of elements). This is
2463 // advantageous because shorter filters are more likely to match, speeding up
2464 // unwinding, but mostly because it increases the effectiveness of the other
2465 // filter optimizations below.
2466 for (unsigned i = 0, e = NewClauses.size(); i + 1 < e; ) {
2468 // Find the maximal 'j' s.t. the range [i, j) consists entirely of filters.
2469 for (j = i; j != e; ++j)
2470 if (!isa<ArrayType>(NewClauses[j]->getType()))
2473 // Check whether the filters are already sorted by length. We need to know
2474 // if sorting them is actually going to do anything so that we only make a
2475 // new landingpad instruction if it does.
2476 for (unsigned k = i; k + 1 < j; ++k)
2477 if (shorter_filter(NewClauses[k+1], NewClauses[k])) {
2478 // Not sorted, so sort the filters now. Doing an unstable sort would be
2479 // correct too but reordering filters pointlessly might confuse users.
2480 std::stable_sort(NewClauses.begin() + i, NewClauses.begin() + j,
2482 MakeNewInstruction = true;
2486 // Look for the next batch of filters.
2490 // If typeinfos matched if and only if equal, then the elements of a filter L
2491 // that occurs later than a filter F could be replaced by the intersection of
2492 // the elements of F and L. In reality two typeinfos can match without being
2493 // equal (for example if one represents a C++ class, and the other some class
2494 // derived from it) so it would be wrong to perform this transform in general.
2495 // However the transform is correct and useful if F is a subset of L. In that
2496 // case L can be replaced by F, and thus removed altogether since repeating a
2497 // filter is pointless. So here we look at all pairs of filters F and L where
2498 // L follows F in the list of clauses, and remove L if every element of F is
2499 // an element of L. This can occur when inlining C++ functions with exception
2501 for (unsigned i = 0; i + 1 < NewClauses.size(); ++i) {
2502 // Examine each filter in turn.
2503 Value *Filter = NewClauses[i];
2504 ArrayType *FTy = dyn_cast<ArrayType>(Filter->getType());
2506 // Not a filter - skip it.
2508 unsigned FElts = FTy->getNumElements();
2509 // Examine each filter following this one. Doing this backwards means that
2510 // we don't have to worry about filters disappearing under us when removed.
2511 for (unsigned j = NewClauses.size() - 1; j != i; --j) {
2512 Value *LFilter = NewClauses[j];
2513 ArrayType *LTy = dyn_cast<ArrayType>(LFilter->getType());
2515 // Not a filter - skip it.
2517 // If Filter is a subset of LFilter, i.e. every element of Filter is also
2518 // an element of LFilter, then discard LFilter.
2519 SmallVectorImpl<Constant *>::iterator J = NewClauses.begin() + j;
2520 // If Filter is empty then it is a subset of LFilter.
2523 NewClauses.erase(J);
2524 MakeNewInstruction = true;
2525 // Move on to the next filter.
2528 unsigned LElts = LTy->getNumElements();
2529 // If Filter is longer than LFilter then it cannot be a subset of it.
2531 // Move on to the next filter.
2533 // At this point we know that LFilter has at least one element.
2534 if (isa<ConstantAggregateZero>(LFilter)) { // LFilter only contains zeros.
2535 // Filter is a subset of LFilter iff Filter contains only zeros (as we
2536 // already know that Filter is not longer than LFilter).
2537 if (isa<ConstantAggregateZero>(Filter)) {
2538 assert(FElts <= LElts && "Should have handled this case earlier!");
2540 NewClauses.erase(J);
2541 MakeNewInstruction = true;
2543 // Move on to the next filter.
2546 ConstantArray *LArray = cast<ConstantArray>(LFilter);
2547 if (isa<ConstantAggregateZero>(Filter)) { // Filter only contains zeros.
2548 // Since Filter is non-empty and contains only zeros, it is a subset of
2549 // LFilter iff LFilter contains a zero.
2550 assert(FElts > 0 && "Should have eliminated the empty filter earlier!");
2551 for (unsigned l = 0; l != LElts; ++l)
2552 if (LArray->getOperand(l)->isNullValue()) {
2553 // LFilter contains a zero - discard it.
2554 NewClauses.erase(J);
2555 MakeNewInstruction = true;
2558 // Move on to the next filter.
2561 // At this point we know that both filters are ConstantArrays. Loop over
2562 // operands to see whether every element of Filter is also an element of
2563 // LFilter. Since filters tend to be short this is probably faster than
2564 // using a method that scales nicely.
2565 ConstantArray *FArray = cast<ConstantArray>(Filter);
2566 bool AllFound = true;
2567 for (unsigned f = 0; f != FElts; ++f) {
2568 Value *FTypeInfo = FArray->getOperand(f)->stripPointerCasts();
2570 for (unsigned l = 0; l != LElts; ++l) {
2571 Value *LTypeInfo = LArray->getOperand(l)->stripPointerCasts();
2572 if (LTypeInfo == FTypeInfo) {
2582 NewClauses.erase(J);
2583 MakeNewInstruction = true;
2585 // Move on to the next filter.
2589 // If we changed any of the clauses, replace the old landingpad instruction
2591 if (MakeNewInstruction) {
2592 LandingPadInst *NLI = LandingPadInst::Create(LI.getType(),
2593 LI.getPersonalityFn(),
2595 for (unsigned i = 0, e = NewClauses.size(); i != e; ++i)
2596 NLI->addClause(NewClauses[i]);
2597 // A landing pad with no clauses must have the cleanup flag set. It is
2598 // theoretically possible, though highly unlikely, that we eliminated all
2599 // clauses. If so, force the cleanup flag to true.
2600 if (NewClauses.empty())
2602 NLI->setCleanup(CleanupFlag);
2606 // Even if none of the clauses changed, we may nonetheless have understood
2607 // that the cleanup flag is pointless. Clear it if so.
2608 if (LI.isCleanup() != CleanupFlag) {
2609 assert(!CleanupFlag && "Adding a cleanup, not removing one?!");
2610 LI.setCleanup(CleanupFlag);
2620 /// TryToSinkInstruction - Try to move the specified instruction from its
2621 /// current block into the beginning of DestBlock, which can only happen if it's
2622 /// safe to move the instruction past all of the instructions between it and the
2623 /// end of its block.
2624 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
2625 assert(I->hasOneUse() && "Invariants didn't hold!");
2627 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
2628 if (isa<PHINode>(I) || isa<LandingPadInst>(I) || I->mayHaveSideEffects() ||
2629 isa<TerminatorInst>(I))
2632 // Do not sink alloca instructions out of the entry block.
2633 if (isa<AllocaInst>(I) && I->getParent() ==
2634 &DestBlock->getParent()->getEntryBlock())
2637 // We can only sink load instructions if there is nothing between the load and
2638 // the end of block that could change the value.
2639 if (I->mayReadFromMemory()) {
2640 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
2642 if (Scan->mayWriteToMemory())
2646 BasicBlock::iterator InsertPos = DestBlock->getFirstInsertionPt();
2647 I->moveBefore(InsertPos);
2653 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
2654 /// all reachable code to the worklist.
2656 /// This has a couple of tricks to make the code faster and more powerful. In
2657 /// particular, we constant fold and DCE instructions as we go, to avoid adding
2658 /// them to the worklist (this significantly speeds up instcombine on code where
2659 /// many instructions are dead or constant). Additionally, if we find a branch
2660 /// whose condition is a known constant, we only visit the reachable successors.
2662 static bool AddReachableCodeToWorklist(BasicBlock *BB,
2663 SmallPtrSetImpl<BasicBlock*> &Visited,
2665 const DataLayout *DL,
2666 const TargetLibraryInfo *TLI) {
2667 bool MadeIRChange = false;
2668 SmallVector<BasicBlock*, 256> Worklist;
2669 Worklist.push_back(BB);
2671 SmallVector<Instruction*, 128> InstrsForInstCombineWorklist;
2672 DenseMap<ConstantExpr*, Constant*> FoldedConstants;
2675 BB = Worklist.pop_back_val();
2677 // We have now visited this block! If we've already been here, ignore it.
2678 if (!Visited.insert(BB)) continue;
2680 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
2681 Instruction *Inst = BBI++;
2683 // DCE instruction if trivially dead.
2684 if (isInstructionTriviallyDead(Inst, TLI)) {
2686 DEBUG(dbgs() << "IC: DCE: " << *Inst << '\n');
2687 Inst->eraseFromParent();
2691 // ConstantProp instruction if trivially constant.
2692 if (!Inst->use_empty() && isa<Constant>(Inst->getOperand(0)))
2693 if (Constant *C = ConstantFoldInstruction(Inst, DL, TLI)) {
2694 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: "
2696 Inst->replaceAllUsesWith(C);
2698 Inst->eraseFromParent();
2703 // See if we can constant fold its operands.
2704 for (User::op_iterator i = Inst->op_begin(), e = Inst->op_end();
2706 ConstantExpr *CE = dyn_cast<ConstantExpr>(i);
2707 if (CE == nullptr) continue;
2709 Constant*& FoldRes = FoldedConstants[CE];
2711 FoldRes = ConstantFoldConstantExpression(CE, DL, TLI);
2715 if (FoldRes != CE) {
2717 MadeIRChange = true;
2722 InstrsForInstCombineWorklist.push_back(Inst);
2725 // Recursively visit successors. If this is a branch or switch on a
2726 // constant, only visit the reachable successor.
2727 TerminatorInst *TI = BB->getTerminator();
2728 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
2729 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
2730 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
2731 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
2732 Worklist.push_back(ReachableBB);
2735 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
2736 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
2737 // See if this is an explicit destination.
2738 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2740 if (i.getCaseValue() == Cond) {
2741 BasicBlock *ReachableBB = i.getCaseSuccessor();
2742 Worklist.push_back(ReachableBB);
2746 // Otherwise it is the default destination.
2747 Worklist.push_back(SI->getDefaultDest());
2752 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
2753 Worklist.push_back(TI->getSuccessor(i));
2754 } while (!Worklist.empty());
2756 // Once we've found all of the instructions to add to instcombine's worklist,
2757 // add them in reverse order. This way instcombine will visit from the top
2758 // of the function down. This jives well with the way that it adds all uses
2759 // of instructions to the worklist after doing a transformation, thus avoiding
2760 // some N^2 behavior in pathological cases.
2761 IC.Worklist.AddInitialGroup(&InstrsForInstCombineWorklist[0],
2762 InstrsForInstCombineWorklist.size());
2764 return MadeIRChange;
2767 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
2768 MadeIRChange = false;
2770 DEBUG(dbgs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
2771 << F.getName() << "\n");
2774 // Do a depth-first traversal of the function, populate the worklist with
2775 // the reachable instructions. Ignore blocks that are not reachable. Keep
2776 // track of which blocks we visit.
2777 SmallPtrSet<BasicBlock*, 64> Visited;
2778 MadeIRChange |= AddReachableCodeToWorklist(F.begin(), Visited, *this, DL,
2781 // Do a quick scan over the function. If we find any blocks that are
2782 // unreachable, remove any instructions inside of them. This prevents
2783 // the instcombine code from having to deal with some bad special cases.
2784 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
2785 if (Visited.count(BB)) continue;
2787 // Delete the instructions backwards, as it has a reduced likelihood of
2788 // having to update as many def-use and use-def chains.
2789 Instruction *EndInst = BB->getTerminator(); // Last not to be deleted.
2790 while (EndInst != BB->begin()) {
2791 // Delete the next to last instruction.
2792 BasicBlock::iterator I = EndInst;
2793 Instruction *Inst = --I;
2794 if (!Inst->use_empty())
2795 Inst->replaceAllUsesWith(UndefValue::get(Inst->getType()));
2796 if (isa<LandingPadInst>(Inst)) {
2800 if (!isa<DbgInfoIntrinsic>(Inst)) {
2802 MadeIRChange = true;
2804 Inst->eraseFromParent();
2809 while (!Worklist.isEmpty()) {
2810 Instruction *I = Worklist.RemoveOne();
2811 if (I == nullptr) continue; // skip null values.
2813 // Check to see if we can DCE the instruction.
2814 if (isInstructionTriviallyDead(I, TLI)) {
2815 DEBUG(dbgs() << "IC: DCE: " << *I << '\n');
2816 EraseInstFromFunction(*I);
2818 MadeIRChange = true;
2822 // Instruction isn't dead, see if we can constant propagate it.
2823 if (!I->use_empty() && isa<Constant>(I->getOperand(0)))
2824 if (Constant *C = ConstantFoldInstruction(I, DL, TLI)) {
2825 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n');
2827 // Add operands to the worklist.
2828 ReplaceInstUsesWith(*I, C);
2830 EraseInstFromFunction(*I);
2831 MadeIRChange = true;
2835 // See if we can trivially sink this instruction to a successor basic block.
2836 if (I->hasOneUse()) {
2837 BasicBlock *BB = I->getParent();
2838 Instruction *UserInst = cast<Instruction>(*I->user_begin());
2839 BasicBlock *UserParent;
2841 // Get the block the use occurs in.
2842 if (PHINode *PN = dyn_cast<PHINode>(UserInst))
2843 UserParent = PN->getIncomingBlock(*I->use_begin());
2845 UserParent = UserInst->getParent();
2847 if (UserParent != BB) {
2848 bool UserIsSuccessor = false;
2849 // See if the user is one of our successors.
2850 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
2851 if (*SI == UserParent) {
2852 UserIsSuccessor = true;
2856 // If the user is one of our immediate successors, and if that successor
2857 // only has us as a predecessors (we'd have to split the critical edge
2858 // otherwise), we can keep going.
2859 if (UserIsSuccessor && UserParent->getSinglePredecessor()) {
2860 // Okay, the CFG is simple enough, try to sink this instruction.
2861 if (TryToSinkInstruction(I, UserParent)) {
2862 MadeIRChange = true;
2863 // We'll add uses of the sunk instruction below, but since sinking
2864 // can expose opportunities for it's *operands* add them to the
2866 for (Use &U : I->operands())
2867 if (Instruction *OpI = dyn_cast<Instruction>(U.get()))
2874 // Now that we have an instruction, try combining it to simplify it.
2875 Builder->SetInsertPoint(I->getParent(), I);
2876 Builder->SetCurrentDebugLocation(I->getDebugLoc());
2881 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str(););
2882 DEBUG(dbgs() << "IC: Visiting: " << OrigI << '\n');
2884 if (Instruction *Result = visit(*I)) {
2886 // Should we replace the old instruction with a new one?
2888 DEBUG(dbgs() << "IC: Old = " << *I << '\n'
2889 << " New = " << *Result << '\n');
2891 if (!I->getDebugLoc().isUnknown())
2892 Result->setDebugLoc(I->getDebugLoc());
2893 // Everything uses the new instruction now.
2894 I->replaceAllUsesWith(Result);
2896 // Move the name to the new instruction first.
2897 Result->takeName(I);
2899 // Push the new instruction and any users onto the worklist.
2900 Worklist.Add(Result);
2901 Worklist.AddUsersToWorkList(*Result);
2903 // Insert the new instruction into the basic block...
2904 BasicBlock *InstParent = I->getParent();
2905 BasicBlock::iterator InsertPos = I;
2907 // If we replace a PHI with something that isn't a PHI, fix up the
2909 if (!isa<PHINode>(Result) && isa<PHINode>(InsertPos))
2910 InsertPos = InstParent->getFirstInsertionPt();
2912 InstParent->getInstList().insert(InsertPos, Result);
2914 EraseInstFromFunction(*I);
2917 DEBUG(dbgs() << "IC: Mod = " << OrigI << '\n'
2918 << " New = " << *I << '\n');
2921 // If the instruction was modified, it's possible that it is now dead.
2922 // if so, remove it.
2923 if (isInstructionTriviallyDead(I, TLI)) {
2924 EraseInstFromFunction(*I);
2927 Worklist.AddUsersToWorkList(*I);
2930 MadeIRChange = true;
2935 return MadeIRChange;
2939 class InstCombinerLibCallSimplifier final : public LibCallSimplifier {
2942 InstCombinerLibCallSimplifier(const DataLayout *DL,
2943 const TargetLibraryInfo *TLI,
2945 : LibCallSimplifier(DL, TLI) {
2949 /// replaceAllUsesWith - override so that instruction replacement
2950 /// can be defined in terms of the instruction combiner framework.
2951 void replaceAllUsesWith(Instruction *I, Value *With) const override {
2952 IC->ReplaceInstUsesWith(*I, With);
2957 bool InstCombiner::runOnFunction(Function &F) {
2958 if (skipOptnoneFunction(F))
2961 AT = &getAnalysis<AssumptionTracker>();
2962 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
2963 DL = DLP ? &DLP->getDataLayout() : nullptr;
2964 TLI = &getAnalysis<TargetLibraryInfo>();
2966 DominatorTreeWrapperPass *DTWP =
2967 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
2968 DT = DTWP ? &DTWP->getDomTree() : nullptr;
2971 MinimizeSize = F.getAttributes().hasAttribute(AttributeSet::FunctionIndex,
2972 Attribute::MinSize);
2974 /// Builder - This is an IRBuilder that automatically inserts new
2975 /// instructions into the worklist when they are created.
2976 IRBuilder<true, TargetFolder, InstCombineIRInserter>
2977 TheBuilder(F.getContext(), TargetFolder(DL),
2978 InstCombineIRInserter(Worklist, AT));
2979 Builder = &TheBuilder;
2981 InstCombinerLibCallSimplifier TheSimplifier(DL, TLI, this);
2982 Simplifier = &TheSimplifier;
2984 bool EverMadeChange = false;
2986 // Lower dbg.declare intrinsics otherwise their value may be clobbered
2988 EverMadeChange = LowerDbgDeclare(F);
2990 // Iterate while there is work to do.
2991 unsigned Iteration = 0;
2992 while (DoOneIteration(F, Iteration++))
2993 EverMadeChange = true;
2996 return EverMadeChange;
2999 FunctionPass *llvm::createInstructionCombiningPass() {
3000 return new InstCombiner();