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/InstCombine/InstCombine.h"
37 #include "InstCombineInternal.h"
38 #include "llvm-c/Initialization.h"
39 #include "llvm/ADT/SmallPtrSet.h"
40 #include "llvm/ADT/Statistic.h"
41 #include "llvm/ADT/StringSwitch.h"
42 #include "llvm/Analysis/AssumptionCache.h"
43 #include "llvm/Analysis/CFG.h"
44 #include "llvm/Analysis/ConstantFolding.h"
45 #include "llvm/Analysis/GlobalsModRef.h"
46 #include "llvm/Analysis/InstructionSimplify.h"
47 #include "llvm/Analysis/LibCallSemantics.h"
48 #include "llvm/Analysis/LoopInfo.h"
49 #include "llvm/Analysis/MemoryBuiltins.h"
50 #include "llvm/Analysis/TargetLibraryInfo.h"
51 #include "llvm/Analysis/ValueTracking.h"
52 #include "llvm/IR/CFG.h"
53 #include "llvm/IR/DataLayout.h"
54 #include "llvm/IR/Dominators.h"
55 #include "llvm/IR/GetElementPtrTypeIterator.h"
56 #include "llvm/IR/IntrinsicInst.h"
57 #include "llvm/IR/PatternMatch.h"
58 #include "llvm/IR/ValueHandle.h"
59 #include "llvm/Support/CommandLine.h"
60 #include "llvm/Support/Debug.h"
61 #include "llvm/Support/raw_ostream.h"
62 #include "llvm/Transforms/Scalar.h"
63 #include "llvm/Transforms/Utils/Local.h"
67 using namespace llvm::PatternMatch;
69 #define DEBUG_TYPE "instcombine"
71 STATISTIC(NumCombined , "Number of insts combined");
72 STATISTIC(NumConstProp, "Number of constant folds");
73 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
74 STATISTIC(NumSunkInst , "Number of instructions sunk");
75 STATISTIC(NumExpand, "Number of expansions");
76 STATISTIC(NumFactor , "Number of factorizations");
77 STATISTIC(NumReassoc , "Number of reassociations");
79 Value *InstCombiner::EmitGEPOffset(User *GEP) {
80 return llvm::EmitGEPOffset(Builder, DL, GEP);
83 /// Return true if it is desirable to convert a computation from 'From' to 'To'.
84 /// We don't want to convert from a legal to an illegal type for example or from
85 /// a smaller to a larger illegal type.
86 bool InstCombiner::ShouldChangeType(Type *From, Type *To) const {
87 assert(From->isIntegerTy() && To->isIntegerTy());
89 unsigned FromWidth = From->getPrimitiveSizeInBits();
90 unsigned ToWidth = To->getPrimitiveSizeInBits();
91 bool FromLegal = DL.isLegalInteger(FromWidth);
92 bool ToLegal = DL.isLegalInteger(ToWidth);
94 // If this is a legal integer from type, and the result would be an illegal
95 // type, don't do the transformation.
96 if (FromLegal && !ToLegal)
99 // Otherwise, if both are illegal, do not increase the size of the result. We
100 // do allow things like i160 -> i64, but not i64 -> i160.
101 if (!FromLegal && !ToLegal && ToWidth > FromWidth)
107 // Return true, if No Signed Wrap should be maintained for I.
108 // The No Signed Wrap flag can be kept if the operation "B (I.getOpcode) C",
109 // where both B and C should be ConstantInts, results in a constant that does
110 // not overflow. This function only handles the Add and Sub opcodes. For
111 // all other opcodes, the function conservatively returns false.
112 static bool MaintainNoSignedWrap(BinaryOperator &I, Value *B, Value *C) {
113 OverflowingBinaryOperator *OBO = dyn_cast<OverflowingBinaryOperator>(&I);
114 if (!OBO || !OBO->hasNoSignedWrap()) {
118 // We reason about Add and Sub Only.
119 Instruction::BinaryOps Opcode = I.getOpcode();
120 if (Opcode != Instruction::Add &&
121 Opcode != Instruction::Sub) {
125 ConstantInt *CB = dyn_cast<ConstantInt>(B);
126 ConstantInt *CC = dyn_cast<ConstantInt>(C);
132 const APInt &BVal = CB->getValue();
133 const APInt &CVal = CC->getValue();
134 bool Overflow = false;
136 if (Opcode == Instruction::Add) {
137 BVal.sadd_ov(CVal, Overflow);
139 BVal.ssub_ov(CVal, Overflow);
145 /// Conservatively clears subclassOptionalData after a reassociation or
146 /// commutation. We preserve fast-math flags when applicable as they can be
148 static void ClearSubclassDataAfterReassociation(BinaryOperator &I) {
149 FPMathOperator *FPMO = dyn_cast<FPMathOperator>(&I);
151 I.clearSubclassOptionalData();
155 FastMathFlags FMF = I.getFastMathFlags();
156 I.clearSubclassOptionalData();
157 I.setFastMathFlags(FMF);
160 /// This performs a few simplifications for operators that are associative or
163 /// Commutative operators:
165 /// 1. Order operands such that they are listed from right (least complex) to
166 /// left (most complex). This puts constants before unary operators before
167 /// binary operators.
169 /// Associative operators:
171 /// 2. Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
172 /// 3. Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
174 /// Associative and commutative operators:
176 /// 4. Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
177 /// 5. Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
178 /// 6. Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
179 /// if C1 and C2 are constants.
180 bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) {
181 Instruction::BinaryOps Opcode = I.getOpcode();
182 bool Changed = false;
185 // Order operands such that they are listed from right (least complex) to
186 // left (most complex). This puts constants before unary operators before
188 if (I.isCommutative() && getComplexity(I.getOperand(0)) <
189 getComplexity(I.getOperand(1)))
190 Changed = !I.swapOperands();
192 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
193 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
195 if (I.isAssociative()) {
196 // Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
197 if (Op0 && Op0->getOpcode() == Opcode) {
198 Value *A = Op0->getOperand(0);
199 Value *B = Op0->getOperand(1);
200 Value *C = I.getOperand(1);
202 // Does "B op C" simplify?
203 if (Value *V = SimplifyBinOp(Opcode, B, C, DL)) {
204 // It simplifies to V. Form "A op V".
207 // Conservatively clear the optional flags, since they may not be
208 // preserved by the reassociation.
209 if (MaintainNoSignedWrap(I, B, C) &&
210 (!Op0 || (isa<BinaryOperator>(Op0) && Op0->hasNoSignedWrap()))) {
211 // Note: this is only valid because SimplifyBinOp doesn't look at
212 // the operands to Op0.
213 I.clearSubclassOptionalData();
214 I.setHasNoSignedWrap(true);
216 ClearSubclassDataAfterReassociation(I);
225 // Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
226 if (Op1 && Op1->getOpcode() == Opcode) {
227 Value *A = I.getOperand(0);
228 Value *B = Op1->getOperand(0);
229 Value *C = Op1->getOperand(1);
231 // Does "A op B" simplify?
232 if (Value *V = SimplifyBinOp(Opcode, A, B, DL)) {
233 // It simplifies to V. Form "V op C".
236 // Conservatively clear the optional flags, since they may not be
237 // preserved by the reassociation.
238 ClearSubclassDataAfterReassociation(I);
246 if (I.isAssociative() && I.isCommutative()) {
247 // Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
248 if (Op0 && Op0->getOpcode() == Opcode) {
249 Value *A = Op0->getOperand(0);
250 Value *B = Op0->getOperand(1);
251 Value *C = I.getOperand(1);
253 // Does "C op A" simplify?
254 if (Value *V = SimplifyBinOp(Opcode, C, A, DL)) {
255 // It simplifies to V. Form "V op B".
258 // Conservatively clear the optional flags, since they may not be
259 // preserved by the reassociation.
260 ClearSubclassDataAfterReassociation(I);
267 // Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
268 if (Op1 && Op1->getOpcode() == Opcode) {
269 Value *A = I.getOperand(0);
270 Value *B = Op1->getOperand(0);
271 Value *C = Op1->getOperand(1);
273 // Does "C op A" simplify?
274 if (Value *V = SimplifyBinOp(Opcode, C, A, DL)) {
275 // It simplifies to V. Form "B op V".
278 // Conservatively clear the optional flags, since they may not be
279 // preserved by the reassociation.
280 ClearSubclassDataAfterReassociation(I);
287 // Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
288 // if C1 and C2 are constants.
290 Op0->getOpcode() == Opcode && Op1->getOpcode() == Opcode &&
291 isa<Constant>(Op0->getOperand(1)) &&
292 isa<Constant>(Op1->getOperand(1)) &&
293 Op0->hasOneUse() && Op1->hasOneUse()) {
294 Value *A = Op0->getOperand(0);
295 Constant *C1 = cast<Constant>(Op0->getOperand(1));
296 Value *B = Op1->getOperand(0);
297 Constant *C2 = cast<Constant>(Op1->getOperand(1));
299 Constant *Folded = ConstantExpr::get(Opcode, C1, C2);
300 BinaryOperator *New = BinaryOperator::Create(Opcode, A, B);
301 if (isa<FPMathOperator>(New)) {
302 FastMathFlags Flags = I.getFastMathFlags();
303 Flags &= Op0->getFastMathFlags();
304 Flags &= Op1->getFastMathFlags();
305 New->setFastMathFlags(Flags);
307 InsertNewInstWith(New, I);
309 I.setOperand(0, New);
310 I.setOperand(1, Folded);
311 // Conservatively clear the optional flags, since they may not be
312 // preserved by the reassociation.
313 ClearSubclassDataAfterReassociation(I);
320 // No further simplifications.
325 /// Return whether "X LOp (Y ROp Z)" is always equal to
326 /// "(X LOp Y) ROp (X LOp Z)".
327 static bool LeftDistributesOverRight(Instruction::BinaryOps LOp,
328 Instruction::BinaryOps ROp) {
333 case Instruction::And:
334 // And distributes over Or and Xor.
338 case Instruction::Or:
339 case Instruction::Xor:
343 case Instruction::Mul:
344 // Multiplication distributes over addition and subtraction.
348 case Instruction::Add:
349 case Instruction::Sub:
353 case Instruction::Or:
354 // Or distributes over And.
358 case Instruction::And:
364 /// Return whether "(X LOp Y) ROp Z" is always equal to
365 /// "(X ROp Z) LOp (Y ROp Z)".
366 static bool RightDistributesOverLeft(Instruction::BinaryOps LOp,
367 Instruction::BinaryOps ROp) {
368 if (Instruction::isCommutative(ROp))
369 return LeftDistributesOverRight(ROp, LOp);
374 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
375 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
376 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
377 case Instruction::And:
378 case Instruction::Or:
379 case Instruction::Xor:
383 case Instruction::Shl:
384 case Instruction::LShr:
385 case Instruction::AShr:
389 // TODO: It would be nice to handle division, aka "(X + Y)/Z = X/Z + Y/Z",
390 // but this requires knowing that the addition does not overflow and other
395 /// This function returns identity value for given opcode, which can be used to
396 /// factor patterns like (X * 2) + X ==> (X * 2) + (X * 1) ==> X * (2 + 1).
397 static Value *getIdentityValue(Instruction::BinaryOps OpCode, Value *V) {
398 if (isa<Constant>(V))
401 if (OpCode == Instruction::Mul)
402 return ConstantInt::get(V->getType(), 1);
404 // TODO: We can handle other cases e.g. Instruction::And, Instruction::Or etc.
409 /// This function factors binary ops which can be combined using distributive
410 /// laws. This function tries to transform 'Op' based TopLevelOpcode to enable
411 /// factorization e.g for ADD(SHL(X , 2), MUL(X, 5)), When this function called
412 /// with TopLevelOpcode == Instruction::Add and Op = SHL(X, 2), transforms
413 /// SHL(X, 2) to MUL(X, 4) i.e. returns Instruction::Mul with LHS set to 'X' and
415 static Instruction::BinaryOps
416 getBinOpsForFactorization(Instruction::BinaryOps TopLevelOpcode,
417 BinaryOperator *Op, Value *&LHS, Value *&RHS) {
419 return Instruction::BinaryOpsEnd;
421 LHS = Op->getOperand(0);
422 RHS = Op->getOperand(1);
424 switch (TopLevelOpcode) {
426 return Op->getOpcode();
428 case Instruction::Add:
429 case Instruction::Sub:
430 if (Op->getOpcode() == Instruction::Shl) {
431 if (Constant *CST = dyn_cast<Constant>(Op->getOperand(1))) {
432 // The multiplier is really 1 << CST.
433 RHS = ConstantExpr::getShl(ConstantInt::get(Op->getType(), 1), CST);
434 return Instruction::Mul;
437 return Op->getOpcode();
440 // TODO: We can add other conversions e.g. shr => div etc.
443 /// This tries to simplify binary operations by factorizing out common terms
444 /// (e. g. "(A*B)+(A*C)" -> "A*(B+C)").
445 static Value *tryFactorization(InstCombiner::BuilderTy *Builder,
446 const DataLayout &DL, BinaryOperator &I,
447 Instruction::BinaryOps InnerOpcode, Value *A,
448 Value *B, Value *C, Value *D) {
450 // If any of A, B, C, D are null, we can not factor I, return early.
451 // Checking A and C should be enough.
452 if (!A || !C || !B || !D)
456 Value *SimplifiedInst = nullptr;
457 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
458 Instruction::BinaryOps TopLevelOpcode = I.getOpcode();
460 // Does "X op' Y" always equal "Y op' X"?
461 bool InnerCommutative = Instruction::isCommutative(InnerOpcode);
463 // Does "X op' (Y op Z)" always equal "(X op' Y) op (X op' Z)"?
464 if (LeftDistributesOverRight(InnerOpcode, TopLevelOpcode))
465 // Does the instruction have the form "(A op' B) op (A op' D)" or, in the
466 // commutative case, "(A op' B) op (C op' A)"?
467 if (A == C || (InnerCommutative && A == D)) {
470 // Consider forming "A op' (B op D)".
471 // If "B op D" simplifies then it can be formed with no cost.
472 V = SimplifyBinOp(TopLevelOpcode, B, D, DL);
473 // If "B op D" doesn't simplify then only go on if both of the existing
474 // operations "A op' B" and "C op' D" will be zapped as no longer used.
475 if (!V && LHS->hasOneUse() && RHS->hasOneUse())
476 V = Builder->CreateBinOp(TopLevelOpcode, B, D, RHS->getName());
478 SimplifiedInst = Builder->CreateBinOp(InnerOpcode, A, V);
482 // Does "(X op Y) op' Z" always equal "(X op' Z) op (Y op' Z)"?
483 if (!SimplifiedInst && RightDistributesOverLeft(TopLevelOpcode, InnerOpcode))
484 // Does the instruction have the form "(A op' B) op (C op' B)" or, in the
485 // commutative case, "(A op' B) op (B op' D)"?
486 if (B == D || (InnerCommutative && B == C)) {
489 // Consider forming "(A op C) op' B".
490 // If "A op C" simplifies then it can be formed with no cost.
491 V = SimplifyBinOp(TopLevelOpcode, A, C, DL);
493 // If "A op C" doesn't simplify then only go on if both of the existing
494 // operations "A op' B" and "C op' D" will be zapped as no longer used.
495 if (!V && LHS->hasOneUse() && RHS->hasOneUse())
496 V = Builder->CreateBinOp(TopLevelOpcode, A, C, LHS->getName());
498 SimplifiedInst = Builder->CreateBinOp(InnerOpcode, V, B);
502 if (SimplifiedInst) {
504 SimplifiedInst->takeName(&I);
506 // Check if we can add NSW flag to SimplifiedInst. If so, set NSW flag.
507 // TODO: Check for NUW.
508 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(SimplifiedInst)) {
509 if (isa<OverflowingBinaryOperator>(SimplifiedInst)) {
511 if (isa<OverflowingBinaryOperator>(&I))
512 HasNSW = I.hasNoSignedWrap();
514 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS))
515 if (isa<OverflowingBinaryOperator>(Op0))
516 HasNSW &= Op0->hasNoSignedWrap();
518 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS))
519 if (isa<OverflowingBinaryOperator>(Op1))
520 HasNSW &= Op1->hasNoSignedWrap();
522 // We can propagate 'nsw' if we know that
523 // %Y = mul nsw i16 %X, C
524 // %Z = add nsw i16 %Y, %X
526 // %Z = mul nsw i16 %X, C+1
528 // iff C+1 isn't INT_MIN
530 if (TopLevelOpcode == Instruction::Add &&
531 InnerOpcode == Instruction::Mul)
532 if (match(V, m_APInt(CInt)) && !CInt->isMinSignedValue())
533 BO->setHasNoSignedWrap(HasNSW);
537 return SimplifiedInst;
540 /// This tries to simplify binary operations which some other binary operation
541 /// distributes over either by factorizing out common terms
542 /// (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this results in
543 /// simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is a win).
544 /// Returns the simplified value, or null if it didn't simplify.
545 Value *InstCombiner::SimplifyUsingDistributiveLaws(BinaryOperator &I) {
546 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
547 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
548 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
551 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
552 auto TopLevelOpcode = I.getOpcode();
553 auto LHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op0, A, B);
554 auto RHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op1, C, D);
556 // The instruction has the form "(A op' B) op (C op' D)". Try to factorize
558 if (LHSOpcode == RHSOpcode) {
559 if (Value *V = tryFactorization(Builder, DL, I, LHSOpcode, A, B, C, D))
563 // The instruction has the form "(A op' B) op (C)". Try to factorize common
565 if (Value *V = tryFactorization(Builder, DL, I, LHSOpcode, A, B, RHS,
566 getIdentityValue(LHSOpcode, RHS)))
569 // The instruction has the form "(B) op (C op' D)". Try to factorize common
571 if (Value *V = tryFactorization(Builder, DL, I, RHSOpcode, LHS,
572 getIdentityValue(RHSOpcode, LHS), C, D))
576 if (Op0 && RightDistributesOverLeft(Op0->getOpcode(), TopLevelOpcode)) {
577 // The instruction has the form "(A op' B) op C". See if expanding it out
578 // to "(A op C) op' (B op C)" results in simplifications.
579 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
580 Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op'
582 // Do "A op C" and "B op C" both simplify?
583 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, C, DL))
584 if (Value *R = SimplifyBinOp(TopLevelOpcode, B, C, DL)) {
585 // They do! Return "L op' R".
587 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
588 if ((L == A && R == B) ||
589 (Instruction::isCommutative(InnerOpcode) && L == B && R == A))
591 // Otherwise return "L op' R" if it simplifies.
592 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, DL))
594 // Otherwise, create a new instruction.
595 C = Builder->CreateBinOp(InnerOpcode, L, R);
601 if (Op1 && LeftDistributesOverRight(TopLevelOpcode, Op1->getOpcode())) {
602 // The instruction has the form "A op (B op' C)". See if expanding it out
603 // to "(A op B) op' (A op C)" results in simplifications.
604 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
605 Instruction::BinaryOps InnerOpcode = Op1->getOpcode(); // op'
607 // Do "A op B" and "A op C" both simplify?
608 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, B, DL))
609 if (Value *R = SimplifyBinOp(TopLevelOpcode, A, C, DL)) {
610 // They do! Return "L op' R".
612 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
613 if ((L == B && R == C) ||
614 (Instruction::isCommutative(InnerOpcode) && L == C && R == B))
616 // Otherwise return "L op' R" if it simplifies.
617 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, DL))
619 // Otherwise, create a new instruction.
620 A = Builder->CreateBinOp(InnerOpcode, L, R);
626 // (op (select (a, c, b)), (select (a, d, b))) -> (select (a, (op c, d), 0))
627 // (op (select (a, b, c)), (select (a, b, d))) -> (select (a, 0, (op c, d)))
628 if (auto *SI0 = dyn_cast<SelectInst>(LHS)) {
629 if (auto *SI1 = dyn_cast<SelectInst>(RHS)) {
630 if (SI0->getCondition() == SI1->getCondition()) {
632 if (Value *V = SimplifyBinOp(TopLevelOpcode, SI0->getFalseValue(),
633 SI1->getFalseValue(), DL, TLI, DT, AC))
634 SI = Builder->CreateSelect(SI0->getCondition(),
635 Builder->CreateBinOp(TopLevelOpcode,
637 SI1->getTrueValue()),
639 if (Value *V = SimplifyBinOp(TopLevelOpcode, SI0->getTrueValue(),
640 SI1->getTrueValue(), DL, TLI, DT, AC))
641 SI = Builder->CreateSelect(
642 SI0->getCondition(), V,
643 Builder->CreateBinOp(TopLevelOpcode, SI0->getFalseValue(),
644 SI1->getFalseValue()));
656 /// Given a 'sub' instruction, return the RHS of the instruction if the LHS is a
657 /// constant zero (which is the 'negate' form).
658 Value *InstCombiner::dyn_castNegVal(Value *V) const {
659 if (BinaryOperator::isNeg(V))
660 return BinaryOperator::getNegArgument(V);
662 // Constants can be considered to be negated values if they can be folded.
663 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
664 return ConstantExpr::getNeg(C);
666 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
667 if (C->getType()->getElementType()->isIntegerTy())
668 return ConstantExpr::getNeg(C);
673 /// Given a 'fsub' instruction, return the RHS of the instruction if the LHS is
674 /// a constant negative zero (which is the 'negate' form).
675 Value *InstCombiner::dyn_castFNegVal(Value *V, bool IgnoreZeroSign) const {
676 if (BinaryOperator::isFNeg(V, IgnoreZeroSign))
677 return BinaryOperator::getFNegArgument(V);
679 // Constants can be considered to be negated values if they can be folded.
680 if (ConstantFP *C = dyn_cast<ConstantFP>(V))
681 return ConstantExpr::getFNeg(C);
683 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
684 if (C->getType()->getElementType()->isFloatingPointTy())
685 return ConstantExpr::getFNeg(C);
690 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
692 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
693 return IC->Builder->CreateCast(CI->getOpcode(), SO, I.getType());
696 // Figure out if the constant is the left or the right argument.
697 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
698 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
700 if (Constant *SOC = dyn_cast<Constant>(SO)) {
702 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
703 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
706 Value *Op0 = SO, *Op1 = ConstOperand;
710 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I)) {
711 Value *RI = IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1,
712 SO->getName()+".op");
713 Instruction *FPInst = dyn_cast<Instruction>(RI);
714 if (FPInst && isa<FPMathOperator>(FPInst))
715 FPInst->copyFastMathFlags(BO);
718 if (ICmpInst *CI = dyn_cast<ICmpInst>(&I))
719 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
720 SO->getName()+".cmp");
721 if (FCmpInst *CI = dyn_cast<FCmpInst>(&I))
722 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
723 SO->getName()+".cmp");
724 llvm_unreachable("Unknown binary instruction type!");
727 /// Given an instruction with a select as one operand and a constant as the
728 /// other operand, try to fold the binary operator into the select arguments.
729 /// This also works for Cast instructions, which obviously do not have a second
731 Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) {
732 // Don't modify shared select instructions
733 if (!SI->hasOneUse()) return nullptr;
734 Value *TV = SI->getOperand(1);
735 Value *FV = SI->getOperand(2);
737 if (isa<Constant>(TV) || isa<Constant>(FV)) {
738 // Bool selects with constant operands can be folded to logical ops.
739 if (SI->getType()->isIntegerTy(1)) return nullptr;
741 // If it's a bitcast involving vectors, make sure it has the same number of
742 // elements on both sides.
743 if (BitCastInst *BC = dyn_cast<BitCastInst>(&Op)) {
744 VectorType *DestTy = dyn_cast<VectorType>(BC->getDestTy());
745 VectorType *SrcTy = dyn_cast<VectorType>(BC->getSrcTy());
747 // Verify that either both or neither are vectors.
748 if ((SrcTy == nullptr) != (DestTy == nullptr)) return nullptr;
749 // If vectors, verify that they have the same number of elements.
750 if (SrcTy && SrcTy->getNumElements() != DestTy->getNumElements())
754 // Test if a CmpInst instruction is used exclusively by a select as
755 // part of a minimum or maximum operation. If so, refrain from doing
756 // any other folding. This helps out other analyses which understand
757 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
758 // and CodeGen. And in this case, at least one of the comparison
759 // operands has at least one user besides the compare (the select),
760 // which would often largely negate the benefit of folding anyway.
761 if (auto *CI = dyn_cast<CmpInst>(SI->getCondition())) {
762 if (CI->hasOneUse()) {
763 Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1);
764 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
765 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
770 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, this);
771 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, this);
773 return SelectInst::Create(SI->getCondition(),
774 SelectTrueVal, SelectFalseVal);
779 /// Given a binary operator, cast instruction, or select which has a PHI node as
780 /// operand #0, see if we can fold the instruction into the PHI (which is only
781 /// possible if all operands to the PHI are constants).
782 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
783 PHINode *PN = cast<PHINode>(I.getOperand(0));
784 unsigned NumPHIValues = PN->getNumIncomingValues();
785 if (NumPHIValues == 0)
788 // We normally only transform phis with a single use. However, if a PHI has
789 // multiple uses and they are all the same operation, we can fold *all* of the
790 // uses into the PHI.
791 if (!PN->hasOneUse()) {
792 // Walk the use list for the instruction, comparing them to I.
793 for (User *U : PN->users()) {
794 Instruction *UI = cast<Instruction>(U);
795 if (UI != &I && !I.isIdenticalTo(UI))
798 // Otherwise, we can replace *all* users with the new PHI we form.
801 // Check to see if all of the operands of the PHI are simple constants
802 // (constantint/constantfp/undef). If there is one non-constant value,
803 // remember the BB it is in. If there is more than one or if *it* is a PHI,
804 // bail out. We don't do arbitrary constant expressions here because moving
805 // their computation can be expensive without a cost model.
806 BasicBlock *NonConstBB = nullptr;
807 for (unsigned i = 0; i != NumPHIValues; ++i) {
808 Value *InVal = PN->getIncomingValue(i);
809 if (isa<Constant>(InVal) && !isa<ConstantExpr>(InVal))
812 if (isa<PHINode>(InVal)) return nullptr; // Itself a phi.
813 if (NonConstBB) return nullptr; // More than one non-const value.
815 NonConstBB = PN->getIncomingBlock(i);
817 // If the InVal is an invoke at the end of the pred block, then we can't
818 // insert a computation after it without breaking the edge.
819 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal))
820 if (II->getParent() == NonConstBB)
823 // If the incoming non-constant value is in I's block, we will remove one
824 // instruction, but insert another equivalent one, leading to infinite
826 if (isPotentiallyReachable(I.getParent(), NonConstBB, DT, LI))
830 // If there is exactly one non-constant value, we can insert a copy of the
831 // operation in that block. However, if this is a critical edge, we would be
832 // inserting the computation on some other paths (e.g. inside a loop). Only
833 // do this if the pred block is unconditionally branching into the phi block.
834 if (NonConstBB != nullptr) {
835 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
836 if (!BI || !BI->isUnconditional()) return nullptr;
839 // Okay, we can do the transformation: create the new PHI node.
840 PHINode *NewPN = PHINode::Create(I.getType(), PN->getNumIncomingValues());
841 InsertNewInstBefore(NewPN, *PN);
844 // If we are going to have to insert a new computation, do so right before the
845 // predecessor's terminator.
847 Builder->SetInsertPoint(NonConstBB->getTerminator());
849 // Next, add all of the operands to the PHI.
850 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) {
851 // We only currently try to fold the condition of a select when it is a phi,
852 // not the true/false values.
853 Value *TrueV = SI->getTrueValue();
854 Value *FalseV = SI->getFalseValue();
855 BasicBlock *PhiTransBB = PN->getParent();
856 for (unsigned i = 0; i != NumPHIValues; ++i) {
857 BasicBlock *ThisBB = PN->getIncomingBlock(i);
858 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB);
859 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB);
860 Value *InV = nullptr;
861 // Beware of ConstantExpr: it may eventually evaluate to getNullValue,
862 // even if currently isNullValue gives false.
863 Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i));
864 if (InC && !isa<ConstantExpr>(InC))
865 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred;
867 InV = Builder->CreateSelect(PN->getIncomingValue(i),
868 TrueVInPred, FalseVInPred, "phitmp");
869 NewPN->addIncoming(InV, ThisBB);
871 } else if (CmpInst *CI = dyn_cast<CmpInst>(&I)) {
872 Constant *C = cast<Constant>(I.getOperand(1));
873 for (unsigned i = 0; i != NumPHIValues; ++i) {
874 Value *InV = nullptr;
875 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
876 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
877 else if (isa<ICmpInst>(CI))
878 InV = Builder->CreateICmp(CI->getPredicate(), PN->getIncomingValue(i),
881 InV = Builder->CreateFCmp(CI->getPredicate(), PN->getIncomingValue(i),
883 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
885 } else if (I.getNumOperands() == 2) {
886 Constant *C = cast<Constant>(I.getOperand(1));
887 for (unsigned i = 0; i != NumPHIValues; ++i) {
888 Value *InV = nullptr;
889 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
890 InV = ConstantExpr::get(I.getOpcode(), InC, C);
892 InV = Builder->CreateBinOp(cast<BinaryOperator>(I).getOpcode(),
893 PN->getIncomingValue(i), C, "phitmp");
894 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
897 CastInst *CI = cast<CastInst>(&I);
898 Type *RetTy = CI->getType();
899 for (unsigned i = 0; i != NumPHIValues; ++i) {
901 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
902 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
904 InV = Builder->CreateCast(CI->getOpcode(),
905 PN->getIncomingValue(i), I.getType(), "phitmp");
906 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
910 for (auto UI = PN->user_begin(), E = PN->user_end(); UI != E;) {
911 Instruction *User = cast<Instruction>(*UI++);
912 if (User == &I) continue;
913 ReplaceInstUsesWith(*User, NewPN);
914 EraseInstFromFunction(*User);
916 return ReplaceInstUsesWith(I, NewPN);
919 /// Given a pointer type and a constant offset, determine whether or not there
920 /// is a sequence of GEP indices into the pointed type that will land us at the
921 /// specified offset. If so, fill them into NewIndices and return the resultant
922 /// element type, otherwise return null.
923 Type *InstCombiner::FindElementAtOffset(PointerType *PtrTy, int64_t Offset,
924 SmallVectorImpl<Value *> &NewIndices) {
925 Type *Ty = PtrTy->getElementType();
929 // Start with the index over the outer type. Note that the type size
930 // might be zero (even if the offset isn't zero) if the indexed type
931 // is something like [0 x {int, int}]
932 Type *IntPtrTy = DL.getIntPtrType(PtrTy);
933 int64_t FirstIdx = 0;
934 if (int64_t TySize = DL.getTypeAllocSize(Ty)) {
935 FirstIdx = Offset/TySize;
936 Offset -= FirstIdx*TySize;
938 // Handle hosts where % returns negative instead of values [0..TySize).
944 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
947 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
949 // Index into the types. If we fail, set OrigBase to null.
951 // Indexing into tail padding between struct/array elements.
952 if (uint64_t(Offset * 8) >= DL.getTypeSizeInBits(Ty))
955 if (StructType *STy = dyn_cast<StructType>(Ty)) {
956 const StructLayout *SL = DL.getStructLayout(STy);
957 assert(Offset < (int64_t)SL->getSizeInBytes() &&
958 "Offset must stay within the indexed type");
960 unsigned Elt = SL->getElementContainingOffset(Offset);
961 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
964 Offset -= SL->getElementOffset(Elt);
965 Ty = STy->getElementType(Elt);
966 } else if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
967 uint64_t EltSize = DL.getTypeAllocSize(AT->getElementType());
968 assert(EltSize && "Cannot index into a zero-sized array");
969 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
971 Ty = AT->getElementType();
973 // Otherwise, we can't index into the middle of this atomic type, bail.
981 static bool shouldMergeGEPs(GEPOperator &GEP, GEPOperator &Src) {
982 // If this GEP has only 0 indices, it is the same pointer as
983 // Src. If Src is not a trivial GEP too, don't combine
985 if (GEP.hasAllZeroIndices() && !Src.hasAllZeroIndices() &&
991 /// Return a value X such that Val = X * Scale, or null if none.
992 /// If the multiplication is known not to overflow, then NoSignedWrap is set.
993 Value *InstCombiner::Descale(Value *Val, APInt Scale, bool &NoSignedWrap) {
994 assert(isa<IntegerType>(Val->getType()) && "Can only descale integers!");
995 assert(cast<IntegerType>(Val->getType())->getBitWidth() ==
996 Scale.getBitWidth() && "Scale not compatible with value!");
998 // If Val is zero or Scale is one then Val = Val * Scale.
999 if (match(Val, m_Zero()) || Scale == 1) {
1000 NoSignedWrap = true;
1004 // If Scale is zero then it does not divide Val.
1005 if (Scale.isMinValue())
1008 // Look through chains of multiplications, searching for a constant that is
1009 // divisible by Scale. For example, descaling X*(Y*(Z*4)) by a factor of 4
1010 // will find the constant factor 4 and produce X*(Y*Z). Descaling X*(Y*8) by
1011 // a factor of 4 will produce X*(Y*2). The principle of operation is to bore
1014 // Val = M1 * X || Analysis starts here and works down
1015 // M1 = M2 * Y || Doesn't descend into terms with more
1016 // M2 = Z * 4 \/ than one use
1018 // Then to modify a term at the bottom:
1021 // M1 = Z * Y || Replaced M2 with Z
1023 // Then to work back up correcting nsw flags.
1025 // Op - the term we are currently analyzing. Starts at Val then drills down.
1026 // Replaced with its descaled value before exiting from the drill down loop.
1029 // Parent - initially null, but after drilling down notes where Op came from.
1030 // In the example above, Parent is (Val, 0) when Op is M1, because M1 is the
1031 // 0'th operand of Val.
1032 std::pair<Instruction*, unsigned> Parent;
1034 // Set if the transform requires a descaling at deeper levels that doesn't
1036 bool RequireNoSignedWrap = false;
1038 // Log base 2 of the scale. Negative if not a power of 2.
1039 int32_t logScale = Scale.exactLogBase2();
1041 for (;; Op = Parent.first->getOperand(Parent.second)) { // Drill down
1043 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
1044 // If Op is a constant divisible by Scale then descale to the quotient.
1045 APInt Quotient(Scale), Remainder(Scale); // Init ensures right bitwidth.
1046 APInt::sdivrem(CI->getValue(), Scale, Quotient, Remainder);
1047 if (!Remainder.isMinValue())
1048 // Not divisible by Scale.
1050 // Replace with the quotient in the parent.
1051 Op = ConstantInt::get(CI->getType(), Quotient);
1052 NoSignedWrap = true;
1056 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op)) {
1058 if (BO->getOpcode() == Instruction::Mul) {
1060 NoSignedWrap = BO->hasNoSignedWrap();
1061 if (RequireNoSignedWrap && !NoSignedWrap)
1064 // There are three cases for multiplication: multiplication by exactly
1065 // the scale, multiplication by a constant different to the scale, and
1066 // multiplication by something else.
1067 Value *LHS = BO->getOperand(0);
1068 Value *RHS = BO->getOperand(1);
1070 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
1071 // Multiplication by a constant.
1072 if (CI->getValue() == Scale) {
1073 // Multiplication by exactly the scale, replace the multiplication
1074 // by its left-hand side in the parent.
1079 // Otherwise drill down into the constant.
1080 if (!Op->hasOneUse())
1083 Parent = std::make_pair(BO, 1);
1087 // Multiplication by something else. Drill down into the left-hand side
1088 // since that's where the reassociate pass puts the good stuff.
1089 if (!Op->hasOneUse())
1092 Parent = std::make_pair(BO, 0);
1096 if (logScale > 0 && BO->getOpcode() == Instruction::Shl &&
1097 isa<ConstantInt>(BO->getOperand(1))) {
1098 // Multiplication by a power of 2.
1099 NoSignedWrap = BO->hasNoSignedWrap();
1100 if (RequireNoSignedWrap && !NoSignedWrap)
1103 Value *LHS = BO->getOperand(0);
1104 int32_t Amt = cast<ConstantInt>(BO->getOperand(1))->
1105 getLimitedValue(Scale.getBitWidth());
1108 if (Amt == logScale) {
1109 // Multiplication by exactly the scale, replace the multiplication
1110 // by its left-hand side in the parent.
1114 if (Amt < logScale || !Op->hasOneUse())
1117 // Multiplication by more than the scale. Reduce the multiplying amount
1118 // by the scale in the parent.
1119 Parent = std::make_pair(BO, 1);
1120 Op = ConstantInt::get(BO->getType(), Amt - logScale);
1125 if (!Op->hasOneUse())
1128 if (CastInst *Cast = dyn_cast<CastInst>(Op)) {
1129 if (Cast->getOpcode() == Instruction::SExt) {
1130 // Op is sign-extended from a smaller type, descale in the smaller type.
1131 unsigned SmallSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1132 APInt SmallScale = Scale.trunc(SmallSize);
1133 // Suppose Op = sext X, and we descale X as Y * SmallScale. We want to
1134 // descale Op as (sext Y) * Scale. In order to have
1135 // sext (Y * SmallScale) = (sext Y) * Scale
1136 // some conditions need to hold however: SmallScale must sign-extend to
1137 // Scale and the multiplication Y * SmallScale should not overflow.
1138 if (SmallScale.sext(Scale.getBitWidth()) != Scale)
1139 // SmallScale does not sign-extend to Scale.
1141 assert(SmallScale.exactLogBase2() == logScale);
1142 // Require that Y * SmallScale must not overflow.
1143 RequireNoSignedWrap = true;
1145 // Drill down through the cast.
1146 Parent = std::make_pair(Cast, 0);
1151 if (Cast->getOpcode() == Instruction::Trunc) {
1152 // Op is truncated from a larger type, descale in the larger type.
1153 // Suppose Op = trunc X, and we descale X as Y * sext Scale. Then
1154 // trunc (Y * sext Scale) = (trunc Y) * Scale
1155 // always holds. However (trunc Y) * Scale may overflow even if
1156 // trunc (Y * sext Scale) does not, so nsw flags need to be cleared
1157 // from this point up in the expression (see later).
1158 if (RequireNoSignedWrap)
1161 // Drill down through the cast.
1162 unsigned LargeSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1163 Parent = std::make_pair(Cast, 0);
1164 Scale = Scale.sext(LargeSize);
1165 if (logScale + 1 == (int32_t)Cast->getType()->getPrimitiveSizeInBits())
1167 assert(Scale.exactLogBase2() == logScale);
1172 // Unsupported expression, bail out.
1176 // If Op is zero then Val = Op * Scale.
1177 if (match(Op, m_Zero())) {
1178 NoSignedWrap = true;
1182 // We know that we can successfully descale, so from here on we can safely
1183 // modify the IR. Op holds the descaled version of the deepest term in the
1184 // expression. NoSignedWrap is 'true' if multiplying Op by Scale is known
1188 // The expression only had one term.
1191 // Rewrite the parent using the descaled version of its operand.
1192 assert(Parent.first->hasOneUse() && "Drilled down when more than one use!");
1193 assert(Op != Parent.first->getOperand(Parent.second) &&
1194 "Descaling was a no-op?");
1195 Parent.first->setOperand(Parent.second, Op);
1196 Worklist.Add(Parent.first);
1198 // Now work back up the expression correcting nsw flags. The logic is based
1199 // on the following observation: if X * Y is known not to overflow as a signed
1200 // multiplication, and Y is replaced by a value Z with smaller absolute value,
1201 // then X * Z will not overflow as a signed multiplication either. As we work
1202 // our way up, having NoSignedWrap 'true' means that the descaled value at the
1203 // current level has strictly smaller absolute value than the original.
1204 Instruction *Ancestor = Parent.first;
1206 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Ancestor)) {
1207 // If the multiplication wasn't nsw then we can't say anything about the
1208 // value of the descaled multiplication, and we have to clear nsw flags
1209 // from this point on up.
1210 bool OpNoSignedWrap = BO->hasNoSignedWrap();
1211 NoSignedWrap &= OpNoSignedWrap;
1212 if (NoSignedWrap != OpNoSignedWrap) {
1213 BO->setHasNoSignedWrap(NoSignedWrap);
1214 Worklist.Add(Ancestor);
1216 } else if (Ancestor->getOpcode() == Instruction::Trunc) {
1217 // The fact that the descaled input to the trunc has smaller absolute
1218 // value than the original input doesn't tell us anything useful about
1219 // the absolute values of the truncations.
1220 NoSignedWrap = false;
1222 assert((Ancestor->getOpcode() != Instruction::SExt || NoSignedWrap) &&
1223 "Failed to keep proper track of nsw flags while drilling down?");
1225 if (Ancestor == Val)
1226 // Got to the top, all done!
1229 // Move up one level in the expression.
1230 assert(Ancestor->hasOneUse() && "Drilled down when more than one use!");
1231 Ancestor = Ancestor->user_back();
1235 /// \brief Creates node of binary operation with the same attributes as the
1236 /// specified one but with other operands.
1237 static Value *CreateBinOpAsGiven(BinaryOperator &Inst, Value *LHS, Value *RHS,
1238 InstCombiner::BuilderTy *B) {
1239 Value *BORes = B->CreateBinOp(Inst.getOpcode(), LHS, RHS);
1240 if (BinaryOperator *NewBO = dyn_cast<BinaryOperator>(BORes)) {
1241 if (isa<OverflowingBinaryOperator>(NewBO)) {
1242 NewBO->setHasNoSignedWrap(Inst.hasNoSignedWrap());
1243 NewBO->setHasNoUnsignedWrap(Inst.hasNoUnsignedWrap());
1245 if (isa<PossiblyExactOperator>(NewBO))
1246 NewBO->setIsExact(Inst.isExact());
1251 /// \brief Makes transformation of binary operation specific for vector types.
1252 /// \param Inst Binary operator to transform.
1253 /// \return Pointer to node that must replace the original binary operator, or
1254 /// null pointer if no transformation was made.
1255 Value *InstCombiner::SimplifyVectorOp(BinaryOperator &Inst) {
1256 if (!Inst.getType()->isVectorTy()) return nullptr;
1258 // It may not be safe to reorder shuffles and things like div, urem, etc.
1259 // because we may trap when executing those ops on unknown vector elements.
1261 if (!isSafeToSpeculativelyExecute(&Inst))
1264 unsigned VWidth = cast<VectorType>(Inst.getType())->getNumElements();
1265 Value *LHS = Inst.getOperand(0), *RHS = Inst.getOperand(1);
1266 assert(cast<VectorType>(LHS->getType())->getNumElements() == VWidth);
1267 assert(cast<VectorType>(RHS->getType())->getNumElements() == VWidth);
1269 // If both arguments of binary operation are shuffles, which use the same
1270 // mask and shuffle within a single vector, it is worthwhile to move the
1271 // shuffle after binary operation:
1272 // Op(shuffle(v1, m), shuffle(v2, m)) -> shuffle(Op(v1, v2), m)
1273 if (isa<ShuffleVectorInst>(LHS) && isa<ShuffleVectorInst>(RHS)) {
1274 ShuffleVectorInst *LShuf = cast<ShuffleVectorInst>(LHS);
1275 ShuffleVectorInst *RShuf = cast<ShuffleVectorInst>(RHS);
1276 if (isa<UndefValue>(LShuf->getOperand(1)) &&
1277 isa<UndefValue>(RShuf->getOperand(1)) &&
1278 LShuf->getOperand(0)->getType() == RShuf->getOperand(0)->getType() &&
1279 LShuf->getMask() == RShuf->getMask()) {
1280 Value *NewBO = CreateBinOpAsGiven(Inst, LShuf->getOperand(0),
1281 RShuf->getOperand(0), Builder);
1282 Value *Res = Builder->CreateShuffleVector(NewBO,
1283 UndefValue::get(NewBO->getType()), LShuf->getMask());
1288 // If one argument is a shuffle within one vector, the other is a constant,
1289 // try moving the shuffle after the binary operation.
1290 ShuffleVectorInst *Shuffle = nullptr;
1291 Constant *C1 = nullptr;
1292 if (isa<ShuffleVectorInst>(LHS)) Shuffle = cast<ShuffleVectorInst>(LHS);
1293 if (isa<ShuffleVectorInst>(RHS)) Shuffle = cast<ShuffleVectorInst>(RHS);
1294 if (isa<Constant>(LHS)) C1 = cast<Constant>(LHS);
1295 if (isa<Constant>(RHS)) C1 = cast<Constant>(RHS);
1296 if (Shuffle && C1 &&
1297 (isa<ConstantVector>(C1) || isa<ConstantDataVector>(C1)) &&
1298 isa<UndefValue>(Shuffle->getOperand(1)) &&
1299 Shuffle->getType() == Shuffle->getOperand(0)->getType()) {
1300 SmallVector<int, 16> ShMask = Shuffle->getShuffleMask();
1301 // Find constant C2 that has property:
1302 // shuffle(C2, ShMask) = C1
1303 // If such constant does not exist (example: ShMask=<0,0> and C1=<1,2>)
1304 // reorder is not possible.
1305 SmallVector<Constant*, 16> C2M(VWidth,
1306 UndefValue::get(C1->getType()->getScalarType()));
1307 bool MayChange = true;
1308 for (unsigned I = 0; I < VWidth; ++I) {
1309 if (ShMask[I] >= 0) {
1310 assert(ShMask[I] < (int)VWidth);
1311 if (!isa<UndefValue>(C2M[ShMask[I]])) {
1315 C2M[ShMask[I]] = C1->getAggregateElement(I);
1319 Constant *C2 = ConstantVector::get(C2M);
1320 Value *NewLHS, *NewRHS;
1321 if (isa<Constant>(LHS)) {
1323 NewRHS = Shuffle->getOperand(0);
1325 NewLHS = Shuffle->getOperand(0);
1328 Value *NewBO = CreateBinOpAsGiven(Inst, NewLHS, NewRHS, Builder);
1329 Value *Res = Builder->CreateShuffleVector(NewBO,
1330 UndefValue::get(Inst.getType()), Shuffle->getMask());
1338 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
1339 SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end());
1341 if (Value *V = SimplifyGEPInst(Ops, DL, TLI, DT, AC))
1342 return ReplaceInstUsesWith(GEP, V);
1344 Value *PtrOp = GEP.getOperand(0);
1346 // Eliminate unneeded casts for indices, and replace indices which displace
1347 // by multiples of a zero size type with zero.
1348 bool MadeChange = false;
1349 Type *IntPtrTy = DL.getIntPtrType(GEP.getPointerOperandType());
1351 gep_type_iterator GTI = gep_type_begin(GEP);
1352 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end(); I != E;
1354 // Skip indices into struct types.
1355 SequentialType *SeqTy = dyn_cast<SequentialType>(*GTI);
1359 // If the element type has zero size then any index over it is equivalent
1360 // to an index of zero, so replace it with zero if it is not zero already.
1361 if (SeqTy->getElementType()->isSized() &&
1362 DL.getTypeAllocSize(SeqTy->getElementType()) == 0)
1363 if (!isa<Constant>(*I) || !cast<Constant>(*I)->isNullValue()) {
1364 *I = Constant::getNullValue(IntPtrTy);
1368 Type *IndexTy = (*I)->getType();
1369 if (IndexTy != IntPtrTy) {
1370 // If we are using a wider index than needed for this platform, shrink
1371 // it to what we need. If narrower, sign-extend it to what we need.
1372 // This explicit cast can make subsequent optimizations more obvious.
1373 *I = Builder->CreateIntCast(*I, IntPtrTy, true);
1380 // Check to see if the inputs to the PHI node are getelementptr instructions.
1381 if (PHINode *PN = dyn_cast<PHINode>(PtrOp)) {
1382 GetElementPtrInst *Op1 = dyn_cast<GetElementPtrInst>(PN->getOperand(0));
1386 // Don't fold a GEP into itself through a PHI node. This can only happen
1387 // through the back-edge of a loop. Folding a GEP into itself means that
1388 // the value of the previous iteration needs to be stored in the meantime,
1389 // thus requiring an additional register variable to be live, but not
1390 // actually achieving anything (the GEP still needs to be executed once per
1397 for (auto I = PN->op_begin()+1, E = PN->op_end(); I !=E; ++I) {
1398 GetElementPtrInst *Op2 = dyn_cast<GetElementPtrInst>(*I);
1399 if (!Op2 || Op1->getNumOperands() != Op2->getNumOperands())
1402 // As for Op1 above, don't try to fold a GEP into itself.
1406 // Keep track of the type as we walk the GEP.
1407 Type *CurTy = Op1->getOperand(0)->getType()->getScalarType();
1409 for (unsigned J = 0, F = Op1->getNumOperands(); J != F; ++J) {
1410 if (Op1->getOperand(J)->getType() != Op2->getOperand(J)->getType())
1413 if (Op1->getOperand(J) != Op2->getOperand(J)) {
1415 // We have not seen any differences yet in the GEPs feeding the
1416 // PHI yet, so we record this one if it is allowed to be a
1419 // The first two arguments can vary for any GEP, the rest have to be
1420 // static for struct slots
1421 if (J > 1 && CurTy->isStructTy())
1426 // The GEP is different by more than one input. While this could be
1427 // extended to support GEPs that vary by more than one variable it
1428 // doesn't make sense since it greatly increases the complexity and
1429 // would result in an R+R+R addressing mode which no backend
1430 // directly supports and would need to be broken into several
1431 // simpler instructions anyway.
1436 // Sink down a layer of the type for the next iteration.
1438 if (CompositeType *CT = dyn_cast<CompositeType>(CurTy)) {
1439 CurTy = CT->getTypeAtIndex(Op1->getOperand(J));
1447 GetElementPtrInst *NewGEP = cast<GetElementPtrInst>(Op1->clone());
1450 // All the GEPs feeding the PHI are identical. Clone one down into our
1451 // BB so that it can be merged with the current GEP.
1452 GEP.getParent()->getInstList().insert(
1453 GEP.getParent()->getFirstInsertionPt(), NewGEP);
1455 // All the GEPs feeding the PHI differ at a single offset. Clone a GEP
1456 // into the current block so it can be merged, and create a new PHI to
1458 Instruction *InsertPt = Builder->GetInsertPoint();
1459 Builder->SetInsertPoint(PN);
1460 PHINode *NewPN = Builder->CreatePHI(Op1->getOperand(DI)->getType(),
1461 PN->getNumOperands());
1462 Builder->SetInsertPoint(InsertPt);
1464 for (auto &I : PN->operands())
1465 NewPN->addIncoming(cast<GEPOperator>(I)->getOperand(DI),
1466 PN->getIncomingBlock(I));
1468 NewGEP->setOperand(DI, NewPN);
1469 GEP.getParent()->getInstList().insert(
1470 GEP.getParent()->getFirstInsertionPt(), NewGEP);
1471 NewGEP->setOperand(DI, NewPN);
1474 GEP.setOperand(0, NewGEP);
1478 // Combine Indices - If the source pointer to this getelementptr instruction
1479 // is a getelementptr instruction, combine the indices of the two
1480 // getelementptr instructions into a single instruction.
1482 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) {
1483 if (!shouldMergeGEPs(*cast<GEPOperator>(&GEP), *Src))
1486 // Note that if our source is a gep chain itself then we wait for that
1487 // chain to be resolved before we perform this transformation. This
1488 // avoids us creating a TON of code in some cases.
1489 if (GEPOperator *SrcGEP =
1490 dyn_cast<GEPOperator>(Src->getOperand(0)))
1491 if (SrcGEP->getNumOperands() == 2 && shouldMergeGEPs(*Src, *SrcGEP))
1492 return nullptr; // Wait until our source is folded to completion.
1494 SmallVector<Value*, 8> Indices;
1496 // Find out whether the last index in the source GEP is a sequential idx.
1497 bool EndsWithSequential = false;
1498 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src);
1500 EndsWithSequential = !(*I)->isStructTy();
1502 // Can we combine the two pointer arithmetics offsets?
1503 if (EndsWithSequential) {
1504 // Replace: gep (gep %P, long B), long A, ...
1505 // With: T = long A+B; gep %P, T, ...
1508 Value *SO1 = Src->getOperand(Src->getNumOperands()-1);
1509 Value *GO1 = GEP.getOperand(1);
1510 if (SO1 == Constant::getNullValue(SO1->getType())) {
1512 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
1515 // If they aren't the same type, then the input hasn't been processed
1516 // by the loop above yet (which canonicalizes sequential index types to
1517 // intptr_t). Just avoid transforming this until the input has been
1519 if (SO1->getType() != GO1->getType())
1521 // Only do the combine when GO1 and SO1 are both constants. Only in
1522 // this case, we are sure the cost after the merge is never more than
1523 // that before the merge.
1524 if (!isa<Constant>(GO1) || !isa<Constant>(SO1))
1526 Sum = Builder->CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
1529 // Update the GEP in place if possible.
1530 if (Src->getNumOperands() == 2) {
1531 GEP.setOperand(0, Src->getOperand(0));
1532 GEP.setOperand(1, Sum);
1535 Indices.append(Src->op_begin()+1, Src->op_end()-1);
1536 Indices.push_back(Sum);
1537 Indices.append(GEP.op_begin()+2, GEP.op_end());
1538 } else if (isa<Constant>(*GEP.idx_begin()) &&
1539 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
1540 Src->getNumOperands() != 1) {
1541 // Otherwise we can do the fold if the first index of the GEP is a zero
1542 Indices.append(Src->op_begin()+1, Src->op_end());
1543 Indices.append(GEP.idx_begin()+1, GEP.idx_end());
1546 if (!Indices.empty())
1547 return GEP.isInBounds() && Src->isInBounds()
1548 ? GetElementPtrInst::CreateInBounds(
1549 Src->getSourceElementType(), Src->getOperand(0), Indices,
1551 : GetElementPtrInst::Create(Src->getSourceElementType(),
1552 Src->getOperand(0), Indices,
1556 if (GEP.getNumIndices() == 1) {
1557 unsigned AS = GEP.getPointerAddressSpace();
1558 if (GEP.getOperand(1)->getType()->getScalarSizeInBits() ==
1559 DL.getPointerSizeInBits(AS)) {
1560 Type *PtrTy = GEP.getPointerOperandType();
1561 Type *Ty = PtrTy->getPointerElementType();
1562 uint64_t TyAllocSize = DL.getTypeAllocSize(Ty);
1564 bool Matched = false;
1567 if (TyAllocSize == 1) {
1568 V = GEP.getOperand(1);
1570 } else if (match(GEP.getOperand(1),
1571 m_AShr(m_Value(V), m_ConstantInt(C)))) {
1572 if (TyAllocSize == 1ULL << C)
1574 } else if (match(GEP.getOperand(1),
1575 m_SDiv(m_Value(V), m_ConstantInt(C)))) {
1576 if (TyAllocSize == C)
1581 // Canonicalize (gep i8* X, -(ptrtoint Y))
1582 // to (inttoptr (sub (ptrtoint X), (ptrtoint Y)))
1583 // The GEP pattern is emitted by the SCEV expander for certain kinds of
1584 // pointer arithmetic.
1585 if (match(V, m_Neg(m_PtrToInt(m_Value())))) {
1586 Operator *Index = cast<Operator>(V);
1587 Value *PtrToInt = Builder->CreatePtrToInt(PtrOp, Index->getType());
1588 Value *NewSub = Builder->CreateSub(PtrToInt, Index->getOperand(1));
1589 return CastInst::Create(Instruction::IntToPtr, NewSub, GEP.getType());
1591 // Canonicalize (gep i8* X, (ptrtoint Y)-(ptrtoint X))
1594 if (match(V, m_Sub(m_PtrToInt(m_Value(Y)),
1595 m_PtrToInt(m_Specific(GEP.getOperand(0)))))) {
1596 return CastInst::CreatePointerBitCastOrAddrSpaceCast(Y,
1603 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0).
1604 Value *StrippedPtr = PtrOp->stripPointerCasts();
1605 PointerType *StrippedPtrTy = dyn_cast<PointerType>(StrippedPtr->getType());
1607 // We do not handle pointer-vector geps here.
1611 if (StrippedPtr != PtrOp) {
1612 bool HasZeroPointerIndex = false;
1613 if (ConstantInt *C = dyn_cast<ConstantInt>(GEP.getOperand(1)))
1614 HasZeroPointerIndex = C->isZero();
1616 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
1617 // into : GEP [10 x i8]* X, i32 0, ...
1619 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
1620 // into : GEP i8* X, ...
1622 // This occurs when the program declares an array extern like "int X[];"
1623 if (HasZeroPointerIndex) {
1624 PointerType *CPTy = cast<PointerType>(PtrOp->getType());
1625 if (ArrayType *CATy =
1626 dyn_cast<ArrayType>(CPTy->getElementType())) {
1627 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
1628 if (CATy->getElementType() == StrippedPtrTy->getElementType()) {
1629 // -> GEP i8* X, ...
1630 SmallVector<Value*, 8> Idx(GEP.idx_begin()+1, GEP.idx_end());
1631 GetElementPtrInst *Res = GetElementPtrInst::Create(
1632 StrippedPtrTy->getElementType(), StrippedPtr, Idx, GEP.getName());
1633 Res->setIsInBounds(GEP.isInBounds());
1634 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace())
1636 // Insert Res, and create an addrspacecast.
1638 // GEP (addrspacecast i8 addrspace(1)* X to [0 x i8]*), i32 0, ...
1640 // %0 = GEP i8 addrspace(1)* X, ...
1641 // addrspacecast i8 addrspace(1)* %0 to i8*
1642 return new AddrSpaceCastInst(Builder->Insert(Res), GEP.getType());
1645 if (ArrayType *XATy =
1646 dyn_cast<ArrayType>(StrippedPtrTy->getElementType())){
1647 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
1648 if (CATy->getElementType() == XATy->getElementType()) {
1649 // -> GEP [10 x i8]* X, i32 0, ...
1650 // At this point, we know that the cast source type is a pointer
1651 // to an array of the same type as the destination pointer
1652 // array. Because the array type is never stepped over (there
1653 // is a leading zero) we can fold the cast into this GEP.
1654 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace()) {
1655 GEP.setOperand(0, StrippedPtr);
1656 GEP.setSourceElementType(XATy);
1659 // Cannot replace the base pointer directly because StrippedPtr's
1660 // address space is different. Instead, create a new GEP followed by
1661 // an addrspacecast.
1663 // GEP (addrspacecast [10 x i8] addrspace(1)* X to [0 x i8]*),
1666 // %0 = GEP [10 x i8] addrspace(1)* X, ...
1667 // addrspacecast i8 addrspace(1)* %0 to i8*
1668 SmallVector<Value*, 8> Idx(GEP.idx_begin(), GEP.idx_end());
1669 Value *NewGEP = GEP.isInBounds()
1670 ? Builder->CreateInBoundsGEP(
1671 nullptr, StrippedPtr, Idx, GEP.getName())
1672 : Builder->CreateGEP(nullptr, StrippedPtr, Idx,
1674 return new AddrSpaceCastInst(NewGEP, GEP.getType());
1678 } else if (GEP.getNumOperands() == 2) {
1679 // Transform things like:
1680 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
1681 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
1682 Type *SrcElTy = StrippedPtrTy->getElementType();
1683 Type *ResElTy = PtrOp->getType()->getPointerElementType();
1684 if (SrcElTy->isArrayTy() &&
1685 DL.getTypeAllocSize(SrcElTy->getArrayElementType()) ==
1686 DL.getTypeAllocSize(ResElTy)) {
1687 Type *IdxType = DL.getIntPtrType(GEP.getType());
1688 Value *Idx[2] = { Constant::getNullValue(IdxType), GEP.getOperand(1) };
1691 ? Builder->CreateInBoundsGEP(nullptr, StrippedPtr, Idx,
1693 : Builder->CreateGEP(nullptr, StrippedPtr, Idx, GEP.getName());
1695 // V and GEP are both pointer types --> BitCast
1696 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
1700 // Transform things like:
1701 // %V = mul i64 %N, 4
1702 // %t = getelementptr i8* bitcast (i32* %arr to i8*), i32 %V
1703 // into: %t1 = getelementptr i32* %arr, i32 %N; bitcast
1704 if (ResElTy->isSized() && SrcElTy->isSized()) {
1705 // Check that changing the type amounts to dividing the index by a scale
1707 uint64_t ResSize = DL.getTypeAllocSize(ResElTy);
1708 uint64_t SrcSize = DL.getTypeAllocSize(SrcElTy);
1709 if (ResSize && SrcSize % ResSize == 0) {
1710 Value *Idx = GEP.getOperand(1);
1711 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
1712 uint64_t Scale = SrcSize / ResSize;
1714 // Earlier transforms ensure that the index has type IntPtrType, which
1715 // considerably simplifies the logic by eliminating implicit casts.
1716 assert(Idx->getType() == DL.getIntPtrType(GEP.getType()) &&
1717 "Index not cast to pointer width?");
1720 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
1721 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
1722 // If the multiplication NewIdx * Scale may overflow then the new
1723 // GEP may not be "inbounds".
1725 GEP.isInBounds() && NSW
1726 ? Builder->CreateInBoundsGEP(nullptr, StrippedPtr, NewIdx,
1728 : Builder->CreateGEP(nullptr, StrippedPtr, NewIdx,
1731 // The NewGEP must be pointer typed, so must the old one -> BitCast
1732 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
1738 // Similarly, transform things like:
1739 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
1740 // (where tmp = 8*tmp2) into:
1741 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
1742 if (ResElTy->isSized() && SrcElTy->isSized() && SrcElTy->isArrayTy()) {
1743 // Check that changing to the array element type amounts to dividing the
1744 // index by a scale factor.
1745 uint64_t ResSize = DL.getTypeAllocSize(ResElTy);
1746 uint64_t ArrayEltSize =
1747 DL.getTypeAllocSize(SrcElTy->getArrayElementType());
1748 if (ResSize && ArrayEltSize % ResSize == 0) {
1749 Value *Idx = GEP.getOperand(1);
1750 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
1751 uint64_t Scale = ArrayEltSize / ResSize;
1753 // Earlier transforms ensure that the index has type IntPtrType, which
1754 // considerably simplifies the logic by eliminating implicit casts.
1755 assert(Idx->getType() == DL.getIntPtrType(GEP.getType()) &&
1756 "Index not cast to pointer width?");
1759 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
1760 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
1761 // If the multiplication NewIdx * Scale may overflow then the new
1762 // GEP may not be "inbounds".
1764 Constant::getNullValue(DL.getIntPtrType(GEP.getType())),
1767 Value *NewGEP = GEP.isInBounds() && NSW
1768 ? Builder->CreateInBoundsGEP(
1769 SrcElTy, StrippedPtr, Off, GEP.getName())
1770 : Builder->CreateGEP(SrcElTy, StrippedPtr, Off,
1772 // The NewGEP must be pointer typed, so must the old one -> BitCast
1773 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
1781 // addrspacecast between types is canonicalized as a bitcast, then an
1782 // addrspacecast. To take advantage of the below bitcast + struct GEP, look
1783 // through the addrspacecast.
1784 if (AddrSpaceCastInst *ASC = dyn_cast<AddrSpaceCastInst>(PtrOp)) {
1785 // X = bitcast A addrspace(1)* to B addrspace(1)*
1786 // Y = addrspacecast A addrspace(1)* to B addrspace(2)*
1787 // Z = gep Y, <...constant indices...>
1788 // Into an addrspacecasted GEP of the struct.
1789 if (BitCastInst *BC = dyn_cast<BitCastInst>(ASC->getOperand(0)))
1793 /// See if we can simplify:
1794 /// X = bitcast A* to B*
1795 /// Y = gep X, <...constant indices...>
1796 /// into a gep of the original struct. This is important for SROA and alias
1797 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
1798 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
1799 Value *Operand = BCI->getOperand(0);
1800 PointerType *OpType = cast<PointerType>(Operand->getType());
1801 unsigned OffsetBits = DL.getPointerTypeSizeInBits(GEP.getType());
1802 APInt Offset(OffsetBits, 0);
1803 if (!isa<BitCastInst>(Operand) &&
1804 GEP.accumulateConstantOffset(DL, Offset)) {
1806 // If this GEP instruction doesn't move the pointer, just replace the GEP
1807 // with a bitcast of the real input to the dest type.
1809 // If the bitcast is of an allocation, and the allocation will be
1810 // converted to match the type of the cast, don't touch this.
1811 if (isa<AllocaInst>(Operand) || isAllocationFn(Operand, TLI)) {
1812 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
1813 if (Instruction *I = visitBitCast(*BCI)) {
1816 BCI->getParent()->getInstList().insert(BCI, I);
1817 ReplaceInstUsesWith(*BCI, I);
1823 if (Operand->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
1824 return new AddrSpaceCastInst(Operand, GEP.getType());
1825 return new BitCastInst(Operand, GEP.getType());
1828 // Otherwise, if the offset is non-zero, we need to find out if there is a
1829 // field at Offset in 'A's type. If so, we can pull the cast through the
1831 SmallVector<Value*, 8> NewIndices;
1832 if (FindElementAtOffset(OpType, Offset.getSExtValue(), NewIndices)) {
1835 ? Builder->CreateInBoundsGEP(nullptr, Operand, NewIndices)
1836 : Builder->CreateGEP(nullptr, Operand, NewIndices);
1838 if (NGEP->getType() == GEP.getType())
1839 return ReplaceInstUsesWith(GEP, NGEP);
1840 NGEP->takeName(&GEP);
1842 if (NGEP->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
1843 return new AddrSpaceCastInst(NGEP, GEP.getType());
1844 return new BitCastInst(NGEP, GEP.getType());
1853 isAllocSiteRemovable(Instruction *AI, SmallVectorImpl<WeakVH> &Users,
1854 const TargetLibraryInfo *TLI) {
1855 SmallVector<Instruction*, 4> Worklist;
1856 Worklist.push_back(AI);
1859 Instruction *PI = Worklist.pop_back_val();
1860 for (User *U : PI->users()) {
1861 Instruction *I = cast<Instruction>(U);
1862 switch (I->getOpcode()) {
1864 // Give up the moment we see something we can't handle.
1867 case Instruction::BitCast:
1868 case Instruction::GetElementPtr:
1869 Users.emplace_back(I);
1870 Worklist.push_back(I);
1873 case Instruction::ICmp: {
1874 ICmpInst *ICI = cast<ICmpInst>(I);
1875 // We can fold eq/ne comparisons with null to false/true, respectively.
1876 if (!ICI->isEquality() || !isa<ConstantPointerNull>(ICI->getOperand(1)))
1878 Users.emplace_back(I);
1882 case Instruction::Call:
1883 // Ignore no-op and store intrinsics.
1884 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1885 switch (II->getIntrinsicID()) {
1889 case Intrinsic::memmove:
1890 case Intrinsic::memcpy:
1891 case Intrinsic::memset: {
1892 MemIntrinsic *MI = cast<MemIntrinsic>(II);
1893 if (MI->isVolatile() || MI->getRawDest() != PI)
1897 case Intrinsic::dbg_declare:
1898 case Intrinsic::dbg_value:
1899 case Intrinsic::invariant_start:
1900 case Intrinsic::invariant_end:
1901 case Intrinsic::lifetime_start:
1902 case Intrinsic::lifetime_end:
1903 case Intrinsic::objectsize:
1904 Users.emplace_back(I);
1909 if (isFreeCall(I, TLI)) {
1910 Users.emplace_back(I);
1915 case Instruction::Store: {
1916 StoreInst *SI = cast<StoreInst>(I);
1917 if (SI->isVolatile() || SI->getPointerOperand() != PI)
1919 Users.emplace_back(I);
1923 llvm_unreachable("missing a return?");
1925 } while (!Worklist.empty());
1929 Instruction *InstCombiner::visitAllocSite(Instruction &MI) {
1930 // If we have a malloc call which is only used in any amount of comparisons
1931 // to null and free calls, delete the calls and replace the comparisons with
1932 // true or false as appropriate.
1933 SmallVector<WeakVH, 64> Users;
1934 if (isAllocSiteRemovable(&MI, Users, TLI)) {
1935 for (unsigned i = 0, e = Users.size(); i != e; ++i) {
1936 Instruction *I = cast_or_null<Instruction>(&*Users[i]);
1939 if (ICmpInst *C = dyn_cast<ICmpInst>(I)) {
1940 ReplaceInstUsesWith(*C,
1941 ConstantInt::get(Type::getInt1Ty(C->getContext()),
1942 C->isFalseWhenEqual()));
1943 } else if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) {
1944 ReplaceInstUsesWith(*I, UndefValue::get(I->getType()));
1945 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1946 if (II->getIntrinsicID() == Intrinsic::objectsize) {
1947 ConstantInt *CI = cast<ConstantInt>(II->getArgOperand(1));
1948 uint64_t DontKnow = CI->isZero() ? -1ULL : 0;
1949 ReplaceInstUsesWith(*I, ConstantInt::get(I->getType(), DontKnow));
1952 EraseInstFromFunction(*I);
1955 if (InvokeInst *II = dyn_cast<InvokeInst>(&MI)) {
1956 // Replace invoke with a NOP intrinsic to maintain the original CFG
1957 Module *M = II->getParent()->getParent()->getParent();
1958 Function *F = Intrinsic::getDeclaration(M, Intrinsic::donothing);
1959 InvokeInst::Create(F, II->getNormalDest(), II->getUnwindDest(),
1960 None, "", II->getParent());
1962 return EraseInstFromFunction(MI);
1967 /// \brief Move the call to free before a NULL test.
1969 /// Check if this free is accessed after its argument has been test
1970 /// against NULL (property 0).
1971 /// If yes, it is legal to move this call in its predecessor block.
1973 /// The move is performed only if the block containing the call to free
1974 /// will be removed, i.e.:
1975 /// 1. it has only one predecessor P, and P has two successors
1976 /// 2. it contains the call and an unconditional branch
1977 /// 3. its successor is the same as its predecessor's successor
1979 /// The profitability is out-of concern here and this function should
1980 /// be called only if the caller knows this transformation would be
1981 /// profitable (e.g., for code size).
1982 static Instruction *
1983 tryToMoveFreeBeforeNullTest(CallInst &FI) {
1984 Value *Op = FI.getArgOperand(0);
1985 BasicBlock *FreeInstrBB = FI.getParent();
1986 BasicBlock *PredBB = FreeInstrBB->getSinglePredecessor();
1988 // Validate part of constraint #1: Only one predecessor
1989 // FIXME: We can extend the number of predecessor, but in that case, we
1990 // would duplicate the call to free in each predecessor and it may
1991 // not be profitable even for code size.
1995 // Validate constraint #2: Does this block contains only the call to
1996 // free and an unconditional branch?
1997 // FIXME: We could check if we can speculate everything in the
1998 // predecessor block
1999 if (FreeInstrBB->size() != 2)
2002 if (!match(FreeInstrBB->getTerminator(), m_UnconditionalBr(SuccBB)))
2005 // Validate the rest of constraint #1 by matching on the pred branch.
2006 TerminatorInst *TI = PredBB->getTerminator();
2007 BasicBlock *TrueBB, *FalseBB;
2008 ICmpInst::Predicate Pred;
2009 if (!match(TI, m_Br(m_ICmp(Pred, m_Specific(Op), m_Zero()), TrueBB, FalseBB)))
2011 if (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE)
2014 // Validate constraint #3: Ensure the null case just falls through.
2015 if (SuccBB != (Pred == ICmpInst::ICMP_EQ ? TrueBB : FalseBB))
2017 assert(FreeInstrBB == (Pred == ICmpInst::ICMP_EQ ? FalseBB : TrueBB) &&
2018 "Broken CFG: missing edge from predecessor to successor");
2025 Instruction *InstCombiner::visitFree(CallInst &FI) {
2026 Value *Op = FI.getArgOperand(0);
2028 // free undef -> unreachable.
2029 if (isa<UndefValue>(Op)) {
2030 // Insert a new store to null because we cannot modify the CFG here.
2031 Builder->CreateStore(ConstantInt::getTrue(FI.getContext()),
2032 UndefValue::get(Type::getInt1PtrTy(FI.getContext())));
2033 return EraseInstFromFunction(FI);
2036 // If we have 'free null' delete the instruction. This can happen in stl code
2037 // when lots of inlining happens.
2038 if (isa<ConstantPointerNull>(Op))
2039 return EraseInstFromFunction(FI);
2041 // If we optimize for code size, try to move the call to free before the null
2042 // test so that simplify cfg can remove the empty block and dead code
2043 // elimination the branch. I.e., helps to turn something like:
2044 // if (foo) free(foo);
2048 if (Instruction *I = tryToMoveFreeBeforeNullTest(FI))
2054 Instruction *InstCombiner::visitReturnInst(ReturnInst &RI) {
2055 if (RI.getNumOperands() == 0) // ret void
2058 Value *ResultOp = RI.getOperand(0);
2059 Type *VTy = ResultOp->getType();
2060 if (!VTy->isIntegerTy())
2063 // There might be assume intrinsics dominating this return that completely
2064 // determine the value. If so, constant fold it.
2065 unsigned BitWidth = VTy->getPrimitiveSizeInBits();
2066 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2067 computeKnownBits(ResultOp, KnownZero, KnownOne, 0, &RI);
2068 if ((KnownZero|KnownOne).isAllOnesValue())
2069 RI.setOperand(0, Constant::getIntegerValue(VTy, KnownOne));
2074 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
2075 // Change br (not X), label True, label False to: br X, label False, True
2077 BasicBlock *TrueDest;
2078 BasicBlock *FalseDest;
2079 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
2080 !isa<Constant>(X)) {
2081 // Swap Destinations and condition...
2083 BI.swapSuccessors();
2087 // If the condition is irrelevant, remove the use so that other
2088 // transforms on the condition become more effective.
2089 if (BI.isConditional() &&
2090 BI.getSuccessor(0) == BI.getSuccessor(1) &&
2091 !isa<UndefValue>(BI.getCondition())) {
2092 BI.setCondition(UndefValue::get(BI.getCondition()->getType()));
2096 // Canonicalize fcmp_one -> fcmp_oeq
2097 FCmpInst::Predicate FPred; Value *Y;
2098 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
2099 TrueDest, FalseDest)) &&
2100 BI.getCondition()->hasOneUse())
2101 if (FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
2102 FPred == FCmpInst::FCMP_OGE) {
2103 FCmpInst *Cond = cast<FCmpInst>(BI.getCondition());
2104 Cond->setPredicate(FCmpInst::getInversePredicate(FPred));
2106 // Swap Destinations and condition.
2107 BI.swapSuccessors();
2112 // Canonicalize icmp_ne -> icmp_eq
2113 ICmpInst::Predicate IPred;
2114 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
2115 TrueDest, FalseDest)) &&
2116 BI.getCondition()->hasOneUse())
2117 if (IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
2118 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
2119 IPred == ICmpInst::ICMP_SGE) {
2120 ICmpInst *Cond = cast<ICmpInst>(BI.getCondition());
2121 Cond->setPredicate(ICmpInst::getInversePredicate(IPred));
2122 // Swap Destinations and condition.
2123 BI.swapSuccessors();
2131 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
2132 Value *Cond = SI.getCondition();
2133 unsigned BitWidth = cast<IntegerType>(Cond->getType())->getBitWidth();
2134 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2135 computeKnownBits(Cond, KnownZero, KnownOne, 0, &SI);
2136 unsigned LeadingKnownZeros = KnownZero.countLeadingOnes();
2137 unsigned LeadingKnownOnes = KnownOne.countLeadingOnes();
2139 // Compute the number of leading bits we can ignore.
2140 for (auto &C : SI.cases()) {
2141 LeadingKnownZeros = std::min(
2142 LeadingKnownZeros, C.getCaseValue()->getValue().countLeadingZeros());
2143 LeadingKnownOnes = std::min(
2144 LeadingKnownOnes, C.getCaseValue()->getValue().countLeadingOnes());
2147 unsigned NewWidth = BitWidth - std::max(LeadingKnownZeros, LeadingKnownOnes);
2149 // Truncate the condition operand if the new type is equal to or larger than
2150 // the largest legal integer type. We need to be conservative here since
2151 // x86 generates redundant zero-extension instructions if the operand is
2152 // truncated to i8 or i16.
2153 bool TruncCond = false;
2154 if (NewWidth > 0 && BitWidth > NewWidth &&
2155 NewWidth >= DL.getLargestLegalIntTypeSize()) {
2157 IntegerType *Ty = IntegerType::get(SI.getContext(), NewWidth);
2158 Builder->SetInsertPoint(&SI);
2159 Value *NewCond = Builder->CreateTrunc(SI.getCondition(), Ty, "trunc");
2160 SI.setCondition(NewCond);
2162 for (auto &C : SI.cases())
2163 static_cast<SwitchInst::CaseIt *>(&C)->setValue(ConstantInt::get(
2164 SI.getContext(), C.getCaseValue()->getValue().trunc(NewWidth)));
2167 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
2168 if (I->getOpcode() == Instruction::Add)
2169 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2170 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
2171 // Skip the first item since that's the default case.
2172 for (SwitchInst::CaseIt i = SI.case_begin(), e = SI.case_end();
2174 ConstantInt* CaseVal = i.getCaseValue();
2175 Constant *LHS = CaseVal;
2177 LHS = LeadingKnownZeros
2178 ? ConstantExpr::getZExt(CaseVal, Cond->getType())
2179 : ConstantExpr::getSExt(CaseVal, Cond->getType());
2180 Constant* NewCaseVal = ConstantExpr::getSub(LHS, AddRHS);
2181 assert(isa<ConstantInt>(NewCaseVal) &&
2182 "Result of expression should be constant");
2183 i.setValue(cast<ConstantInt>(NewCaseVal));
2185 SI.setCondition(I->getOperand(0));
2191 return TruncCond ? &SI : nullptr;
2194 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
2195 Value *Agg = EV.getAggregateOperand();
2197 if (!EV.hasIndices())
2198 return ReplaceInstUsesWith(EV, Agg);
2201 SimplifyExtractValueInst(Agg, EV.getIndices(), DL, TLI, DT, AC))
2202 return ReplaceInstUsesWith(EV, V);
2204 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
2205 // We're extracting from an insertvalue instruction, compare the indices
2206 const unsigned *exti, *exte, *insi, *inse;
2207 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
2208 exte = EV.idx_end(), inse = IV->idx_end();
2209 exti != exte && insi != inse;
2212 // The insert and extract both reference distinctly different elements.
2213 // This means the extract is not influenced by the insert, and we can
2214 // replace the aggregate operand of the extract with the aggregate
2215 // operand of the insert. i.e., replace
2216 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
2217 // %E = extractvalue { i32, { i32 } } %I, 0
2219 // %E = extractvalue { i32, { i32 } } %A, 0
2220 return ExtractValueInst::Create(IV->getAggregateOperand(),
2223 if (exti == exte && insi == inse)
2224 // Both iterators are at the end: Index lists are identical. Replace
2225 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
2226 // %C = extractvalue { i32, { i32 } } %B, 1, 0
2228 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
2230 // The extract list is a prefix of the insert list. i.e. replace
2231 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
2232 // %E = extractvalue { i32, { i32 } } %I, 1
2234 // %X = extractvalue { i32, { i32 } } %A, 1
2235 // %E = insertvalue { i32 } %X, i32 42, 0
2236 // by switching the order of the insert and extract (though the
2237 // insertvalue should be left in, since it may have other uses).
2238 Value *NewEV = Builder->CreateExtractValue(IV->getAggregateOperand(),
2240 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
2241 makeArrayRef(insi, inse));
2244 // The insert list is a prefix of the extract list
2245 // We can simply remove the common indices from the extract and make it
2246 // operate on the inserted value instead of the insertvalue result.
2248 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
2249 // %E = extractvalue { i32, { i32 } } %I, 1, 0
2251 // %E extractvalue { i32 } { i32 42 }, 0
2252 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
2253 makeArrayRef(exti, exte));
2255 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) {
2256 // We're extracting from an intrinsic, see if we're the only user, which
2257 // allows us to simplify multiple result intrinsics to simpler things that
2258 // just get one value.
2259 if (II->hasOneUse()) {
2260 // Check if we're grabbing the overflow bit or the result of a 'with
2261 // overflow' intrinsic. If it's the latter we can remove the intrinsic
2262 // and replace it with a traditional binary instruction.
2263 switch (II->getIntrinsicID()) {
2264 case Intrinsic::uadd_with_overflow:
2265 case Intrinsic::sadd_with_overflow:
2266 if (*EV.idx_begin() == 0) { // Normal result.
2267 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2268 ReplaceInstUsesWith(*II, UndefValue::get(II->getType()));
2269 EraseInstFromFunction(*II);
2270 return BinaryOperator::CreateAdd(LHS, RHS);
2273 // If the normal result of the add is dead, and the RHS is a constant,
2274 // we can transform this into a range comparison.
2275 // overflow = uadd a, -4 --> overflow = icmp ugt a, 3
2276 if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow)
2277 if (ConstantInt *CI = dyn_cast<ConstantInt>(II->getArgOperand(1)))
2278 return new ICmpInst(ICmpInst::ICMP_UGT, II->getArgOperand(0),
2279 ConstantExpr::getNot(CI));
2281 case Intrinsic::usub_with_overflow:
2282 case Intrinsic::ssub_with_overflow:
2283 if (*EV.idx_begin() == 0) { // Normal result.
2284 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2285 ReplaceInstUsesWith(*II, UndefValue::get(II->getType()));
2286 EraseInstFromFunction(*II);
2287 return BinaryOperator::CreateSub(LHS, RHS);
2290 case Intrinsic::umul_with_overflow:
2291 case Intrinsic::smul_with_overflow:
2292 if (*EV.idx_begin() == 0) { // Normal result.
2293 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2294 ReplaceInstUsesWith(*II, UndefValue::get(II->getType()));
2295 EraseInstFromFunction(*II);
2296 return BinaryOperator::CreateMul(LHS, RHS);
2304 if (LoadInst *L = dyn_cast<LoadInst>(Agg))
2305 // If the (non-volatile) load only has one use, we can rewrite this to a
2306 // load from a GEP. This reduces the size of the load.
2307 // FIXME: If a load is used only by extractvalue instructions then this
2308 // could be done regardless of having multiple uses.
2309 if (L->isSimple() && L->hasOneUse()) {
2310 // extractvalue has integer indices, getelementptr has Value*s. Convert.
2311 SmallVector<Value*, 4> Indices;
2312 // Prefix an i32 0 since we need the first element.
2313 Indices.push_back(Builder->getInt32(0));
2314 for (ExtractValueInst::idx_iterator I = EV.idx_begin(), E = EV.idx_end();
2316 Indices.push_back(Builder->getInt32(*I));
2318 // We need to insert these at the location of the old load, not at that of
2319 // the extractvalue.
2320 Builder->SetInsertPoint(L->getParent(), L);
2321 Value *GEP = Builder->CreateInBoundsGEP(L->getType(),
2322 L->getPointerOperand(), Indices);
2323 // Returning the load directly will cause the main loop to insert it in
2324 // the wrong spot, so use ReplaceInstUsesWith().
2325 return ReplaceInstUsesWith(EV, Builder->CreateLoad(GEP));
2327 // We could simplify extracts from other values. Note that nested extracts may
2328 // already be simplified implicitly by the above: extract (extract (insert) )
2329 // will be translated into extract ( insert ( extract ) ) first and then just
2330 // the value inserted, if appropriate. Similarly for extracts from single-use
2331 // loads: extract (extract (load)) will be translated to extract (load (gep))
2332 // and if again single-use then via load (gep (gep)) to load (gep).
2333 // However, double extracts from e.g. function arguments or return values
2334 // aren't handled yet.
2338 /// Return 'true' if the given typeinfo will match anything.
2339 static bool isCatchAll(EHPersonality Personality, Constant *TypeInfo) {
2340 switch (Personality) {
2341 case EHPersonality::GNU_C:
2342 // The GCC C EH personality only exists to support cleanups, so it's not
2343 // clear what the semantics of catch clauses are.
2345 case EHPersonality::Unknown:
2347 case EHPersonality::GNU_Ada:
2348 // While __gnat_all_others_value will match any Ada exception, it doesn't
2349 // match foreign exceptions (or didn't, before gcc-4.7).
2351 case EHPersonality::GNU_CXX:
2352 case EHPersonality::GNU_ObjC:
2353 case EHPersonality::MSVC_X86SEH:
2354 case EHPersonality::MSVC_Win64SEH:
2355 case EHPersonality::MSVC_CXX:
2356 return TypeInfo->isNullValue();
2358 llvm_unreachable("invalid enum");
2361 static bool shorter_filter(const Value *LHS, const Value *RHS) {
2363 cast<ArrayType>(LHS->getType())->getNumElements()
2365 cast<ArrayType>(RHS->getType())->getNumElements();
2368 Instruction *InstCombiner::visitLandingPadInst(LandingPadInst &LI) {
2369 // The logic here should be correct for any real-world personality function.
2370 // However if that turns out not to be true, the offending logic can always
2371 // be conditioned on the personality function, like the catch-all logic is.
2372 EHPersonality Personality =
2373 classifyEHPersonality(LI.getParent()->getParent()->getPersonalityFn());
2375 // Simplify the list of clauses, eg by removing repeated catch clauses
2376 // (these are often created by inlining).
2377 bool MakeNewInstruction = false; // If true, recreate using the following:
2378 SmallVector<Constant *, 16> NewClauses; // - Clauses for the new instruction;
2379 bool CleanupFlag = LI.isCleanup(); // - The new instruction is a cleanup.
2381 SmallPtrSet<Value *, 16> AlreadyCaught; // Typeinfos known caught already.
2382 for (unsigned i = 0, e = LI.getNumClauses(); i != e; ++i) {
2383 bool isLastClause = i + 1 == e;
2384 if (LI.isCatch(i)) {
2386 Constant *CatchClause = LI.getClause(i);
2387 Constant *TypeInfo = CatchClause->stripPointerCasts();
2389 // If we already saw this clause, there is no point in having a second
2391 if (AlreadyCaught.insert(TypeInfo).second) {
2392 // This catch clause was not already seen.
2393 NewClauses.push_back(CatchClause);
2395 // Repeated catch clause - drop the redundant copy.
2396 MakeNewInstruction = true;
2399 // If this is a catch-all then there is no point in keeping any following
2400 // clauses or marking the landingpad as having a cleanup.
2401 if (isCatchAll(Personality, TypeInfo)) {
2403 MakeNewInstruction = true;
2404 CleanupFlag = false;
2408 // A filter clause. If any of the filter elements were already caught
2409 // then they can be dropped from the filter. It is tempting to try to
2410 // exploit the filter further by saying that any typeinfo that does not
2411 // occur in the filter can't be caught later (and thus can be dropped).
2412 // However this would be wrong, since typeinfos can match without being
2413 // equal (for example if one represents a C++ class, and the other some
2414 // class derived from it).
2415 assert(LI.isFilter(i) && "Unsupported landingpad clause!");
2416 Constant *FilterClause = LI.getClause(i);
2417 ArrayType *FilterType = cast<ArrayType>(FilterClause->getType());
2418 unsigned NumTypeInfos = FilterType->getNumElements();
2420 // An empty filter catches everything, so there is no point in keeping any
2421 // following clauses or marking the landingpad as having a cleanup. By
2422 // dealing with this case here the following code is made a bit simpler.
2423 if (!NumTypeInfos) {
2424 NewClauses.push_back(FilterClause);
2426 MakeNewInstruction = true;
2427 CleanupFlag = false;
2431 bool MakeNewFilter = false; // If true, make a new filter.
2432 SmallVector<Constant *, 16> NewFilterElts; // New elements.
2433 if (isa<ConstantAggregateZero>(FilterClause)) {
2434 // Not an empty filter - it contains at least one null typeinfo.
2435 assert(NumTypeInfos > 0 && "Should have handled empty filter already!");
2436 Constant *TypeInfo =
2437 Constant::getNullValue(FilterType->getElementType());
2438 // If this typeinfo is a catch-all then the filter can never match.
2439 if (isCatchAll(Personality, TypeInfo)) {
2440 // Throw the filter away.
2441 MakeNewInstruction = true;
2445 // There is no point in having multiple copies of this typeinfo, so
2446 // discard all but the first copy if there is more than one.
2447 NewFilterElts.push_back(TypeInfo);
2448 if (NumTypeInfos > 1)
2449 MakeNewFilter = true;
2451 ConstantArray *Filter = cast<ConstantArray>(FilterClause);
2452 SmallPtrSet<Value *, 16> SeenInFilter; // For uniquing the elements.
2453 NewFilterElts.reserve(NumTypeInfos);
2455 // Remove any filter elements that were already caught or that already
2456 // occurred in the filter. While there, see if any of the elements are
2457 // catch-alls. If so, the filter can be discarded.
2458 bool SawCatchAll = false;
2459 for (unsigned j = 0; j != NumTypeInfos; ++j) {
2460 Constant *Elt = Filter->getOperand(j);
2461 Constant *TypeInfo = Elt->stripPointerCasts();
2462 if (isCatchAll(Personality, TypeInfo)) {
2463 // This element is a catch-all. Bail out, noting this fact.
2467 if (AlreadyCaught.count(TypeInfo))
2468 // Already caught by an earlier clause, so having it in the filter
2471 // There is no point in having multiple copies of the same typeinfo in
2472 // a filter, so only add it if we didn't already.
2473 if (SeenInFilter.insert(TypeInfo).second)
2474 NewFilterElts.push_back(cast<Constant>(Elt));
2476 // A filter containing a catch-all cannot match anything by definition.
2478 // Throw the filter away.
2479 MakeNewInstruction = true;
2483 // If we dropped something from the filter, make a new one.
2484 if (NewFilterElts.size() < NumTypeInfos)
2485 MakeNewFilter = true;
2487 if (MakeNewFilter) {
2488 FilterType = ArrayType::get(FilterType->getElementType(),
2489 NewFilterElts.size());
2490 FilterClause = ConstantArray::get(FilterType, NewFilterElts);
2491 MakeNewInstruction = true;
2494 NewClauses.push_back(FilterClause);
2496 // If the new filter is empty then it will catch everything so there is
2497 // no point in keeping any following clauses or marking the landingpad
2498 // as having a cleanup. The case of the original filter being empty was
2499 // already handled above.
2500 if (MakeNewFilter && !NewFilterElts.size()) {
2501 assert(MakeNewInstruction && "New filter but not a new instruction!");
2502 CleanupFlag = false;
2508 // If several filters occur in a row then reorder them so that the shortest
2509 // filters come first (those with the smallest number of elements). This is
2510 // advantageous because shorter filters are more likely to match, speeding up
2511 // unwinding, but mostly because it increases the effectiveness of the other
2512 // filter optimizations below.
2513 for (unsigned i = 0, e = NewClauses.size(); i + 1 < e; ) {
2515 // Find the maximal 'j' s.t. the range [i, j) consists entirely of filters.
2516 for (j = i; j != e; ++j)
2517 if (!isa<ArrayType>(NewClauses[j]->getType()))
2520 // Check whether the filters are already sorted by length. We need to know
2521 // if sorting them is actually going to do anything so that we only make a
2522 // new landingpad instruction if it does.
2523 for (unsigned k = i; k + 1 < j; ++k)
2524 if (shorter_filter(NewClauses[k+1], NewClauses[k])) {
2525 // Not sorted, so sort the filters now. Doing an unstable sort would be
2526 // correct too but reordering filters pointlessly might confuse users.
2527 std::stable_sort(NewClauses.begin() + i, NewClauses.begin() + j,
2529 MakeNewInstruction = true;
2533 // Look for the next batch of filters.
2537 // If typeinfos matched if and only if equal, then the elements of a filter L
2538 // that occurs later than a filter F could be replaced by the intersection of
2539 // the elements of F and L. In reality two typeinfos can match without being
2540 // equal (for example if one represents a C++ class, and the other some class
2541 // derived from it) so it would be wrong to perform this transform in general.
2542 // However the transform is correct and useful if F is a subset of L. In that
2543 // case L can be replaced by F, and thus removed altogether since repeating a
2544 // filter is pointless. So here we look at all pairs of filters F and L where
2545 // L follows F in the list of clauses, and remove L if every element of F is
2546 // an element of L. This can occur when inlining C++ functions with exception
2548 for (unsigned i = 0; i + 1 < NewClauses.size(); ++i) {
2549 // Examine each filter in turn.
2550 Value *Filter = NewClauses[i];
2551 ArrayType *FTy = dyn_cast<ArrayType>(Filter->getType());
2553 // Not a filter - skip it.
2555 unsigned FElts = FTy->getNumElements();
2556 // Examine each filter following this one. Doing this backwards means that
2557 // we don't have to worry about filters disappearing under us when removed.
2558 for (unsigned j = NewClauses.size() - 1; j != i; --j) {
2559 Value *LFilter = NewClauses[j];
2560 ArrayType *LTy = dyn_cast<ArrayType>(LFilter->getType());
2562 // Not a filter - skip it.
2564 // If Filter is a subset of LFilter, i.e. every element of Filter is also
2565 // an element of LFilter, then discard LFilter.
2566 SmallVectorImpl<Constant *>::iterator J = NewClauses.begin() + j;
2567 // If Filter is empty then it is a subset of LFilter.
2570 NewClauses.erase(J);
2571 MakeNewInstruction = true;
2572 // Move on to the next filter.
2575 unsigned LElts = LTy->getNumElements();
2576 // If Filter is longer than LFilter then it cannot be a subset of it.
2578 // Move on to the next filter.
2580 // At this point we know that LFilter has at least one element.
2581 if (isa<ConstantAggregateZero>(LFilter)) { // LFilter only contains zeros.
2582 // Filter is a subset of LFilter iff Filter contains only zeros (as we
2583 // already know that Filter is not longer than LFilter).
2584 if (isa<ConstantAggregateZero>(Filter)) {
2585 assert(FElts <= LElts && "Should have handled this case earlier!");
2587 NewClauses.erase(J);
2588 MakeNewInstruction = true;
2590 // Move on to the next filter.
2593 ConstantArray *LArray = cast<ConstantArray>(LFilter);
2594 if (isa<ConstantAggregateZero>(Filter)) { // Filter only contains zeros.
2595 // Since Filter is non-empty and contains only zeros, it is a subset of
2596 // LFilter iff LFilter contains a zero.
2597 assert(FElts > 0 && "Should have eliminated the empty filter earlier!");
2598 for (unsigned l = 0; l != LElts; ++l)
2599 if (LArray->getOperand(l)->isNullValue()) {
2600 // LFilter contains a zero - discard it.
2601 NewClauses.erase(J);
2602 MakeNewInstruction = true;
2605 // Move on to the next filter.
2608 // At this point we know that both filters are ConstantArrays. Loop over
2609 // operands to see whether every element of Filter is also an element of
2610 // LFilter. Since filters tend to be short this is probably faster than
2611 // using a method that scales nicely.
2612 ConstantArray *FArray = cast<ConstantArray>(Filter);
2613 bool AllFound = true;
2614 for (unsigned f = 0; f != FElts; ++f) {
2615 Value *FTypeInfo = FArray->getOperand(f)->stripPointerCasts();
2617 for (unsigned l = 0; l != LElts; ++l) {
2618 Value *LTypeInfo = LArray->getOperand(l)->stripPointerCasts();
2619 if (LTypeInfo == FTypeInfo) {
2629 NewClauses.erase(J);
2630 MakeNewInstruction = true;
2632 // Move on to the next filter.
2636 // If we changed any of the clauses, replace the old landingpad instruction
2638 if (MakeNewInstruction) {
2639 LandingPadInst *NLI = LandingPadInst::Create(LI.getType(),
2641 for (unsigned i = 0, e = NewClauses.size(); i != e; ++i)
2642 NLI->addClause(NewClauses[i]);
2643 // A landing pad with no clauses must have the cleanup flag set. It is
2644 // theoretically possible, though highly unlikely, that we eliminated all
2645 // clauses. If so, force the cleanup flag to true.
2646 if (NewClauses.empty())
2648 NLI->setCleanup(CleanupFlag);
2652 // Even if none of the clauses changed, we may nonetheless have understood
2653 // that the cleanup flag is pointless. Clear it if so.
2654 if (LI.isCleanup() != CleanupFlag) {
2655 assert(!CleanupFlag && "Adding a cleanup, not removing one?!");
2656 LI.setCleanup(CleanupFlag);
2663 /// Try to move the specified instruction from its current block into the
2664 /// beginning of DestBlock, which can only happen if it's safe to move the
2665 /// instruction past all of the instructions between it and the end of its
2667 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
2668 assert(I->hasOneUse() && "Invariants didn't hold!");
2670 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
2671 if (isa<PHINode>(I) || I->isEHPad() || I->mayHaveSideEffects() ||
2672 isa<TerminatorInst>(I))
2675 // Do not sink alloca instructions out of the entry block.
2676 if (isa<AllocaInst>(I) && I->getParent() ==
2677 &DestBlock->getParent()->getEntryBlock())
2680 // We can only sink load instructions if there is nothing between the load and
2681 // the end of block that could change the value.
2682 if (I->mayReadFromMemory()) {
2683 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
2685 if (Scan->mayWriteToMemory())
2689 BasicBlock::iterator InsertPos = DestBlock->getFirstInsertionPt();
2690 I->moveBefore(InsertPos);
2695 bool InstCombiner::run() {
2696 while (!Worklist.isEmpty()) {
2697 Instruction *I = Worklist.RemoveOne();
2698 if (I == nullptr) continue; // skip null values.
2700 // Check to see if we can DCE the instruction.
2701 if (isInstructionTriviallyDead(I, TLI)) {
2702 DEBUG(dbgs() << "IC: DCE: " << *I << '\n');
2703 EraseInstFromFunction(*I);
2705 MadeIRChange = true;
2709 // Instruction isn't dead, see if we can constant propagate it.
2710 if (!I->use_empty() &&
2711 (I->getNumOperands() == 0 || isa<Constant>(I->getOperand(0)))) {
2712 if (Constant *C = ConstantFoldInstruction(I, DL, TLI)) {
2713 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n');
2715 // Add operands to the worklist.
2716 ReplaceInstUsesWith(*I, C);
2718 EraseInstFromFunction(*I);
2719 MadeIRChange = true;
2724 // See if we can trivially sink this instruction to a successor basic block.
2725 if (I->hasOneUse()) {
2726 BasicBlock *BB = I->getParent();
2727 Instruction *UserInst = cast<Instruction>(*I->user_begin());
2728 BasicBlock *UserParent;
2730 // Get the block the use occurs in.
2731 if (PHINode *PN = dyn_cast<PHINode>(UserInst))
2732 UserParent = PN->getIncomingBlock(*I->use_begin());
2734 UserParent = UserInst->getParent();
2736 if (UserParent != BB) {
2737 bool UserIsSuccessor = false;
2738 // See if the user is one of our successors.
2739 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
2740 if (*SI == UserParent) {
2741 UserIsSuccessor = true;
2745 // If the user is one of our immediate successors, and if that successor
2746 // only has us as a predecessors (we'd have to split the critical edge
2747 // otherwise), we can keep going.
2748 if (UserIsSuccessor && UserParent->getSinglePredecessor()) {
2749 // Okay, the CFG is simple enough, try to sink this instruction.
2750 if (TryToSinkInstruction(I, UserParent)) {
2751 MadeIRChange = true;
2752 // We'll add uses of the sunk instruction below, but since sinking
2753 // can expose opportunities for it's *operands* add them to the
2755 for (Use &U : I->operands())
2756 if (Instruction *OpI = dyn_cast<Instruction>(U.get()))
2763 // Now that we have an instruction, try combining it to simplify it.
2764 Builder->SetInsertPoint(I->getParent(), I);
2765 Builder->SetCurrentDebugLocation(I->getDebugLoc());
2770 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str(););
2771 DEBUG(dbgs() << "IC: Visiting: " << OrigI << '\n');
2773 if (Instruction *Result = visit(*I)) {
2775 // Should we replace the old instruction with a new one?
2777 DEBUG(dbgs() << "IC: Old = " << *I << '\n'
2778 << " New = " << *Result << '\n');
2780 if (I->getDebugLoc())
2781 Result->setDebugLoc(I->getDebugLoc());
2782 // Everything uses the new instruction now.
2783 I->replaceAllUsesWith(Result);
2785 // Move the name to the new instruction first.
2786 Result->takeName(I);
2788 // Push the new instruction and any users onto the worklist.
2789 Worklist.Add(Result);
2790 Worklist.AddUsersToWorkList(*Result);
2792 // Insert the new instruction into the basic block...
2793 BasicBlock *InstParent = I->getParent();
2794 BasicBlock::iterator InsertPos = I;
2796 // If we replace a PHI with something that isn't a PHI, fix up the
2798 if (!isa<PHINode>(Result) && isa<PHINode>(InsertPos))
2799 InsertPos = InstParent->getFirstInsertionPt();
2801 InstParent->getInstList().insert(InsertPos, Result);
2803 EraseInstFromFunction(*I);
2806 DEBUG(dbgs() << "IC: Mod = " << OrigI << '\n'
2807 << " New = " << *I << '\n');
2810 // If the instruction was modified, it's possible that it is now dead.
2811 // if so, remove it.
2812 if (isInstructionTriviallyDead(I, TLI)) {
2813 EraseInstFromFunction(*I);
2816 Worklist.AddUsersToWorkList(*I);
2819 MadeIRChange = true;
2824 return MadeIRChange;
2827 /// Walk the function in depth-first order, adding all reachable code to the
2830 /// This has a couple of tricks to make the code faster and more powerful. In
2831 /// particular, we constant fold and DCE instructions as we go, to avoid adding
2832 /// them to the worklist (this significantly speeds up instcombine on code where
2833 /// many instructions are dead or constant). Additionally, if we find a branch
2834 /// whose condition is a known constant, we only visit the reachable successors.
2836 static bool AddReachableCodeToWorklist(BasicBlock *BB, const DataLayout &DL,
2837 SmallPtrSetImpl<BasicBlock *> &Visited,
2838 InstCombineWorklist &ICWorklist,
2839 const TargetLibraryInfo *TLI) {
2840 bool MadeIRChange = false;
2841 SmallVector<BasicBlock*, 256> Worklist;
2842 Worklist.push_back(BB);
2844 SmallVector<Instruction*, 128> InstrsForInstCombineWorklist;
2845 DenseMap<ConstantExpr*, Constant*> FoldedConstants;
2848 BB = Worklist.pop_back_val();
2850 // We have now visited this block! If we've already been here, ignore it.
2851 if (!Visited.insert(BB).second)
2854 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
2855 Instruction *Inst = BBI++;
2857 // DCE instruction if trivially dead.
2858 if (isInstructionTriviallyDead(Inst, TLI)) {
2860 DEBUG(dbgs() << "IC: DCE: " << *Inst << '\n');
2861 Inst->eraseFromParent();
2865 // ConstantProp instruction if trivially constant.
2866 if (!Inst->use_empty() &&
2867 (Inst->getNumOperands() == 0 || isa<Constant>(Inst->getOperand(0))))
2868 if (Constant *C = ConstantFoldInstruction(Inst, DL, TLI)) {
2869 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: "
2871 Inst->replaceAllUsesWith(C);
2873 Inst->eraseFromParent();
2877 // See if we can constant fold its operands.
2878 for (User::op_iterator i = Inst->op_begin(), e = Inst->op_end(); i != e;
2880 ConstantExpr *CE = dyn_cast<ConstantExpr>(i);
2884 Constant *&FoldRes = FoldedConstants[CE];
2886 FoldRes = ConstantFoldConstantExpression(CE, DL, TLI);
2890 if (FoldRes != CE) {
2892 MadeIRChange = true;
2896 InstrsForInstCombineWorklist.push_back(Inst);
2899 // Recursively visit successors. If this is a branch or switch on a
2900 // constant, only visit the reachable successor.
2901 TerminatorInst *TI = BB->getTerminator();
2902 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
2903 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
2904 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
2905 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
2906 Worklist.push_back(ReachableBB);
2909 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
2910 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
2911 // See if this is an explicit destination.
2912 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2914 if (i.getCaseValue() == Cond) {
2915 BasicBlock *ReachableBB = i.getCaseSuccessor();
2916 Worklist.push_back(ReachableBB);
2920 // Otherwise it is the default destination.
2921 Worklist.push_back(SI->getDefaultDest());
2926 for (BasicBlock *SuccBB : TI->successors())
2927 Worklist.push_back(SuccBB);
2928 } while (!Worklist.empty());
2930 // Once we've found all of the instructions to add to instcombine's worklist,
2931 // add them in reverse order. This way instcombine will visit from the top
2932 // of the function down. This jives well with the way that it adds all uses
2933 // of instructions to the worklist after doing a transformation, thus avoiding
2934 // some N^2 behavior in pathological cases.
2935 ICWorklist.AddInitialGroup(&InstrsForInstCombineWorklist[0],
2936 InstrsForInstCombineWorklist.size());
2938 return MadeIRChange;
2941 /// \brief Populate the IC worklist from a function, and prune any dead basic
2942 /// blocks discovered in the process.
2944 /// This also does basic constant propagation and other forward fixing to make
2945 /// the combiner itself run much faster.
2946 static bool prepareICWorklistFromFunction(Function &F, const DataLayout &DL,
2947 TargetLibraryInfo *TLI,
2948 InstCombineWorklist &ICWorklist) {
2949 bool MadeIRChange = false;
2951 // Do a depth-first traversal of the function, populate the worklist with
2952 // the reachable instructions. Ignore blocks that are not reachable. Keep
2953 // track of which blocks we visit.
2954 SmallPtrSet<BasicBlock *, 64> Visited;
2956 AddReachableCodeToWorklist(F.begin(), DL, Visited, ICWorklist, TLI);
2958 // Do a quick scan over the function. If we find any blocks that are
2959 // unreachable, remove any instructions inside of them. This prevents
2960 // the instcombine code from having to deal with some bad special cases.
2961 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
2962 if (Visited.count(BB))
2965 // Delete the instructions backwards, as it has a reduced likelihood of
2966 // having to update as many def-use and use-def chains.
2967 Instruction *EndInst = BB->getTerminator(); // Last not to be deleted.
2968 while (EndInst != BB->begin()) {
2969 // Delete the next to last instruction.
2970 BasicBlock::iterator I = EndInst;
2971 Instruction *Inst = --I;
2972 if (!Inst->use_empty())
2973 Inst->replaceAllUsesWith(UndefValue::get(Inst->getType()));
2974 if (Inst->isEHPad()) {
2978 if (!isa<DbgInfoIntrinsic>(Inst)) {
2980 MadeIRChange = true;
2982 Inst->eraseFromParent();
2986 return MadeIRChange;
2990 combineInstructionsOverFunction(Function &F, InstCombineWorklist &Worklist,
2991 AliasAnalysis *AA, AssumptionCache &AC,
2992 TargetLibraryInfo &TLI, DominatorTree &DT,
2993 LoopInfo *LI = nullptr) {
2994 auto &DL = F.getParent()->getDataLayout();
2996 /// Builder - This is an IRBuilder that automatically inserts new
2997 /// instructions into the worklist when they are created.
2998 IRBuilder<true, TargetFolder, InstCombineIRInserter> Builder(
2999 F.getContext(), TargetFolder(DL), InstCombineIRInserter(Worklist, &AC));
3001 // Lower dbg.declare intrinsics otherwise their value may be clobbered
3003 bool DbgDeclaresChanged = LowerDbgDeclare(F);
3005 // Iterate while there is work to do.
3009 DEBUG(dbgs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
3010 << F.getName() << "\n");
3012 bool Changed = false;
3013 if (prepareICWorklistFromFunction(F, DL, &TLI, Worklist))
3016 InstCombiner IC(Worklist, &Builder, F.optForMinSize(),
3017 AA, &AC, &TLI, &DT, DL, LI);
3025 return DbgDeclaresChanged || Iteration > 1;
3028 PreservedAnalyses InstCombinePass::run(Function &F,
3029 AnalysisManager<Function> *AM) {
3030 auto &AC = AM->getResult<AssumptionAnalysis>(F);
3031 auto &DT = AM->getResult<DominatorTreeAnalysis>(F);
3032 auto &TLI = AM->getResult<TargetLibraryAnalysis>(F);
3034 auto *LI = AM->getCachedResult<LoopAnalysis>(F);
3036 // FIXME: The AliasAnalysis is not yet supported in the new pass manager
3037 if (!combineInstructionsOverFunction(F, Worklist, nullptr, AC, TLI, DT, LI))
3038 // No changes, all analyses are preserved.
3039 return PreservedAnalyses::all();
3041 // Mark all the analyses that instcombine updates as preserved.
3042 // FIXME: Need a way to preserve CFG analyses here!
3043 PreservedAnalyses PA;
3044 PA.preserve<DominatorTreeAnalysis>();
3049 /// \brief The legacy pass manager's instcombine pass.
3051 /// This is a basic whole-function wrapper around the instcombine utility. It
3052 /// will try to combine all instructions in the function.
3053 class InstructionCombiningPass : public FunctionPass {
3054 InstCombineWorklist Worklist;
3057 static char ID; // Pass identification, replacement for typeid
3059 InstructionCombiningPass() : FunctionPass(ID) {
3060 initializeInstructionCombiningPassPass(*PassRegistry::getPassRegistry());
3063 void getAnalysisUsage(AnalysisUsage &AU) const override;
3064 bool runOnFunction(Function &F) override;
3068 void InstructionCombiningPass::getAnalysisUsage(AnalysisUsage &AU) const {
3069 AU.setPreservesCFG();
3070 AU.addRequired<AAResultsWrapperPass>();
3071 AU.addRequired<AssumptionCacheTracker>();
3072 AU.addRequired<TargetLibraryInfoWrapperPass>();
3073 AU.addRequired<DominatorTreeWrapperPass>();
3074 AU.addPreserved<DominatorTreeWrapperPass>();
3075 AU.addPreserved<GlobalsAAWrapperPass>();
3078 bool InstructionCombiningPass::runOnFunction(Function &F) {
3079 if (skipOptnoneFunction(F))
3082 // Required analyses.
3083 auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
3084 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
3085 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
3086 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
3088 // Optional analyses.
3089 auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>();
3090 auto *LI = LIWP ? &LIWP->getLoopInfo() : nullptr;
3092 return combineInstructionsOverFunction(F, Worklist, AA, AC, TLI, DT, LI);
3095 char InstructionCombiningPass::ID = 0;
3096 INITIALIZE_PASS_BEGIN(InstructionCombiningPass, "instcombine",
3097 "Combine redundant instructions", false, false)
3098 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
3099 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
3100 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
3101 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
3102 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
3103 INITIALIZE_PASS_END(InstructionCombiningPass, "instcombine",
3104 "Combine redundant instructions", false, false)
3106 // Initialization Routines
3107 void llvm::initializeInstCombine(PassRegistry &Registry) {
3108 initializeInstructionCombiningPassPass(Registry);
3111 void LLVMInitializeInstCombine(LLVMPassRegistryRef R) {
3112 initializeInstructionCombiningPassPass(*unwrap(R));
3115 FunctionPass *llvm::createInstructionCombiningPass() {
3116 return new InstructionCombiningPass();