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/InstructionSimplify.h"
46 #include "llvm/Analysis/LibCallSemantics.h"
47 #include "llvm/Analysis/LoopInfo.h"
48 #include "llvm/Analysis/MemoryBuiltins.h"
49 #include "llvm/Analysis/TargetLibraryInfo.h"
50 #include "llvm/Analysis/ValueTracking.h"
51 #include "llvm/IR/CFG.h"
52 #include "llvm/IR/DataLayout.h"
53 #include "llvm/IR/Dominators.h"
54 #include "llvm/IR/GetElementPtrTypeIterator.h"
55 #include "llvm/IR/IntrinsicInst.h"
56 #include "llvm/IR/PatternMatch.h"
57 #include "llvm/IR/ValueHandle.h"
58 #include "llvm/Support/CommandLine.h"
59 #include "llvm/Support/Debug.h"
60 #include "llvm/Transforms/Scalar.h"
61 #include "llvm/Transforms/Utils/Local.h"
65 using namespace llvm::PatternMatch;
67 #define DEBUG_TYPE "instcombine"
69 STATISTIC(NumCombined , "Number of insts combined");
70 STATISTIC(NumConstProp, "Number of constant folds");
71 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
72 STATISTIC(NumSunkInst , "Number of instructions sunk");
73 STATISTIC(NumExpand, "Number of expansions");
74 STATISTIC(NumFactor , "Number of factorizations");
75 STATISTIC(NumReassoc , "Number of reassociations");
77 Value *InstCombiner::EmitGEPOffset(User *GEP) {
78 return llvm::EmitGEPOffset(Builder, DL, GEP);
81 /// ShouldChangeType - Return true if it is desirable to convert a computation
82 /// from 'From' to 'To'. We don't want to convert from a legal to an illegal
83 /// type for example, or from a smaller to a larger illegal type.
84 bool InstCombiner::ShouldChangeType(Type *From, Type *To) const {
85 assert(From->isIntegerTy() && To->isIntegerTy());
87 unsigned FromWidth = From->getPrimitiveSizeInBits();
88 unsigned ToWidth = To->getPrimitiveSizeInBits();
89 bool FromLegal = DL.isLegalInteger(FromWidth);
90 bool ToLegal = DL.isLegalInteger(ToWidth);
92 // If this is a legal integer from type, and the result would be an illegal
93 // type, don't do the transformation.
94 if (FromLegal && !ToLegal)
97 // Otherwise, if both are illegal, do not increase the size of the result. We
98 // do allow things like i160 -> i64, but not i64 -> i160.
99 if (!FromLegal && !ToLegal && ToWidth > FromWidth)
105 // Return true, if No Signed Wrap should be maintained for I.
106 // The No Signed Wrap flag can be kept if the operation "B (I.getOpcode) C",
107 // where both B and C should be ConstantInts, results in a constant that does
108 // not overflow. This function only handles the Add and Sub opcodes. For
109 // all other opcodes, the function conservatively returns false.
110 static bool MaintainNoSignedWrap(BinaryOperator &I, Value *B, Value *C) {
111 OverflowingBinaryOperator *OBO = dyn_cast<OverflowingBinaryOperator>(&I);
112 if (!OBO || !OBO->hasNoSignedWrap()) {
116 // We reason about Add and Sub Only.
117 Instruction::BinaryOps Opcode = I.getOpcode();
118 if (Opcode != Instruction::Add &&
119 Opcode != Instruction::Sub) {
123 ConstantInt *CB = dyn_cast<ConstantInt>(B);
124 ConstantInt *CC = dyn_cast<ConstantInt>(C);
130 const APInt &BVal = CB->getValue();
131 const APInt &CVal = CC->getValue();
132 bool Overflow = false;
134 if (Opcode == Instruction::Add) {
135 BVal.sadd_ov(CVal, Overflow);
137 BVal.ssub_ov(CVal, Overflow);
143 /// Conservatively clears subclassOptionalData after a reassociation or
144 /// commutation. We preserve fast-math flags when applicable as they can be
146 static void ClearSubclassDataAfterReassociation(BinaryOperator &I) {
147 FPMathOperator *FPMO = dyn_cast<FPMathOperator>(&I);
149 I.clearSubclassOptionalData();
153 FastMathFlags FMF = I.getFastMathFlags();
154 I.clearSubclassOptionalData();
155 I.setFastMathFlags(FMF);
158 /// SimplifyAssociativeOrCommutative - This performs a few simplifications for
159 /// operators which are associative or commutative:
161 // Commutative operators:
163 // 1. Order operands such that they are listed from right (least complex) to
164 // left (most complex). This puts constants before unary operators before
167 // Associative operators:
169 // 2. Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
170 // 3. Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
172 // Associative and commutative operators:
174 // 4. Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
175 // 5. Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
176 // 6. Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
177 // if C1 and C2 are constants.
179 bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) {
180 Instruction::BinaryOps Opcode = I.getOpcode();
181 bool Changed = false;
184 // Order operands such that they are listed from right (least complex) to
185 // left (most complex). This puts constants before unary operators before
187 if (I.isCommutative() && getComplexity(I.getOperand(0)) <
188 getComplexity(I.getOperand(1)))
189 Changed = !I.swapOperands();
191 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
192 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
194 if (I.isAssociative()) {
195 // Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
196 if (Op0 && Op0->getOpcode() == Opcode) {
197 Value *A = Op0->getOperand(0);
198 Value *B = Op0->getOperand(1);
199 Value *C = I.getOperand(1);
201 // Does "B op C" simplify?
202 if (Value *V = SimplifyBinOp(Opcode, B, C, DL)) {
203 // It simplifies to V. Form "A op V".
206 // Conservatively clear the optional flags, since they may not be
207 // preserved by the reassociation.
208 if (MaintainNoSignedWrap(I, B, C) &&
209 (!Op0 || (isa<BinaryOperator>(Op0) && Op0->hasNoSignedWrap()))) {
210 // Note: this is only valid because SimplifyBinOp doesn't look at
211 // the operands to Op0.
212 I.clearSubclassOptionalData();
213 I.setHasNoSignedWrap(true);
215 ClearSubclassDataAfterReassociation(I);
224 // Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
225 if (Op1 && Op1->getOpcode() == Opcode) {
226 Value *A = I.getOperand(0);
227 Value *B = Op1->getOperand(0);
228 Value *C = Op1->getOperand(1);
230 // Does "A op B" simplify?
231 if (Value *V = SimplifyBinOp(Opcode, A, B, DL)) {
232 // It simplifies to V. Form "V op C".
235 // Conservatively clear the optional flags, since they may not be
236 // preserved by the reassociation.
237 ClearSubclassDataAfterReassociation(I);
245 if (I.isAssociative() && I.isCommutative()) {
246 // Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
247 if (Op0 && Op0->getOpcode() == Opcode) {
248 Value *A = Op0->getOperand(0);
249 Value *B = Op0->getOperand(1);
250 Value *C = I.getOperand(1);
252 // Does "C op A" simplify?
253 if (Value *V = SimplifyBinOp(Opcode, C, A, DL)) {
254 // It simplifies to V. Form "V op B".
257 // Conservatively clear the optional flags, since they may not be
258 // preserved by the reassociation.
259 ClearSubclassDataAfterReassociation(I);
266 // Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
267 if (Op1 && Op1->getOpcode() == Opcode) {
268 Value *A = I.getOperand(0);
269 Value *B = Op1->getOperand(0);
270 Value *C = Op1->getOperand(1);
272 // Does "C op A" simplify?
273 if (Value *V = SimplifyBinOp(Opcode, C, A, DL)) {
274 // It simplifies to V. Form "B op V".
277 // Conservatively clear the optional flags, since they may not be
278 // preserved by the reassociation.
279 ClearSubclassDataAfterReassociation(I);
286 // Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
287 // if C1 and C2 are constants.
289 Op0->getOpcode() == Opcode && Op1->getOpcode() == Opcode &&
290 isa<Constant>(Op0->getOperand(1)) &&
291 isa<Constant>(Op1->getOperand(1)) &&
292 Op0->hasOneUse() && Op1->hasOneUse()) {
293 Value *A = Op0->getOperand(0);
294 Constant *C1 = cast<Constant>(Op0->getOperand(1));
295 Value *B = Op1->getOperand(0);
296 Constant *C2 = cast<Constant>(Op1->getOperand(1));
298 Constant *Folded = ConstantExpr::get(Opcode, C1, C2);
299 BinaryOperator *New = BinaryOperator::Create(Opcode, A, B);
300 if (isa<FPMathOperator>(New)) {
301 FastMathFlags Flags = I.getFastMathFlags();
302 Flags &= Op0->getFastMathFlags();
303 Flags &= Op1->getFastMathFlags();
304 New->setFastMathFlags(Flags);
306 InsertNewInstWith(New, I);
308 I.setOperand(0, New);
309 I.setOperand(1, Folded);
310 // Conservatively clear the optional flags, since they may not be
311 // preserved by the reassociation.
312 ClearSubclassDataAfterReassociation(I);
319 // No further simplifications.
324 /// LeftDistributesOverRight - Whether "X LOp (Y ROp Z)" is always equal to
325 /// "(X LOp Y) ROp (X LOp Z)".
326 static bool LeftDistributesOverRight(Instruction::BinaryOps LOp,
327 Instruction::BinaryOps ROp) {
332 case Instruction::And:
333 // And distributes over Or and Xor.
337 case Instruction::Or:
338 case Instruction::Xor:
342 case Instruction::Mul:
343 // Multiplication distributes over addition and subtraction.
347 case Instruction::Add:
348 case Instruction::Sub:
352 case Instruction::Or:
353 // Or distributes over And.
357 case Instruction::And:
363 /// RightDistributesOverLeft - Whether "(X LOp Y) ROp Z" is always equal to
364 /// "(X ROp Z) LOp (Y ROp Z)".
365 static bool RightDistributesOverLeft(Instruction::BinaryOps LOp,
366 Instruction::BinaryOps ROp) {
367 if (Instruction::isCommutative(ROp))
368 return LeftDistributesOverRight(ROp, LOp);
373 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
374 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
375 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
376 case Instruction::And:
377 case Instruction::Or:
378 case Instruction::Xor:
382 case Instruction::Shl:
383 case Instruction::LShr:
384 case Instruction::AShr:
388 // TODO: It would be nice to handle division, aka "(X + Y)/Z = X/Z + Y/Z",
389 // but this requires knowing that the addition does not overflow and other
394 /// This function returns identity value for given opcode, which can be used to
395 /// factor patterns like (X * 2) + X ==> (X * 2) + (X * 1) ==> X * (2 + 1).
396 static Value *getIdentityValue(Instruction::BinaryOps OpCode, Value *V) {
397 if (isa<Constant>(V))
400 if (OpCode == Instruction::Mul)
401 return ConstantInt::get(V->getType(), 1);
403 // TODO: We can handle other cases e.g. Instruction::And, Instruction::Or etc.
408 /// This function factors binary ops which can be combined using distributive
409 /// laws. This function tries to transform 'Op' based TopLevelOpcode to enable
410 /// factorization e.g for ADD(SHL(X , 2), MUL(X, 5)), When this function called
411 /// with TopLevelOpcode == Instruction::Add and Op = SHL(X, 2), transforms
412 /// SHL(X, 2) to MUL(X, 4) i.e. returns Instruction::Mul with LHS set to 'X' and
414 static Instruction::BinaryOps
415 getBinOpsForFactorization(Instruction::BinaryOps TopLevelOpcode,
416 BinaryOperator *Op, Value *&LHS, Value *&RHS) {
418 return Instruction::BinaryOpsEnd;
420 LHS = Op->getOperand(0);
421 RHS = Op->getOperand(1);
423 switch (TopLevelOpcode) {
425 return Op->getOpcode();
427 case Instruction::Add:
428 case Instruction::Sub:
429 if (Op->getOpcode() == Instruction::Shl) {
430 if (Constant *CST = dyn_cast<Constant>(Op->getOperand(1))) {
431 // The multiplier is really 1 << CST.
432 RHS = ConstantExpr::getShl(ConstantInt::get(Op->getType(), 1), CST);
433 return Instruction::Mul;
436 return Op->getOpcode();
439 // TODO: We can add other conversions e.g. shr => div etc.
442 /// This tries to simplify binary operations by factorizing out common terms
443 /// (e. g. "(A*B)+(A*C)" -> "A*(B+C)").
444 static Value *tryFactorization(InstCombiner::BuilderTy *Builder,
445 const DataLayout &DL, BinaryOperator &I,
446 Instruction::BinaryOps InnerOpcode, Value *A,
447 Value *B, Value *C, Value *D) {
449 // If any of A, B, C, D are null, we can not factor I, return early.
450 // Checking A and C should be enough.
451 if (!A || !C || !B || !D)
454 Value *SimplifiedInst = nullptr;
455 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
456 Instruction::BinaryOps TopLevelOpcode = I.getOpcode();
458 // Does "X op' Y" always equal "Y op' X"?
459 bool InnerCommutative = Instruction::isCommutative(InnerOpcode);
461 // Does "X op' (Y op Z)" always equal "(X op' Y) op (X op' Z)"?
462 if (LeftDistributesOverRight(InnerOpcode, TopLevelOpcode))
463 // Does the instruction have the form "(A op' B) op (A op' D)" or, in the
464 // commutative case, "(A op' B) op (C op' A)"?
465 if (A == C || (InnerCommutative && A == D)) {
468 // Consider forming "A op' (B op D)".
469 // If "B op D" simplifies then it can be formed with no cost.
470 Value *V = SimplifyBinOp(TopLevelOpcode, B, D, DL);
471 // If "B op D" doesn't simplify then only go on if both of the existing
472 // operations "A op' B" and "C op' D" will be zapped as no longer used.
473 if (!V && LHS->hasOneUse() && RHS->hasOneUse())
474 V = Builder->CreateBinOp(TopLevelOpcode, B, D, RHS->getName());
476 SimplifiedInst = Builder->CreateBinOp(InnerOpcode, A, V);
480 // Does "(X op Y) op' Z" always equal "(X op' Z) op (Y op' Z)"?
481 if (!SimplifiedInst && RightDistributesOverLeft(TopLevelOpcode, InnerOpcode))
482 // Does the instruction have the form "(A op' B) op (C op' B)" or, in the
483 // commutative case, "(A op' B) op (B op' D)"?
484 if (B == D || (InnerCommutative && B == C)) {
487 // Consider forming "(A op C) op' B".
488 // If "A op C" simplifies then it can be formed with no cost.
489 Value *V = SimplifyBinOp(TopLevelOpcode, A, C, DL);
491 // If "A op C" doesn't simplify then only go on if both of the existing
492 // operations "A op' B" and "C op' D" will be zapped as no longer used.
493 if (!V && LHS->hasOneUse() && RHS->hasOneUse())
494 V = Builder->CreateBinOp(TopLevelOpcode, A, C, LHS->getName());
496 SimplifiedInst = Builder->CreateBinOp(InnerOpcode, V, B);
500 if (SimplifiedInst) {
502 SimplifiedInst->takeName(&I);
504 // Check if we can add NSW flag to SimplifiedInst. If so, set NSW flag.
505 // TODO: Check for NUW.
506 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(SimplifiedInst)) {
507 if (isa<OverflowingBinaryOperator>(SimplifiedInst)) {
509 if (isa<OverflowingBinaryOperator>(&I))
510 HasNSW = I.hasNoSignedWrap();
512 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS))
513 if (isa<OverflowingBinaryOperator>(Op0))
514 HasNSW &= Op0->hasNoSignedWrap();
516 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS))
517 if (isa<OverflowingBinaryOperator>(Op1))
518 HasNSW &= Op1->hasNoSignedWrap();
519 BO->setHasNoSignedWrap(HasNSW);
523 return SimplifiedInst;
526 /// SimplifyUsingDistributiveLaws - This tries to simplify binary operations
527 /// which some other binary operation distributes over either by factorizing
528 /// out common terms (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this
529 /// results in simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is
530 /// a win). Returns the simplified value, or null if it didn't simplify.
531 Value *InstCombiner::SimplifyUsingDistributiveLaws(BinaryOperator &I) {
532 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
533 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
534 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
537 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
538 auto TopLevelOpcode = I.getOpcode();
539 auto LHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op0, A, B);
540 auto RHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op1, C, D);
542 // The instruction has the form "(A op' B) op (C op' D)". Try to factorize
544 if (LHSOpcode == RHSOpcode) {
545 if (Value *V = tryFactorization(Builder, DL, I, LHSOpcode, A, B, C, D))
549 // The instruction has the form "(A op' B) op (C)". Try to factorize common
551 if (Value *V = tryFactorization(Builder, DL, I, LHSOpcode, A, B, RHS,
552 getIdentityValue(LHSOpcode, RHS)))
555 // The instruction has the form "(B) op (C op' D)". Try to factorize common
557 if (Value *V = tryFactorization(Builder, DL, I, RHSOpcode, LHS,
558 getIdentityValue(RHSOpcode, LHS), C, D))
562 if (Op0 && RightDistributesOverLeft(Op0->getOpcode(), TopLevelOpcode)) {
563 // The instruction has the form "(A op' B) op C". See if expanding it out
564 // to "(A op C) op' (B op C)" results in simplifications.
565 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
566 Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op'
568 // Do "A op C" and "B op C" both simplify?
569 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, C, DL))
570 if (Value *R = SimplifyBinOp(TopLevelOpcode, B, C, DL)) {
571 // They do! Return "L op' R".
573 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
574 if ((L == A && R == B) ||
575 (Instruction::isCommutative(InnerOpcode) && L == B && R == A))
577 // Otherwise return "L op' R" if it simplifies.
578 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, DL))
580 // Otherwise, create a new instruction.
581 C = Builder->CreateBinOp(InnerOpcode, L, R);
587 if (Op1 && LeftDistributesOverRight(TopLevelOpcode, Op1->getOpcode())) {
588 // The instruction has the form "A op (B op' C)". See if expanding it out
589 // to "(A op B) op' (A op C)" results in simplifications.
590 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
591 Instruction::BinaryOps InnerOpcode = Op1->getOpcode(); // op'
593 // Do "A op B" and "A op C" both simplify?
594 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, B, DL))
595 if (Value *R = SimplifyBinOp(TopLevelOpcode, A, C, DL)) {
596 // They do! Return "L op' R".
598 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
599 if ((L == B && R == C) ||
600 (Instruction::isCommutative(InnerOpcode) && L == C && R == B))
602 // Otherwise return "L op' R" if it simplifies.
603 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, DL))
605 // Otherwise, create a new instruction.
606 A = Builder->CreateBinOp(InnerOpcode, L, R);
615 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
616 // if the LHS is a constant zero (which is the 'negate' form).
618 Value *InstCombiner::dyn_castNegVal(Value *V) const {
619 if (BinaryOperator::isNeg(V))
620 return BinaryOperator::getNegArgument(V);
622 // Constants can be considered to be negated values if they can be folded.
623 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
624 return ConstantExpr::getNeg(C);
626 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
627 if (C->getType()->getElementType()->isIntegerTy())
628 return ConstantExpr::getNeg(C);
633 // dyn_castFNegVal - Given a 'fsub' instruction, return the RHS of the
634 // instruction if the LHS is a constant negative zero (which is the 'negate'
637 Value *InstCombiner::dyn_castFNegVal(Value *V, bool IgnoreZeroSign) const {
638 if (BinaryOperator::isFNeg(V, IgnoreZeroSign))
639 return BinaryOperator::getFNegArgument(V);
641 // Constants can be considered to be negated values if they can be folded.
642 if (ConstantFP *C = dyn_cast<ConstantFP>(V))
643 return ConstantExpr::getFNeg(C);
645 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
646 if (C->getType()->getElementType()->isFloatingPointTy())
647 return ConstantExpr::getFNeg(C);
652 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
654 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
655 return IC->Builder->CreateCast(CI->getOpcode(), SO, I.getType());
658 // Figure out if the constant is the left or the right argument.
659 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
660 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
662 if (Constant *SOC = dyn_cast<Constant>(SO)) {
664 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
665 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
668 Value *Op0 = SO, *Op1 = ConstOperand;
672 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I)) {
673 Value *RI = IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1,
674 SO->getName()+".op");
675 Instruction *FPInst = dyn_cast<Instruction>(RI);
676 if (FPInst && isa<FPMathOperator>(FPInst))
677 FPInst->copyFastMathFlags(BO);
680 if (ICmpInst *CI = dyn_cast<ICmpInst>(&I))
681 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
682 SO->getName()+".cmp");
683 if (FCmpInst *CI = dyn_cast<FCmpInst>(&I))
684 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
685 SO->getName()+".cmp");
686 llvm_unreachable("Unknown binary instruction type!");
689 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
690 // constant as the other operand, try to fold the binary operator into the
691 // select arguments. This also works for Cast instructions, which obviously do
692 // not have a second operand.
693 Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) {
694 // Don't modify shared select instructions
695 if (!SI->hasOneUse()) return nullptr;
696 Value *TV = SI->getOperand(1);
697 Value *FV = SI->getOperand(2);
699 if (isa<Constant>(TV) || isa<Constant>(FV)) {
700 // Bool selects with constant operands can be folded to logical ops.
701 if (SI->getType()->isIntegerTy(1)) return nullptr;
703 // If it's a bitcast involving vectors, make sure it has the same number of
704 // elements on both sides.
705 if (BitCastInst *BC = dyn_cast<BitCastInst>(&Op)) {
706 VectorType *DestTy = dyn_cast<VectorType>(BC->getDestTy());
707 VectorType *SrcTy = dyn_cast<VectorType>(BC->getSrcTy());
709 // Verify that either both or neither are vectors.
710 if ((SrcTy == nullptr) != (DestTy == nullptr)) return nullptr;
711 // If vectors, verify that they have the same number of elements.
712 if (SrcTy && SrcTy->getNumElements() != DestTy->getNumElements())
716 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, this);
717 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, this);
719 return SelectInst::Create(SI->getCondition(),
720 SelectTrueVal, SelectFalseVal);
726 /// FoldOpIntoPhi - Given a binary operator, cast instruction, or select which
727 /// has a PHI node as operand #0, see if we can fold the instruction into the
728 /// PHI (which is only possible if all operands to the PHI are constants).
730 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
731 PHINode *PN = cast<PHINode>(I.getOperand(0));
732 unsigned NumPHIValues = PN->getNumIncomingValues();
733 if (NumPHIValues == 0)
736 // We normally only transform phis with a single use. However, if a PHI has
737 // multiple uses and they are all the same operation, we can fold *all* of the
738 // uses into the PHI.
739 if (!PN->hasOneUse()) {
740 // Walk the use list for the instruction, comparing them to I.
741 for (User *U : PN->users()) {
742 Instruction *UI = cast<Instruction>(U);
743 if (UI != &I && !I.isIdenticalTo(UI))
746 // Otherwise, we can replace *all* users with the new PHI we form.
749 // Check to see if all of the operands of the PHI are simple constants
750 // (constantint/constantfp/undef). If there is one non-constant value,
751 // remember the BB it is in. If there is more than one or if *it* is a PHI,
752 // bail out. We don't do arbitrary constant expressions here because moving
753 // their computation can be expensive without a cost model.
754 BasicBlock *NonConstBB = nullptr;
755 for (unsigned i = 0; i != NumPHIValues; ++i) {
756 Value *InVal = PN->getIncomingValue(i);
757 if (isa<Constant>(InVal) && !isa<ConstantExpr>(InVal))
760 if (isa<PHINode>(InVal)) return nullptr; // Itself a phi.
761 if (NonConstBB) return nullptr; // More than one non-const value.
763 NonConstBB = PN->getIncomingBlock(i);
765 // If the InVal is an invoke at the end of the pred block, then we can't
766 // insert a computation after it without breaking the edge.
767 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal))
768 if (II->getParent() == NonConstBB)
771 // If the incoming non-constant value is in I's block, we will remove one
772 // instruction, but insert another equivalent one, leading to infinite
774 if (isPotentiallyReachable(I.getParent(), NonConstBB, DT, LI))
778 // If there is exactly one non-constant value, we can insert a copy of the
779 // operation in that block. However, if this is a critical edge, we would be
780 // inserting the computation on some other paths (e.g. inside a loop). Only
781 // do this if the pred block is unconditionally branching into the phi block.
782 if (NonConstBB != nullptr) {
783 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
784 if (!BI || !BI->isUnconditional()) return nullptr;
787 // Okay, we can do the transformation: create the new PHI node.
788 PHINode *NewPN = PHINode::Create(I.getType(), PN->getNumIncomingValues());
789 InsertNewInstBefore(NewPN, *PN);
792 // If we are going to have to insert a new computation, do so right before the
793 // predecessors terminator.
795 Builder->SetInsertPoint(NonConstBB->getTerminator());
797 // Next, add all of the operands to the PHI.
798 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) {
799 // We only currently try to fold the condition of a select when it is a phi,
800 // not the true/false values.
801 Value *TrueV = SI->getTrueValue();
802 Value *FalseV = SI->getFalseValue();
803 BasicBlock *PhiTransBB = PN->getParent();
804 for (unsigned i = 0; i != NumPHIValues; ++i) {
805 BasicBlock *ThisBB = PN->getIncomingBlock(i);
806 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB);
807 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB);
808 Value *InV = nullptr;
809 // Beware of ConstantExpr: it may eventually evaluate to getNullValue,
810 // even if currently isNullValue gives false.
811 Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i));
812 if (InC && !isa<ConstantExpr>(InC))
813 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred;
815 InV = Builder->CreateSelect(PN->getIncomingValue(i),
816 TrueVInPred, FalseVInPred, "phitmp");
817 NewPN->addIncoming(InV, ThisBB);
819 } else if (CmpInst *CI = dyn_cast<CmpInst>(&I)) {
820 Constant *C = cast<Constant>(I.getOperand(1));
821 for (unsigned i = 0; i != NumPHIValues; ++i) {
822 Value *InV = nullptr;
823 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
824 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
825 else if (isa<ICmpInst>(CI))
826 InV = Builder->CreateICmp(CI->getPredicate(), PN->getIncomingValue(i),
829 InV = Builder->CreateFCmp(CI->getPredicate(), PN->getIncomingValue(i),
831 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
833 } else if (I.getNumOperands() == 2) {
834 Constant *C = cast<Constant>(I.getOperand(1));
835 for (unsigned i = 0; i != NumPHIValues; ++i) {
836 Value *InV = nullptr;
837 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
838 InV = ConstantExpr::get(I.getOpcode(), InC, C);
840 InV = Builder->CreateBinOp(cast<BinaryOperator>(I).getOpcode(),
841 PN->getIncomingValue(i), C, "phitmp");
842 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
845 CastInst *CI = cast<CastInst>(&I);
846 Type *RetTy = CI->getType();
847 for (unsigned i = 0; i != NumPHIValues; ++i) {
849 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
850 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
852 InV = Builder->CreateCast(CI->getOpcode(),
853 PN->getIncomingValue(i), I.getType(), "phitmp");
854 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
858 for (auto UI = PN->user_begin(), E = PN->user_end(); UI != E;) {
859 Instruction *User = cast<Instruction>(*UI++);
860 if (User == &I) continue;
861 ReplaceInstUsesWith(*User, NewPN);
862 EraseInstFromFunction(*User);
864 return ReplaceInstUsesWith(I, NewPN);
867 /// FindElementAtOffset - Given a pointer type and a constant offset, determine
868 /// whether or not there is a sequence of GEP indices into the pointed type that
869 /// will land us at the specified offset. If so, fill them into NewIndices and
870 /// return the resultant element type, otherwise return null.
871 Type *InstCombiner::FindElementAtOffset(Type *PtrTy, int64_t Offset,
872 SmallVectorImpl<Value *> &NewIndices) {
873 assert(PtrTy->isPtrOrPtrVectorTy());
875 Type *Ty = PtrTy->getPointerElementType();
879 // Start with the index over the outer type. Note that the type size
880 // might be zero (even if the offset isn't zero) if the indexed type
881 // is something like [0 x {int, int}]
882 Type *IntPtrTy = DL.getIntPtrType(PtrTy);
883 int64_t FirstIdx = 0;
884 if (int64_t TySize = DL.getTypeAllocSize(Ty)) {
885 FirstIdx = Offset/TySize;
886 Offset -= FirstIdx*TySize;
888 // Handle hosts where % returns negative instead of values [0..TySize).
894 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
897 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
899 // Index into the types. If we fail, set OrigBase to null.
901 // Indexing into tail padding between struct/array elements.
902 if (uint64_t(Offset * 8) >= DL.getTypeSizeInBits(Ty))
905 if (StructType *STy = dyn_cast<StructType>(Ty)) {
906 const StructLayout *SL = DL.getStructLayout(STy);
907 assert(Offset < (int64_t)SL->getSizeInBytes() &&
908 "Offset must stay within the indexed type");
910 unsigned Elt = SL->getElementContainingOffset(Offset);
911 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
914 Offset -= SL->getElementOffset(Elt);
915 Ty = STy->getElementType(Elt);
916 } else if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
917 uint64_t EltSize = DL.getTypeAllocSize(AT->getElementType());
918 assert(EltSize && "Cannot index into a zero-sized array");
919 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
921 Ty = AT->getElementType();
923 // Otherwise, we can't index into the middle of this atomic type, bail.
931 static bool shouldMergeGEPs(GEPOperator &GEP, GEPOperator &Src) {
932 // If this GEP has only 0 indices, it is the same pointer as
933 // Src. If Src is not a trivial GEP too, don't combine
935 if (GEP.hasAllZeroIndices() && !Src.hasAllZeroIndices() &&
941 /// Descale - Return a value X such that Val = X * Scale, or null if none. If
942 /// the multiplication is known not to overflow then NoSignedWrap is set.
943 Value *InstCombiner::Descale(Value *Val, APInt Scale, bool &NoSignedWrap) {
944 assert(isa<IntegerType>(Val->getType()) && "Can only descale integers!");
945 assert(cast<IntegerType>(Val->getType())->getBitWidth() ==
946 Scale.getBitWidth() && "Scale not compatible with value!");
948 // If Val is zero or Scale is one then Val = Val * Scale.
949 if (match(Val, m_Zero()) || Scale == 1) {
954 // If Scale is zero then it does not divide Val.
955 if (Scale.isMinValue())
958 // Look through chains of multiplications, searching for a constant that is
959 // divisible by Scale. For example, descaling X*(Y*(Z*4)) by a factor of 4
960 // will find the constant factor 4 and produce X*(Y*Z). Descaling X*(Y*8) by
961 // a factor of 4 will produce X*(Y*2). The principle of operation is to bore
964 // Val = M1 * X || Analysis starts here and works down
965 // M1 = M2 * Y || Doesn't descend into terms with more
966 // M2 = Z * 4 \/ than one use
968 // Then to modify a term at the bottom:
971 // M1 = Z * Y || Replaced M2 with Z
973 // Then to work back up correcting nsw flags.
975 // Op - the term we are currently analyzing. Starts at Val then drills down.
976 // Replaced with its descaled value before exiting from the drill down loop.
979 // Parent - initially null, but after drilling down notes where Op came from.
980 // In the example above, Parent is (Val, 0) when Op is M1, because M1 is the
981 // 0'th operand of Val.
982 std::pair<Instruction*, unsigned> Parent;
984 // RequireNoSignedWrap - Set if the transform requires a descaling at deeper
985 // levels that doesn't overflow.
986 bool RequireNoSignedWrap = false;
988 // logScale - log base 2 of the scale. Negative if not a power of 2.
989 int32_t logScale = Scale.exactLogBase2();
991 for (;; Op = Parent.first->getOperand(Parent.second)) { // Drill down
993 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
994 // If Op is a constant divisible by Scale then descale to the quotient.
995 APInt Quotient(Scale), Remainder(Scale); // Init ensures right bitwidth.
996 APInt::sdivrem(CI->getValue(), Scale, Quotient, Remainder);
997 if (!Remainder.isMinValue())
998 // Not divisible by Scale.
1000 // Replace with the quotient in the parent.
1001 Op = ConstantInt::get(CI->getType(), Quotient);
1002 NoSignedWrap = true;
1006 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op)) {
1008 if (BO->getOpcode() == Instruction::Mul) {
1010 NoSignedWrap = BO->hasNoSignedWrap();
1011 if (RequireNoSignedWrap && !NoSignedWrap)
1014 // There are three cases for multiplication: multiplication by exactly
1015 // the scale, multiplication by a constant different to the scale, and
1016 // multiplication by something else.
1017 Value *LHS = BO->getOperand(0);
1018 Value *RHS = BO->getOperand(1);
1020 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
1021 // Multiplication by a constant.
1022 if (CI->getValue() == Scale) {
1023 // Multiplication by exactly the scale, replace the multiplication
1024 // by its left-hand side in the parent.
1029 // Otherwise drill down into the constant.
1030 if (!Op->hasOneUse())
1033 Parent = std::make_pair(BO, 1);
1037 // Multiplication by something else. Drill down into the left-hand side
1038 // since that's where the reassociate pass puts the good stuff.
1039 if (!Op->hasOneUse())
1042 Parent = std::make_pair(BO, 0);
1046 if (logScale > 0 && BO->getOpcode() == Instruction::Shl &&
1047 isa<ConstantInt>(BO->getOperand(1))) {
1048 // Multiplication by a power of 2.
1049 NoSignedWrap = BO->hasNoSignedWrap();
1050 if (RequireNoSignedWrap && !NoSignedWrap)
1053 Value *LHS = BO->getOperand(0);
1054 int32_t Amt = cast<ConstantInt>(BO->getOperand(1))->
1055 getLimitedValue(Scale.getBitWidth());
1058 if (Amt == logScale) {
1059 // Multiplication by exactly the scale, replace the multiplication
1060 // by its left-hand side in the parent.
1064 if (Amt < logScale || !Op->hasOneUse())
1067 // Multiplication by more than the scale. Reduce the multiplying amount
1068 // by the scale in the parent.
1069 Parent = std::make_pair(BO, 1);
1070 Op = ConstantInt::get(BO->getType(), Amt - logScale);
1075 if (!Op->hasOneUse())
1078 if (CastInst *Cast = dyn_cast<CastInst>(Op)) {
1079 if (Cast->getOpcode() == Instruction::SExt) {
1080 // Op is sign-extended from a smaller type, descale in the smaller type.
1081 unsigned SmallSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1082 APInt SmallScale = Scale.trunc(SmallSize);
1083 // Suppose Op = sext X, and we descale X as Y * SmallScale. We want to
1084 // descale Op as (sext Y) * Scale. In order to have
1085 // sext (Y * SmallScale) = (sext Y) * Scale
1086 // some conditions need to hold however: SmallScale must sign-extend to
1087 // Scale and the multiplication Y * SmallScale should not overflow.
1088 if (SmallScale.sext(Scale.getBitWidth()) != Scale)
1089 // SmallScale does not sign-extend to Scale.
1091 assert(SmallScale.exactLogBase2() == logScale);
1092 // Require that Y * SmallScale must not overflow.
1093 RequireNoSignedWrap = true;
1095 // Drill down through the cast.
1096 Parent = std::make_pair(Cast, 0);
1101 if (Cast->getOpcode() == Instruction::Trunc) {
1102 // Op is truncated from a larger type, descale in the larger type.
1103 // Suppose Op = trunc X, and we descale X as Y * sext Scale. Then
1104 // trunc (Y * sext Scale) = (trunc Y) * Scale
1105 // always holds. However (trunc Y) * Scale may overflow even if
1106 // trunc (Y * sext Scale) does not, so nsw flags need to be cleared
1107 // from this point up in the expression (see later).
1108 if (RequireNoSignedWrap)
1111 // Drill down through the cast.
1112 unsigned LargeSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1113 Parent = std::make_pair(Cast, 0);
1114 Scale = Scale.sext(LargeSize);
1115 if (logScale + 1 == (int32_t)Cast->getType()->getPrimitiveSizeInBits())
1117 assert(Scale.exactLogBase2() == logScale);
1122 // Unsupported expression, bail out.
1126 // If Op is zero then Val = Op * Scale.
1127 if (match(Op, m_Zero())) {
1128 NoSignedWrap = true;
1132 // We know that we can successfully descale, so from here on we can safely
1133 // modify the IR. Op holds the descaled version of the deepest term in the
1134 // expression. NoSignedWrap is 'true' if multiplying Op by Scale is known
1138 // The expression only had one term.
1141 // Rewrite the parent using the descaled version of its operand.
1142 assert(Parent.first->hasOneUse() && "Drilled down when more than one use!");
1143 assert(Op != Parent.first->getOperand(Parent.second) &&
1144 "Descaling was a no-op?");
1145 Parent.first->setOperand(Parent.second, Op);
1146 Worklist.Add(Parent.first);
1148 // Now work back up the expression correcting nsw flags. The logic is based
1149 // on the following observation: if X * Y is known not to overflow as a signed
1150 // multiplication, and Y is replaced by a value Z with smaller absolute value,
1151 // then X * Z will not overflow as a signed multiplication either. As we work
1152 // our way up, having NoSignedWrap 'true' means that the descaled value at the
1153 // current level has strictly smaller absolute value than the original.
1154 Instruction *Ancestor = Parent.first;
1156 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Ancestor)) {
1157 // If the multiplication wasn't nsw then we can't say anything about the
1158 // value of the descaled multiplication, and we have to clear nsw flags
1159 // from this point on up.
1160 bool OpNoSignedWrap = BO->hasNoSignedWrap();
1161 NoSignedWrap &= OpNoSignedWrap;
1162 if (NoSignedWrap != OpNoSignedWrap) {
1163 BO->setHasNoSignedWrap(NoSignedWrap);
1164 Worklist.Add(Ancestor);
1166 } else if (Ancestor->getOpcode() == Instruction::Trunc) {
1167 // The fact that the descaled input to the trunc has smaller absolute
1168 // value than the original input doesn't tell us anything useful about
1169 // the absolute values of the truncations.
1170 NoSignedWrap = false;
1172 assert((Ancestor->getOpcode() != Instruction::SExt || NoSignedWrap) &&
1173 "Failed to keep proper track of nsw flags while drilling down?");
1175 if (Ancestor == Val)
1176 // Got to the top, all done!
1179 // Move up one level in the expression.
1180 assert(Ancestor->hasOneUse() && "Drilled down when more than one use!");
1181 Ancestor = Ancestor->user_back();
1185 /// \brief Creates node of binary operation with the same attributes as the
1186 /// specified one but with other operands.
1187 static Value *CreateBinOpAsGiven(BinaryOperator &Inst, Value *LHS, Value *RHS,
1188 InstCombiner::BuilderTy *B) {
1189 Value *BORes = B->CreateBinOp(Inst.getOpcode(), LHS, RHS);
1190 if (BinaryOperator *NewBO = dyn_cast<BinaryOperator>(BORes)) {
1191 if (isa<OverflowingBinaryOperator>(NewBO)) {
1192 NewBO->setHasNoSignedWrap(Inst.hasNoSignedWrap());
1193 NewBO->setHasNoUnsignedWrap(Inst.hasNoUnsignedWrap());
1195 if (isa<PossiblyExactOperator>(NewBO))
1196 NewBO->setIsExact(Inst.isExact());
1201 /// \brief Makes transformation of binary operation specific for vector types.
1202 /// \param Inst Binary operator to transform.
1203 /// \return Pointer to node that must replace the original binary operator, or
1204 /// null pointer if no transformation was made.
1205 Value *InstCombiner::SimplifyVectorOp(BinaryOperator &Inst) {
1206 if (!Inst.getType()->isVectorTy()) return nullptr;
1208 // It may not be safe to reorder shuffles and things like div, urem, etc.
1209 // because we may trap when executing those ops on unknown vector elements.
1211 if (!isSafeToSpeculativelyExecute(&Inst))
1214 unsigned VWidth = cast<VectorType>(Inst.getType())->getNumElements();
1215 Value *LHS = Inst.getOperand(0), *RHS = Inst.getOperand(1);
1216 assert(cast<VectorType>(LHS->getType())->getNumElements() == VWidth);
1217 assert(cast<VectorType>(RHS->getType())->getNumElements() == VWidth);
1219 // If both arguments of binary operation are shuffles, which use the same
1220 // mask and shuffle within a single vector, it is worthwhile to move the
1221 // shuffle after binary operation:
1222 // Op(shuffle(v1, m), shuffle(v2, m)) -> shuffle(Op(v1, v2), m)
1223 if (isa<ShuffleVectorInst>(LHS) && isa<ShuffleVectorInst>(RHS)) {
1224 ShuffleVectorInst *LShuf = cast<ShuffleVectorInst>(LHS);
1225 ShuffleVectorInst *RShuf = cast<ShuffleVectorInst>(RHS);
1226 if (isa<UndefValue>(LShuf->getOperand(1)) &&
1227 isa<UndefValue>(RShuf->getOperand(1)) &&
1228 LShuf->getOperand(0)->getType() == RShuf->getOperand(0)->getType() &&
1229 LShuf->getMask() == RShuf->getMask()) {
1230 Value *NewBO = CreateBinOpAsGiven(Inst, LShuf->getOperand(0),
1231 RShuf->getOperand(0), Builder);
1232 Value *Res = Builder->CreateShuffleVector(NewBO,
1233 UndefValue::get(NewBO->getType()), LShuf->getMask());
1238 // If one argument is a shuffle within one vector, the other is a constant,
1239 // try moving the shuffle after the binary operation.
1240 ShuffleVectorInst *Shuffle = nullptr;
1241 Constant *C1 = nullptr;
1242 if (isa<ShuffleVectorInst>(LHS)) Shuffle = cast<ShuffleVectorInst>(LHS);
1243 if (isa<ShuffleVectorInst>(RHS)) Shuffle = cast<ShuffleVectorInst>(RHS);
1244 if (isa<Constant>(LHS)) C1 = cast<Constant>(LHS);
1245 if (isa<Constant>(RHS)) C1 = cast<Constant>(RHS);
1246 if (Shuffle && C1 &&
1247 (isa<ConstantVector>(C1) || isa<ConstantDataVector>(C1)) &&
1248 isa<UndefValue>(Shuffle->getOperand(1)) &&
1249 Shuffle->getType() == Shuffle->getOperand(0)->getType()) {
1250 SmallVector<int, 16> ShMask = Shuffle->getShuffleMask();
1251 // Find constant C2 that has property:
1252 // shuffle(C2, ShMask) = C1
1253 // If such constant does not exist (example: ShMask=<0,0> and C1=<1,2>)
1254 // reorder is not possible.
1255 SmallVector<Constant*, 16> C2M(VWidth,
1256 UndefValue::get(C1->getType()->getScalarType()));
1257 bool MayChange = true;
1258 for (unsigned I = 0; I < VWidth; ++I) {
1259 if (ShMask[I] >= 0) {
1260 assert(ShMask[I] < (int)VWidth);
1261 if (!isa<UndefValue>(C2M[ShMask[I]])) {
1265 C2M[ShMask[I]] = C1->getAggregateElement(I);
1269 Constant *C2 = ConstantVector::get(C2M);
1270 Value *NewLHS, *NewRHS;
1271 if (isa<Constant>(LHS)) {
1273 NewRHS = Shuffle->getOperand(0);
1275 NewLHS = Shuffle->getOperand(0);
1278 Value *NewBO = CreateBinOpAsGiven(Inst, NewLHS, NewRHS, Builder);
1279 Value *Res = Builder->CreateShuffleVector(NewBO,
1280 UndefValue::get(Inst.getType()), Shuffle->getMask());
1288 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
1289 SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end());
1291 if (Value *V = SimplifyGEPInst(Ops, DL, TLI, DT, AC))
1292 return ReplaceInstUsesWith(GEP, V);
1294 Value *PtrOp = GEP.getOperand(0);
1296 // Eliminate unneeded casts for indices, and replace indices which displace
1297 // by multiples of a zero size type with zero.
1298 bool MadeChange = false;
1299 Type *IntPtrTy = DL.getIntPtrType(GEP.getPointerOperandType());
1301 gep_type_iterator GTI = gep_type_begin(GEP);
1302 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end(); I != E;
1304 // Skip indices into struct types.
1305 SequentialType *SeqTy = dyn_cast<SequentialType>(*GTI);
1309 // If the element type has zero size then any index over it is equivalent
1310 // to an index of zero, so replace it with zero if it is not zero already.
1311 if (SeqTy->getElementType()->isSized() &&
1312 DL.getTypeAllocSize(SeqTy->getElementType()) == 0)
1313 if (!isa<Constant>(*I) || !cast<Constant>(*I)->isNullValue()) {
1314 *I = Constant::getNullValue(IntPtrTy);
1318 Type *IndexTy = (*I)->getType();
1319 if (IndexTy != IntPtrTy) {
1320 // If we are using a wider index than needed for this platform, shrink
1321 // it to what we need. If narrower, sign-extend it to what we need.
1322 // This explicit cast can make subsequent optimizations more obvious.
1323 *I = Builder->CreateIntCast(*I, IntPtrTy, true);
1330 // Check to see if the inputs to the PHI node are getelementptr instructions.
1331 if (PHINode *PN = dyn_cast<PHINode>(PtrOp)) {
1332 GetElementPtrInst *Op1 = dyn_cast<GetElementPtrInst>(PN->getOperand(0));
1338 for (auto I = PN->op_begin()+1, E = PN->op_end(); I !=E; ++I) {
1339 GetElementPtrInst *Op2 = dyn_cast<GetElementPtrInst>(*I);
1340 if (!Op2 || Op1->getNumOperands() != Op2->getNumOperands())
1343 // Keep track of the type as we walk the GEP.
1344 Type *CurTy = Op1->getOperand(0)->getType()->getScalarType();
1346 for (unsigned J = 0, F = Op1->getNumOperands(); J != F; ++J) {
1347 if (Op1->getOperand(J)->getType() != Op2->getOperand(J)->getType())
1350 if (Op1->getOperand(J) != Op2->getOperand(J)) {
1352 // We have not seen any differences yet in the GEPs feeding the
1353 // PHI yet, so we record this one if it is allowed to be a
1356 // The first two arguments can vary for any GEP, the rest have to be
1357 // static for struct slots
1358 if (J > 1 && CurTy->isStructTy())
1363 // The GEP is different by more than one input. While this could be
1364 // extended to support GEPs that vary by more than one variable it
1365 // doesn't make sense since it greatly increases the complexity and
1366 // would result in an R+R+R addressing mode which no backend
1367 // directly supports and would need to be broken into several
1368 // simpler instructions anyway.
1373 // Sink down a layer of the type for the next iteration.
1375 if (CompositeType *CT = dyn_cast<CompositeType>(CurTy)) {
1376 CurTy = CT->getTypeAtIndex(Op1->getOperand(J));
1384 GetElementPtrInst *NewGEP = cast<GetElementPtrInst>(Op1->clone());
1387 // All the GEPs feeding the PHI are identical. Clone one down into our
1388 // BB so that it can be merged with the current GEP.
1389 GEP.getParent()->getInstList().insert(
1390 GEP.getParent()->getFirstInsertionPt(), NewGEP);
1392 // All the GEPs feeding the PHI differ at a single offset. Clone a GEP
1393 // into the current block so it can be merged, and create a new PHI to
1395 Instruction *InsertPt = Builder->GetInsertPoint();
1396 Builder->SetInsertPoint(PN);
1397 PHINode *NewPN = Builder->CreatePHI(Op1->getOperand(DI)->getType(),
1398 PN->getNumOperands());
1399 Builder->SetInsertPoint(InsertPt);
1401 for (auto &I : PN->operands())
1402 NewPN->addIncoming(cast<GEPOperator>(I)->getOperand(DI),
1403 PN->getIncomingBlock(I));
1405 NewGEP->setOperand(DI, NewPN);
1406 GEP.getParent()->getInstList().insert(
1407 GEP.getParent()->getFirstInsertionPt(), NewGEP);
1408 NewGEP->setOperand(DI, NewPN);
1411 GEP.setOperand(0, NewGEP);
1415 // Combine Indices - If the source pointer to this getelementptr instruction
1416 // is a getelementptr instruction, combine the indices of the two
1417 // getelementptr instructions into a single instruction.
1419 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) {
1420 if (!shouldMergeGEPs(*cast<GEPOperator>(&GEP), *Src))
1423 // Note that if our source is a gep chain itself then we wait for that
1424 // chain to be resolved before we perform this transformation. This
1425 // avoids us creating a TON of code in some cases.
1426 if (GEPOperator *SrcGEP =
1427 dyn_cast<GEPOperator>(Src->getOperand(0)))
1428 if (SrcGEP->getNumOperands() == 2 && shouldMergeGEPs(*Src, *SrcGEP))
1429 return nullptr; // Wait until our source is folded to completion.
1431 SmallVector<Value*, 8> Indices;
1433 // Find out whether the last index in the source GEP is a sequential idx.
1434 bool EndsWithSequential = false;
1435 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src);
1437 EndsWithSequential = !(*I)->isStructTy();
1439 // Can we combine the two pointer arithmetics offsets?
1440 if (EndsWithSequential) {
1441 // Replace: gep (gep %P, long B), long A, ...
1442 // With: T = long A+B; gep %P, T, ...
1445 Value *SO1 = Src->getOperand(Src->getNumOperands()-1);
1446 Value *GO1 = GEP.getOperand(1);
1447 if (SO1 == Constant::getNullValue(SO1->getType())) {
1449 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
1452 // If they aren't the same type, then the input hasn't been processed
1453 // by the loop above yet (which canonicalizes sequential index types to
1454 // intptr_t). Just avoid transforming this until the input has been
1456 if (SO1->getType() != GO1->getType())
1458 Sum = Builder->CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
1461 // Update the GEP in place if possible.
1462 if (Src->getNumOperands() == 2) {
1463 GEP.setOperand(0, Src->getOperand(0));
1464 GEP.setOperand(1, Sum);
1467 Indices.append(Src->op_begin()+1, Src->op_end()-1);
1468 Indices.push_back(Sum);
1469 Indices.append(GEP.op_begin()+2, GEP.op_end());
1470 } else if (isa<Constant>(*GEP.idx_begin()) &&
1471 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
1472 Src->getNumOperands() != 1) {
1473 // Otherwise we can do the fold if the first index of the GEP is a zero
1474 Indices.append(Src->op_begin()+1, Src->op_end());
1475 Indices.append(GEP.idx_begin()+1, GEP.idx_end());
1478 if (!Indices.empty())
1479 return (GEP.isInBounds() && Src->isInBounds()) ?
1480 GetElementPtrInst::CreateInBounds(Src->getOperand(0), Indices,
1482 GetElementPtrInst::Create(Src->getOperand(0), Indices, GEP.getName());
1485 if (GEP.getNumIndices() == 1) {
1486 unsigned AS = GEP.getPointerAddressSpace();
1487 if (GEP.getOperand(1)->getType()->getScalarSizeInBits() ==
1488 DL.getPointerSizeInBits(AS)) {
1489 Type *PtrTy = GEP.getPointerOperandType();
1490 Type *Ty = PtrTy->getPointerElementType();
1491 uint64_t TyAllocSize = DL.getTypeAllocSize(Ty);
1493 bool Matched = false;
1496 if (TyAllocSize == 1) {
1497 V = GEP.getOperand(1);
1499 } else if (match(GEP.getOperand(1),
1500 m_AShr(m_Value(V), m_ConstantInt(C)))) {
1501 if (TyAllocSize == 1ULL << C)
1503 } else if (match(GEP.getOperand(1),
1504 m_SDiv(m_Value(V), m_ConstantInt(C)))) {
1505 if (TyAllocSize == C)
1510 // Canonicalize (gep i8* X, -(ptrtoint Y))
1511 // to (inttoptr (sub (ptrtoint X), (ptrtoint Y)))
1512 // The GEP pattern is emitted by the SCEV expander for certain kinds of
1513 // pointer arithmetic.
1514 if (match(V, m_Neg(m_PtrToInt(m_Value())))) {
1515 Operator *Index = cast<Operator>(V);
1516 Value *PtrToInt = Builder->CreatePtrToInt(PtrOp, Index->getType());
1517 Value *NewSub = Builder->CreateSub(PtrToInt, Index->getOperand(1));
1518 return CastInst::Create(Instruction::IntToPtr, NewSub, GEP.getType());
1520 // Canonicalize (gep i8* X, (ptrtoint Y)-(ptrtoint X))
1523 if (match(V, m_Sub(m_PtrToInt(m_Value(Y)),
1524 m_PtrToInt(m_Specific(GEP.getOperand(0)))))) {
1525 return CastInst::CreatePointerBitCastOrAddrSpaceCast(Y,
1532 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0).
1533 Value *StrippedPtr = PtrOp->stripPointerCasts();
1534 PointerType *StrippedPtrTy = dyn_cast<PointerType>(StrippedPtr->getType());
1536 // We do not handle pointer-vector geps here.
1540 if (StrippedPtr != PtrOp) {
1541 bool HasZeroPointerIndex = false;
1542 if (ConstantInt *C = dyn_cast<ConstantInt>(GEP.getOperand(1)))
1543 HasZeroPointerIndex = C->isZero();
1545 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
1546 // into : GEP [10 x i8]* X, i32 0, ...
1548 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
1549 // into : GEP i8* X, ...
1551 // This occurs when the program declares an array extern like "int X[];"
1552 if (HasZeroPointerIndex) {
1553 PointerType *CPTy = cast<PointerType>(PtrOp->getType());
1554 if (ArrayType *CATy =
1555 dyn_cast<ArrayType>(CPTy->getElementType())) {
1556 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
1557 if (CATy->getElementType() == StrippedPtrTy->getElementType()) {
1558 // -> GEP i8* X, ...
1559 SmallVector<Value*, 8> Idx(GEP.idx_begin()+1, GEP.idx_end());
1560 GetElementPtrInst *Res =
1561 GetElementPtrInst::Create(StrippedPtr, Idx, GEP.getName());
1562 Res->setIsInBounds(GEP.isInBounds());
1563 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace())
1565 // Insert Res, and create an addrspacecast.
1567 // GEP (addrspacecast i8 addrspace(1)* X to [0 x i8]*), i32 0, ...
1569 // %0 = GEP i8 addrspace(1)* X, ...
1570 // addrspacecast i8 addrspace(1)* %0 to i8*
1571 return new AddrSpaceCastInst(Builder->Insert(Res), GEP.getType());
1574 if (ArrayType *XATy =
1575 dyn_cast<ArrayType>(StrippedPtrTy->getElementType())){
1576 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
1577 if (CATy->getElementType() == XATy->getElementType()) {
1578 // -> GEP [10 x i8]* X, i32 0, ...
1579 // At this point, we know that the cast source type is a pointer
1580 // to an array of the same type as the destination pointer
1581 // array. Because the array type is never stepped over (there
1582 // is a leading zero) we can fold the cast into this GEP.
1583 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace()) {
1584 GEP.setOperand(0, StrippedPtr);
1587 // Cannot replace the base pointer directly because StrippedPtr's
1588 // address space is different. Instead, create a new GEP followed by
1589 // an addrspacecast.
1591 // GEP (addrspacecast [10 x i8] addrspace(1)* X to [0 x i8]*),
1594 // %0 = GEP [10 x i8] addrspace(1)* X, ...
1595 // addrspacecast i8 addrspace(1)* %0 to i8*
1596 SmallVector<Value*, 8> Idx(GEP.idx_begin(), GEP.idx_end());
1597 Value *NewGEP = GEP.isInBounds() ?
1598 Builder->CreateInBoundsGEP(StrippedPtr, Idx, GEP.getName()) :
1599 Builder->CreateGEP(StrippedPtr, Idx, GEP.getName());
1600 return new AddrSpaceCastInst(NewGEP, GEP.getType());
1604 } else if (GEP.getNumOperands() == 2) {
1605 // Transform things like:
1606 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
1607 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
1608 Type *SrcElTy = StrippedPtrTy->getElementType();
1609 Type *ResElTy = PtrOp->getType()->getPointerElementType();
1610 if (SrcElTy->isArrayTy() &&
1611 DL.getTypeAllocSize(SrcElTy->getArrayElementType()) ==
1612 DL.getTypeAllocSize(ResElTy)) {
1613 Type *IdxType = DL.getIntPtrType(GEP.getType());
1614 Value *Idx[2] = { Constant::getNullValue(IdxType), GEP.getOperand(1) };
1615 Value *NewGEP = GEP.isInBounds() ?
1616 Builder->CreateInBoundsGEP(StrippedPtr, Idx, GEP.getName()) :
1617 Builder->CreateGEP(StrippedPtr, Idx, GEP.getName());
1619 // V and GEP are both pointer types --> BitCast
1620 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
1624 // Transform things like:
1625 // %V = mul i64 %N, 4
1626 // %t = getelementptr i8* bitcast (i32* %arr to i8*), i32 %V
1627 // into: %t1 = getelementptr i32* %arr, i32 %N; bitcast
1628 if (ResElTy->isSized() && SrcElTy->isSized()) {
1629 // Check that changing the type amounts to dividing the index by a scale
1631 uint64_t ResSize = DL.getTypeAllocSize(ResElTy);
1632 uint64_t SrcSize = DL.getTypeAllocSize(SrcElTy);
1633 if (ResSize && SrcSize % ResSize == 0) {
1634 Value *Idx = GEP.getOperand(1);
1635 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
1636 uint64_t Scale = SrcSize / ResSize;
1638 // Earlier transforms ensure that the index has type IntPtrType, which
1639 // considerably simplifies the logic by eliminating implicit casts.
1640 assert(Idx->getType() == DL.getIntPtrType(GEP.getType()) &&
1641 "Index not cast to pointer width?");
1644 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
1645 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
1646 // If the multiplication NewIdx * Scale may overflow then the new
1647 // GEP may not be "inbounds".
1648 Value *NewGEP = GEP.isInBounds() && NSW ?
1649 Builder->CreateInBoundsGEP(StrippedPtr, NewIdx, GEP.getName()) :
1650 Builder->CreateGEP(StrippedPtr, NewIdx, GEP.getName());
1652 // The NewGEP must be pointer typed, so must the old one -> BitCast
1653 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
1659 // Similarly, transform things like:
1660 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
1661 // (where tmp = 8*tmp2) into:
1662 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
1663 if (ResElTy->isSized() && SrcElTy->isSized() && SrcElTy->isArrayTy()) {
1664 // Check that changing to the array element type amounts to dividing the
1665 // index by a scale factor.
1666 uint64_t ResSize = DL.getTypeAllocSize(ResElTy);
1667 uint64_t ArrayEltSize =
1668 DL.getTypeAllocSize(SrcElTy->getArrayElementType());
1669 if (ResSize && ArrayEltSize % ResSize == 0) {
1670 Value *Idx = GEP.getOperand(1);
1671 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
1672 uint64_t Scale = ArrayEltSize / ResSize;
1674 // Earlier transforms ensure that the index has type IntPtrType, which
1675 // considerably simplifies the logic by eliminating implicit casts.
1676 assert(Idx->getType() == DL.getIntPtrType(GEP.getType()) &&
1677 "Index not cast to pointer width?");
1680 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
1681 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
1682 // If the multiplication NewIdx * Scale may overflow then the new
1683 // GEP may not be "inbounds".
1685 Constant::getNullValue(DL.getIntPtrType(GEP.getType())),
1688 Value *NewGEP = GEP.isInBounds() && NSW ?
1689 Builder->CreateInBoundsGEP(StrippedPtr, Off, GEP.getName()) :
1690 Builder->CreateGEP(StrippedPtr, Off, GEP.getName());
1691 // The NewGEP must be pointer typed, so must the old one -> BitCast
1692 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
1700 // addrspacecast between types is canonicalized as a bitcast, then an
1701 // addrspacecast. To take advantage of the below bitcast + struct GEP, look
1702 // through the addrspacecast.
1703 if (AddrSpaceCastInst *ASC = dyn_cast<AddrSpaceCastInst>(PtrOp)) {
1704 // X = bitcast A addrspace(1)* to B addrspace(1)*
1705 // Y = addrspacecast A addrspace(1)* to B addrspace(2)*
1706 // Z = gep Y, <...constant indices...>
1707 // Into an addrspacecasted GEP of the struct.
1708 if (BitCastInst *BC = dyn_cast<BitCastInst>(ASC->getOperand(0)))
1712 /// See if we can simplify:
1713 /// X = bitcast A* to B*
1714 /// Y = gep X, <...constant indices...>
1715 /// into a gep of the original struct. This is important for SROA and alias
1716 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
1717 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
1718 Value *Operand = BCI->getOperand(0);
1719 PointerType *OpType = cast<PointerType>(Operand->getType());
1720 unsigned OffsetBits = DL.getPointerTypeSizeInBits(GEP.getType());
1721 APInt Offset(OffsetBits, 0);
1722 if (!isa<BitCastInst>(Operand) &&
1723 GEP.accumulateConstantOffset(DL, Offset)) {
1725 // If this GEP instruction doesn't move the pointer, just replace the GEP
1726 // with a bitcast of the real input to the dest type.
1728 // If the bitcast is of an allocation, and the allocation will be
1729 // converted to match the type of the cast, don't touch this.
1730 if (isa<AllocaInst>(Operand) || isAllocationFn(Operand, TLI)) {
1731 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
1732 if (Instruction *I = visitBitCast(*BCI)) {
1735 BCI->getParent()->getInstList().insert(BCI, I);
1736 ReplaceInstUsesWith(*BCI, I);
1742 if (Operand->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
1743 return new AddrSpaceCastInst(Operand, GEP.getType());
1744 return new BitCastInst(Operand, GEP.getType());
1747 // Otherwise, if the offset is non-zero, we need to find out if there is a
1748 // field at Offset in 'A's type. If so, we can pull the cast through the
1750 SmallVector<Value*, 8> NewIndices;
1751 if (FindElementAtOffset(OpType, Offset.getSExtValue(), NewIndices)) {
1752 Value *NGEP = GEP.isInBounds() ?
1753 Builder->CreateInBoundsGEP(Operand, NewIndices) :
1754 Builder->CreateGEP(Operand, NewIndices);
1756 if (NGEP->getType() == GEP.getType())
1757 return ReplaceInstUsesWith(GEP, NGEP);
1758 NGEP->takeName(&GEP);
1760 if (NGEP->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
1761 return new AddrSpaceCastInst(NGEP, GEP.getType());
1762 return new BitCastInst(NGEP, GEP.getType());
1771 isAllocSiteRemovable(Instruction *AI, SmallVectorImpl<WeakVH> &Users,
1772 const TargetLibraryInfo *TLI) {
1773 SmallVector<Instruction*, 4> Worklist;
1774 Worklist.push_back(AI);
1777 Instruction *PI = Worklist.pop_back_val();
1778 for (User *U : PI->users()) {
1779 Instruction *I = cast<Instruction>(U);
1780 switch (I->getOpcode()) {
1782 // Give up the moment we see something we can't handle.
1785 case Instruction::BitCast:
1786 case Instruction::GetElementPtr:
1788 Worklist.push_back(I);
1791 case Instruction::ICmp: {
1792 ICmpInst *ICI = cast<ICmpInst>(I);
1793 // We can fold eq/ne comparisons with null to false/true, respectively.
1794 if (!ICI->isEquality() || !isa<ConstantPointerNull>(ICI->getOperand(1)))
1800 case Instruction::Call:
1801 // Ignore no-op and store intrinsics.
1802 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1803 switch (II->getIntrinsicID()) {
1807 case Intrinsic::memmove:
1808 case Intrinsic::memcpy:
1809 case Intrinsic::memset: {
1810 MemIntrinsic *MI = cast<MemIntrinsic>(II);
1811 if (MI->isVolatile() || MI->getRawDest() != PI)
1815 case Intrinsic::dbg_declare:
1816 case Intrinsic::dbg_value:
1817 case Intrinsic::invariant_start:
1818 case Intrinsic::invariant_end:
1819 case Intrinsic::lifetime_start:
1820 case Intrinsic::lifetime_end:
1821 case Intrinsic::objectsize:
1827 if (isFreeCall(I, TLI)) {
1833 case Instruction::Store: {
1834 StoreInst *SI = cast<StoreInst>(I);
1835 if (SI->isVolatile() || SI->getPointerOperand() != PI)
1841 llvm_unreachable("missing a return?");
1843 } while (!Worklist.empty());
1847 Instruction *InstCombiner::visitAllocSite(Instruction &MI) {
1848 // If we have a malloc call which is only used in any amount of comparisons
1849 // to null and free calls, delete the calls and replace the comparisons with
1850 // true or false as appropriate.
1851 SmallVector<WeakVH, 64> Users;
1852 if (isAllocSiteRemovable(&MI, Users, TLI)) {
1853 for (unsigned i = 0, e = Users.size(); i != e; ++i) {
1854 Instruction *I = cast_or_null<Instruction>(&*Users[i]);
1857 if (ICmpInst *C = dyn_cast<ICmpInst>(I)) {
1858 ReplaceInstUsesWith(*C,
1859 ConstantInt::get(Type::getInt1Ty(C->getContext()),
1860 C->isFalseWhenEqual()));
1861 } else if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) {
1862 ReplaceInstUsesWith(*I, UndefValue::get(I->getType()));
1863 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1864 if (II->getIntrinsicID() == Intrinsic::objectsize) {
1865 ConstantInt *CI = cast<ConstantInt>(II->getArgOperand(1));
1866 uint64_t DontKnow = CI->isZero() ? -1ULL : 0;
1867 ReplaceInstUsesWith(*I, ConstantInt::get(I->getType(), DontKnow));
1870 EraseInstFromFunction(*I);
1873 if (InvokeInst *II = dyn_cast<InvokeInst>(&MI)) {
1874 // Replace invoke with a NOP intrinsic to maintain the original CFG
1875 Module *M = II->getParent()->getParent()->getParent();
1876 Function *F = Intrinsic::getDeclaration(M, Intrinsic::donothing);
1877 InvokeInst::Create(F, II->getNormalDest(), II->getUnwindDest(),
1878 None, "", II->getParent());
1880 return EraseInstFromFunction(MI);
1885 /// \brief Move the call to free before a NULL test.
1887 /// Check if this free is accessed after its argument has been test
1888 /// against NULL (property 0).
1889 /// If yes, it is legal to move this call in its predecessor block.
1891 /// The move is performed only if the block containing the call to free
1892 /// will be removed, i.e.:
1893 /// 1. it has only one predecessor P, and P has two successors
1894 /// 2. it contains the call and an unconditional branch
1895 /// 3. its successor is the same as its predecessor's successor
1897 /// The profitability is out-of concern here and this function should
1898 /// be called only if the caller knows this transformation would be
1899 /// profitable (e.g., for code size).
1900 static Instruction *
1901 tryToMoveFreeBeforeNullTest(CallInst &FI) {
1902 Value *Op = FI.getArgOperand(0);
1903 BasicBlock *FreeInstrBB = FI.getParent();
1904 BasicBlock *PredBB = FreeInstrBB->getSinglePredecessor();
1906 // Validate part of constraint #1: Only one predecessor
1907 // FIXME: We can extend the number of predecessor, but in that case, we
1908 // would duplicate the call to free in each predecessor and it may
1909 // not be profitable even for code size.
1913 // Validate constraint #2: Does this block contains only the call to
1914 // free and an unconditional branch?
1915 // FIXME: We could check if we can speculate everything in the
1916 // predecessor block
1917 if (FreeInstrBB->size() != 2)
1920 if (!match(FreeInstrBB->getTerminator(), m_UnconditionalBr(SuccBB)))
1923 // Validate the rest of constraint #1 by matching on the pred branch.
1924 TerminatorInst *TI = PredBB->getTerminator();
1925 BasicBlock *TrueBB, *FalseBB;
1926 ICmpInst::Predicate Pred;
1927 if (!match(TI, m_Br(m_ICmp(Pred, m_Specific(Op), m_Zero()), TrueBB, FalseBB)))
1929 if (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE)
1932 // Validate constraint #3: Ensure the null case just falls through.
1933 if (SuccBB != (Pred == ICmpInst::ICMP_EQ ? TrueBB : FalseBB))
1935 assert(FreeInstrBB == (Pred == ICmpInst::ICMP_EQ ? FalseBB : TrueBB) &&
1936 "Broken CFG: missing edge from predecessor to successor");
1943 Instruction *InstCombiner::visitFree(CallInst &FI) {
1944 Value *Op = FI.getArgOperand(0);
1946 // free undef -> unreachable.
1947 if (isa<UndefValue>(Op)) {
1948 // Insert a new store to null because we cannot modify the CFG here.
1949 Builder->CreateStore(ConstantInt::getTrue(FI.getContext()),
1950 UndefValue::get(Type::getInt1PtrTy(FI.getContext())));
1951 return EraseInstFromFunction(FI);
1954 // If we have 'free null' delete the instruction. This can happen in stl code
1955 // when lots of inlining happens.
1956 if (isa<ConstantPointerNull>(Op))
1957 return EraseInstFromFunction(FI);
1959 // If we optimize for code size, try to move the call to free before the null
1960 // test so that simplify cfg can remove the empty block and dead code
1961 // elimination the branch. I.e., helps to turn something like:
1962 // if (foo) free(foo);
1966 if (Instruction *I = tryToMoveFreeBeforeNullTest(FI))
1972 Instruction *InstCombiner::visitReturnInst(ReturnInst &RI) {
1973 if (RI.getNumOperands() == 0) // ret void
1976 Value *ResultOp = RI.getOperand(0);
1977 Type *VTy = ResultOp->getType();
1978 if (!VTy->isIntegerTy())
1981 // There might be assume intrinsics dominating this return that completely
1982 // determine the value. If so, constant fold it.
1983 unsigned BitWidth = VTy->getPrimitiveSizeInBits();
1984 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
1985 computeKnownBits(ResultOp, KnownZero, KnownOne, 0, &RI);
1986 if ((KnownZero|KnownOne).isAllOnesValue())
1987 RI.setOperand(0, Constant::getIntegerValue(VTy, KnownOne));
1992 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
1993 // Change br (not X), label True, label False to: br X, label False, True
1995 BasicBlock *TrueDest;
1996 BasicBlock *FalseDest;
1997 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
1998 !isa<Constant>(X)) {
1999 // Swap Destinations and condition...
2001 BI.swapSuccessors();
2005 // Canonicalize fcmp_one -> fcmp_oeq
2006 FCmpInst::Predicate FPred; Value *Y;
2007 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
2008 TrueDest, FalseDest)) &&
2009 BI.getCondition()->hasOneUse())
2010 if (FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
2011 FPred == FCmpInst::FCMP_OGE) {
2012 FCmpInst *Cond = cast<FCmpInst>(BI.getCondition());
2013 Cond->setPredicate(FCmpInst::getInversePredicate(FPred));
2015 // Swap Destinations and condition.
2016 BI.swapSuccessors();
2021 // Canonicalize icmp_ne -> icmp_eq
2022 ICmpInst::Predicate IPred;
2023 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
2024 TrueDest, FalseDest)) &&
2025 BI.getCondition()->hasOneUse())
2026 if (IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
2027 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
2028 IPred == ICmpInst::ICMP_SGE) {
2029 ICmpInst *Cond = cast<ICmpInst>(BI.getCondition());
2030 Cond->setPredicate(ICmpInst::getInversePredicate(IPred));
2031 // Swap Destinations and condition.
2032 BI.swapSuccessors();
2040 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
2041 Value *Cond = SI.getCondition();
2042 unsigned BitWidth = cast<IntegerType>(Cond->getType())->getBitWidth();
2043 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2044 computeKnownBits(Cond, KnownZero, KnownOne, 0, &SI);
2045 unsigned LeadingKnownZeros = KnownZero.countLeadingOnes();
2046 unsigned LeadingKnownOnes = KnownOne.countLeadingOnes();
2048 // Compute the number of leading bits we can ignore.
2049 for (auto &C : SI.cases()) {
2050 LeadingKnownZeros = std::min(
2051 LeadingKnownZeros, C.getCaseValue()->getValue().countLeadingZeros());
2052 LeadingKnownOnes = std::min(
2053 LeadingKnownOnes, C.getCaseValue()->getValue().countLeadingOnes());
2056 unsigned NewWidth = BitWidth - std::max(LeadingKnownZeros, LeadingKnownOnes);
2058 // Truncate the condition operand if the new type is equal to or larger than
2059 // the largest legal integer type. We need to be conservative here since
2060 // x86 generates redundant zero-extenstion instructions if the operand is
2061 // truncated to i8 or i16.
2062 bool TruncCond = false;
2063 if (NewWidth > 0 && BitWidth > NewWidth &&
2064 NewWidth >= DL.getLargestLegalIntTypeSize()) {
2066 IntegerType *Ty = IntegerType::get(SI.getContext(), NewWidth);
2067 Builder->SetInsertPoint(&SI);
2068 Value *NewCond = Builder->CreateTrunc(SI.getCondition(), Ty, "trunc");
2069 SI.setCondition(NewCond);
2071 for (auto &C : SI.cases())
2072 static_cast<SwitchInst::CaseIt *>(&C)->setValue(ConstantInt::get(
2073 SI.getContext(), C.getCaseValue()->getValue().trunc(NewWidth)));
2076 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
2077 if (I->getOpcode() == Instruction::Add)
2078 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2079 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
2080 // Skip the first item since that's the default case.
2081 for (SwitchInst::CaseIt i = SI.case_begin(), e = SI.case_end();
2083 ConstantInt* CaseVal = i.getCaseValue();
2084 Constant *LHS = CaseVal;
2086 LHS = LeadingKnownZeros
2087 ? ConstantExpr::getZExt(CaseVal, Cond->getType())
2088 : ConstantExpr::getSExt(CaseVal, Cond->getType());
2089 Constant* NewCaseVal = ConstantExpr::getSub(LHS, AddRHS);
2090 assert(isa<ConstantInt>(NewCaseVal) &&
2091 "Result of expression should be constant");
2092 i.setValue(cast<ConstantInt>(NewCaseVal));
2094 SI.setCondition(I->getOperand(0));
2100 return TruncCond ? &SI : nullptr;
2103 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
2104 Value *Agg = EV.getAggregateOperand();
2106 if (!EV.hasIndices())
2107 return ReplaceInstUsesWith(EV, Agg);
2109 if (Constant *C = dyn_cast<Constant>(Agg)) {
2110 if (Constant *C2 = C->getAggregateElement(*EV.idx_begin())) {
2111 if (EV.getNumIndices() == 0)
2112 return ReplaceInstUsesWith(EV, C2);
2113 // Extract the remaining indices out of the constant indexed by the
2115 return ExtractValueInst::Create(C2, EV.getIndices().slice(1));
2117 return nullptr; // Can't handle other constants
2120 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
2121 // We're extracting from an insertvalue instruction, compare the indices
2122 const unsigned *exti, *exte, *insi, *inse;
2123 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
2124 exte = EV.idx_end(), inse = IV->idx_end();
2125 exti != exte && insi != inse;
2128 // The insert and extract both reference distinctly different elements.
2129 // This means the extract is not influenced by the insert, and we can
2130 // replace the aggregate operand of the extract with the aggregate
2131 // operand of the insert. i.e., replace
2132 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
2133 // %E = extractvalue { i32, { i32 } } %I, 0
2135 // %E = extractvalue { i32, { i32 } } %A, 0
2136 return ExtractValueInst::Create(IV->getAggregateOperand(),
2139 if (exti == exte && insi == inse)
2140 // Both iterators are at the end: Index lists are identical. Replace
2141 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
2142 // %C = extractvalue { i32, { i32 } } %B, 1, 0
2144 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
2146 // The extract list is a prefix of the insert list. i.e. replace
2147 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
2148 // %E = extractvalue { i32, { i32 } } %I, 1
2150 // %X = extractvalue { i32, { i32 } } %A, 1
2151 // %E = insertvalue { i32 } %X, i32 42, 0
2152 // by switching the order of the insert and extract (though the
2153 // insertvalue should be left in, since it may have other uses).
2154 Value *NewEV = Builder->CreateExtractValue(IV->getAggregateOperand(),
2156 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
2157 makeArrayRef(insi, inse));
2160 // The insert list is a prefix of the extract list
2161 // We can simply remove the common indices from the extract and make it
2162 // operate on the inserted value instead of the insertvalue result.
2164 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
2165 // %E = extractvalue { i32, { i32 } } %I, 1, 0
2167 // %E extractvalue { i32 } { i32 42 }, 0
2168 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
2169 makeArrayRef(exti, exte));
2171 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) {
2172 // We're extracting from an intrinsic, see if we're the only user, which
2173 // allows us to simplify multiple result intrinsics to simpler things that
2174 // just get one value.
2175 if (II->hasOneUse()) {
2176 // Check if we're grabbing the overflow bit or the result of a 'with
2177 // overflow' intrinsic. If it's the latter we can remove the intrinsic
2178 // and replace it with a traditional binary instruction.
2179 switch (II->getIntrinsicID()) {
2180 case Intrinsic::uadd_with_overflow:
2181 case Intrinsic::sadd_with_overflow:
2182 if (*EV.idx_begin() == 0) { // Normal result.
2183 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2184 ReplaceInstUsesWith(*II, UndefValue::get(II->getType()));
2185 EraseInstFromFunction(*II);
2186 return BinaryOperator::CreateAdd(LHS, RHS);
2189 // If the normal result of the add is dead, and the RHS is a constant,
2190 // we can transform this into a range comparison.
2191 // overflow = uadd a, -4 --> overflow = icmp ugt a, 3
2192 if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow)
2193 if (ConstantInt *CI = dyn_cast<ConstantInt>(II->getArgOperand(1)))
2194 return new ICmpInst(ICmpInst::ICMP_UGT, II->getArgOperand(0),
2195 ConstantExpr::getNot(CI));
2197 case Intrinsic::usub_with_overflow:
2198 case Intrinsic::ssub_with_overflow:
2199 if (*EV.idx_begin() == 0) { // Normal result.
2200 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2201 ReplaceInstUsesWith(*II, UndefValue::get(II->getType()));
2202 EraseInstFromFunction(*II);
2203 return BinaryOperator::CreateSub(LHS, RHS);
2206 case Intrinsic::umul_with_overflow:
2207 case Intrinsic::smul_with_overflow:
2208 if (*EV.idx_begin() == 0) { // Normal result.
2209 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2210 ReplaceInstUsesWith(*II, UndefValue::get(II->getType()));
2211 EraseInstFromFunction(*II);
2212 return BinaryOperator::CreateMul(LHS, RHS);
2220 if (LoadInst *L = dyn_cast<LoadInst>(Agg))
2221 // If the (non-volatile) load only has one use, we can rewrite this to a
2222 // load from a GEP. This reduces the size of the load.
2223 // FIXME: If a load is used only by extractvalue instructions then this
2224 // could be done regardless of having multiple uses.
2225 if (L->isSimple() && L->hasOneUse()) {
2226 // extractvalue has integer indices, getelementptr has Value*s. Convert.
2227 SmallVector<Value*, 4> Indices;
2228 // Prefix an i32 0 since we need the first element.
2229 Indices.push_back(Builder->getInt32(0));
2230 for (ExtractValueInst::idx_iterator I = EV.idx_begin(), E = EV.idx_end();
2232 Indices.push_back(Builder->getInt32(*I));
2234 // We need to insert these at the location of the old load, not at that of
2235 // the extractvalue.
2236 Builder->SetInsertPoint(L->getParent(), L);
2237 Value *GEP = Builder->CreateInBoundsGEP(L->getPointerOperand(), Indices);
2238 // Returning the load directly will cause the main loop to insert it in
2239 // the wrong spot, so use ReplaceInstUsesWith().
2240 return ReplaceInstUsesWith(EV, Builder->CreateLoad(GEP));
2242 // We could simplify extracts from other values. Note that nested extracts may
2243 // already be simplified implicitly by the above: extract (extract (insert) )
2244 // will be translated into extract ( insert ( extract ) ) first and then just
2245 // the value inserted, if appropriate. Similarly for extracts from single-use
2246 // loads: extract (extract (load)) will be translated to extract (load (gep))
2247 // and if again single-use then via load (gep (gep)) to load (gep).
2248 // However, double extracts from e.g. function arguments or return values
2249 // aren't handled yet.
2253 /// isCatchAll - Return 'true' if the given typeinfo will match anything.
2254 static bool isCatchAll(EHPersonality Personality, Constant *TypeInfo) {
2255 switch (Personality) {
2256 case EHPersonality::GNU_C:
2257 // The GCC C EH personality only exists to support cleanups, so it's not
2258 // clear what the semantics of catch clauses are.
2260 case EHPersonality::Unknown:
2262 case EHPersonality::GNU_Ada:
2263 // While __gnat_all_others_value will match any Ada exception, it doesn't
2264 // match foreign exceptions (or didn't, before gcc-4.7).
2266 case EHPersonality::GNU_CXX:
2267 case EHPersonality::GNU_ObjC:
2268 case EHPersonality::MSVC_X86SEH:
2269 case EHPersonality::MSVC_Win64SEH:
2270 case EHPersonality::MSVC_CXX:
2271 return TypeInfo->isNullValue();
2273 llvm_unreachable("invalid enum");
2276 static bool shorter_filter(const Value *LHS, const Value *RHS) {
2278 cast<ArrayType>(LHS->getType())->getNumElements()
2280 cast<ArrayType>(RHS->getType())->getNumElements();
2283 Instruction *InstCombiner::visitLandingPadInst(LandingPadInst &LI) {
2284 // The logic here should be correct for any real-world personality function.
2285 // However if that turns out not to be true, the offending logic can always
2286 // be conditioned on the personality function, like the catch-all logic is.
2287 EHPersonality Personality = classifyEHPersonality(LI.getPersonalityFn());
2289 // Simplify the list of clauses, eg by removing repeated catch clauses
2290 // (these are often created by inlining).
2291 bool MakeNewInstruction = false; // If true, recreate using the following:
2292 SmallVector<Constant *, 16> NewClauses; // - Clauses for the new instruction;
2293 bool CleanupFlag = LI.isCleanup(); // - The new instruction is a cleanup.
2295 SmallPtrSet<Value *, 16> AlreadyCaught; // Typeinfos known caught already.
2296 for (unsigned i = 0, e = LI.getNumClauses(); i != e; ++i) {
2297 bool isLastClause = i + 1 == e;
2298 if (LI.isCatch(i)) {
2300 Constant *CatchClause = LI.getClause(i);
2301 Constant *TypeInfo = CatchClause->stripPointerCasts();
2303 // If we already saw this clause, there is no point in having a second
2305 if (AlreadyCaught.insert(TypeInfo).second) {
2306 // This catch clause was not already seen.
2307 NewClauses.push_back(CatchClause);
2309 // Repeated catch clause - drop the redundant copy.
2310 MakeNewInstruction = true;
2313 // If this is a catch-all then there is no point in keeping any following
2314 // clauses or marking the landingpad as having a cleanup.
2315 if (isCatchAll(Personality, TypeInfo)) {
2317 MakeNewInstruction = true;
2318 CleanupFlag = false;
2322 // A filter clause. If any of the filter elements were already caught
2323 // then they can be dropped from the filter. It is tempting to try to
2324 // exploit the filter further by saying that any typeinfo that does not
2325 // occur in the filter can't be caught later (and thus can be dropped).
2326 // However this would be wrong, since typeinfos can match without being
2327 // equal (for example if one represents a C++ class, and the other some
2328 // class derived from it).
2329 assert(LI.isFilter(i) && "Unsupported landingpad clause!");
2330 Constant *FilterClause = LI.getClause(i);
2331 ArrayType *FilterType = cast<ArrayType>(FilterClause->getType());
2332 unsigned NumTypeInfos = FilterType->getNumElements();
2334 // An empty filter catches everything, so there is no point in keeping any
2335 // following clauses or marking the landingpad as having a cleanup. By
2336 // dealing with this case here the following code is made a bit simpler.
2337 if (!NumTypeInfos) {
2338 NewClauses.push_back(FilterClause);
2340 MakeNewInstruction = true;
2341 CleanupFlag = false;
2345 bool MakeNewFilter = false; // If true, make a new filter.
2346 SmallVector<Constant *, 16> NewFilterElts; // New elements.
2347 if (isa<ConstantAggregateZero>(FilterClause)) {
2348 // Not an empty filter - it contains at least one null typeinfo.
2349 assert(NumTypeInfos > 0 && "Should have handled empty filter already!");
2350 Constant *TypeInfo =
2351 Constant::getNullValue(FilterType->getElementType());
2352 // If this typeinfo is a catch-all then the filter can never match.
2353 if (isCatchAll(Personality, TypeInfo)) {
2354 // Throw the filter away.
2355 MakeNewInstruction = true;
2359 // There is no point in having multiple copies of this typeinfo, so
2360 // discard all but the first copy if there is more than one.
2361 NewFilterElts.push_back(TypeInfo);
2362 if (NumTypeInfos > 1)
2363 MakeNewFilter = true;
2365 ConstantArray *Filter = cast<ConstantArray>(FilterClause);
2366 SmallPtrSet<Value *, 16> SeenInFilter; // For uniquing the elements.
2367 NewFilterElts.reserve(NumTypeInfos);
2369 // Remove any filter elements that were already caught or that already
2370 // occurred in the filter. While there, see if any of the elements are
2371 // catch-alls. If so, the filter can be discarded.
2372 bool SawCatchAll = false;
2373 for (unsigned j = 0; j != NumTypeInfos; ++j) {
2374 Constant *Elt = Filter->getOperand(j);
2375 Constant *TypeInfo = Elt->stripPointerCasts();
2376 if (isCatchAll(Personality, TypeInfo)) {
2377 // This element is a catch-all. Bail out, noting this fact.
2381 if (AlreadyCaught.count(TypeInfo))
2382 // Already caught by an earlier clause, so having it in the filter
2385 // There is no point in having multiple copies of the same typeinfo in
2386 // a filter, so only add it if we didn't already.
2387 if (SeenInFilter.insert(TypeInfo).second)
2388 NewFilterElts.push_back(cast<Constant>(Elt));
2390 // A filter containing a catch-all cannot match anything by definition.
2392 // Throw the filter away.
2393 MakeNewInstruction = true;
2397 // If we dropped something from the filter, make a new one.
2398 if (NewFilterElts.size() < NumTypeInfos)
2399 MakeNewFilter = true;
2401 if (MakeNewFilter) {
2402 FilterType = ArrayType::get(FilterType->getElementType(),
2403 NewFilterElts.size());
2404 FilterClause = ConstantArray::get(FilterType, NewFilterElts);
2405 MakeNewInstruction = true;
2408 NewClauses.push_back(FilterClause);
2410 // If the new filter is empty then it will catch everything so there is
2411 // no point in keeping any following clauses or marking the landingpad
2412 // as having a cleanup. The case of the original filter being empty was
2413 // already handled above.
2414 if (MakeNewFilter && !NewFilterElts.size()) {
2415 assert(MakeNewInstruction && "New filter but not a new instruction!");
2416 CleanupFlag = false;
2422 // If several filters occur in a row then reorder them so that the shortest
2423 // filters come first (those with the smallest number of elements). This is
2424 // advantageous because shorter filters are more likely to match, speeding up
2425 // unwinding, but mostly because it increases the effectiveness of the other
2426 // filter optimizations below.
2427 for (unsigned i = 0, e = NewClauses.size(); i + 1 < e; ) {
2429 // Find the maximal 'j' s.t. the range [i, j) consists entirely of filters.
2430 for (j = i; j != e; ++j)
2431 if (!isa<ArrayType>(NewClauses[j]->getType()))
2434 // Check whether the filters are already sorted by length. We need to know
2435 // if sorting them is actually going to do anything so that we only make a
2436 // new landingpad instruction if it does.
2437 for (unsigned k = i; k + 1 < j; ++k)
2438 if (shorter_filter(NewClauses[k+1], NewClauses[k])) {
2439 // Not sorted, so sort the filters now. Doing an unstable sort would be
2440 // correct too but reordering filters pointlessly might confuse users.
2441 std::stable_sort(NewClauses.begin() + i, NewClauses.begin() + j,
2443 MakeNewInstruction = true;
2447 // Look for the next batch of filters.
2451 // If typeinfos matched if and only if equal, then the elements of a filter L
2452 // that occurs later than a filter F could be replaced by the intersection of
2453 // the elements of F and L. In reality two typeinfos can match without being
2454 // equal (for example if one represents a C++ class, and the other some class
2455 // derived from it) so it would be wrong to perform this transform in general.
2456 // However the transform is correct and useful if F is a subset of L. In that
2457 // case L can be replaced by F, and thus removed altogether since repeating a
2458 // filter is pointless. So here we look at all pairs of filters F and L where
2459 // L follows F in the list of clauses, and remove L if every element of F is
2460 // an element of L. This can occur when inlining C++ functions with exception
2462 for (unsigned i = 0; i + 1 < NewClauses.size(); ++i) {
2463 // Examine each filter in turn.
2464 Value *Filter = NewClauses[i];
2465 ArrayType *FTy = dyn_cast<ArrayType>(Filter->getType());
2467 // Not a filter - skip it.
2469 unsigned FElts = FTy->getNumElements();
2470 // Examine each filter following this one. Doing this backwards means that
2471 // we don't have to worry about filters disappearing under us when removed.
2472 for (unsigned j = NewClauses.size() - 1; j != i; --j) {
2473 Value *LFilter = NewClauses[j];
2474 ArrayType *LTy = dyn_cast<ArrayType>(LFilter->getType());
2476 // Not a filter - skip it.
2478 // If Filter is a subset of LFilter, i.e. every element of Filter is also
2479 // an element of LFilter, then discard LFilter.
2480 SmallVectorImpl<Constant *>::iterator J = NewClauses.begin() + j;
2481 // If Filter is empty then it is a subset of LFilter.
2484 NewClauses.erase(J);
2485 MakeNewInstruction = true;
2486 // Move on to the next filter.
2489 unsigned LElts = LTy->getNumElements();
2490 // If Filter is longer than LFilter then it cannot be a subset of it.
2492 // Move on to the next filter.
2494 // At this point we know that LFilter has at least one element.
2495 if (isa<ConstantAggregateZero>(LFilter)) { // LFilter only contains zeros.
2496 // Filter is a subset of LFilter iff Filter contains only zeros (as we
2497 // already know that Filter is not longer than LFilter).
2498 if (isa<ConstantAggregateZero>(Filter)) {
2499 assert(FElts <= LElts && "Should have handled this case earlier!");
2501 NewClauses.erase(J);
2502 MakeNewInstruction = true;
2504 // Move on to the next filter.
2507 ConstantArray *LArray = cast<ConstantArray>(LFilter);
2508 if (isa<ConstantAggregateZero>(Filter)) { // Filter only contains zeros.
2509 // Since Filter is non-empty and contains only zeros, it is a subset of
2510 // LFilter iff LFilter contains a zero.
2511 assert(FElts > 0 && "Should have eliminated the empty filter earlier!");
2512 for (unsigned l = 0; l != LElts; ++l)
2513 if (LArray->getOperand(l)->isNullValue()) {
2514 // LFilter contains a zero - discard it.
2515 NewClauses.erase(J);
2516 MakeNewInstruction = true;
2519 // Move on to the next filter.
2522 // At this point we know that both filters are ConstantArrays. Loop over
2523 // operands to see whether every element of Filter is also an element of
2524 // LFilter. Since filters tend to be short this is probably faster than
2525 // using a method that scales nicely.
2526 ConstantArray *FArray = cast<ConstantArray>(Filter);
2527 bool AllFound = true;
2528 for (unsigned f = 0; f != FElts; ++f) {
2529 Value *FTypeInfo = FArray->getOperand(f)->stripPointerCasts();
2531 for (unsigned l = 0; l != LElts; ++l) {
2532 Value *LTypeInfo = LArray->getOperand(l)->stripPointerCasts();
2533 if (LTypeInfo == FTypeInfo) {
2543 NewClauses.erase(J);
2544 MakeNewInstruction = true;
2546 // Move on to the next filter.
2550 // If we changed any of the clauses, replace the old landingpad instruction
2552 if (MakeNewInstruction) {
2553 LandingPadInst *NLI = LandingPadInst::Create(LI.getType(),
2554 LI.getPersonalityFn(),
2556 for (unsigned i = 0, e = NewClauses.size(); i != e; ++i)
2557 NLI->addClause(NewClauses[i]);
2558 // A landing pad with no clauses must have the cleanup flag set. It is
2559 // theoretically possible, though highly unlikely, that we eliminated all
2560 // clauses. If so, force the cleanup flag to true.
2561 if (NewClauses.empty())
2563 NLI->setCleanup(CleanupFlag);
2567 // Even if none of the clauses changed, we may nonetheless have understood
2568 // that the cleanup flag is pointless. Clear it if so.
2569 if (LI.isCleanup() != CleanupFlag) {
2570 assert(!CleanupFlag && "Adding a cleanup, not removing one?!");
2571 LI.setCleanup(CleanupFlag);
2578 /// TryToSinkInstruction - Try to move the specified instruction from its
2579 /// current block into the beginning of DestBlock, which can only happen if it's
2580 /// safe to move the instruction past all of the instructions between it and the
2581 /// end of its block.
2582 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
2583 assert(I->hasOneUse() && "Invariants didn't hold!");
2585 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
2586 if (isa<PHINode>(I) || isa<LandingPadInst>(I) || I->mayHaveSideEffects() ||
2587 isa<TerminatorInst>(I))
2590 // Do not sink alloca instructions out of the entry block.
2591 if (isa<AllocaInst>(I) && I->getParent() ==
2592 &DestBlock->getParent()->getEntryBlock())
2595 // We can only sink load instructions if there is nothing between the load and
2596 // the end of block that could change the value.
2597 if (I->mayReadFromMemory()) {
2598 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
2600 if (Scan->mayWriteToMemory())
2604 BasicBlock::iterator InsertPos = DestBlock->getFirstInsertionPt();
2605 I->moveBefore(InsertPos);
2610 bool InstCombiner::run() {
2611 while (!Worklist.isEmpty()) {
2612 Instruction *I = Worklist.RemoveOne();
2613 if (I == nullptr) continue; // skip null values.
2615 // Check to see if we can DCE the instruction.
2616 if (isInstructionTriviallyDead(I, TLI)) {
2617 DEBUG(dbgs() << "IC: DCE: " << *I << '\n');
2618 EraseInstFromFunction(*I);
2620 MadeIRChange = true;
2624 // Instruction isn't dead, see if we can constant propagate it.
2625 if (!I->use_empty() && isa<Constant>(I->getOperand(0))) {
2626 if (Constant *C = ConstantFoldInstruction(I, DL, TLI)) {
2627 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n');
2629 // Add operands to the worklist.
2630 ReplaceInstUsesWith(*I, C);
2632 EraseInstFromFunction(*I);
2633 MadeIRChange = true;
2638 // See if we can trivially sink this instruction to a successor basic block.
2639 if (I->hasOneUse()) {
2640 BasicBlock *BB = I->getParent();
2641 Instruction *UserInst = cast<Instruction>(*I->user_begin());
2642 BasicBlock *UserParent;
2644 // Get the block the use occurs in.
2645 if (PHINode *PN = dyn_cast<PHINode>(UserInst))
2646 UserParent = PN->getIncomingBlock(*I->use_begin());
2648 UserParent = UserInst->getParent();
2650 if (UserParent != BB) {
2651 bool UserIsSuccessor = false;
2652 // See if the user is one of our successors.
2653 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
2654 if (*SI == UserParent) {
2655 UserIsSuccessor = true;
2659 // If the user is one of our immediate successors, and if that successor
2660 // only has us as a predecessors (we'd have to split the critical edge
2661 // otherwise), we can keep going.
2662 if (UserIsSuccessor && UserParent->getSinglePredecessor()) {
2663 // Okay, the CFG is simple enough, try to sink this instruction.
2664 if (TryToSinkInstruction(I, UserParent)) {
2665 MadeIRChange = true;
2666 // We'll add uses of the sunk instruction below, but since sinking
2667 // can expose opportunities for it's *operands* add them to the
2669 for (Use &U : I->operands())
2670 if (Instruction *OpI = dyn_cast<Instruction>(U.get()))
2677 // Now that we have an instruction, try combining it to simplify it.
2678 Builder->SetInsertPoint(I->getParent(), I);
2679 Builder->SetCurrentDebugLocation(I->getDebugLoc());
2684 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str(););
2685 DEBUG(dbgs() << "IC: Visiting: " << OrigI << '\n');
2687 if (Instruction *Result = visit(*I)) {
2689 // Should we replace the old instruction with a new one?
2691 DEBUG(dbgs() << "IC: Old = " << *I << '\n'
2692 << " New = " << *Result << '\n');
2694 if (!I->getDebugLoc().isUnknown())
2695 Result->setDebugLoc(I->getDebugLoc());
2696 // Everything uses the new instruction now.
2697 I->replaceAllUsesWith(Result);
2699 // Move the name to the new instruction first.
2700 Result->takeName(I);
2702 // Push the new instruction and any users onto the worklist.
2703 Worklist.Add(Result);
2704 Worklist.AddUsersToWorkList(*Result);
2706 // Insert the new instruction into the basic block...
2707 BasicBlock *InstParent = I->getParent();
2708 BasicBlock::iterator InsertPos = I;
2710 // If we replace a PHI with something that isn't a PHI, fix up the
2712 if (!isa<PHINode>(Result) && isa<PHINode>(InsertPos))
2713 InsertPos = InstParent->getFirstInsertionPt();
2715 InstParent->getInstList().insert(InsertPos, Result);
2717 EraseInstFromFunction(*I);
2720 DEBUG(dbgs() << "IC: Mod = " << OrigI << '\n'
2721 << " New = " << *I << '\n');
2724 // If the instruction was modified, it's possible that it is now dead.
2725 // if so, remove it.
2726 if (isInstructionTriviallyDead(I, TLI)) {
2727 EraseInstFromFunction(*I);
2730 Worklist.AddUsersToWorkList(*I);
2733 MadeIRChange = true;
2738 return MadeIRChange;
2741 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
2742 /// all reachable code to the worklist.
2744 /// This has a couple of tricks to make the code faster and more powerful. In
2745 /// particular, we constant fold and DCE instructions as we go, to avoid adding
2746 /// them to the worklist (this significantly speeds up instcombine on code where
2747 /// many instructions are dead or constant). Additionally, if we find a branch
2748 /// whose condition is a known constant, we only visit the reachable successors.
2750 static bool AddReachableCodeToWorklist(BasicBlock *BB, const DataLayout &DL,
2751 SmallPtrSetImpl<BasicBlock *> &Visited,
2752 InstCombineWorklist &ICWorklist,
2753 const TargetLibraryInfo *TLI) {
2754 bool MadeIRChange = false;
2755 SmallVector<BasicBlock*, 256> Worklist;
2756 Worklist.push_back(BB);
2758 SmallVector<Instruction*, 128> InstrsForInstCombineWorklist;
2759 DenseMap<ConstantExpr*, Constant*> FoldedConstants;
2762 BB = Worklist.pop_back_val();
2764 // We have now visited this block! If we've already been here, ignore it.
2765 if (!Visited.insert(BB).second)
2768 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
2769 Instruction *Inst = BBI++;
2771 // DCE instruction if trivially dead.
2772 if (isInstructionTriviallyDead(Inst, TLI)) {
2774 DEBUG(dbgs() << "IC: DCE: " << *Inst << '\n');
2775 Inst->eraseFromParent();
2779 // ConstantProp instruction if trivially constant.
2780 if (!Inst->use_empty() && isa<Constant>(Inst->getOperand(0)))
2781 if (Constant *C = ConstantFoldInstruction(Inst, DL, TLI)) {
2782 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: "
2784 Inst->replaceAllUsesWith(C);
2786 Inst->eraseFromParent();
2790 // See if we can constant fold its operands.
2791 for (User::op_iterator i = Inst->op_begin(), e = Inst->op_end(); i != e;
2793 ConstantExpr *CE = dyn_cast<ConstantExpr>(i);
2797 Constant *&FoldRes = FoldedConstants[CE];
2799 FoldRes = ConstantFoldConstantExpression(CE, DL, TLI);
2803 if (FoldRes != CE) {
2805 MadeIRChange = true;
2809 InstrsForInstCombineWorklist.push_back(Inst);
2812 // Recursively visit successors. If this is a branch or switch on a
2813 // constant, only visit the reachable successor.
2814 TerminatorInst *TI = BB->getTerminator();
2815 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
2816 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
2817 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
2818 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
2819 Worklist.push_back(ReachableBB);
2822 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
2823 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
2824 // See if this is an explicit destination.
2825 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2827 if (i.getCaseValue() == Cond) {
2828 BasicBlock *ReachableBB = i.getCaseSuccessor();
2829 Worklist.push_back(ReachableBB);
2833 // Otherwise it is the default destination.
2834 Worklist.push_back(SI->getDefaultDest());
2839 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
2840 Worklist.push_back(TI->getSuccessor(i));
2841 } while (!Worklist.empty());
2843 // Once we've found all of the instructions to add to instcombine's worklist,
2844 // add them in reverse order. This way instcombine will visit from the top
2845 // of the function down. This jives well with the way that it adds all uses
2846 // of instructions to the worklist after doing a transformation, thus avoiding
2847 // some N^2 behavior in pathological cases.
2848 ICWorklist.AddInitialGroup(&InstrsForInstCombineWorklist[0],
2849 InstrsForInstCombineWorklist.size());
2851 return MadeIRChange;
2854 /// \brief Populate the IC worklist from a function, and prune any dead basic
2855 /// blocks discovered in the process.
2857 /// This also does basic constant propagation and other forward fixing to make
2858 /// the combiner itself run much faster.
2859 static bool prepareICWorklistFromFunction(Function &F, const DataLayout &DL,
2860 TargetLibraryInfo *TLI,
2861 InstCombineWorklist &ICWorklist) {
2862 bool MadeIRChange = false;
2864 // Do a depth-first traversal of the function, populate the worklist with
2865 // the reachable instructions. Ignore blocks that are not reachable. Keep
2866 // track of which blocks we visit.
2867 SmallPtrSet<BasicBlock *, 64> Visited;
2869 AddReachableCodeToWorklist(F.begin(), DL, Visited, ICWorklist, TLI);
2871 // Do a quick scan over the function. If we find any blocks that are
2872 // unreachable, remove any instructions inside of them. This prevents
2873 // the instcombine code from having to deal with some bad special cases.
2874 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
2875 if (Visited.count(BB))
2878 // Delete the instructions backwards, as it has a reduced likelihood of
2879 // having to update as many def-use and use-def chains.
2880 Instruction *EndInst = BB->getTerminator(); // Last not to be deleted.
2881 while (EndInst != BB->begin()) {
2882 // Delete the next to last instruction.
2883 BasicBlock::iterator I = EndInst;
2884 Instruction *Inst = --I;
2885 if (!Inst->use_empty())
2886 Inst->replaceAllUsesWith(UndefValue::get(Inst->getType()));
2887 if (isa<LandingPadInst>(Inst)) {
2891 if (!isa<DbgInfoIntrinsic>(Inst)) {
2893 MadeIRChange = true;
2895 Inst->eraseFromParent();
2899 return MadeIRChange;
2903 combineInstructionsOverFunction(Function &F, InstCombineWorklist &Worklist,
2904 AssumptionCache &AC, TargetLibraryInfo &TLI,
2905 DominatorTree &DT, LoopInfo *LI = nullptr) {
2907 bool MinimizeSize = F.hasFnAttribute(Attribute::MinSize);
2908 auto &DL = F.getParent()->getDataLayout();
2910 /// Builder - This is an IRBuilder that automatically inserts new
2911 /// instructions into the worklist when they are created.
2912 IRBuilder<true, TargetFolder, InstCombineIRInserter> Builder(
2913 F.getContext(), TargetFolder(DL), InstCombineIRInserter(Worklist, &AC));
2915 // Lower dbg.declare intrinsics otherwise their value may be clobbered
2917 bool DbgDeclaresChanged = LowerDbgDeclare(F);
2919 // Iterate while there is work to do.
2923 DEBUG(dbgs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
2924 << F.getName() << "\n");
2926 bool Changed = false;
2927 if (prepareICWorklistFromFunction(F, DL, &TLI, Worklist))
2930 InstCombiner IC(Worklist, &Builder, MinimizeSize, &AC, &TLI, &DT, DL, LI);
2938 return DbgDeclaresChanged || Iteration > 1;
2941 PreservedAnalyses InstCombinePass::run(Function &F,
2942 AnalysisManager<Function> *AM) {
2943 auto &AC = AM->getResult<AssumptionAnalysis>(F);
2944 auto &DT = AM->getResult<DominatorTreeAnalysis>(F);
2945 auto &TLI = AM->getResult<TargetLibraryAnalysis>(F);
2947 auto *LI = AM->getCachedResult<LoopAnalysis>(F);
2949 if (!combineInstructionsOverFunction(F, Worklist, AC, TLI, DT, LI))
2950 // No changes, all analyses are preserved.
2951 return PreservedAnalyses::all();
2953 // Mark all the analyses that instcombine updates as preserved.
2954 // FIXME: Need a way to preserve CFG analyses here!
2955 PreservedAnalyses PA;
2956 PA.preserve<DominatorTreeAnalysis>();
2961 /// \brief The legacy pass manager's instcombine pass.
2963 /// This is a basic whole-function wrapper around the instcombine utility. It
2964 /// will try to combine all instructions in the function.
2965 class InstructionCombiningPass : public FunctionPass {
2966 InstCombineWorklist Worklist;
2969 static char ID; // Pass identification, replacement for typeid
2971 InstructionCombiningPass() : FunctionPass(ID) {
2972 initializeInstructionCombiningPassPass(*PassRegistry::getPassRegistry());
2975 void getAnalysisUsage(AnalysisUsage &AU) const override;
2976 bool runOnFunction(Function &F) override;
2980 void InstructionCombiningPass::getAnalysisUsage(AnalysisUsage &AU) const {
2981 AU.setPreservesCFG();
2982 AU.addRequired<AssumptionCacheTracker>();
2983 AU.addRequired<TargetLibraryInfoWrapperPass>();
2984 AU.addRequired<DominatorTreeWrapperPass>();
2985 AU.addPreserved<DominatorTreeWrapperPass>();
2988 bool InstructionCombiningPass::runOnFunction(Function &F) {
2989 if (skipOptnoneFunction(F))
2992 // Required analyses.
2993 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
2994 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
2995 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2997 // Optional analyses.
2998 auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>();
2999 auto *LI = LIWP ? &LIWP->getLoopInfo() : nullptr;
3001 return combineInstructionsOverFunction(F, Worklist, AC, TLI, DT, LI);
3004 char InstructionCombiningPass::ID = 0;
3005 INITIALIZE_PASS_BEGIN(InstructionCombiningPass, "instcombine",
3006 "Combine redundant instructions", false, false)
3007 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
3008 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
3009 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
3010 INITIALIZE_PASS_END(InstructionCombiningPass, "instcombine",
3011 "Combine redundant instructions", false, false)
3013 // Initialization Routines
3014 void llvm::initializeInstCombine(PassRegistry &Registry) {
3015 initializeInstructionCombiningPassPass(Registry);
3018 void LLVMInitializeInstCombine(LLVMPassRegistryRef R) {
3019 initializeInstructionCombiningPassPass(*unwrap(R));
3022 FunctionPass *llvm::createInstructionCombiningPass() {
3023 return new InstructionCombiningPass();