1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // InstructionCombining - Combine instructions to form fewer, simple
11 // instructions. This pass does not modify the CFG. This pass is where
12 // algebraic simplification happens.
14 // This pass combines things like:
20 // This is a simple worklist driven algorithm.
22 // This pass guarantees that the following canonicalizations are performed on
24 // 1. If a binary operator has a constant operand, it is moved to the RHS
25 // 2. Bitwise operators with constant operands are always grouped so that
26 // shifts are performed first, then or's, then and's, then xor's.
27 // 3. Compare instructions are converted from <,>,<=,>= to ==,!= if possible
28 // 4. All cmp instructions on boolean values are replaced with logical ops
29 // 5. add X, X is represented as (X*2) => (X << 1)
30 // 6. Multiplies with a power-of-two constant argument are transformed into
34 //===----------------------------------------------------------------------===//
36 #include "llvm/Transforms/Scalar.h"
37 #include "InstCombine.h"
38 #include "llvm-c/Initialization.h"
39 #include "llvm/ADT/SmallPtrSet.h"
40 #include "llvm/ADT/Statistic.h"
41 #include "llvm/ADT/StringSwitch.h"
42 #include "llvm/Analysis/ConstantFolding.h"
43 #include "llvm/Analysis/InstructionSimplify.h"
44 #include "llvm/Analysis/MemoryBuiltins.h"
45 #include "llvm/Analysis/ValueTracking.h"
46 #include "llvm/IR/CFG.h"
47 #include "llvm/IR/DataLayout.h"
48 #include "llvm/IR/GetElementPtrTypeIterator.h"
49 #include "llvm/IR/IntrinsicInst.h"
50 #include "llvm/IR/PatternMatch.h"
51 #include "llvm/IR/ValueHandle.h"
52 #include "llvm/Support/CommandLine.h"
53 #include "llvm/Support/Debug.h"
54 #include "llvm/Target/TargetLibraryInfo.h"
55 #include "llvm/Transforms/Utils/Local.h"
59 using namespace llvm::PatternMatch;
61 #define DEBUG_TYPE "instcombine"
63 STATISTIC(NumCombined , "Number of insts combined");
64 STATISTIC(NumConstProp, "Number of constant folds");
65 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
66 STATISTIC(NumSunkInst , "Number of instructions sunk");
67 STATISTIC(NumExpand, "Number of expansions");
68 STATISTIC(NumFactor , "Number of factorizations");
69 STATISTIC(NumReassoc , "Number of reassociations");
71 static cl::opt<bool> UnsafeFPShrink("enable-double-float-shrink", cl::Hidden,
73 cl::desc("Enable unsafe double to float "
74 "shrinking for math lib calls"));
76 // Initialization Routines
77 void llvm::initializeInstCombine(PassRegistry &Registry) {
78 initializeInstCombinerPass(Registry);
81 void LLVMInitializeInstCombine(LLVMPassRegistryRef R) {
82 initializeInstCombine(*unwrap(R));
85 char InstCombiner::ID = 0;
86 INITIALIZE_PASS_BEGIN(InstCombiner, "instcombine",
87 "Combine redundant instructions", false, false)
88 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
89 INITIALIZE_PASS_END(InstCombiner, "instcombine",
90 "Combine redundant instructions", false, false)
92 void InstCombiner::getAnalysisUsage(AnalysisUsage &AU) const {
94 AU.addRequired<TargetLibraryInfo>();
98 Value *InstCombiner::EmitGEPOffset(User *GEP) {
99 return llvm::EmitGEPOffset(Builder, *getDataLayout(), GEP);
102 /// ShouldChangeType - Return true if it is desirable to convert a computation
103 /// from 'From' to 'To'. We don't want to convert from a legal to an illegal
104 /// type for example, or from a smaller to a larger illegal type.
105 bool InstCombiner::ShouldChangeType(Type *From, Type *To) const {
106 assert(From->isIntegerTy() && To->isIntegerTy());
108 // If we don't have DL, we don't know if the source/dest are legal.
109 if (!DL) return false;
111 unsigned FromWidth = From->getPrimitiveSizeInBits();
112 unsigned ToWidth = To->getPrimitiveSizeInBits();
113 bool FromLegal = DL->isLegalInteger(FromWidth);
114 bool ToLegal = DL->isLegalInteger(ToWidth);
116 // If this is a legal integer from type, and the result would be an illegal
117 // type, don't do the transformation.
118 if (FromLegal && !ToLegal)
121 // Otherwise, if both are illegal, do not increase the size of the result. We
122 // do allow things like i160 -> i64, but not i64 -> i160.
123 if (!FromLegal && !ToLegal && ToWidth > FromWidth)
129 // Return true, if No Signed Wrap should be maintained for I.
130 // The No Signed Wrap flag can be kept if the operation "B (I.getOpcode) C",
131 // where both B and C should be ConstantInts, results in a constant that does
132 // not overflow. This function only handles the Add and Sub opcodes. For
133 // all other opcodes, the function conservatively returns false.
134 static bool MaintainNoSignedWrap(BinaryOperator &I, Value *B, Value *C) {
135 OverflowingBinaryOperator *OBO = dyn_cast<OverflowingBinaryOperator>(&I);
136 if (!OBO || !OBO->hasNoSignedWrap()) {
140 // We reason about Add and Sub Only.
141 Instruction::BinaryOps Opcode = I.getOpcode();
142 if (Opcode != Instruction::Add &&
143 Opcode != Instruction::Sub) {
147 ConstantInt *CB = dyn_cast<ConstantInt>(B);
148 ConstantInt *CC = dyn_cast<ConstantInt>(C);
154 const APInt &BVal = CB->getValue();
155 const APInt &CVal = CC->getValue();
156 bool Overflow = false;
158 if (Opcode == Instruction::Add) {
159 BVal.sadd_ov(CVal, Overflow);
161 BVal.ssub_ov(CVal, Overflow);
167 /// Conservatively clears subclassOptionalData after a reassociation or
168 /// commutation. We preserve fast-math flags when applicable as they can be
170 static void ClearSubclassDataAfterReassociation(BinaryOperator &I) {
171 FPMathOperator *FPMO = dyn_cast<FPMathOperator>(&I);
173 I.clearSubclassOptionalData();
177 FastMathFlags FMF = I.getFastMathFlags();
178 I.clearSubclassOptionalData();
179 I.setFastMathFlags(FMF);
182 /// SimplifyAssociativeOrCommutative - This performs a few simplifications for
183 /// operators which are associative or commutative:
185 // Commutative operators:
187 // 1. Order operands such that they are listed from right (least complex) to
188 // left (most complex). This puts constants before unary operators before
191 // Associative operators:
193 // 2. Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
194 // 3. Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
196 // Associative and commutative operators:
198 // 4. Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
199 // 5. Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
200 // 6. Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
201 // if C1 and C2 are constants.
203 bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) {
204 Instruction::BinaryOps Opcode = I.getOpcode();
205 bool Changed = false;
208 // Order operands such that they are listed from right (least complex) to
209 // left (most complex). This puts constants before unary operators before
211 if (I.isCommutative() && getComplexity(I.getOperand(0)) <
212 getComplexity(I.getOperand(1)))
213 Changed = !I.swapOperands();
215 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
216 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
218 if (I.isAssociative()) {
219 // Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
220 if (Op0 && Op0->getOpcode() == Opcode) {
221 Value *A = Op0->getOperand(0);
222 Value *B = Op0->getOperand(1);
223 Value *C = I.getOperand(1);
225 // Does "B op C" simplify?
226 if (Value *V = SimplifyBinOp(Opcode, B, C, DL)) {
227 // It simplifies to V. Form "A op V".
230 // Conservatively clear the optional flags, since they may not be
231 // preserved by the reassociation.
232 if (MaintainNoSignedWrap(I, B, C) &&
233 (!Op0 || (isa<BinaryOperator>(Op0) && Op0->hasNoSignedWrap()))) {
234 // Note: this is only valid because SimplifyBinOp doesn't look at
235 // the operands to Op0.
236 I.clearSubclassOptionalData();
237 I.setHasNoSignedWrap(true);
239 ClearSubclassDataAfterReassociation(I);
248 // Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
249 if (Op1 && Op1->getOpcode() == Opcode) {
250 Value *A = I.getOperand(0);
251 Value *B = Op1->getOperand(0);
252 Value *C = Op1->getOperand(1);
254 // Does "A op B" simplify?
255 if (Value *V = SimplifyBinOp(Opcode, A, B, DL)) {
256 // It simplifies to V. Form "V op C".
259 // Conservatively clear the optional flags, since they may not be
260 // preserved by the reassociation.
261 ClearSubclassDataAfterReassociation(I);
269 if (I.isAssociative() && I.isCommutative()) {
270 // Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
271 if (Op0 && Op0->getOpcode() == Opcode) {
272 Value *A = Op0->getOperand(0);
273 Value *B = Op0->getOperand(1);
274 Value *C = I.getOperand(1);
276 // Does "C op A" simplify?
277 if (Value *V = SimplifyBinOp(Opcode, C, A, DL)) {
278 // It simplifies to V. Form "V op B".
281 // Conservatively clear the optional flags, since they may not be
282 // preserved by the reassociation.
283 ClearSubclassDataAfterReassociation(I);
290 // Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
291 if (Op1 && Op1->getOpcode() == Opcode) {
292 Value *A = I.getOperand(0);
293 Value *B = Op1->getOperand(0);
294 Value *C = Op1->getOperand(1);
296 // Does "C op A" simplify?
297 if (Value *V = SimplifyBinOp(Opcode, C, A, DL)) {
298 // It simplifies to V. Form "B op V".
301 // Conservatively clear the optional flags, since they may not be
302 // preserved by the reassociation.
303 ClearSubclassDataAfterReassociation(I);
310 // Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
311 // if C1 and C2 are constants.
313 Op0->getOpcode() == Opcode && Op1->getOpcode() == Opcode &&
314 isa<Constant>(Op0->getOperand(1)) &&
315 isa<Constant>(Op1->getOperand(1)) &&
316 Op0->hasOneUse() && Op1->hasOneUse()) {
317 Value *A = Op0->getOperand(0);
318 Constant *C1 = cast<Constant>(Op0->getOperand(1));
319 Value *B = Op1->getOperand(0);
320 Constant *C2 = cast<Constant>(Op1->getOperand(1));
322 Constant *Folded = ConstantExpr::get(Opcode, C1, C2);
323 BinaryOperator *New = BinaryOperator::Create(Opcode, A, B);
324 if (isa<FPMathOperator>(New)) {
325 FastMathFlags Flags = I.getFastMathFlags();
326 Flags &= Op0->getFastMathFlags();
327 Flags &= Op1->getFastMathFlags();
328 New->setFastMathFlags(Flags);
330 InsertNewInstWith(New, I);
332 I.setOperand(0, New);
333 I.setOperand(1, Folded);
334 // Conservatively clear the optional flags, since they may not be
335 // preserved by the reassociation.
336 ClearSubclassDataAfterReassociation(I);
343 // No further simplifications.
348 /// LeftDistributesOverRight - Whether "X LOp (Y ROp Z)" is always equal to
349 /// "(X LOp Y) ROp (X LOp Z)".
350 static bool LeftDistributesOverRight(Instruction::BinaryOps LOp,
351 Instruction::BinaryOps ROp) {
356 case Instruction::And:
357 // And distributes over Or and Xor.
361 case Instruction::Or:
362 case Instruction::Xor:
366 case Instruction::Mul:
367 // Multiplication distributes over addition and subtraction.
371 case Instruction::Add:
372 case Instruction::Sub:
376 case Instruction::Or:
377 // Or distributes over And.
381 case Instruction::And:
387 /// RightDistributesOverLeft - Whether "(X LOp Y) ROp Z" is always equal to
388 /// "(X ROp Z) LOp (Y ROp Z)".
389 static bool RightDistributesOverLeft(Instruction::BinaryOps LOp,
390 Instruction::BinaryOps ROp) {
391 if (Instruction::isCommutative(ROp))
392 return LeftDistributesOverRight(ROp, LOp);
397 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
398 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
399 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
400 case Instruction::And:
401 case Instruction::Or:
402 case Instruction::Xor:
406 case Instruction::Shl:
407 case Instruction::LShr:
408 case Instruction::AShr:
412 // TODO: It would be nice to handle division, aka "(X + Y)/Z = X/Z + Y/Z",
413 // but this requires knowing that the addition does not overflow and other
418 /// This function returns identity value for given opcode, which can be used to
419 /// factor patterns like (X * 2) + X ==> (X * 2) + (X * 1) ==> X * (2 + 1).
420 static Value *getIdentityValue(Instruction::BinaryOps OpCode, Value *V) {
421 if (isa<Constant>(V))
424 if (OpCode == Instruction::Mul)
425 return ConstantInt::get(V->getType(), 1);
427 // TODO: We can handle other cases e.g. Instruction::And, Instruction::Or etc.
432 /// This function factors binary ops which can be combined using distributive
433 /// laws. This function tries to transform 'Op' based TopLevelOpcode to enable
434 /// factorization e.g for ADD(SHL(X , 2), MUL(X, 5)), When this function called
435 /// with TopLevelOpcode == Instruction::Add and Op = SHL(X, 2), transforms
436 /// SHL(X, 2) to MUL(X, 4) i.e. returns Instruction::Mul with LHS set to 'X' and
438 static Instruction::BinaryOps
439 getBinOpsForFactorization(Instruction::BinaryOps TopLevelOpcode,
440 BinaryOperator *Op, Value *&LHS, Value *&RHS) {
442 return Instruction::BinaryOpsEnd;
444 LHS = Op->getOperand(0);
445 RHS = Op->getOperand(1);
447 switch (TopLevelOpcode) {
449 return Op->getOpcode();
451 case Instruction::Add:
452 case Instruction::Sub:
453 if (Op->getOpcode() == Instruction::Shl) {
454 if (Constant *CST = dyn_cast<Constant>(Op->getOperand(1))) {
455 // The multiplier is really 1 << CST.
456 RHS = ConstantExpr::getShl(ConstantInt::get(Op->getType(), 1), CST);
457 return Instruction::Mul;
460 return Op->getOpcode();
463 // TODO: We can add other conversions e.g. shr => div etc.
466 /// This tries to simplify binary operations by factorizing out common terms
467 /// (e. g. "(A*B)+(A*C)" -> "A*(B+C)").
468 static Value *tryFactorization(InstCombiner::BuilderTy *Builder,
469 const DataLayout *DL, BinaryOperator &I,
470 Instruction::BinaryOps InnerOpcode, Value *A,
471 Value *B, Value *C, Value *D) {
473 // If any of A, B, C, D are null, we can not factor I, return early.
474 // Checking A and C should be enough.
475 if (!A || !C || !B || !D)
478 Value *SimplifiedInst = nullptr;
479 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
480 Instruction::BinaryOps TopLevelOpcode = I.getOpcode();
482 // Does "X op' Y" always equal "Y op' X"?
483 bool InnerCommutative = Instruction::isCommutative(InnerOpcode);
485 // Does "X op' (Y op Z)" always equal "(X op' Y) op (X op' Z)"?
486 if (LeftDistributesOverRight(InnerOpcode, TopLevelOpcode))
487 // Does the instruction have the form "(A op' B) op (A op' D)" or, in the
488 // commutative case, "(A op' B) op (C op' A)"?
489 if (A == C || (InnerCommutative && A == D)) {
492 // Consider forming "A op' (B op D)".
493 // If "B op D" simplifies then it can be formed with no cost.
494 Value *V = SimplifyBinOp(TopLevelOpcode, B, D, DL);
495 // If "B op D" doesn't simplify then only go on if both of the existing
496 // operations "A op' B" and "C op' D" will be zapped as no longer used.
497 if (!V && LHS->hasOneUse() && RHS->hasOneUse())
498 V = Builder->CreateBinOp(TopLevelOpcode, B, D, RHS->getName());
500 SimplifiedInst = Builder->CreateBinOp(InnerOpcode, A, V);
504 // Does "(X op Y) op' Z" always equal "(X op' Z) op (Y op' Z)"?
505 if (!SimplifiedInst && RightDistributesOverLeft(TopLevelOpcode, InnerOpcode))
506 // Does the instruction have the form "(A op' B) op (C op' B)" or, in the
507 // commutative case, "(A op' B) op (B op' D)"?
508 if (B == D || (InnerCommutative && B == C)) {
511 // Consider forming "(A op C) op' B".
512 // If "A op C" simplifies then it can be formed with no cost.
513 Value *V = SimplifyBinOp(TopLevelOpcode, A, C, DL);
515 // If "A op C" doesn't simplify then only go on if both of the existing
516 // operations "A op' B" and "C op' D" will be zapped as no longer used.
517 if (!V && LHS->hasOneUse() && RHS->hasOneUse())
518 V = Builder->CreateBinOp(TopLevelOpcode, A, C, LHS->getName());
520 SimplifiedInst = Builder->CreateBinOp(InnerOpcode, V, B);
524 if (SimplifiedInst) {
526 SimplifiedInst->takeName(&I);
528 // Check if we can add NSW flag to SimplifiedInst. If so, set NSW flag.
529 // TODO: Check for NUW.
530 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(SimplifiedInst)) {
531 if (isa<OverflowingBinaryOperator>(SimplifiedInst)) {
533 if (isa<OverflowingBinaryOperator>(&I))
534 HasNSW = I.hasNoSignedWrap();
536 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS))
537 if (isa<OverflowingBinaryOperator>(Op0))
538 HasNSW &= Op0->hasNoSignedWrap();
540 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS))
541 if (isa<OverflowingBinaryOperator>(Op1))
542 HasNSW &= Op1->hasNoSignedWrap();
543 BO->setHasNoSignedWrap(HasNSW);
547 return SimplifiedInst;
550 /// SimplifyUsingDistributiveLaws - This tries to simplify binary operations
551 /// which some other binary operation distributes over either by factorizing
552 /// out common terms (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this
553 /// results in simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is
554 /// a win). Returns the simplified value, or null if it didn't simplify.
555 Value *InstCombiner::SimplifyUsingDistributiveLaws(BinaryOperator &I) {
556 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
557 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
558 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
561 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
562 auto TopLevelOpcode = I.getOpcode();
563 auto LHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op0, A, B);
564 auto RHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op1, C, D);
566 // The instruction has the form "(A op' B) op (C op' D)". Try to factorize
568 if (LHSOpcode == RHSOpcode) {
569 if (Value *V = tryFactorization(Builder, DL, I, LHSOpcode, A, B, C, D))
573 // The instruction has the form "(A op' B) op (C)". Try to factorize common
575 if (Value *V = tryFactorization(Builder, DL, I, LHSOpcode, A, B, RHS,
576 getIdentityValue(LHSOpcode, RHS)))
579 // The instruction has the form "(B) op (C op' D)". Try to factorize common
581 if (Value *V = tryFactorization(Builder, DL, I, RHSOpcode, LHS,
582 getIdentityValue(RHSOpcode, LHS), C, D))
586 if (Op0 && RightDistributesOverLeft(Op0->getOpcode(), TopLevelOpcode)) {
587 // The instruction has the form "(A op' B) op C". See if expanding it out
588 // to "(A op C) op' (B op C)" results in simplifications.
589 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
590 Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op'
592 // Do "A op C" and "B op C" both simplify?
593 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, C, DL))
594 if (Value *R = SimplifyBinOp(TopLevelOpcode, B, C, DL)) {
595 // They do! Return "L op' R".
597 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
598 if ((L == A && R == B) ||
599 (Instruction::isCommutative(InnerOpcode) && L == B && R == A))
601 // Otherwise return "L op' R" if it simplifies.
602 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, DL))
604 // Otherwise, create a new instruction.
605 C = Builder->CreateBinOp(InnerOpcode, L, R);
611 if (Op1 && LeftDistributesOverRight(TopLevelOpcode, Op1->getOpcode())) {
612 // The instruction has the form "A op (B op' C)". See if expanding it out
613 // to "(A op B) op' (A op C)" results in simplifications.
614 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
615 Instruction::BinaryOps InnerOpcode = Op1->getOpcode(); // op'
617 // Do "A op B" and "A op C" both simplify?
618 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, B, DL))
619 if (Value *R = SimplifyBinOp(TopLevelOpcode, A, C, DL)) {
620 // They do! Return "L op' R".
622 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
623 if ((L == B && R == C) ||
624 (Instruction::isCommutative(InnerOpcode) && L == C && R == B))
626 // Otherwise return "L op' R" if it simplifies.
627 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, DL))
629 // Otherwise, create a new instruction.
630 A = Builder->CreateBinOp(InnerOpcode, L, R);
639 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
640 // if the LHS is a constant zero (which is the 'negate' form).
642 Value *InstCombiner::dyn_castNegVal(Value *V) const {
643 if (BinaryOperator::isNeg(V))
644 return BinaryOperator::getNegArgument(V);
646 // Constants can be considered to be negated values if they can be folded.
647 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
648 return ConstantExpr::getNeg(C);
650 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
651 if (C->getType()->getElementType()->isIntegerTy())
652 return ConstantExpr::getNeg(C);
657 // dyn_castFNegVal - Given a 'fsub' instruction, return the RHS of the
658 // instruction if the LHS is a constant negative zero (which is the 'negate'
661 Value *InstCombiner::dyn_castFNegVal(Value *V, bool IgnoreZeroSign) const {
662 if (BinaryOperator::isFNeg(V, IgnoreZeroSign))
663 return BinaryOperator::getFNegArgument(V);
665 // Constants can be considered to be negated values if they can be folded.
666 if (ConstantFP *C = dyn_cast<ConstantFP>(V))
667 return ConstantExpr::getFNeg(C);
669 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
670 if (C->getType()->getElementType()->isFloatingPointTy())
671 return ConstantExpr::getFNeg(C);
676 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
678 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
679 return IC->Builder->CreateCast(CI->getOpcode(), SO, I.getType());
682 // Figure out if the constant is the left or the right argument.
683 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
684 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
686 if (Constant *SOC = dyn_cast<Constant>(SO)) {
688 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
689 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
692 Value *Op0 = SO, *Op1 = ConstOperand;
696 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I)) {
697 Value *RI = IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1,
698 SO->getName()+".op");
699 Instruction *FPInst = dyn_cast<Instruction>(RI);
700 if (FPInst && isa<FPMathOperator>(FPInst))
701 FPInst->copyFastMathFlags(BO);
704 if (ICmpInst *CI = dyn_cast<ICmpInst>(&I))
705 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
706 SO->getName()+".cmp");
707 if (FCmpInst *CI = dyn_cast<FCmpInst>(&I))
708 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
709 SO->getName()+".cmp");
710 llvm_unreachable("Unknown binary instruction type!");
713 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
714 // constant as the other operand, try to fold the binary operator into the
715 // select arguments. This also works for Cast instructions, which obviously do
716 // not have a second operand.
717 Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) {
718 // Don't modify shared select instructions
719 if (!SI->hasOneUse()) return nullptr;
720 Value *TV = SI->getOperand(1);
721 Value *FV = SI->getOperand(2);
723 if (isa<Constant>(TV) || isa<Constant>(FV)) {
724 // Bool selects with constant operands can be folded to logical ops.
725 if (SI->getType()->isIntegerTy(1)) return nullptr;
727 // If it's a bitcast involving vectors, make sure it has the same number of
728 // elements on both sides.
729 if (BitCastInst *BC = dyn_cast<BitCastInst>(&Op)) {
730 VectorType *DestTy = dyn_cast<VectorType>(BC->getDestTy());
731 VectorType *SrcTy = dyn_cast<VectorType>(BC->getSrcTy());
733 // Verify that either both or neither are vectors.
734 if ((SrcTy == nullptr) != (DestTy == nullptr)) return nullptr;
735 // If vectors, verify that they have the same number of elements.
736 if (SrcTy && SrcTy->getNumElements() != DestTy->getNumElements())
740 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, this);
741 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, this);
743 return SelectInst::Create(SI->getCondition(),
744 SelectTrueVal, SelectFalseVal);
750 /// FoldOpIntoPhi - Given a binary operator, cast instruction, or select which
751 /// has a PHI node as operand #0, see if we can fold the instruction into the
752 /// PHI (which is only possible if all operands to the PHI are constants).
754 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
755 PHINode *PN = cast<PHINode>(I.getOperand(0));
756 unsigned NumPHIValues = PN->getNumIncomingValues();
757 if (NumPHIValues == 0)
760 // We normally only transform phis with a single use. However, if a PHI has
761 // multiple uses and they are all the same operation, we can fold *all* of the
762 // uses into the PHI.
763 if (!PN->hasOneUse()) {
764 // Walk the use list for the instruction, comparing them to I.
765 for (User *U : PN->users()) {
766 Instruction *UI = cast<Instruction>(U);
767 if (UI != &I && !I.isIdenticalTo(UI))
770 // Otherwise, we can replace *all* users with the new PHI we form.
773 // Check to see if all of the operands of the PHI are simple constants
774 // (constantint/constantfp/undef). If there is one non-constant value,
775 // remember the BB it is in. If there is more than one or if *it* is a PHI,
776 // bail out. We don't do arbitrary constant expressions here because moving
777 // their computation can be expensive without a cost model.
778 BasicBlock *NonConstBB = nullptr;
779 for (unsigned i = 0; i != NumPHIValues; ++i) {
780 Value *InVal = PN->getIncomingValue(i);
781 if (isa<Constant>(InVal) && !isa<ConstantExpr>(InVal))
784 if (isa<PHINode>(InVal)) return nullptr; // Itself a phi.
785 if (NonConstBB) return nullptr; // More than one non-const value.
787 NonConstBB = PN->getIncomingBlock(i);
789 // If the InVal is an invoke at the end of the pred block, then we can't
790 // insert a computation after it without breaking the edge.
791 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal))
792 if (II->getParent() == NonConstBB)
795 // If the incoming non-constant value is in I's block, we will remove one
796 // instruction, but insert another equivalent one, leading to infinite
798 if (NonConstBB == I.getParent())
802 // If there is exactly one non-constant value, we can insert a copy of the
803 // operation in that block. However, if this is a critical edge, we would be
804 // inserting the computation one some other paths (e.g. inside a loop). Only
805 // do this if the pred block is unconditionally branching into the phi block.
806 if (NonConstBB != nullptr) {
807 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
808 if (!BI || !BI->isUnconditional()) return nullptr;
811 // Okay, we can do the transformation: create the new PHI node.
812 PHINode *NewPN = PHINode::Create(I.getType(), PN->getNumIncomingValues());
813 InsertNewInstBefore(NewPN, *PN);
816 // If we are going to have to insert a new computation, do so right before the
817 // predecessors terminator.
819 Builder->SetInsertPoint(NonConstBB->getTerminator());
821 // Next, add all of the operands to the PHI.
822 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) {
823 // We only currently try to fold the condition of a select when it is a phi,
824 // not the true/false values.
825 Value *TrueV = SI->getTrueValue();
826 Value *FalseV = SI->getFalseValue();
827 BasicBlock *PhiTransBB = PN->getParent();
828 for (unsigned i = 0; i != NumPHIValues; ++i) {
829 BasicBlock *ThisBB = PN->getIncomingBlock(i);
830 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB);
831 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB);
832 Value *InV = nullptr;
833 // Beware of ConstantExpr: it may eventually evaluate to getNullValue,
834 // even if currently isNullValue gives false.
835 Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i));
836 if (InC && !isa<ConstantExpr>(InC))
837 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred;
839 InV = Builder->CreateSelect(PN->getIncomingValue(i),
840 TrueVInPred, FalseVInPred, "phitmp");
841 NewPN->addIncoming(InV, ThisBB);
843 } else if (CmpInst *CI = dyn_cast<CmpInst>(&I)) {
844 Constant *C = cast<Constant>(I.getOperand(1));
845 for (unsigned i = 0; i != NumPHIValues; ++i) {
846 Value *InV = nullptr;
847 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
848 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
849 else if (isa<ICmpInst>(CI))
850 InV = Builder->CreateICmp(CI->getPredicate(), PN->getIncomingValue(i),
853 InV = Builder->CreateFCmp(CI->getPredicate(), PN->getIncomingValue(i),
855 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
857 } else if (I.getNumOperands() == 2) {
858 Constant *C = cast<Constant>(I.getOperand(1));
859 for (unsigned i = 0; i != NumPHIValues; ++i) {
860 Value *InV = nullptr;
861 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
862 InV = ConstantExpr::get(I.getOpcode(), InC, C);
864 InV = Builder->CreateBinOp(cast<BinaryOperator>(I).getOpcode(),
865 PN->getIncomingValue(i), C, "phitmp");
866 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
869 CastInst *CI = cast<CastInst>(&I);
870 Type *RetTy = CI->getType();
871 for (unsigned i = 0; i != NumPHIValues; ++i) {
873 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
874 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
876 InV = Builder->CreateCast(CI->getOpcode(),
877 PN->getIncomingValue(i), I.getType(), "phitmp");
878 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
882 for (auto UI = PN->user_begin(), E = PN->user_end(); UI != E;) {
883 Instruction *User = cast<Instruction>(*UI++);
884 if (User == &I) continue;
885 ReplaceInstUsesWith(*User, NewPN);
886 EraseInstFromFunction(*User);
888 return ReplaceInstUsesWith(I, NewPN);
891 /// FindElementAtOffset - Given a pointer type and a constant offset, determine
892 /// whether or not there is a sequence of GEP indices into the pointed type that
893 /// will land us at the specified offset. If so, fill them into NewIndices and
894 /// return the resultant element type, otherwise return null.
895 Type *InstCombiner::FindElementAtOffset(Type *PtrTy, int64_t Offset,
896 SmallVectorImpl<Value*> &NewIndices) {
897 assert(PtrTy->isPtrOrPtrVectorTy());
902 Type *Ty = PtrTy->getPointerElementType();
906 // Start with the index over the outer type. Note that the type size
907 // might be zero (even if the offset isn't zero) if the indexed type
908 // is something like [0 x {int, int}]
909 Type *IntPtrTy = DL->getIntPtrType(PtrTy);
910 int64_t FirstIdx = 0;
911 if (int64_t TySize = DL->getTypeAllocSize(Ty)) {
912 FirstIdx = Offset/TySize;
913 Offset -= FirstIdx*TySize;
915 // Handle hosts where % returns negative instead of values [0..TySize).
921 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
924 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
926 // Index into the types. If we fail, set OrigBase to null.
928 // Indexing into tail padding between struct/array elements.
929 if (uint64_t(Offset*8) >= DL->getTypeSizeInBits(Ty))
932 if (StructType *STy = dyn_cast<StructType>(Ty)) {
933 const StructLayout *SL = DL->getStructLayout(STy);
934 assert(Offset < (int64_t)SL->getSizeInBytes() &&
935 "Offset must stay within the indexed type");
937 unsigned Elt = SL->getElementContainingOffset(Offset);
938 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
941 Offset -= SL->getElementOffset(Elt);
942 Ty = STy->getElementType(Elt);
943 } else if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
944 uint64_t EltSize = DL->getTypeAllocSize(AT->getElementType());
945 assert(EltSize && "Cannot index into a zero-sized array");
946 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
948 Ty = AT->getElementType();
950 // Otherwise, we can't index into the middle of this atomic type, bail.
958 static bool shouldMergeGEPs(GEPOperator &GEP, GEPOperator &Src) {
959 // If this GEP has only 0 indices, it is the same pointer as
960 // Src. If Src is not a trivial GEP too, don't combine
962 if (GEP.hasAllZeroIndices() && !Src.hasAllZeroIndices() &&
968 /// Descale - Return a value X such that Val = X * Scale, or null if none. If
969 /// the multiplication is known not to overflow then NoSignedWrap is set.
970 Value *InstCombiner::Descale(Value *Val, APInt Scale, bool &NoSignedWrap) {
971 assert(isa<IntegerType>(Val->getType()) && "Can only descale integers!");
972 assert(cast<IntegerType>(Val->getType())->getBitWidth() ==
973 Scale.getBitWidth() && "Scale not compatible with value!");
975 // If Val is zero or Scale is one then Val = Val * Scale.
976 if (match(Val, m_Zero()) || Scale == 1) {
981 // If Scale is zero then it does not divide Val.
982 if (Scale.isMinValue())
985 // Look through chains of multiplications, searching for a constant that is
986 // divisible by Scale. For example, descaling X*(Y*(Z*4)) by a factor of 4
987 // will find the constant factor 4 and produce X*(Y*Z). Descaling X*(Y*8) by
988 // a factor of 4 will produce X*(Y*2). The principle of operation is to bore
991 // Val = M1 * X || Analysis starts here and works down
992 // M1 = M2 * Y || Doesn't descend into terms with more
993 // M2 = Z * 4 \/ than one use
995 // Then to modify a term at the bottom:
998 // M1 = Z * Y || Replaced M2 with Z
1000 // Then to work back up correcting nsw flags.
1002 // Op - the term we are currently analyzing. Starts at Val then drills down.
1003 // Replaced with its descaled value before exiting from the drill down loop.
1006 // Parent - initially null, but after drilling down notes where Op came from.
1007 // In the example above, Parent is (Val, 0) when Op is M1, because M1 is the
1008 // 0'th operand of Val.
1009 std::pair<Instruction*, unsigned> Parent;
1011 // RequireNoSignedWrap - Set if the transform requires a descaling at deeper
1012 // levels that doesn't overflow.
1013 bool RequireNoSignedWrap = false;
1015 // logScale - log base 2 of the scale. Negative if not a power of 2.
1016 int32_t logScale = Scale.exactLogBase2();
1018 for (;; Op = Parent.first->getOperand(Parent.second)) { // Drill down
1020 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
1021 // If Op is a constant divisible by Scale then descale to the quotient.
1022 APInt Quotient(Scale), Remainder(Scale); // Init ensures right bitwidth.
1023 APInt::sdivrem(CI->getValue(), Scale, Quotient, Remainder);
1024 if (!Remainder.isMinValue())
1025 // Not divisible by Scale.
1027 // Replace with the quotient in the parent.
1028 Op = ConstantInt::get(CI->getType(), Quotient);
1029 NoSignedWrap = true;
1033 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op)) {
1035 if (BO->getOpcode() == Instruction::Mul) {
1037 NoSignedWrap = BO->hasNoSignedWrap();
1038 if (RequireNoSignedWrap && !NoSignedWrap)
1041 // There are three cases for multiplication: multiplication by exactly
1042 // the scale, multiplication by a constant different to the scale, and
1043 // multiplication by something else.
1044 Value *LHS = BO->getOperand(0);
1045 Value *RHS = BO->getOperand(1);
1047 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
1048 // Multiplication by a constant.
1049 if (CI->getValue() == Scale) {
1050 // Multiplication by exactly the scale, replace the multiplication
1051 // by its left-hand side in the parent.
1056 // Otherwise drill down into the constant.
1057 if (!Op->hasOneUse())
1060 Parent = std::make_pair(BO, 1);
1064 // Multiplication by something else. Drill down into the left-hand side
1065 // since that's where the reassociate pass puts the good stuff.
1066 if (!Op->hasOneUse())
1069 Parent = std::make_pair(BO, 0);
1073 if (logScale > 0 && BO->getOpcode() == Instruction::Shl &&
1074 isa<ConstantInt>(BO->getOperand(1))) {
1075 // Multiplication by a power of 2.
1076 NoSignedWrap = BO->hasNoSignedWrap();
1077 if (RequireNoSignedWrap && !NoSignedWrap)
1080 Value *LHS = BO->getOperand(0);
1081 int32_t Amt = cast<ConstantInt>(BO->getOperand(1))->
1082 getLimitedValue(Scale.getBitWidth());
1085 if (Amt == logScale) {
1086 // Multiplication by exactly the scale, replace the multiplication
1087 // by its left-hand side in the parent.
1091 if (Amt < logScale || !Op->hasOneUse())
1094 // Multiplication by more than the scale. Reduce the multiplying amount
1095 // by the scale in the parent.
1096 Parent = std::make_pair(BO, 1);
1097 Op = ConstantInt::get(BO->getType(), Amt - logScale);
1102 if (!Op->hasOneUse())
1105 if (CastInst *Cast = dyn_cast<CastInst>(Op)) {
1106 if (Cast->getOpcode() == Instruction::SExt) {
1107 // Op is sign-extended from a smaller type, descale in the smaller type.
1108 unsigned SmallSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1109 APInt SmallScale = Scale.trunc(SmallSize);
1110 // Suppose Op = sext X, and we descale X as Y * SmallScale. We want to
1111 // descale Op as (sext Y) * Scale. In order to have
1112 // sext (Y * SmallScale) = (sext Y) * Scale
1113 // some conditions need to hold however: SmallScale must sign-extend to
1114 // Scale and the multiplication Y * SmallScale should not overflow.
1115 if (SmallScale.sext(Scale.getBitWidth()) != Scale)
1116 // SmallScale does not sign-extend to Scale.
1118 assert(SmallScale.exactLogBase2() == logScale);
1119 // Require that Y * SmallScale must not overflow.
1120 RequireNoSignedWrap = true;
1122 // Drill down through the cast.
1123 Parent = std::make_pair(Cast, 0);
1128 if (Cast->getOpcode() == Instruction::Trunc) {
1129 // Op is truncated from a larger type, descale in the larger type.
1130 // Suppose Op = trunc X, and we descale X as Y * sext Scale. Then
1131 // trunc (Y * sext Scale) = (trunc Y) * Scale
1132 // always holds. However (trunc Y) * Scale may overflow even if
1133 // trunc (Y * sext Scale) does not, so nsw flags need to be cleared
1134 // from this point up in the expression (see later).
1135 if (RequireNoSignedWrap)
1138 // Drill down through the cast.
1139 unsigned LargeSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1140 Parent = std::make_pair(Cast, 0);
1141 Scale = Scale.sext(LargeSize);
1142 if (logScale + 1 == (int32_t)Cast->getType()->getPrimitiveSizeInBits())
1144 assert(Scale.exactLogBase2() == logScale);
1149 // Unsupported expression, bail out.
1153 // If Op is zero then Val = Op * Scale.
1154 if (match(Op, m_Zero())) {
1155 NoSignedWrap = true;
1159 // We know that we can successfully descale, so from here on we can safely
1160 // modify the IR. Op holds the descaled version of the deepest term in the
1161 // expression. NoSignedWrap is 'true' if multiplying Op by Scale is known
1165 // The expression only had one term.
1168 // Rewrite the parent using the descaled version of its operand.
1169 assert(Parent.first->hasOneUse() && "Drilled down when more than one use!");
1170 assert(Op != Parent.first->getOperand(Parent.second) &&
1171 "Descaling was a no-op?");
1172 Parent.first->setOperand(Parent.second, Op);
1173 Worklist.Add(Parent.first);
1175 // Now work back up the expression correcting nsw flags. The logic is based
1176 // on the following observation: if X * Y is known not to overflow as a signed
1177 // multiplication, and Y is replaced by a value Z with smaller absolute value,
1178 // then X * Z will not overflow as a signed multiplication either. As we work
1179 // our way up, having NoSignedWrap 'true' means that the descaled value at the
1180 // current level has strictly smaller absolute value than the original.
1181 Instruction *Ancestor = Parent.first;
1183 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Ancestor)) {
1184 // If the multiplication wasn't nsw then we can't say anything about the
1185 // value of the descaled multiplication, and we have to clear nsw flags
1186 // from this point on up.
1187 bool OpNoSignedWrap = BO->hasNoSignedWrap();
1188 NoSignedWrap &= OpNoSignedWrap;
1189 if (NoSignedWrap != OpNoSignedWrap) {
1190 BO->setHasNoSignedWrap(NoSignedWrap);
1191 Worklist.Add(Ancestor);
1193 } else if (Ancestor->getOpcode() == Instruction::Trunc) {
1194 // The fact that the descaled input to the trunc has smaller absolute
1195 // value than the original input doesn't tell us anything useful about
1196 // the absolute values of the truncations.
1197 NoSignedWrap = false;
1199 assert((Ancestor->getOpcode() != Instruction::SExt || NoSignedWrap) &&
1200 "Failed to keep proper track of nsw flags while drilling down?");
1202 if (Ancestor == Val)
1203 // Got to the top, all done!
1206 // Move up one level in the expression.
1207 assert(Ancestor->hasOneUse() && "Drilled down when more than one use!");
1208 Ancestor = Ancestor->user_back();
1212 /// \brief Creates node of binary operation with the same attributes as the
1213 /// specified one but with other operands.
1214 static Value *CreateBinOpAsGiven(BinaryOperator &Inst, Value *LHS, Value *RHS,
1215 InstCombiner::BuilderTy *B) {
1216 Value *BORes = B->CreateBinOp(Inst.getOpcode(), LHS, RHS);
1217 if (BinaryOperator *NewBO = dyn_cast<BinaryOperator>(BORes)) {
1218 if (isa<OverflowingBinaryOperator>(NewBO)) {
1219 NewBO->setHasNoSignedWrap(Inst.hasNoSignedWrap());
1220 NewBO->setHasNoUnsignedWrap(Inst.hasNoUnsignedWrap());
1222 if (isa<PossiblyExactOperator>(NewBO))
1223 NewBO->setIsExact(Inst.isExact());
1228 /// \brief Makes transformation of binary operation specific for vector types.
1229 /// \param Inst Binary operator to transform.
1230 /// \return Pointer to node that must replace the original binary operator, or
1231 /// null pointer if no transformation was made.
1232 Value *InstCombiner::SimplifyVectorOp(BinaryOperator &Inst) {
1233 if (!Inst.getType()->isVectorTy()) return nullptr;
1235 // It may not be safe to reorder shuffles and things like div, urem, etc.
1236 // because we may trap when executing those ops on unknown vector elements.
1238 if (!isSafeToSpeculativelyExecute(&Inst, DL)) return nullptr;
1240 unsigned VWidth = cast<VectorType>(Inst.getType())->getNumElements();
1241 Value *LHS = Inst.getOperand(0), *RHS = Inst.getOperand(1);
1242 assert(cast<VectorType>(LHS->getType())->getNumElements() == VWidth);
1243 assert(cast<VectorType>(RHS->getType())->getNumElements() == VWidth);
1245 // If both arguments of binary operation are shuffles, which use the same
1246 // mask and shuffle within a single vector, it is worthwhile to move the
1247 // shuffle after binary operation:
1248 // Op(shuffle(v1, m), shuffle(v2, m)) -> shuffle(Op(v1, v2), m)
1249 if (isa<ShuffleVectorInst>(LHS) && isa<ShuffleVectorInst>(RHS)) {
1250 ShuffleVectorInst *LShuf = cast<ShuffleVectorInst>(LHS);
1251 ShuffleVectorInst *RShuf = cast<ShuffleVectorInst>(RHS);
1252 if (isa<UndefValue>(LShuf->getOperand(1)) &&
1253 isa<UndefValue>(RShuf->getOperand(1)) &&
1254 LShuf->getOperand(0)->getType() == RShuf->getOperand(0)->getType() &&
1255 LShuf->getMask() == RShuf->getMask()) {
1256 Value *NewBO = CreateBinOpAsGiven(Inst, LShuf->getOperand(0),
1257 RShuf->getOperand(0), Builder);
1258 Value *Res = Builder->CreateShuffleVector(NewBO,
1259 UndefValue::get(NewBO->getType()), LShuf->getMask());
1264 // If one argument is a shuffle within one vector, the other is a constant,
1265 // try moving the shuffle after the binary operation.
1266 ShuffleVectorInst *Shuffle = nullptr;
1267 Constant *C1 = nullptr;
1268 if (isa<ShuffleVectorInst>(LHS)) Shuffle = cast<ShuffleVectorInst>(LHS);
1269 if (isa<ShuffleVectorInst>(RHS)) Shuffle = cast<ShuffleVectorInst>(RHS);
1270 if (isa<Constant>(LHS)) C1 = cast<Constant>(LHS);
1271 if (isa<Constant>(RHS)) C1 = cast<Constant>(RHS);
1272 if (Shuffle && C1 &&
1273 (isa<ConstantVector>(C1) || isa<ConstantDataVector>(C1)) &&
1274 isa<UndefValue>(Shuffle->getOperand(1)) &&
1275 Shuffle->getType() == Shuffle->getOperand(0)->getType()) {
1276 SmallVector<int, 16> ShMask = Shuffle->getShuffleMask();
1277 // Find constant C2 that has property:
1278 // shuffle(C2, ShMask) = C1
1279 // If such constant does not exist (example: ShMask=<0,0> and C1=<1,2>)
1280 // reorder is not possible.
1281 SmallVector<Constant*, 16> C2M(VWidth,
1282 UndefValue::get(C1->getType()->getScalarType()));
1283 bool MayChange = true;
1284 for (unsigned I = 0; I < VWidth; ++I) {
1285 if (ShMask[I] >= 0) {
1286 assert(ShMask[I] < (int)VWidth);
1287 if (!isa<UndefValue>(C2M[ShMask[I]])) {
1291 C2M[ShMask[I]] = C1->getAggregateElement(I);
1295 Constant *C2 = ConstantVector::get(C2M);
1296 Value *NewLHS, *NewRHS;
1297 if (isa<Constant>(LHS)) {
1299 NewRHS = Shuffle->getOperand(0);
1301 NewLHS = Shuffle->getOperand(0);
1304 Value *NewBO = CreateBinOpAsGiven(Inst, NewLHS, NewRHS, Builder);
1305 Value *Res = Builder->CreateShuffleVector(NewBO,
1306 UndefValue::get(Inst.getType()), Shuffle->getMask());
1314 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
1315 SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end());
1317 if (Value *V = SimplifyGEPInst(Ops, DL))
1318 return ReplaceInstUsesWith(GEP, V);
1320 Value *PtrOp = GEP.getOperand(0);
1322 // Eliminate unneeded casts for indices, and replace indices which displace
1323 // by multiples of a zero size type with zero.
1325 bool MadeChange = false;
1326 Type *IntPtrTy = DL->getIntPtrType(GEP.getPointerOperandType());
1328 gep_type_iterator GTI = gep_type_begin(GEP);
1329 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end();
1330 I != E; ++I, ++GTI) {
1331 // Skip indices into struct types.
1332 SequentialType *SeqTy = dyn_cast<SequentialType>(*GTI);
1333 if (!SeqTy) continue;
1335 // If the element type has zero size then any index over it is equivalent
1336 // to an index of zero, so replace it with zero if it is not zero already.
1337 if (SeqTy->getElementType()->isSized() &&
1338 DL->getTypeAllocSize(SeqTy->getElementType()) == 0)
1339 if (!isa<Constant>(*I) || !cast<Constant>(*I)->isNullValue()) {
1340 *I = Constant::getNullValue(IntPtrTy);
1344 Type *IndexTy = (*I)->getType();
1345 if (IndexTy != IntPtrTy) {
1346 // If we are using a wider index than needed for this platform, shrink
1347 // it to what we need. If narrower, sign-extend it to what we need.
1348 // This explicit cast can make subsequent optimizations more obvious.
1349 *I = Builder->CreateIntCast(*I, IntPtrTy, true);
1353 if (MadeChange) return &GEP;
1356 // Check to see if the inputs to the PHI node are getelementptr instructions.
1357 if (PHINode *PN = dyn_cast<PHINode>(PtrOp)) {
1358 GetElementPtrInst *Op1 = dyn_cast<GetElementPtrInst>(PN->getOperand(0));
1364 for (auto I = PN->op_begin()+1, E = PN->op_end(); I !=E; ++I) {
1365 GetElementPtrInst *Op2 = dyn_cast<GetElementPtrInst>(*I);
1366 if (!Op2 || Op1->getNumOperands() != Op2->getNumOperands())
1369 // Keep track of the type as we walk the GEP.
1370 Type *CurTy = Op1->getOperand(0)->getType()->getScalarType();
1372 for (unsigned J = 0, F = Op1->getNumOperands(); J != F; ++J) {
1373 if (Op1->getOperand(J)->getType() != Op2->getOperand(J)->getType())
1376 if (Op1->getOperand(J) != Op2->getOperand(J)) {
1378 // We have not seen any differences yet in the GEPs feeding the
1379 // PHI yet, so we record this one if it is allowed to be a
1382 // The first two arguments can vary for any GEP, the rest have to be
1383 // static for struct slots
1384 if (J > 1 && CurTy->isStructTy())
1389 // The GEP is different by more than one input. While this could be
1390 // extended to support GEPs that vary by more than one variable it
1391 // doesn't make sense since it greatly increases the complexity and
1392 // would result in an R+R+R addressing mode which no backend
1393 // directly supports and would need to be broken into several
1394 // simpler instructions anyway.
1399 // Sink down a layer of the type for the next iteration.
1401 if (CompositeType *CT = dyn_cast<CompositeType>(CurTy)) {
1402 CurTy = CT->getTypeAtIndex(Op1->getOperand(J));
1410 GetElementPtrInst *NewGEP = cast<GetElementPtrInst>(Op1->clone());
1413 // All the GEPs feeding the PHI are identical. Clone one down into our
1414 // BB so that it can be merged with the current GEP.
1415 GEP.getParent()->getInstList().insert(GEP.getParent()->getFirstNonPHI(),
1418 // All the GEPs feeding the PHI differ at a single offset. Clone a GEP
1419 // into the current block so it can be merged, and create a new PHI to
1421 Instruction *InsertPt = Builder->GetInsertPoint();
1422 Builder->SetInsertPoint(PN);
1423 PHINode *NewPN = Builder->CreatePHI(Op1->getOperand(DI)->getType(),
1424 PN->getNumOperands());
1425 Builder->SetInsertPoint(InsertPt);
1427 for (auto &I : PN->operands())
1428 NewPN->addIncoming(cast<GEPOperator>(I)->getOperand(DI),
1429 PN->getIncomingBlock(I));
1431 NewGEP->setOperand(DI, NewPN);
1432 GEP.getParent()->getInstList().insert(GEP.getParent()->getFirstNonPHI(),
1434 NewGEP->setOperand(DI, NewPN);
1437 GEP.setOperand(0, NewGEP);
1441 // Combine Indices - If the source pointer to this getelementptr instruction
1442 // is a getelementptr instruction, combine the indices of the two
1443 // getelementptr instructions into a single instruction.
1445 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) {
1446 if (!shouldMergeGEPs(*cast<GEPOperator>(&GEP), *Src))
1449 // Note that if our source is a gep chain itself then we wait for that
1450 // chain to be resolved before we perform this transformation. This
1451 // avoids us creating a TON of code in some cases.
1452 if (GEPOperator *SrcGEP =
1453 dyn_cast<GEPOperator>(Src->getOperand(0)))
1454 if (SrcGEP->getNumOperands() == 2 && shouldMergeGEPs(*Src, *SrcGEP))
1455 return nullptr; // Wait until our source is folded to completion.
1457 SmallVector<Value*, 8> Indices;
1459 // Find out whether the last index in the source GEP is a sequential idx.
1460 bool EndsWithSequential = false;
1461 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src);
1463 EndsWithSequential = !(*I)->isStructTy();
1465 // Can we combine the two pointer arithmetics offsets?
1466 if (EndsWithSequential) {
1467 // Replace: gep (gep %P, long B), long A, ...
1468 // With: T = long A+B; gep %P, T, ...
1471 Value *SO1 = Src->getOperand(Src->getNumOperands()-1);
1472 Value *GO1 = GEP.getOperand(1);
1473 if (SO1 == Constant::getNullValue(SO1->getType())) {
1475 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
1478 // If they aren't the same type, then the input hasn't been processed
1479 // by the loop above yet (which canonicalizes sequential index types to
1480 // intptr_t). Just avoid transforming this until the input has been
1482 if (SO1->getType() != GO1->getType())
1484 Sum = Builder->CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
1487 // Update the GEP in place if possible.
1488 if (Src->getNumOperands() == 2) {
1489 GEP.setOperand(0, Src->getOperand(0));
1490 GEP.setOperand(1, Sum);
1493 Indices.append(Src->op_begin()+1, Src->op_end()-1);
1494 Indices.push_back(Sum);
1495 Indices.append(GEP.op_begin()+2, GEP.op_end());
1496 } else if (isa<Constant>(*GEP.idx_begin()) &&
1497 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
1498 Src->getNumOperands() != 1) {
1499 // Otherwise we can do the fold if the first index of the GEP is a zero
1500 Indices.append(Src->op_begin()+1, Src->op_end());
1501 Indices.append(GEP.idx_begin()+1, GEP.idx_end());
1504 if (!Indices.empty())
1505 return (GEP.isInBounds() && Src->isInBounds()) ?
1506 GetElementPtrInst::CreateInBounds(Src->getOperand(0), Indices,
1508 GetElementPtrInst::Create(Src->getOperand(0), Indices, GEP.getName());
1511 if (DL && GEP.getNumIndices() == 1) {
1512 unsigned AS = GEP.getPointerAddressSpace();
1513 if (GEP.getOperand(1)->getType()->getScalarSizeInBits() ==
1514 DL->getPointerSizeInBits(AS)) {
1515 Type *PtrTy = GEP.getPointerOperandType();
1516 Type *Ty = PtrTy->getPointerElementType();
1517 uint64_t TyAllocSize = DL->getTypeAllocSize(Ty);
1519 bool Matched = false;
1522 if (TyAllocSize == 1) {
1523 V = GEP.getOperand(1);
1525 } else if (match(GEP.getOperand(1),
1526 m_AShr(m_Value(V), m_ConstantInt(C)))) {
1527 if (TyAllocSize == 1ULL << C)
1529 } else if (match(GEP.getOperand(1),
1530 m_SDiv(m_Value(V), m_ConstantInt(C)))) {
1531 if (TyAllocSize == C)
1536 // Canonicalize (gep i8* X, -(ptrtoint Y))
1537 // to (inttoptr (sub (ptrtoint X), (ptrtoint Y)))
1538 // The GEP pattern is emitted by the SCEV expander for certain kinds of
1539 // pointer arithmetic.
1540 if (match(V, m_Neg(m_PtrToInt(m_Value())))) {
1541 Operator *Index = cast<Operator>(V);
1542 Value *PtrToInt = Builder->CreatePtrToInt(PtrOp, Index->getType());
1543 Value *NewSub = Builder->CreateSub(PtrToInt, Index->getOperand(1));
1544 return CastInst::Create(Instruction::IntToPtr, NewSub, GEP.getType());
1546 // Canonicalize (gep i8* X, (ptrtoint Y)-(ptrtoint X))
1549 if (match(V, m_Sub(m_PtrToInt(m_Value(Y)),
1550 m_PtrToInt(m_Specific(GEP.getOperand(0)))))) {
1551 return CastInst::CreatePointerBitCastOrAddrSpaceCast(Y,
1558 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0).
1559 Value *StrippedPtr = PtrOp->stripPointerCasts();
1560 PointerType *StrippedPtrTy = dyn_cast<PointerType>(StrippedPtr->getType());
1562 // We do not handle pointer-vector geps here.
1566 if (StrippedPtr != PtrOp) {
1567 bool HasZeroPointerIndex = false;
1568 if (ConstantInt *C = dyn_cast<ConstantInt>(GEP.getOperand(1)))
1569 HasZeroPointerIndex = C->isZero();
1571 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
1572 // into : GEP [10 x i8]* X, i32 0, ...
1574 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
1575 // into : GEP i8* X, ...
1577 // This occurs when the program declares an array extern like "int X[];"
1578 if (HasZeroPointerIndex) {
1579 PointerType *CPTy = cast<PointerType>(PtrOp->getType());
1580 if (ArrayType *CATy =
1581 dyn_cast<ArrayType>(CPTy->getElementType())) {
1582 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
1583 if (CATy->getElementType() == StrippedPtrTy->getElementType()) {
1584 // -> GEP i8* X, ...
1585 SmallVector<Value*, 8> Idx(GEP.idx_begin()+1, GEP.idx_end());
1586 GetElementPtrInst *Res =
1587 GetElementPtrInst::Create(StrippedPtr, Idx, GEP.getName());
1588 Res->setIsInBounds(GEP.isInBounds());
1589 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace())
1591 // Insert Res, and create an addrspacecast.
1593 // GEP (addrspacecast i8 addrspace(1)* X to [0 x i8]*), i32 0, ...
1595 // %0 = GEP i8 addrspace(1)* X, ...
1596 // addrspacecast i8 addrspace(1)* %0 to i8*
1597 return new AddrSpaceCastInst(Builder->Insert(Res), GEP.getType());
1600 if (ArrayType *XATy =
1601 dyn_cast<ArrayType>(StrippedPtrTy->getElementType())){
1602 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
1603 if (CATy->getElementType() == XATy->getElementType()) {
1604 // -> GEP [10 x i8]* X, i32 0, ...
1605 // At this point, we know that the cast source type is a pointer
1606 // to an array of the same type as the destination pointer
1607 // array. Because the array type is never stepped over (there
1608 // is a leading zero) we can fold the cast into this GEP.
1609 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace()) {
1610 GEP.setOperand(0, StrippedPtr);
1613 // Cannot replace the base pointer directly because StrippedPtr's
1614 // address space is different. Instead, create a new GEP followed by
1615 // an addrspacecast.
1617 // GEP (addrspacecast [10 x i8] addrspace(1)* X to [0 x i8]*),
1620 // %0 = GEP [10 x i8] addrspace(1)* X, ...
1621 // addrspacecast i8 addrspace(1)* %0 to i8*
1622 SmallVector<Value*, 8> Idx(GEP.idx_begin(), GEP.idx_end());
1623 Value *NewGEP = GEP.isInBounds() ?
1624 Builder->CreateInBoundsGEP(StrippedPtr, Idx, GEP.getName()) :
1625 Builder->CreateGEP(StrippedPtr, Idx, GEP.getName());
1626 return new AddrSpaceCastInst(NewGEP, GEP.getType());
1630 } else if (GEP.getNumOperands() == 2) {
1631 // Transform things like:
1632 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
1633 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
1634 Type *SrcElTy = StrippedPtrTy->getElementType();
1635 Type *ResElTy = PtrOp->getType()->getPointerElementType();
1636 if (DL && SrcElTy->isArrayTy() &&
1637 DL->getTypeAllocSize(SrcElTy->getArrayElementType()) ==
1638 DL->getTypeAllocSize(ResElTy)) {
1639 Type *IdxType = DL->getIntPtrType(GEP.getType());
1640 Value *Idx[2] = { Constant::getNullValue(IdxType), GEP.getOperand(1) };
1641 Value *NewGEP = GEP.isInBounds() ?
1642 Builder->CreateInBoundsGEP(StrippedPtr, Idx, GEP.getName()) :
1643 Builder->CreateGEP(StrippedPtr, Idx, GEP.getName());
1645 // V and GEP are both pointer types --> BitCast
1646 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
1650 // Transform things like:
1651 // %V = mul i64 %N, 4
1652 // %t = getelementptr i8* bitcast (i32* %arr to i8*), i32 %V
1653 // into: %t1 = getelementptr i32* %arr, i32 %N; bitcast
1654 if (DL && ResElTy->isSized() && SrcElTy->isSized()) {
1655 // Check that changing the type amounts to dividing the index by a scale
1657 uint64_t ResSize = DL->getTypeAllocSize(ResElTy);
1658 uint64_t SrcSize = DL->getTypeAllocSize(SrcElTy);
1659 if (ResSize && SrcSize % ResSize == 0) {
1660 Value *Idx = GEP.getOperand(1);
1661 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
1662 uint64_t Scale = SrcSize / ResSize;
1664 // Earlier transforms ensure that the index has type IntPtrType, which
1665 // considerably simplifies the logic by eliminating implicit casts.
1666 assert(Idx->getType() == DL->getIntPtrType(GEP.getType()) &&
1667 "Index not cast to pointer width?");
1670 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
1671 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
1672 // If the multiplication NewIdx * Scale may overflow then the new
1673 // GEP may not be "inbounds".
1674 Value *NewGEP = GEP.isInBounds() && NSW ?
1675 Builder->CreateInBoundsGEP(StrippedPtr, NewIdx, GEP.getName()) :
1676 Builder->CreateGEP(StrippedPtr, NewIdx, GEP.getName());
1678 // The NewGEP must be pointer typed, so must the old one -> BitCast
1679 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
1685 // Similarly, transform things like:
1686 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
1687 // (where tmp = 8*tmp2) into:
1688 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
1689 if (DL && ResElTy->isSized() && SrcElTy->isSized() &&
1690 SrcElTy->isArrayTy()) {
1691 // Check that changing to the array element type amounts to dividing the
1692 // index by a scale factor.
1693 uint64_t ResSize = DL->getTypeAllocSize(ResElTy);
1694 uint64_t ArrayEltSize
1695 = DL->getTypeAllocSize(SrcElTy->getArrayElementType());
1696 if (ResSize && ArrayEltSize % ResSize == 0) {
1697 Value *Idx = GEP.getOperand(1);
1698 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
1699 uint64_t Scale = ArrayEltSize / ResSize;
1701 // Earlier transforms ensure that the index has type IntPtrType, which
1702 // considerably simplifies the logic by eliminating implicit casts.
1703 assert(Idx->getType() == DL->getIntPtrType(GEP.getType()) &&
1704 "Index not cast to pointer width?");
1707 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
1708 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
1709 // If the multiplication NewIdx * Scale may overflow then the new
1710 // GEP may not be "inbounds".
1712 Constant::getNullValue(DL->getIntPtrType(GEP.getType())),
1716 Value *NewGEP = GEP.isInBounds() && NSW ?
1717 Builder->CreateInBoundsGEP(StrippedPtr, Off, GEP.getName()) :
1718 Builder->CreateGEP(StrippedPtr, Off, GEP.getName());
1719 // The NewGEP must be pointer typed, so must the old one -> BitCast
1720 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
1731 // addrspacecast between types is canonicalized as a bitcast, then an
1732 // addrspacecast. To take advantage of the below bitcast + struct GEP, look
1733 // through the addrspacecast.
1734 if (AddrSpaceCastInst *ASC = dyn_cast<AddrSpaceCastInst>(PtrOp)) {
1735 // X = bitcast A addrspace(1)* to B addrspace(1)*
1736 // Y = addrspacecast A addrspace(1)* to B addrspace(2)*
1737 // Z = gep Y, <...constant indices...>
1738 // Into an addrspacecasted GEP of the struct.
1739 if (BitCastInst *BC = dyn_cast<BitCastInst>(ASC->getOperand(0)))
1743 /// See if we can simplify:
1744 /// X = bitcast A* to B*
1745 /// Y = gep X, <...constant indices...>
1746 /// into a gep of the original struct. This is important for SROA and alias
1747 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
1748 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
1749 Value *Operand = BCI->getOperand(0);
1750 PointerType *OpType = cast<PointerType>(Operand->getType());
1751 unsigned OffsetBits = DL->getPointerTypeSizeInBits(GEP.getType());
1752 APInt Offset(OffsetBits, 0);
1753 if (!isa<BitCastInst>(Operand) &&
1754 GEP.accumulateConstantOffset(*DL, Offset)) {
1756 // If this GEP instruction doesn't move the pointer, just replace the GEP
1757 // with a bitcast of the real input to the dest type.
1759 // If the bitcast is of an allocation, and the allocation will be
1760 // converted to match the type of the cast, don't touch this.
1761 if (isa<AllocaInst>(Operand) || isAllocationFn(Operand, TLI)) {
1762 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
1763 if (Instruction *I = visitBitCast(*BCI)) {
1766 BCI->getParent()->getInstList().insert(BCI, I);
1767 ReplaceInstUsesWith(*BCI, I);
1773 if (Operand->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
1774 return new AddrSpaceCastInst(Operand, GEP.getType());
1775 return new BitCastInst(Operand, GEP.getType());
1778 // Otherwise, if the offset is non-zero, we need to find out if there is a
1779 // field at Offset in 'A's type. If so, we can pull the cast through the
1781 SmallVector<Value*, 8> NewIndices;
1782 if (FindElementAtOffset(OpType, Offset.getSExtValue(), NewIndices)) {
1783 Value *NGEP = GEP.isInBounds() ?
1784 Builder->CreateInBoundsGEP(Operand, NewIndices) :
1785 Builder->CreateGEP(Operand, NewIndices);
1787 if (NGEP->getType() == GEP.getType())
1788 return ReplaceInstUsesWith(GEP, NGEP);
1789 NGEP->takeName(&GEP);
1791 if (NGEP->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
1792 return new AddrSpaceCastInst(NGEP, GEP.getType());
1793 return new BitCastInst(NGEP, GEP.getType());
1802 isAllocSiteRemovable(Instruction *AI, SmallVectorImpl<WeakVH> &Users,
1803 const TargetLibraryInfo *TLI) {
1804 SmallVector<Instruction*, 4> Worklist;
1805 Worklist.push_back(AI);
1808 Instruction *PI = Worklist.pop_back_val();
1809 for (User *U : PI->users()) {
1810 Instruction *I = cast<Instruction>(U);
1811 switch (I->getOpcode()) {
1813 // Give up the moment we see something we can't handle.
1816 case Instruction::BitCast:
1817 case Instruction::GetElementPtr:
1819 Worklist.push_back(I);
1822 case Instruction::ICmp: {
1823 ICmpInst *ICI = cast<ICmpInst>(I);
1824 // We can fold eq/ne comparisons with null to false/true, respectively.
1825 if (!ICI->isEquality() || !isa<ConstantPointerNull>(ICI->getOperand(1)))
1831 case Instruction::Call:
1832 // Ignore no-op and store intrinsics.
1833 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1834 switch (II->getIntrinsicID()) {
1838 case Intrinsic::memmove:
1839 case Intrinsic::memcpy:
1840 case Intrinsic::memset: {
1841 MemIntrinsic *MI = cast<MemIntrinsic>(II);
1842 if (MI->isVolatile() || MI->getRawDest() != PI)
1846 case Intrinsic::dbg_declare:
1847 case Intrinsic::dbg_value:
1848 case Intrinsic::invariant_start:
1849 case Intrinsic::invariant_end:
1850 case Intrinsic::lifetime_start:
1851 case Intrinsic::lifetime_end:
1852 case Intrinsic::objectsize:
1858 if (isFreeCall(I, TLI)) {
1864 case Instruction::Store: {
1865 StoreInst *SI = cast<StoreInst>(I);
1866 if (SI->isVolatile() || SI->getPointerOperand() != PI)
1872 llvm_unreachable("missing a return?");
1874 } while (!Worklist.empty());
1878 Instruction *InstCombiner::visitAllocSite(Instruction &MI) {
1879 // If we have a malloc call which is only used in any amount of comparisons
1880 // to null and free calls, delete the calls and replace the comparisons with
1881 // true or false as appropriate.
1882 SmallVector<WeakVH, 64> Users;
1883 if (isAllocSiteRemovable(&MI, Users, TLI)) {
1884 for (unsigned i = 0, e = Users.size(); i != e; ++i) {
1885 Instruction *I = cast_or_null<Instruction>(&*Users[i]);
1888 if (ICmpInst *C = dyn_cast<ICmpInst>(I)) {
1889 ReplaceInstUsesWith(*C,
1890 ConstantInt::get(Type::getInt1Ty(C->getContext()),
1891 C->isFalseWhenEqual()));
1892 } else if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) {
1893 ReplaceInstUsesWith(*I, UndefValue::get(I->getType()));
1894 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1895 if (II->getIntrinsicID() == Intrinsic::objectsize) {
1896 ConstantInt *CI = cast<ConstantInt>(II->getArgOperand(1));
1897 uint64_t DontKnow = CI->isZero() ? -1ULL : 0;
1898 ReplaceInstUsesWith(*I, ConstantInt::get(I->getType(), DontKnow));
1901 EraseInstFromFunction(*I);
1904 if (InvokeInst *II = dyn_cast<InvokeInst>(&MI)) {
1905 // Replace invoke with a NOP intrinsic to maintain the original CFG
1906 Module *M = II->getParent()->getParent()->getParent();
1907 Function *F = Intrinsic::getDeclaration(M, Intrinsic::donothing);
1908 InvokeInst::Create(F, II->getNormalDest(), II->getUnwindDest(),
1909 None, "", II->getParent());
1911 return EraseInstFromFunction(MI);
1916 /// \brief Move the call to free before a NULL test.
1918 /// Check if this free is accessed after its argument has been test
1919 /// against NULL (property 0).
1920 /// If yes, it is legal to move this call in its predecessor block.
1922 /// The move is performed only if the block containing the call to free
1923 /// will be removed, i.e.:
1924 /// 1. it has only one predecessor P, and P has two successors
1925 /// 2. it contains the call and an unconditional branch
1926 /// 3. its successor is the same as its predecessor's successor
1928 /// The profitability is out-of concern here and this function should
1929 /// be called only if the caller knows this transformation would be
1930 /// profitable (e.g., for code size).
1931 static Instruction *
1932 tryToMoveFreeBeforeNullTest(CallInst &FI) {
1933 Value *Op = FI.getArgOperand(0);
1934 BasicBlock *FreeInstrBB = FI.getParent();
1935 BasicBlock *PredBB = FreeInstrBB->getSinglePredecessor();
1937 // Validate part of constraint #1: Only one predecessor
1938 // FIXME: We can extend the number of predecessor, but in that case, we
1939 // would duplicate the call to free in each predecessor and it may
1940 // not be profitable even for code size.
1944 // Validate constraint #2: Does this block contains only the call to
1945 // free and an unconditional branch?
1946 // FIXME: We could check if we can speculate everything in the
1947 // predecessor block
1948 if (FreeInstrBB->size() != 2)
1951 if (!match(FreeInstrBB->getTerminator(), m_UnconditionalBr(SuccBB)))
1954 // Validate the rest of constraint #1 by matching on the pred branch.
1955 TerminatorInst *TI = PredBB->getTerminator();
1956 BasicBlock *TrueBB, *FalseBB;
1957 ICmpInst::Predicate Pred;
1958 if (!match(TI, m_Br(m_ICmp(Pred, m_Specific(Op), m_Zero()), TrueBB, FalseBB)))
1960 if (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE)
1963 // Validate constraint #3: Ensure the null case just falls through.
1964 if (SuccBB != (Pred == ICmpInst::ICMP_EQ ? TrueBB : FalseBB))
1966 assert(FreeInstrBB == (Pred == ICmpInst::ICMP_EQ ? FalseBB : TrueBB) &&
1967 "Broken CFG: missing edge from predecessor to successor");
1974 Instruction *InstCombiner::visitFree(CallInst &FI) {
1975 Value *Op = FI.getArgOperand(0);
1977 // free undef -> unreachable.
1978 if (isa<UndefValue>(Op)) {
1979 // Insert a new store to null because we cannot modify the CFG here.
1980 Builder->CreateStore(ConstantInt::getTrue(FI.getContext()),
1981 UndefValue::get(Type::getInt1PtrTy(FI.getContext())));
1982 return EraseInstFromFunction(FI);
1985 // If we have 'free null' delete the instruction. This can happen in stl code
1986 // when lots of inlining happens.
1987 if (isa<ConstantPointerNull>(Op))
1988 return EraseInstFromFunction(FI);
1990 // If we optimize for code size, try to move the call to free before the null
1991 // test so that simplify cfg can remove the empty block and dead code
1992 // elimination the branch. I.e., helps to turn something like:
1993 // if (foo) free(foo);
1997 if (Instruction *I = tryToMoveFreeBeforeNullTest(FI))
2005 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
2006 // Change br (not X), label True, label False to: br X, label False, True
2008 BasicBlock *TrueDest;
2009 BasicBlock *FalseDest;
2010 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
2011 !isa<Constant>(X)) {
2012 // Swap Destinations and condition...
2014 BI.swapSuccessors();
2018 // Canonicalize fcmp_one -> fcmp_oeq
2019 FCmpInst::Predicate FPred; Value *Y;
2020 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
2021 TrueDest, FalseDest)) &&
2022 BI.getCondition()->hasOneUse())
2023 if (FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
2024 FPred == FCmpInst::FCMP_OGE) {
2025 FCmpInst *Cond = cast<FCmpInst>(BI.getCondition());
2026 Cond->setPredicate(FCmpInst::getInversePredicate(FPred));
2028 // Swap Destinations and condition.
2029 BI.swapSuccessors();
2034 // Canonicalize icmp_ne -> icmp_eq
2035 ICmpInst::Predicate IPred;
2036 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
2037 TrueDest, FalseDest)) &&
2038 BI.getCondition()->hasOneUse())
2039 if (IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
2040 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
2041 IPred == ICmpInst::ICMP_SGE) {
2042 ICmpInst *Cond = cast<ICmpInst>(BI.getCondition());
2043 Cond->setPredicate(ICmpInst::getInversePredicate(IPred));
2044 // Swap Destinations and condition.
2045 BI.swapSuccessors();
2053 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
2054 Value *Cond = SI.getCondition();
2055 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
2056 if (I->getOpcode() == Instruction::Add)
2057 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2058 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
2059 // Skip the first item since that's the default case.
2060 for (SwitchInst::CaseIt i = SI.case_begin(), e = SI.case_end();
2062 ConstantInt* CaseVal = i.getCaseValue();
2063 Constant* NewCaseVal = ConstantExpr::getSub(cast<Constant>(CaseVal),
2065 assert(isa<ConstantInt>(NewCaseVal) &&
2066 "Result of expression should be constant");
2067 i.setValue(cast<ConstantInt>(NewCaseVal));
2069 SI.setCondition(I->getOperand(0));
2077 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
2078 Value *Agg = EV.getAggregateOperand();
2080 if (!EV.hasIndices())
2081 return ReplaceInstUsesWith(EV, Agg);
2083 if (Constant *C = dyn_cast<Constant>(Agg)) {
2084 if (Constant *C2 = C->getAggregateElement(*EV.idx_begin())) {
2085 if (EV.getNumIndices() == 0)
2086 return ReplaceInstUsesWith(EV, C2);
2087 // Extract the remaining indices out of the constant indexed by the
2089 return ExtractValueInst::Create(C2, EV.getIndices().slice(1));
2091 return nullptr; // Can't handle other constants
2094 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
2095 // We're extracting from an insertvalue instruction, compare the indices
2096 const unsigned *exti, *exte, *insi, *inse;
2097 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
2098 exte = EV.idx_end(), inse = IV->idx_end();
2099 exti != exte && insi != inse;
2102 // The insert and extract both reference distinctly different elements.
2103 // This means the extract is not influenced by the insert, and we can
2104 // replace the aggregate operand of the extract with the aggregate
2105 // operand of the insert. i.e., replace
2106 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
2107 // %E = extractvalue { i32, { i32 } } %I, 0
2109 // %E = extractvalue { i32, { i32 } } %A, 0
2110 return ExtractValueInst::Create(IV->getAggregateOperand(),
2113 if (exti == exte && insi == inse)
2114 // Both iterators are at the end: Index lists are identical. Replace
2115 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
2116 // %C = extractvalue { i32, { i32 } } %B, 1, 0
2118 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
2120 // The extract list is a prefix of the insert list. i.e. replace
2121 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
2122 // %E = extractvalue { i32, { i32 } } %I, 1
2124 // %X = extractvalue { i32, { i32 } } %A, 1
2125 // %E = insertvalue { i32 } %X, i32 42, 0
2126 // by switching the order of the insert and extract (though the
2127 // insertvalue should be left in, since it may have other uses).
2128 Value *NewEV = Builder->CreateExtractValue(IV->getAggregateOperand(),
2130 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
2131 makeArrayRef(insi, inse));
2134 // The insert list is a prefix of the extract list
2135 // We can simply remove the common indices from the extract and make it
2136 // operate on the inserted value instead of the insertvalue result.
2138 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
2139 // %E = extractvalue { i32, { i32 } } %I, 1, 0
2141 // %E extractvalue { i32 } { i32 42 }, 0
2142 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
2143 makeArrayRef(exti, exte));
2145 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) {
2146 // We're extracting from an intrinsic, see if we're the only user, which
2147 // allows us to simplify multiple result intrinsics to simpler things that
2148 // just get one value.
2149 if (II->hasOneUse()) {
2150 // Check if we're grabbing the overflow bit or the result of a 'with
2151 // overflow' intrinsic. If it's the latter we can remove the intrinsic
2152 // and replace it with a traditional binary instruction.
2153 switch (II->getIntrinsicID()) {
2154 case Intrinsic::uadd_with_overflow:
2155 case Intrinsic::sadd_with_overflow:
2156 if (*EV.idx_begin() == 0) { // Normal result.
2157 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2158 ReplaceInstUsesWith(*II, UndefValue::get(II->getType()));
2159 EraseInstFromFunction(*II);
2160 return BinaryOperator::CreateAdd(LHS, RHS);
2163 // If the normal result of the add is dead, and the RHS is a constant,
2164 // we can transform this into a range comparison.
2165 // overflow = uadd a, -4 --> overflow = icmp ugt a, 3
2166 if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow)
2167 if (ConstantInt *CI = dyn_cast<ConstantInt>(II->getArgOperand(1)))
2168 return new ICmpInst(ICmpInst::ICMP_UGT, II->getArgOperand(0),
2169 ConstantExpr::getNot(CI));
2171 case Intrinsic::usub_with_overflow:
2172 case Intrinsic::ssub_with_overflow:
2173 if (*EV.idx_begin() == 0) { // Normal result.
2174 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2175 ReplaceInstUsesWith(*II, UndefValue::get(II->getType()));
2176 EraseInstFromFunction(*II);
2177 return BinaryOperator::CreateSub(LHS, RHS);
2180 case Intrinsic::umul_with_overflow:
2181 case Intrinsic::smul_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::CreateMul(LHS, RHS);
2194 if (LoadInst *L = dyn_cast<LoadInst>(Agg))
2195 // If the (non-volatile) load only has one use, we can rewrite this to a
2196 // load from a GEP. This reduces the size of the load.
2197 // FIXME: If a load is used only by extractvalue instructions then this
2198 // could be done regardless of having multiple uses.
2199 if (L->isSimple() && L->hasOneUse()) {
2200 // extractvalue has integer indices, getelementptr has Value*s. Convert.
2201 SmallVector<Value*, 4> Indices;
2202 // Prefix an i32 0 since we need the first element.
2203 Indices.push_back(Builder->getInt32(0));
2204 for (ExtractValueInst::idx_iterator I = EV.idx_begin(), E = EV.idx_end();
2206 Indices.push_back(Builder->getInt32(*I));
2208 // We need to insert these at the location of the old load, not at that of
2209 // the extractvalue.
2210 Builder->SetInsertPoint(L->getParent(), L);
2211 Value *GEP = Builder->CreateInBoundsGEP(L->getPointerOperand(), Indices);
2212 // Returning the load directly will cause the main loop to insert it in
2213 // the wrong spot, so use ReplaceInstUsesWith().
2214 return ReplaceInstUsesWith(EV, Builder->CreateLoad(GEP));
2216 // We could simplify extracts from other values. Note that nested extracts may
2217 // already be simplified implicitly by the above: extract (extract (insert) )
2218 // will be translated into extract ( insert ( extract ) ) first and then just
2219 // the value inserted, if appropriate. Similarly for extracts from single-use
2220 // loads: extract (extract (load)) will be translated to extract (load (gep))
2221 // and if again single-use then via load (gep (gep)) to load (gep).
2222 // However, double extracts from e.g. function arguments or return values
2223 // aren't handled yet.
2227 enum Personality_Type {
2228 Unknown_Personality,
2229 GNU_Ada_Personality,
2230 GNU_CXX_Personality,
2231 GNU_ObjC_Personality
2234 /// RecognizePersonality - See if the given exception handling personality
2235 /// function is one that we understand. If so, return a description of it;
2236 /// otherwise return Unknown_Personality.
2237 static Personality_Type RecognizePersonality(Value *Pers) {
2238 Function *F = dyn_cast<Function>(Pers->stripPointerCasts());
2240 return Unknown_Personality;
2241 return StringSwitch<Personality_Type>(F->getName())
2242 .Case("__gnat_eh_personality", GNU_Ada_Personality)
2243 .Case("__gxx_personality_v0", GNU_CXX_Personality)
2244 .Case("__objc_personality_v0", GNU_ObjC_Personality)
2245 .Default(Unknown_Personality);
2248 /// isCatchAll - Return 'true' if the given typeinfo will match anything.
2249 static bool isCatchAll(Personality_Type Personality, Constant *TypeInfo) {
2250 switch (Personality) {
2251 case Unknown_Personality:
2253 case GNU_Ada_Personality:
2254 // While __gnat_all_others_value will match any Ada exception, it doesn't
2255 // match foreign exceptions (or didn't, before gcc-4.7).
2257 case GNU_CXX_Personality:
2258 case GNU_ObjC_Personality:
2259 return TypeInfo->isNullValue();
2261 llvm_unreachable("Unknown personality!");
2264 static bool shorter_filter(const Value *LHS, const Value *RHS) {
2266 cast<ArrayType>(LHS->getType())->getNumElements()
2268 cast<ArrayType>(RHS->getType())->getNumElements();
2271 Instruction *InstCombiner::visitLandingPadInst(LandingPadInst &LI) {
2272 // The logic here should be correct for any real-world personality function.
2273 // However if that turns out not to be true, the offending logic can always
2274 // be conditioned on the personality function, like the catch-all logic is.
2275 Personality_Type Personality = RecognizePersonality(LI.getPersonalityFn());
2277 // Simplify the list of clauses, eg by removing repeated catch clauses
2278 // (these are often created by inlining).
2279 bool MakeNewInstruction = false; // If true, recreate using the following:
2280 SmallVector<Constant *, 16> NewClauses; // - Clauses for the new instruction;
2281 bool CleanupFlag = LI.isCleanup(); // - The new instruction is a cleanup.
2283 SmallPtrSet<Value *, 16> AlreadyCaught; // Typeinfos known caught already.
2284 for (unsigned i = 0, e = LI.getNumClauses(); i != e; ++i) {
2285 bool isLastClause = i + 1 == e;
2286 if (LI.isCatch(i)) {
2288 Constant *CatchClause = LI.getClause(i);
2289 Constant *TypeInfo = CatchClause->stripPointerCasts();
2291 // If we already saw this clause, there is no point in having a second
2293 if (AlreadyCaught.insert(TypeInfo)) {
2294 // This catch clause was not already seen.
2295 NewClauses.push_back(CatchClause);
2297 // Repeated catch clause - drop the redundant copy.
2298 MakeNewInstruction = true;
2301 // If this is a catch-all then there is no point in keeping any following
2302 // clauses or marking the landingpad as having a cleanup.
2303 if (isCatchAll(Personality, TypeInfo)) {
2305 MakeNewInstruction = true;
2306 CleanupFlag = false;
2310 // A filter clause. If any of the filter elements were already caught
2311 // then they can be dropped from the filter. It is tempting to try to
2312 // exploit the filter further by saying that any typeinfo that does not
2313 // occur in the filter can't be caught later (and thus can be dropped).
2314 // However this would be wrong, since typeinfos can match without being
2315 // equal (for example if one represents a C++ class, and the other some
2316 // class derived from it).
2317 assert(LI.isFilter(i) && "Unsupported landingpad clause!");
2318 Constant *FilterClause = LI.getClause(i);
2319 ArrayType *FilterType = cast<ArrayType>(FilterClause->getType());
2320 unsigned NumTypeInfos = FilterType->getNumElements();
2322 // An empty filter catches everything, so there is no point in keeping any
2323 // following clauses or marking the landingpad as having a cleanup. By
2324 // dealing with this case here the following code is made a bit simpler.
2325 if (!NumTypeInfos) {
2326 NewClauses.push_back(FilterClause);
2328 MakeNewInstruction = true;
2329 CleanupFlag = false;
2333 bool MakeNewFilter = false; // If true, make a new filter.
2334 SmallVector<Constant *, 16> NewFilterElts; // New elements.
2335 if (isa<ConstantAggregateZero>(FilterClause)) {
2336 // Not an empty filter - it contains at least one null typeinfo.
2337 assert(NumTypeInfos > 0 && "Should have handled empty filter already!");
2338 Constant *TypeInfo =
2339 Constant::getNullValue(FilterType->getElementType());
2340 // If this typeinfo is a catch-all then the filter can never match.
2341 if (isCatchAll(Personality, TypeInfo)) {
2342 // Throw the filter away.
2343 MakeNewInstruction = true;
2347 // There is no point in having multiple copies of this typeinfo, so
2348 // discard all but the first copy if there is more than one.
2349 NewFilterElts.push_back(TypeInfo);
2350 if (NumTypeInfos > 1)
2351 MakeNewFilter = true;
2353 ConstantArray *Filter = cast<ConstantArray>(FilterClause);
2354 SmallPtrSet<Value *, 16> SeenInFilter; // For uniquing the elements.
2355 NewFilterElts.reserve(NumTypeInfos);
2357 // Remove any filter elements that were already caught or that already
2358 // occurred in the filter. While there, see if any of the elements are
2359 // catch-alls. If so, the filter can be discarded.
2360 bool SawCatchAll = false;
2361 for (unsigned j = 0; j != NumTypeInfos; ++j) {
2362 Constant *Elt = Filter->getOperand(j);
2363 Constant *TypeInfo = Elt->stripPointerCasts();
2364 if (isCatchAll(Personality, TypeInfo)) {
2365 // This element is a catch-all. Bail out, noting this fact.
2369 if (AlreadyCaught.count(TypeInfo))
2370 // Already caught by an earlier clause, so having it in the filter
2373 // There is no point in having multiple copies of the same typeinfo in
2374 // a filter, so only add it if we didn't already.
2375 if (SeenInFilter.insert(TypeInfo))
2376 NewFilterElts.push_back(cast<Constant>(Elt));
2378 // A filter containing a catch-all cannot match anything by definition.
2380 // Throw the filter away.
2381 MakeNewInstruction = true;
2385 // If we dropped something from the filter, make a new one.
2386 if (NewFilterElts.size() < NumTypeInfos)
2387 MakeNewFilter = true;
2389 if (MakeNewFilter) {
2390 FilterType = ArrayType::get(FilterType->getElementType(),
2391 NewFilterElts.size());
2392 FilterClause = ConstantArray::get(FilterType, NewFilterElts);
2393 MakeNewInstruction = true;
2396 NewClauses.push_back(FilterClause);
2398 // If the new filter is empty then it will catch everything so there is
2399 // no point in keeping any following clauses or marking the landingpad
2400 // as having a cleanup. The case of the original filter being empty was
2401 // already handled above.
2402 if (MakeNewFilter && !NewFilterElts.size()) {
2403 assert(MakeNewInstruction && "New filter but not a new instruction!");
2404 CleanupFlag = false;
2410 // If several filters occur in a row then reorder them so that the shortest
2411 // filters come first (those with the smallest number of elements). This is
2412 // advantageous because shorter filters are more likely to match, speeding up
2413 // unwinding, but mostly because it increases the effectiveness of the other
2414 // filter optimizations below.
2415 for (unsigned i = 0, e = NewClauses.size(); i + 1 < e; ) {
2417 // Find the maximal 'j' s.t. the range [i, j) consists entirely of filters.
2418 for (j = i; j != e; ++j)
2419 if (!isa<ArrayType>(NewClauses[j]->getType()))
2422 // Check whether the filters are already sorted by length. We need to know
2423 // if sorting them is actually going to do anything so that we only make a
2424 // new landingpad instruction if it does.
2425 for (unsigned k = i; k + 1 < j; ++k)
2426 if (shorter_filter(NewClauses[k+1], NewClauses[k])) {
2427 // Not sorted, so sort the filters now. Doing an unstable sort would be
2428 // correct too but reordering filters pointlessly might confuse users.
2429 std::stable_sort(NewClauses.begin() + i, NewClauses.begin() + j,
2431 MakeNewInstruction = true;
2435 // Look for the next batch of filters.
2439 // If typeinfos matched if and only if equal, then the elements of a filter L
2440 // that occurs later than a filter F could be replaced by the intersection of
2441 // the elements of F and L. In reality two typeinfos can match without being
2442 // equal (for example if one represents a C++ class, and the other some class
2443 // derived from it) so it would be wrong to perform this transform in general.
2444 // However the transform is correct and useful if F is a subset of L. In that
2445 // case L can be replaced by F, and thus removed altogether since repeating a
2446 // filter is pointless. So here we look at all pairs of filters F and L where
2447 // L follows F in the list of clauses, and remove L if every element of F is
2448 // an element of L. This can occur when inlining C++ functions with exception
2450 for (unsigned i = 0; i + 1 < NewClauses.size(); ++i) {
2451 // Examine each filter in turn.
2452 Value *Filter = NewClauses[i];
2453 ArrayType *FTy = dyn_cast<ArrayType>(Filter->getType());
2455 // Not a filter - skip it.
2457 unsigned FElts = FTy->getNumElements();
2458 // Examine each filter following this one. Doing this backwards means that
2459 // we don't have to worry about filters disappearing under us when removed.
2460 for (unsigned j = NewClauses.size() - 1; j != i; --j) {
2461 Value *LFilter = NewClauses[j];
2462 ArrayType *LTy = dyn_cast<ArrayType>(LFilter->getType());
2464 // Not a filter - skip it.
2466 // If Filter is a subset of LFilter, i.e. every element of Filter is also
2467 // an element of LFilter, then discard LFilter.
2468 SmallVectorImpl<Constant *>::iterator J = NewClauses.begin() + j;
2469 // If Filter is empty then it is a subset of LFilter.
2472 NewClauses.erase(J);
2473 MakeNewInstruction = true;
2474 // Move on to the next filter.
2477 unsigned LElts = LTy->getNumElements();
2478 // If Filter is longer than LFilter then it cannot be a subset of it.
2480 // Move on to the next filter.
2482 // At this point we know that LFilter has at least one element.
2483 if (isa<ConstantAggregateZero>(LFilter)) { // LFilter only contains zeros.
2484 // Filter is a subset of LFilter iff Filter contains only zeros (as we
2485 // already know that Filter is not longer than LFilter).
2486 if (isa<ConstantAggregateZero>(Filter)) {
2487 assert(FElts <= LElts && "Should have handled this case earlier!");
2489 NewClauses.erase(J);
2490 MakeNewInstruction = true;
2492 // Move on to the next filter.
2495 ConstantArray *LArray = cast<ConstantArray>(LFilter);
2496 if (isa<ConstantAggregateZero>(Filter)) { // Filter only contains zeros.
2497 // Since Filter is non-empty and contains only zeros, it is a subset of
2498 // LFilter iff LFilter contains a zero.
2499 assert(FElts > 0 && "Should have eliminated the empty filter earlier!");
2500 for (unsigned l = 0; l != LElts; ++l)
2501 if (LArray->getOperand(l)->isNullValue()) {
2502 // LFilter contains a zero - discard it.
2503 NewClauses.erase(J);
2504 MakeNewInstruction = true;
2507 // Move on to the next filter.
2510 // At this point we know that both filters are ConstantArrays. Loop over
2511 // operands to see whether every element of Filter is also an element of
2512 // LFilter. Since filters tend to be short this is probably faster than
2513 // using a method that scales nicely.
2514 ConstantArray *FArray = cast<ConstantArray>(Filter);
2515 bool AllFound = true;
2516 for (unsigned f = 0; f != FElts; ++f) {
2517 Value *FTypeInfo = FArray->getOperand(f)->stripPointerCasts();
2519 for (unsigned l = 0; l != LElts; ++l) {
2520 Value *LTypeInfo = LArray->getOperand(l)->stripPointerCasts();
2521 if (LTypeInfo == FTypeInfo) {
2531 NewClauses.erase(J);
2532 MakeNewInstruction = true;
2534 // Move on to the next filter.
2538 // If we changed any of the clauses, replace the old landingpad instruction
2540 if (MakeNewInstruction) {
2541 LandingPadInst *NLI = LandingPadInst::Create(LI.getType(),
2542 LI.getPersonalityFn(),
2544 for (unsigned i = 0, e = NewClauses.size(); i != e; ++i)
2545 NLI->addClause(NewClauses[i]);
2546 // A landing pad with no clauses must have the cleanup flag set. It is
2547 // theoretically possible, though highly unlikely, that we eliminated all
2548 // clauses. If so, force the cleanup flag to true.
2549 if (NewClauses.empty())
2551 NLI->setCleanup(CleanupFlag);
2555 // Even if none of the clauses changed, we may nonetheless have understood
2556 // that the cleanup flag is pointless. Clear it if so.
2557 if (LI.isCleanup() != CleanupFlag) {
2558 assert(!CleanupFlag && "Adding a cleanup, not removing one?!");
2559 LI.setCleanup(CleanupFlag);
2569 /// TryToSinkInstruction - Try to move the specified instruction from its
2570 /// current block into the beginning of DestBlock, which can only happen if it's
2571 /// safe to move the instruction past all of the instructions between it and the
2572 /// end of its block.
2573 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
2574 assert(I->hasOneUse() && "Invariants didn't hold!");
2576 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
2577 if (isa<PHINode>(I) || isa<LandingPadInst>(I) || I->mayHaveSideEffects() ||
2578 isa<TerminatorInst>(I))
2581 // Do not sink alloca instructions out of the entry block.
2582 if (isa<AllocaInst>(I) && I->getParent() ==
2583 &DestBlock->getParent()->getEntryBlock())
2586 // We can only sink load instructions if there is nothing between the load and
2587 // the end of block that could change the value.
2588 if (I->mayReadFromMemory()) {
2589 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
2591 if (Scan->mayWriteToMemory())
2595 BasicBlock::iterator InsertPos = DestBlock->getFirstInsertionPt();
2596 I->moveBefore(InsertPos);
2602 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
2603 /// all reachable code to the worklist.
2605 /// This has a couple of tricks to make the code faster and more powerful. In
2606 /// particular, we constant fold and DCE instructions as we go, to avoid adding
2607 /// them to the worklist (this significantly speeds up instcombine on code where
2608 /// many instructions are dead or constant). Additionally, if we find a branch
2609 /// whose condition is a known constant, we only visit the reachable successors.
2611 static bool AddReachableCodeToWorklist(BasicBlock *BB,
2612 SmallPtrSetImpl<BasicBlock*> &Visited,
2614 const DataLayout *DL,
2615 const TargetLibraryInfo *TLI) {
2616 bool MadeIRChange = false;
2617 SmallVector<BasicBlock*, 256> Worklist;
2618 Worklist.push_back(BB);
2620 SmallVector<Instruction*, 128> InstrsForInstCombineWorklist;
2621 DenseMap<ConstantExpr*, Constant*> FoldedConstants;
2624 BB = Worklist.pop_back_val();
2626 // We have now visited this block! If we've already been here, ignore it.
2627 if (!Visited.insert(BB)) continue;
2629 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
2630 Instruction *Inst = BBI++;
2632 // DCE instruction if trivially dead.
2633 if (isInstructionTriviallyDead(Inst, TLI)) {
2635 DEBUG(dbgs() << "IC: DCE: " << *Inst << '\n');
2636 Inst->eraseFromParent();
2640 // ConstantProp instruction if trivially constant.
2641 if (!Inst->use_empty() && isa<Constant>(Inst->getOperand(0)))
2642 if (Constant *C = ConstantFoldInstruction(Inst, DL, TLI)) {
2643 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: "
2645 Inst->replaceAllUsesWith(C);
2647 Inst->eraseFromParent();
2652 // See if we can constant fold its operands.
2653 for (User::op_iterator i = Inst->op_begin(), e = Inst->op_end();
2655 ConstantExpr *CE = dyn_cast<ConstantExpr>(i);
2656 if (CE == nullptr) continue;
2658 Constant*& FoldRes = FoldedConstants[CE];
2660 FoldRes = ConstantFoldConstantExpression(CE, DL, TLI);
2664 if (FoldRes != CE) {
2666 MadeIRChange = true;
2671 InstrsForInstCombineWorklist.push_back(Inst);
2674 // Recursively visit successors. If this is a branch or switch on a
2675 // constant, only visit the reachable successor.
2676 TerminatorInst *TI = BB->getTerminator();
2677 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
2678 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
2679 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
2680 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
2681 Worklist.push_back(ReachableBB);
2684 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
2685 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
2686 // See if this is an explicit destination.
2687 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2689 if (i.getCaseValue() == Cond) {
2690 BasicBlock *ReachableBB = i.getCaseSuccessor();
2691 Worklist.push_back(ReachableBB);
2695 // Otherwise it is the default destination.
2696 Worklist.push_back(SI->getDefaultDest());
2701 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
2702 Worklist.push_back(TI->getSuccessor(i));
2703 } while (!Worklist.empty());
2705 // Once we've found all of the instructions to add to instcombine's worklist,
2706 // add them in reverse order. This way instcombine will visit from the top
2707 // of the function down. This jives well with the way that it adds all uses
2708 // of instructions to the worklist after doing a transformation, thus avoiding
2709 // some N^2 behavior in pathological cases.
2710 IC.Worklist.AddInitialGroup(&InstrsForInstCombineWorklist[0],
2711 InstrsForInstCombineWorklist.size());
2713 return MadeIRChange;
2716 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
2717 MadeIRChange = false;
2719 DEBUG(dbgs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
2720 << F.getName() << "\n");
2723 // Do a depth-first traversal of the function, populate the worklist with
2724 // the reachable instructions. Ignore blocks that are not reachable. Keep
2725 // track of which blocks we visit.
2726 SmallPtrSet<BasicBlock*, 64> Visited;
2727 MadeIRChange |= AddReachableCodeToWorklist(F.begin(), Visited, *this, DL,
2730 // Do a quick scan over the function. If we find any blocks that are
2731 // unreachable, remove any instructions inside of them. This prevents
2732 // the instcombine code from having to deal with some bad special cases.
2733 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
2734 if (Visited.count(BB)) continue;
2736 // Delete the instructions backwards, as it has a reduced likelihood of
2737 // having to update as many def-use and use-def chains.
2738 Instruction *EndInst = BB->getTerminator(); // Last not to be deleted.
2739 while (EndInst != BB->begin()) {
2740 // Delete the next to last instruction.
2741 BasicBlock::iterator I = EndInst;
2742 Instruction *Inst = --I;
2743 if (!Inst->use_empty())
2744 Inst->replaceAllUsesWith(UndefValue::get(Inst->getType()));
2745 if (isa<LandingPadInst>(Inst)) {
2749 if (!isa<DbgInfoIntrinsic>(Inst)) {
2751 MadeIRChange = true;
2753 Inst->eraseFromParent();
2758 while (!Worklist.isEmpty()) {
2759 Instruction *I = Worklist.RemoveOne();
2760 if (I == nullptr) continue; // skip null values.
2762 // Check to see if we can DCE the instruction.
2763 if (isInstructionTriviallyDead(I, TLI)) {
2764 DEBUG(dbgs() << "IC: DCE: " << *I << '\n');
2765 EraseInstFromFunction(*I);
2767 MadeIRChange = true;
2771 // Instruction isn't dead, see if we can constant propagate it.
2772 if (!I->use_empty() && isa<Constant>(I->getOperand(0)))
2773 if (Constant *C = ConstantFoldInstruction(I, DL, TLI)) {
2774 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n');
2776 // Add operands to the worklist.
2777 ReplaceInstUsesWith(*I, C);
2779 EraseInstFromFunction(*I);
2780 MadeIRChange = true;
2784 // See if we can trivially sink this instruction to a successor basic block.
2785 if (I->hasOneUse()) {
2786 BasicBlock *BB = I->getParent();
2787 Instruction *UserInst = cast<Instruction>(*I->user_begin());
2788 BasicBlock *UserParent;
2790 // Get the block the use occurs in.
2791 if (PHINode *PN = dyn_cast<PHINode>(UserInst))
2792 UserParent = PN->getIncomingBlock(*I->use_begin());
2794 UserParent = UserInst->getParent();
2796 if (UserParent != BB) {
2797 bool UserIsSuccessor = false;
2798 // See if the user is one of our successors.
2799 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
2800 if (*SI == UserParent) {
2801 UserIsSuccessor = true;
2805 // If the user is one of our immediate successors, and if that successor
2806 // only has us as a predecessors (we'd have to split the critical edge
2807 // otherwise), we can keep going.
2808 if (UserIsSuccessor && UserParent->getSinglePredecessor()) {
2809 // Okay, the CFG is simple enough, try to sink this instruction.
2810 if (TryToSinkInstruction(I, UserParent)) {
2811 MadeIRChange = true;
2812 // We'll add uses of the sunk instruction below, but since sinking
2813 // can expose opportunities for it's *operands* add them to the
2815 for (Use &U : I->operands())
2816 if (Instruction *OpI = dyn_cast<Instruction>(U.get()))
2823 // Now that we have an instruction, try combining it to simplify it.
2824 Builder->SetInsertPoint(I->getParent(), I);
2825 Builder->SetCurrentDebugLocation(I->getDebugLoc());
2830 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str(););
2831 DEBUG(dbgs() << "IC: Visiting: " << OrigI << '\n');
2833 if (Instruction *Result = visit(*I)) {
2835 // Should we replace the old instruction with a new one?
2837 DEBUG(dbgs() << "IC: Old = " << *I << '\n'
2838 << " New = " << *Result << '\n');
2840 if (!I->getDebugLoc().isUnknown())
2841 Result->setDebugLoc(I->getDebugLoc());
2842 // Everything uses the new instruction now.
2843 I->replaceAllUsesWith(Result);
2845 // Move the name to the new instruction first.
2846 Result->takeName(I);
2848 // Push the new instruction and any users onto the worklist.
2849 Worklist.Add(Result);
2850 Worklist.AddUsersToWorkList(*Result);
2852 // Insert the new instruction into the basic block...
2853 BasicBlock *InstParent = I->getParent();
2854 BasicBlock::iterator InsertPos = I;
2856 // If we replace a PHI with something that isn't a PHI, fix up the
2858 if (!isa<PHINode>(Result) && isa<PHINode>(InsertPos))
2859 InsertPos = InstParent->getFirstInsertionPt();
2861 InstParent->getInstList().insert(InsertPos, Result);
2863 EraseInstFromFunction(*I);
2866 DEBUG(dbgs() << "IC: Mod = " << OrigI << '\n'
2867 << " New = " << *I << '\n');
2870 // If the instruction was modified, it's possible that it is now dead.
2871 // if so, remove it.
2872 if (isInstructionTriviallyDead(I, TLI)) {
2873 EraseInstFromFunction(*I);
2876 Worklist.AddUsersToWorkList(*I);
2879 MadeIRChange = true;
2884 return MadeIRChange;
2888 class InstCombinerLibCallSimplifier : public LibCallSimplifier {
2891 InstCombinerLibCallSimplifier(const DataLayout *DL,
2892 const TargetLibraryInfo *TLI,
2894 : LibCallSimplifier(DL, TLI, UnsafeFPShrink) {
2898 /// replaceAllUsesWith - override so that instruction replacement
2899 /// can be defined in terms of the instruction combiner framework.
2900 void replaceAllUsesWith(Instruction *I, Value *With) const override {
2901 IC->ReplaceInstUsesWith(*I, With);
2906 bool InstCombiner::runOnFunction(Function &F) {
2907 if (skipOptnoneFunction(F))
2910 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
2911 DL = DLP ? &DLP->getDataLayout() : nullptr;
2912 TLI = &getAnalysis<TargetLibraryInfo>();
2914 MinimizeSize = F.getAttributes().hasAttribute(AttributeSet::FunctionIndex,
2915 Attribute::MinSize);
2917 /// Builder - This is an IRBuilder that automatically inserts new
2918 /// instructions into the worklist when they are created.
2919 IRBuilder<true, TargetFolder, InstCombineIRInserter>
2920 TheBuilder(F.getContext(), TargetFolder(DL),
2921 InstCombineIRInserter(Worklist));
2922 Builder = &TheBuilder;
2924 InstCombinerLibCallSimplifier TheSimplifier(DL, TLI, this);
2925 Simplifier = &TheSimplifier;
2927 bool EverMadeChange = false;
2929 // Lower dbg.declare intrinsics otherwise their value may be clobbered
2931 EverMadeChange = LowerDbgDeclare(F);
2933 // Iterate while there is work to do.
2934 unsigned Iteration = 0;
2935 while (DoOneIteration(F, Iteration++))
2936 EverMadeChange = true;
2939 return EverMadeChange;
2942 FunctionPass *llvm::createInstructionCombiningPass() {
2943 return new InstCombiner();