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/IR/CFG.h"
46 #include "llvm/IR/DataLayout.h"
47 #include "llvm/IR/GetElementPtrTypeIterator.h"
48 #include "llvm/IR/IntrinsicInst.h"
49 #include "llvm/IR/PatternMatch.h"
50 #include "llvm/IR/ValueHandle.h"
51 #include "llvm/Support/CommandLine.h"
52 #include "llvm/Support/Debug.h"
53 #include "llvm/Target/TargetLibraryInfo.h"
54 #include "llvm/Transforms/Utils/Local.h"
58 using namespace llvm::PatternMatch;
60 #define DEBUG_TYPE "instcombine"
62 STATISTIC(NumCombined , "Number of insts combined");
63 STATISTIC(NumConstProp, "Number of constant folds");
64 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
65 STATISTIC(NumSunkInst , "Number of instructions sunk");
66 STATISTIC(NumExpand, "Number of expansions");
67 STATISTIC(NumFactor , "Number of factorizations");
68 STATISTIC(NumReassoc , "Number of reassociations");
70 static cl::opt<bool> UnsafeFPShrink("enable-double-float-shrink", cl::Hidden,
72 cl::desc("Enable unsafe double to float "
73 "shrinking for math lib calls"));
75 // Initialization Routines
76 void llvm::initializeInstCombine(PassRegistry &Registry) {
77 initializeInstCombinerPass(Registry);
80 void LLVMInitializeInstCombine(LLVMPassRegistryRef R) {
81 initializeInstCombine(*unwrap(R));
84 char InstCombiner::ID = 0;
85 INITIALIZE_PASS_BEGIN(InstCombiner, "instcombine",
86 "Combine redundant instructions", false, false)
87 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
88 INITIALIZE_PASS_END(InstCombiner, "instcombine",
89 "Combine redundant instructions", false, false)
91 void InstCombiner::getAnalysisUsage(AnalysisUsage &AU) const {
93 AU.addRequired<TargetLibraryInfo>();
97 Value *InstCombiner::EmitGEPOffset(User *GEP) {
98 return llvm::EmitGEPOffset(Builder, *getDataLayout(), GEP);
101 /// ShouldChangeType - Return true if it is desirable to convert a computation
102 /// from 'From' to 'To'. We don't want to convert from a legal to an illegal
103 /// type for example, or from a smaller to a larger illegal type.
104 bool InstCombiner::ShouldChangeType(Type *From, Type *To) const {
105 assert(From->isIntegerTy() && To->isIntegerTy());
107 // If we don't have DL, we don't know if the source/dest are legal.
108 if (!DL) return false;
110 unsigned FromWidth = From->getPrimitiveSizeInBits();
111 unsigned ToWidth = To->getPrimitiveSizeInBits();
112 bool FromLegal = DL->isLegalInteger(FromWidth);
113 bool ToLegal = DL->isLegalInteger(ToWidth);
115 // If this is a legal integer from type, and the result would be an illegal
116 // type, don't do the transformation.
117 if (FromLegal && !ToLegal)
120 // Otherwise, if both are illegal, do not increase the size of the result. We
121 // do allow things like i160 -> i64, but not i64 -> i160.
122 if (!FromLegal && !ToLegal && ToWidth > FromWidth)
128 // Return true, if No Signed Wrap should be maintained for I.
129 // The No Signed Wrap flag can be kept if the operation "B (I.getOpcode) C",
130 // where both B and C should be ConstantInts, results in a constant that does
131 // not overflow. This function only handles the Add and Sub opcodes. For
132 // all other opcodes, the function conservatively returns false.
133 static bool MaintainNoSignedWrap(BinaryOperator &I, Value *B, Value *C) {
134 OverflowingBinaryOperator *OBO = dyn_cast<OverflowingBinaryOperator>(&I);
135 if (!OBO || !OBO->hasNoSignedWrap()) {
139 // We reason about Add and Sub Only.
140 Instruction::BinaryOps Opcode = I.getOpcode();
141 if (Opcode != Instruction::Add &&
142 Opcode != Instruction::Sub) {
146 ConstantInt *CB = dyn_cast<ConstantInt>(B);
147 ConstantInt *CC = dyn_cast<ConstantInt>(C);
153 const APInt &BVal = CB->getValue();
154 const APInt &CVal = CC->getValue();
155 bool Overflow = false;
157 if (Opcode == Instruction::Add) {
158 BVal.sadd_ov(CVal, Overflow);
160 BVal.ssub_ov(CVal, Overflow);
166 /// Conservatively clears subclassOptionalData after a reassociation or
167 /// commutation. We preserve fast-math flags when applicable as they can be
169 static void ClearSubclassDataAfterReassociation(BinaryOperator &I) {
170 FPMathOperator *FPMO = dyn_cast<FPMathOperator>(&I);
172 I.clearSubclassOptionalData();
176 FastMathFlags FMF = I.getFastMathFlags();
177 I.clearSubclassOptionalData();
178 I.setFastMathFlags(FMF);
181 /// SimplifyAssociativeOrCommutative - This performs a few simplifications for
182 /// operators which are associative or commutative:
184 // Commutative operators:
186 // 1. Order operands such that they are listed from right (least complex) to
187 // left (most complex). This puts constants before unary operators before
190 // Associative operators:
192 // 2. Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
193 // 3. Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
195 // Associative and commutative operators:
197 // 4. Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
198 // 5. Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
199 // 6. Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
200 // if C1 and C2 are constants.
202 bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) {
203 Instruction::BinaryOps Opcode = I.getOpcode();
204 bool Changed = false;
207 // Order operands such that they are listed from right (least complex) to
208 // left (most complex). This puts constants before unary operators before
210 if (I.isCommutative() && getComplexity(I.getOperand(0)) <
211 getComplexity(I.getOperand(1)))
212 Changed = !I.swapOperands();
214 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
215 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
217 if (I.isAssociative()) {
218 // Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
219 if (Op0 && Op0->getOpcode() == Opcode) {
220 Value *A = Op0->getOperand(0);
221 Value *B = Op0->getOperand(1);
222 Value *C = I.getOperand(1);
224 // Does "B op C" simplify?
225 if (Value *V = SimplifyBinOp(Opcode, B, C, DL)) {
226 // It simplifies to V. Form "A op V".
229 // Conservatively clear the optional flags, since they may not be
230 // preserved by the reassociation.
231 if (MaintainNoSignedWrap(I, B, C) &&
232 (!Op0 || (isa<BinaryOperator>(Op0) && Op0->hasNoSignedWrap()))) {
233 // Note: this is only valid because SimplifyBinOp doesn't look at
234 // the operands to Op0.
235 I.clearSubclassOptionalData();
236 I.setHasNoSignedWrap(true);
238 ClearSubclassDataAfterReassociation(I);
247 // Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
248 if (Op1 && Op1->getOpcode() == Opcode) {
249 Value *A = I.getOperand(0);
250 Value *B = Op1->getOperand(0);
251 Value *C = Op1->getOperand(1);
253 // Does "A op B" simplify?
254 if (Value *V = SimplifyBinOp(Opcode, A, B, DL)) {
255 // It simplifies to V. Form "V op C".
258 // Conservatively clear the optional flags, since they may not be
259 // preserved by the reassociation.
260 ClearSubclassDataAfterReassociation(I);
268 if (I.isAssociative() && I.isCommutative()) {
269 // Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
270 if (Op0 && Op0->getOpcode() == Opcode) {
271 Value *A = Op0->getOperand(0);
272 Value *B = Op0->getOperand(1);
273 Value *C = I.getOperand(1);
275 // Does "C op A" simplify?
276 if (Value *V = SimplifyBinOp(Opcode, C, A, DL)) {
277 // It simplifies to V. Form "V op B".
280 // Conservatively clear the optional flags, since they may not be
281 // preserved by the reassociation.
282 ClearSubclassDataAfterReassociation(I);
289 // Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
290 if (Op1 && Op1->getOpcode() == Opcode) {
291 Value *A = I.getOperand(0);
292 Value *B = Op1->getOperand(0);
293 Value *C = Op1->getOperand(1);
295 // Does "C op A" simplify?
296 if (Value *V = SimplifyBinOp(Opcode, C, A, DL)) {
297 // It simplifies to V. Form "B op V".
300 // Conservatively clear the optional flags, since they may not be
301 // preserved by the reassociation.
302 ClearSubclassDataAfterReassociation(I);
309 // Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
310 // if C1 and C2 are constants.
312 Op0->getOpcode() == Opcode && Op1->getOpcode() == Opcode &&
313 isa<Constant>(Op0->getOperand(1)) &&
314 isa<Constant>(Op1->getOperand(1)) &&
315 Op0->hasOneUse() && Op1->hasOneUse()) {
316 Value *A = Op0->getOperand(0);
317 Constant *C1 = cast<Constant>(Op0->getOperand(1));
318 Value *B = Op1->getOperand(0);
319 Constant *C2 = cast<Constant>(Op1->getOperand(1));
321 Constant *Folded = ConstantExpr::get(Opcode, C1, C2);
322 BinaryOperator *New = BinaryOperator::Create(Opcode, A, B);
323 if (isa<FPMathOperator>(New)) {
324 FastMathFlags Flags = I.getFastMathFlags();
325 Flags &= Op0->getFastMathFlags();
326 Flags &= Op1->getFastMathFlags();
327 New->setFastMathFlags(Flags);
329 InsertNewInstWith(New, I);
331 I.setOperand(0, New);
332 I.setOperand(1, Folded);
333 // Conservatively clear the optional flags, since they may not be
334 // preserved by the reassociation.
335 ClearSubclassDataAfterReassociation(I);
342 // No further simplifications.
347 /// LeftDistributesOverRight - Whether "X LOp (Y ROp Z)" is always equal to
348 /// "(X LOp Y) ROp (X LOp Z)".
349 static bool LeftDistributesOverRight(Instruction::BinaryOps LOp,
350 Instruction::BinaryOps ROp) {
355 case Instruction::And:
356 // And distributes over Or and Xor.
360 case Instruction::Or:
361 case Instruction::Xor:
365 case Instruction::Mul:
366 // Multiplication distributes over addition and subtraction.
370 case Instruction::Add:
371 case Instruction::Sub:
375 case Instruction::Or:
376 // Or distributes over And.
380 case Instruction::And:
386 /// RightDistributesOverLeft - Whether "(X LOp Y) ROp Z" is always equal to
387 /// "(X ROp Z) LOp (Y ROp Z)".
388 static bool RightDistributesOverLeft(Instruction::BinaryOps LOp,
389 Instruction::BinaryOps ROp) {
390 if (Instruction::isCommutative(ROp))
391 return LeftDistributesOverRight(ROp, LOp);
392 // TODO: It would be nice to handle division, aka "(X + Y)/Z = X/Z + Y/Z",
393 // but this requires knowing that the addition does not overflow and other
398 /// This function returns identity value for given opcode, which can be used to
399 /// factor patterns like (X * 2) + X ==> (X * 2) + (X * 1) ==> X * (2 + 1).
400 static Value *getIdentityValue(Instruction::BinaryOps OpCode, Value *V) {
401 if (isa<Constant>(V))
404 if (OpCode == Instruction::Mul)
405 return ConstantInt::get(V->getType(), 1);
407 // TODO: We can handle other cases e.g. Instruction::And, Instruction::Or etc.
412 /// This function factors binary ops which can be combined using distributive
413 /// laws. This also factor SHL as MUL e.g. SHL(X, 2) ==> MUL(X, 4).
414 Instruction::BinaryOps getBinOpsForFactorization(BinaryOperator *Op,
415 Value *&LHS, Value *&RHS) {
417 return Instruction::BinaryOpsEnd;
419 if (Op->getOpcode() == Instruction::Shl) {
420 if (Constant *CST = dyn_cast<Constant>(Op->getOperand(1))) {
421 // The multiplier is really 1 << CST.
422 RHS = ConstantExpr::getShl(ConstantInt::get(Op->getType(), 1), CST);
423 LHS = Op->getOperand(0);
424 return Instruction::Mul;
428 // TODO: We can add other conversions e.g. shr => div etc.
430 LHS = Op->getOperand(0);
431 RHS = Op->getOperand(1);
432 return Op->getOpcode();
435 /// This tries to simplify binary operations by factorizing out common terms
436 /// (e. g. "(A*B)+(A*C)" -> "A*(B+C)").
437 static Value *tryFactorization(InstCombiner::BuilderTy *Builder,
438 const DataLayout *DL, BinaryOperator &I,
439 Instruction::BinaryOps InnerOpcode, Value *A,
440 Value *B, Value *C, Value *D) {
442 // If any of A, B, C, D are null, we can not factor I, return early.
443 // Checking A and C should be enough.
444 if (!A || !C || !B || !D)
447 Value *SimplifiedInst = nullptr;
448 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
449 Instruction::BinaryOps TopLevelOpcode = I.getOpcode();
451 // Does "X op' Y" always equal "Y op' X"?
452 bool InnerCommutative = Instruction::isCommutative(InnerOpcode);
454 // Does "X op' (Y op Z)" always equal "(X op' Y) op (X op' Z)"?
455 if (LeftDistributesOverRight(InnerOpcode, TopLevelOpcode))
456 // Does the instruction have the form "(A op' B) op (A op' D)" or, in the
457 // commutative case, "(A op' B) op (C op' A)"?
458 if (A == C || (InnerCommutative && A == D)) {
461 // Consider forming "A op' (B op D)".
462 // If "B op D" simplifies then it can be formed with no cost.
463 Value *V = SimplifyBinOp(TopLevelOpcode, B, D, DL);
464 // If "B op D" doesn't simplify then only go on if both of the existing
465 // operations "A op' B" and "C op' D" will be zapped as no longer used.
466 if (!V && LHS->hasOneUse() && RHS->hasOneUse())
467 V = Builder->CreateBinOp(TopLevelOpcode, B, D, RHS->getName());
469 SimplifiedInst = Builder->CreateBinOp(InnerOpcode, A, V);
473 // Does "(X op Y) op' Z" always equal "(X op' Z) op (Y op' Z)"?
474 if (!SimplifiedInst && RightDistributesOverLeft(TopLevelOpcode, InnerOpcode))
475 // Does the instruction have the form "(A op' B) op (C op' B)" or, in the
476 // commutative case, "(A op' B) op (B op' D)"?
477 if (B == D || (InnerCommutative && B == C)) {
480 // Consider forming "(A op C) op' B".
481 // If "A op C" simplifies then it can be formed with no cost.
482 Value *V = SimplifyBinOp(TopLevelOpcode, A, C, DL);
484 // If "A op C" doesn't simplify then only go on if both of the existing
485 // operations "A op' B" and "C op' D" will be zapped as no longer used.
486 if (!V && LHS->hasOneUse() && RHS->hasOneUse())
487 V = Builder->CreateBinOp(TopLevelOpcode, A, C, LHS->getName());
489 SimplifiedInst = Builder->CreateBinOp(InnerOpcode, V, B);
493 if (SimplifiedInst) {
495 SimplifiedInst->takeName(&I);
497 // Check if we can add NSW flag to SimplifiedInst. If so, set NSW flag.
498 // TODO: Check for NUW.
499 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(SimplifiedInst)) {
500 if (isa<OverflowingBinaryOperator>(SimplifiedInst)) {
502 if (isa<OverflowingBinaryOperator>(&I))
503 HasNSW = I.hasNoSignedWrap();
505 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS))
506 if (isa<OverflowingBinaryOperator>(Op0))
507 HasNSW &= Op0->hasNoSignedWrap();
509 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS))
510 if (isa<OverflowingBinaryOperator>(Op1))
511 HasNSW &= Op1->hasNoSignedWrap();
512 BO->setHasNoSignedWrap(HasNSW);
516 return SimplifiedInst;
519 /// SimplifyUsingDistributiveLaws - This tries to simplify binary operations
520 /// which some other binary operation distributes over either by factorizing
521 /// out common terms (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this
522 /// results in simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is
523 /// a win). Returns the simplified value, or null if it didn't simplify.
524 Value *InstCombiner::SimplifyUsingDistributiveLaws(BinaryOperator &I) {
525 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
526 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
527 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
530 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
531 Instruction::BinaryOps LHSOpcode = getBinOpsForFactorization(Op0, A, B);
532 Instruction::BinaryOps RHSOpcode = getBinOpsForFactorization(Op1, C, D);
534 // The instruction has the form "(A op' B) op (C op' D)". Try to factorize
536 if (LHSOpcode == RHSOpcode) {
537 if (Value *V = tryFactorization(Builder, DL, I, LHSOpcode, A, B, C, D))
541 // The instruction has the form "(A op' B) op (C)". Try to factorize common
543 if (Value *V = tryFactorization(Builder, DL, I, LHSOpcode, A, B, RHS,
544 getIdentityValue(LHSOpcode, RHS)))
547 // The instruction has the form "(B) op (C op' D)". Try to factorize common
549 if (Value *V = tryFactorization(Builder, DL, I, RHSOpcode, LHS,
550 getIdentityValue(RHSOpcode, LHS), C, D))
554 Instruction::BinaryOps TopLevelOpcode = I.getOpcode();
555 if (Op0 && RightDistributesOverLeft(Op0->getOpcode(), TopLevelOpcode)) {
556 // The instruction has the form "(A op' B) op C". See if expanding it out
557 // to "(A op C) op' (B op C)" results in simplifications.
558 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
559 Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op'
561 // Do "A op C" and "B op C" both simplify?
562 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, C, DL))
563 if (Value *R = SimplifyBinOp(TopLevelOpcode, B, C, DL)) {
564 // They do! Return "L op' R".
566 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
567 if ((L == A && R == B) ||
568 (Instruction::isCommutative(InnerOpcode) && L == B && R == A))
570 // Otherwise return "L op' R" if it simplifies.
571 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, DL))
573 // Otherwise, create a new instruction.
574 C = Builder->CreateBinOp(InnerOpcode, L, R);
580 if (Op1 && LeftDistributesOverRight(TopLevelOpcode, Op1->getOpcode())) {
581 // The instruction has the form "A op (B op' C)". See if expanding it out
582 // to "(A op B) op' (A op C)" results in simplifications.
583 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
584 Instruction::BinaryOps InnerOpcode = Op1->getOpcode(); // op'
586 // Do "A op B" and "A op C" both simplify?
587 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, B, DL))
588 if (Value *R = SimplifyBinOp(TopLevelOpcode, A, C, DL)) {
589 // They do! Return "L op' R".
591 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
592 if ((L == B && R == C) ||
593 (Instruction::isCommutative(InnerOpcode) && L == C && R == B))
595 // Otherwise return "L op' R" if it simplifies.
596 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, DL))
598 // Otherwise, create a new instruction.
599 A = Builder->CreateBinOp(InnerOpcode, L, R);
608 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
609 // if the LHS is a constant zero (which is the 'negate' form).
611 Value *InstCombiner::dyn_castNegVal(Value *V) const {
612 if (BinaryOperator::isNeg(V))
613 return BinaryOperator::getNegArgument(V);
615 // Constants can be considered to be negated values if they can be folded.
616 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
617 return ConstantExpr::getNeg(C);
619 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
620 if (C->getType()->getElementType()->isIntegerTy())
621 return ConstantExpr::getNeg(C);
626 // dyn_castFNegVal - Given a 'fsub' instruction, return the RHS of the
627 // instruction if the LHS is a constant negative zero (which is the 'negate'
630 Value *InstCombiner::dyn_castFNegVal(Value *V, bool IgnoreZeroSign) const {
631 if (BinaryOperator::isFNeg(V, IgnoreZeroSign))
632 return BinaryOperator::getFNegArgument(V);
634 // Constants can be considered to be negated values if they can be folded.
635 if (ConstantFP *C = dyn_cast<ConstantFP>(V))
636 return ConstantExpr::getFNeg(C);
638 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
639 if (C->getType()->getElementType()->isFloatingPointTy())
640 return ConstantExpr::getFNeg(C);
645 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
647 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
648 return IC->Builder->CreateCast(CI->getOpcode(), SO, I.getType());
651 // Figure out if the constant is the left or the right argument.
652 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
653 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
655 if (Constant *SOC = dyn_cast<Constant>(SO)) {
657 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
658 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
661 Value *Op0 = SO, *Op1 = ConstOperand;
665 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I)) {
666 Value *RI = IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1,
667 SO->getName()+".op");
668 Instruction *FPInst = dyn_cast<Instruction>(RI);
669 if (FPInst && isa<FPMathOperator>(FPInst))
670 FPInst->copyFastMathFlags(BO);
673 if (ICmpInst *CI = dyn_cast<ICmpInst>(&I))
674 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
675 SO->getName()+".cmp");
676 if (FCmpInst *CI = dyn_cast<FCmpInst>(&I))
677 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
678 SO->getName()+".cmp");
679 llvm_unreachable("Unknown binary instruction type!");
682 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
683 // constant as the other operand, try to fold the binary operator into the
684 // select arguments. This also works for Cast instructions, which obviously do
685 // not have a second operand.
686 Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) {
687 // Don't modify shared select instructions
688 if (!SI->hasOneUse()) return nullptr;
689 Value *TV = SI->getOperand(1);
690 Value *FV = SI->getOperand(2);
692 if (isa<Constant>(TV) || isa<Constant>(FV)) {
693 // Bool selects with constant operands can be folded to logical ops.
694 if (SI->getType()->isIntegerTy(1)) return nullptr;
696 // If it's a bitcast involving vectors, make sure it has the same number of
697 // elements on both sides.
698 if (BitCastInst *BC = dyn_cast<BitCastInst>(&Op)) {
699 VectorType *DestTy = dyn_cast<VectorType>(BC->getDestTy());
700 VectorType *SrcTy = dyn_cast<VectorType>(BC->getSrcTy());
702 // Verify that either both or neither are vectors.
703 if ((SrcTy == nullptr) != (DestTy == nullptr)) return nullptr;
704 // If vectors, verify that they have the same number of elements.
705 if (SrcTy && SrcTy->getNumElements() != DestTy->getNumElements())
709 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, this);
710 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, this);
712 return SelectInst::Create(SI->getCondition(),
713 SelectTrueVal, SelectFalseVal);
719 /// FoldOpIntoPhi - Given a binary operator, cast instruction, or select which
720 /// has a PHI node as operand #0, see if we can fold the instruction into the
721 /// PHI (which is only possible if all operands to the PHI are constants).
723 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
724 PHINode *PN = cast<PHINode>(I.getOperand(0));
725 unsigned NumPHIValues = PN->getNumIncomingValues();
726 if (NumPHIValues == 0)
729 // We normally only transform phis with a single use. However, if a PHI has
730 // multiple uses and they are all the same operation, we can fold *all* of the
731 // uses into the PHI.
732 if (!PN->hasOneUse()) {
733 // Walk the use list for the instruction, comparing them to I.
734 for (User *U : PN->users()) {
735 Instruction *UI = cast<Instruction>(U);
736 if (UI != &I && !I.isIdenticalTo(UI))
739 // Otherwise, we can replace *all* users with the new PHI we form.
742 // Check to see if all of the operands of the PHI are simple constants
743 // (constantint/constantfp/undef). If there is one non-constant value,
744 // remember the BB it is in. If there is more than one or if *it* is a PHI,
745 // bail out. We don't do arbitrary constant expressions here because moving
746 // their computation can be expensive without a cost model.
747 BasicBlock *NonConstBB = nullptr;
748 for (unsigned i = 0; i != NumPHIValues; ++i) {
749 Value *InVal = PN->getIncomingValue(i);
750 if (isa<Constant>(InVal) && !isa<ConstantExpr>(InVal))
753 if (isa<PHINode>(InVal)) return nullptr; // Itself a phi.
754 if (NonConstBB) return nullptr; // More than one non-const value.
756 NonConstBB = PN->getIncomingBlock(i);
758 // If the InVal is an invoke at the end of the pred block, then we can't
759 // insert a computation after it without breaking the edge.
760 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal))
761 if (II->getParent() == NonConstBB)
764 // If the incoming non-constant value is in I's block, we will remove one
765 // instruction, but insert another equivalent one, leading to infinite
767 if (NonConstBB == I.getParent())
771 // If there is exactly one non-constant value, we can insert a copy of the
772 // operation in that block. However, if this is a critical edge, we would be
773 // inserting the computation one some other paths (e.g. inside a loop). Only
774 // do this if the pred block is unconditionally branching into the phi block.
775 if (NonConstBB != nullptr) {
776 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
777 if (!BI || !BI->isUnconditional()) return nullptr;
780 // Okay, we can do the transformation: create the new PHI node.
781 PHINode *NewPN = PHINode::Create(I.getType(), PN->getNumIncomingValues());
782 InsertNewInstBefore(NewPN, *PN);
785 // If we are going to have to insert a new computation, do so right before the
786 // predecessors terminator.
788 Builder->SetInsertPoint(NonConstBB->getTerminator());
790 // Next, add all of the operands to the PHI.
791 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) {
792 // We only currently try to fold the condition of a select when it is a phi,
793 // not the true/false values.
794 Value *TrueV = SI->getTrueValue();
795 Value *FalseV = SI->getFalseValue();
796 BasicBlock *PhiTransBB = PN->getParent();
797 for (unsigned i = 0; i != NumPHIValues; ++i) {
798 BasicBlock *ThisBB = PN->getIncomingBlock(i);
799 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB);
800 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB);
801 Value *InV = nullptr;
802 // Beware of ConstantExpr: it may eventually evaluate to getNullValue,
803 // even if currently isNullValue gives false.
804 Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i));
805 if (InC && !isa<ConstantExpr>(InC))
806 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred;
808 InV = Builder->CreateSelect(PN->getIncomingValue(i),
809 TrueVInPred, FalseVInPred, "phitmp");
810 NewPN->addIncoming(InV, ThisBB);
812 } else if (CmpInst *CI = dyn_cast<CmpInst>(&I)) {
813 Constant *C = cast<Constant>(I.getOperand(1));
814 for (unsigned i = 0; i != NumPHIValues; ++i) {
815 Value *InV = nullptr;
816 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
817 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
818 else if (isa<ICmpInst>(CI))
819 InV = Builder->CreateICmp(CI->getPredicate(), PN->getIncomingValue(i),
822 InV = Builder->CreateFCmp(CI->getPredicate(), PN->getIncomingValue(i),
824 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
826 } else if (I.getNumOperands() == 2) {
827 Constant *C = cast<Constant>(I.getOperand(1));
828 for (unsigned i = 0; i != NumPHIValues; ++i) {
829 Value *InV = nullptr;
830 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
831 InV = ConstantExpr::get(I.getOpcode(), InC, C);
833 InV = Builder->CreateBinOp(cast<BinaryOperator>(I).getOpcode(),
834 PN->getIncomingValue(i), C, "phitmp");
835 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
838 CastInst *CI = cast<CastInst>(&I);
839 Type *RetTy = CI->getType();
840 for (unsigned i = 0; i != NumPHIValues; ++i) {
842 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
843 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
845 InV = Builder->CreateCast(CI->getOpcode(),
846 PN->getIncomingValue(i), I.getType(), "phitmp");
847 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
851 for (auto UI = PN->user_begin(), E = PN->user_end(); UI != E;) {
852 Instruction *User = cast<Instruction>(*UI++);
853 if (User == &I) continue;
854 ReplaceInstUsesWith(*User, NewPN);
855 EraseInstFromFunction(*User);
857 return ReplaceInstUsesWith(I, NewPN);
860 /// FindElementAtOffset - Given a pointer type and a constant offset, determine
861 /// whether or not there is a sequence of GEP indices into the pointed type that
862 /// will land us at the specified offset. If so, fill them into NewIndices and
863 /// return the resultant element type, otherwise return null.
864 Type *InstCombiner::FindElementAtOffset(Type *PtrTy, int64_t Offset,
865 SmallVectorImpl<Value*> &NewIndices) {
866 assert(PtrTy->isPtrOrPtrVectorTy());
871 Type *Ty = PtrTy->getPointerElementType();
875 // Start with the index over the outer type. Note that the type size
876 // might be zero (even if the offset isn't zero) if the indexed type
877 // is something like [0 x {int, int}]
878 Type *IntPtrTy = DL->getIntPtrType(PtrTy);
879 int64_t FirstIdx = 0;
880 if (int64_t TySize = DL->getTypeAllocSize(Ty)) {
881 FirstIdx = Offset/TySize;
882 Offset -= FirstIdx*TySize;
884 // Handle hosts where % returns negative instead of values [0..TySize).
890 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
893 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
895 // Index into the types. If we fail, set OrigBase to null.
897 // Indexing into tail padding between struct/array elements.
898 if (uint64_t(Offset*8) >= DL->getTypeSizeInBits(Ty))
901 if (StructType *STy = dyn_cast<StructType>(Ty)) {
902 const StructLayout *SL = DL->getStructLayout(STy);
903 assert(Offset < (int64_t)SL->getSizeInBytes() &&
904 "Offset must stay within the indexed type");
906 unsigned Elt = SL->getElementContainingOffset(Offset);
907 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
910 Offset -= SL->getElementOffset(Elt);
911 Ty = STy->getElementType(Elt);
912 } else if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
913 uint64_t EltSize = DL->getTypeAllocSize(AT->getElementType());
914 assert(EltSize && "Cannot index into a zero-sized array");
915 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
917 Ty = AT->getElementType();
919 // Otherwise, we can't index into the middle of this atomic type, bail.
927 static bool shouldMergeGEPs(GEPOperator &GEP, GEPOperator &Src) {
928 // If this GEP has only 0 indices, it is the same pointer as
929 // Src. If Src is not a trivial GEP too, don't combine
931 if (GEP.hasAllZeroIndices() && !Src.hasAllZeroIndices() &&
937 /// Descale - Return a value X such that Val = X * Scale, or null if none. If
938 /// the multiplication is known not to overflow then NoSignedWrap is set.
939 Value *InstCombiner::Descale(Value *Val, APInt Scale, bool &NoSignedWrap) {
940 assert(isa<IntegerType>(Val->getType()) && "Can only descale integers!");
941 assert(cast<IntegerType>(Val->getType())->getBitWidth() ==
942 Scale.getBitWidth() && "Scale not compatible with value!");
944 // If Val is zero or Scale is one then Val = Val * Scale.
945 if (match(Val, m_Zero()) || Scale == 1) {
950 // If Scale is zero then it does not divide Val.
951 if (Scale.isMinValue())
954 // Look through chains of multiplications, searching for a constant that is
955 // divisible by Scale. For example, descaling X*(Y*(Z*4)) by a factor of 4
956 // will find the constant factor 4 and produce X*(Y*Z). Descaling X*(Y*8) by
957 // a factor of 4 will produce X*(Y*2). The principle of operation is to bore
960 // Val = M1 * X || Analysis starts here and works down
961 // M1 = M2 * Y || Doesn't descend into terms with more
962 // M2 = Z * 4 \/ than one use
964 // Then to modify a term at the bottom:
967 // M1 = Z * Y || Replaced M2 with Z
969 // Then to work back up correcting nsw flags.
971 // Op - the term we are currently analyzing. Starts at Val then drills down.
972 // Replaced with its descaled value before exiting from the drill down loop.
975 // Parent - initially null, but after drilling down notes where Op came from.
976 // In the example above, Parent is (Val, 0) when Op is M1, because M1 is the
977 // 0'th operand of Val.
978 std::pair<Instruction*, unsigned> Parent;
980 // RequireNoSignedWrap - Set if the transform requires a descaling at deeper
981 // levels that doesn't overflow.
982 bool RequireNoSignedWrap = false;
984 // logScale - log base 2 of the scale. Negative if not a power of 2.
985 int32_t logScale = Scale.exactLogBase2();
987 for (;; Op = Parent.first->getOperand(Parent.second)) { // Drill down
989 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
990 // If Op is a constant divisible by Scale then descale to the quotient.
991 APInt Quotient(Scale), Remainder(Scale); // Init ensures right bitwidth.
992 APInt::sdivrem(CI->getValue(), Scale, Quotient, Remainder);
993 if (!Remainder.isMinValue())
994 // Not divisible by Scale.
996 // Replace with the quotient in the parent.
997 Op = ConstantInt::get(CI->getType(), Quotient);
1002 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op)) {
1004 if (BO->getOpcode() == Instruction::Mul) {
1006 NoSignedWrap = BO->hasNoSignedWrap();
1007 if (RequireNoSignedWrap && !NoSignedWrap)
1010 // There are three cases for multiplication: multiplication by exactly
1011 // the scale, multiplication by a constant different to the scale, and
1012 // multiplication by something else.
1013 Value *LHS = BO->getOperand(0);
1014 Value *RHS = BO->getOperand(1);
1016 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
1017 // Multiplication by a constant.
1018 if (CI->getValue() == Scale) {
1019 // Multiplication by exactly the scale, replace the multiplication
1020 // by its left-hand side in the parent.
1025 // Otherwise drill down into the constant.
1026 if (!Op->hasOneUse())
1029 Parent = std::make_pair(BO, 1);
1033 // Multiplication by something else. Drill down into the left-hand side
1034 // since that's where the reassociate pass puts the good stuff.
1035 if (!Op->hasOneUse())
1038 Parent = std::make_pair(BO, 0);
1042 if (logScale > 0 && BO->getOpcode() == Instruction::Shl &&
1043 isa<ConstantInt>(BO->getOperand(1))) {
1044 // Multiplication by a power of 2.
1045 NoSignedWrap = BO->hasNoSignedWrap();
1046 if (RequireNoSignedWrap && !NoSignedWrap)
1049 Value *LHS = BO->getOperand(0);
1050 int32_t Amt = cast<ConstantInt>(BO->getOperand(1))->
1051 getLimitedValue(Scale.getBitWidth());
1054 if (Amt == logScale) {
1055 // Multiplication by exactly the scale, replace the multiplication
1056 // by its left-hand side in the parent.
1060 if (Amt < logScale || !Op->hasOneUse())
1063 // Multiplication by more than the scale. Reduce the multiplying amount
1064 // by the scale in the parent.
1065 Parent = std::make_pair(BO, 1);
1066 Op = ConstantInt::get(BO->getType(), Amt - logScale);
1071 if (!Op->hasOneUse())
1074 if (CastInst *Cast = dyn_cast<CastInst>(Op)) {
1075 if (Cast->getOpcode() == Instruction::SExt) {
1076 // Op is sign-extended from a smaller type, descale in the smaller type.
1077 unsigned SmallSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1078 APInt SmallScale = Scale.trunc(SmallSize);
1079 // Suppose Op = sext X, and we descale X as Y * SmallScale. We want to
1080 // descale Op as (sext Y) * Scale. In order to have
1081 // sext (Y * SmallScale) = (sext Y) * Scale
1082 // some conditions need to hold however: SmallScale must sign-extend to
1083 // Scale and the multiplication Y * SmallScale should not overflow.
1084 if (SmallScale.sext(Scale.getBitWidth()) != Scale)
1085 // SmallScale does not sign-extend to Scale.
1087 assert(SmallScale.exactLogBase2() == logScale);
1088 // Require that Y * SmallScale must not overflow.
1089 RequireNoSignedWrap = true;
1091 // Drill down through the cast.
1092 Parent = std::make_pair(Cast, 0);
1097 if (Cast->getOpcode() == Instruction::Trunc) {
1098 // Op is truncated from a larger type, descale in the larger type.
1099 // Suppose Op = trunc X, and we descale X as Y * sext Scale. Then
1100 // trunc (Y * sext Scale) = (trunc Y) * Scale
1101 // always holds. However (trunc Y) * Scale may overflow even if
1102 // trunc (Y * sext Scale) does not, so nsw flags need to be cleared
1103 // from this point up in the expression (see later).
1104 if (RequireNoSignedWrap)
1107 // Drill down through the cast.
1108 unsigned LargeSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1109 Parent = std::make_pair(Cast, 0);
1110 Scale = Scale.sext(LargeSize);
1111 if (logScale + 1 == (int32_t)Cast->getType()->getPrimitiveSizeInBits())
1113 assert(Scale.exactLogBase2() == logScale);
1118 // Unsupported expression, bail out.
1122 // We know that we can successfully descale, so from here on we can safely
1123 // modify the IR. Op holds the descaled version of the deepest term in the
1124 // expression. NoSignedWrap is 'true' if multiplying Op by Scale is known
1128 // The expression only had one term.
1131 // Rewrite the parent using the descaled version of its operand.
1132 assert(Parent.first->hasOneUse() && "Drilled down when more than one use!");
1133 assert(Op != Parent.first->getOperand(Parent.second) &&
1134 "Descaling was a no-op?");
1135 Parent.first->setOperand(Parent.second, Op);
1136 Worklist.Add(Parent.first);
1138 // Now work back up the expression correcting nsw flags. The logic is based
1139 // on the following observation: if X * Y is known not to overflow as a signed
1140 // multiplication, and Y is replaced by a value Z with smaller absolute value,
1141 // then X * Z will not overflow as a signed multiplication either. As we work
1142 // our way up, having NoSignedWrap 'true' means that the descaled value at the
1143 // current level has strictly smaller absolute value than the original.
1144 Instruction *Ancestor = Parent.first;
1146 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Ancestor)) {
1147 // If the multiplication wasn't nsw then we can't say anything about the
1148 // value of the descaled multiplication, and we have to clear nsw flags
1149 // from this point on up.
1150 bool OpNoSignedWrap = BO->hasNoSignedWrap();
1151 NoSignedWrap &= OpNoSignedWrap;
1152 if (NoSignedWrap != OpNoSignedWrap) {
1153 BO->setHasNoSignedWrap(NoSignedWrap);
1154 Worklist.Add(Ancestor);
1156 } else if (Ancestor->getOpcode() == Instruction::Trunc) {
1157 // The fact that the descaled input to the trunc has smaller absolute
1158 // value than the original input doesn't tell us anything useful about
1159 // the absolute values of the truncations.
1160 NoSignedWrap = false;
1162 assert((Ancestor->getOpcode() != Instruction::SExt || NoSignedWrap) &&
1163 "Failed to keep proper track of nsw flags while drilling down?");
1165 if (Ancestor == Val)
1166 // Got to the top, all done!
1169 // Move up one level in the expression.
1170 assert(Ancestor->hasOneUse() && "Drilled down when more than one use!");
1171 Ancestor = Ancestor->user_back();
1175 /// \brief Creates node of binary operation with the same attributes as the
1176 /// specified one but with other operands.
1177 static Value *CreateBinOpAsGiven(BinaryOperator &Inst, Value *LHS, Value *RHS,
1178 InstCombiner::BuilderTy *B) {
1179 Value *BORes = B->CreateBinOp(Inst.getOpcode(), LHS, RHS);
1180 if (BinaryOperator *NewBO = dyn_cast<BinaryOperator>(BORes)) {
1181 if (isa<OverflowingBinaryOperator>(NewBO)) {
1182 NewBO->setHasNoSignedWrap(Inst.hasNoSignedWrap());
1183 NewBO->setHasNoUnsignedWrap(Inst.hasNoUnsignedWrap());
1185 if (isa<PossiblyExactOperator>(NewBO))
1186 NewBO->setIsExact(Inst.isExact());
1191 /// \brief Makes transformation of binary operation specific for vector types.
1192 /// \param Inst Binary operator to transform.
1193 /// \return Pointer to node that must replace the original binary operator, or
1194 /// null pointer if no transformation was made.
1195 Value *InstCombiner::SimplifyVectorOp(BinaryOperator &Inst) {
1196 if (!Inst.getType()->isVectorTy()) return nullptr;
1198 unsigned VWidth = cast<VectorType>(Inst.getType())->getNumElements();
1199 Value *LHS = Inst.getOperand(0), *RHS = Inst.getOperand(1);
1200 assert(cast<VectorType>(LHS->getType())->getNumElements() == VWidth);
1201 assert(cast<VectorType>(RHS->getType())->getNumElements() == VWidth);
1203 // If both arguments of binary operation are shuffles, which use the same
1204 // mask and shuffle within a single vector, it is worthwhile to move the
1205 // shuffle after binary operation:
1206 // Op(shuffle(v1, m), shuffle(v2, m)) -> shuffle(Op(v1, v2), m)
1207 if (isa<ShuffleVectorInst>(LHS) && isa<ShuffleVectorInst>(RHS)) {
1208 ShuffleVectorInst *LShuf = cast<ShuffleVectorInst>(LHS);
1209 ShuffleVectorInst *RShuf = cast<ShuffleVectorInst>(RHS);
1210 if (isa<UndefValue>(LShuf->getOperand(1)) &&
1211 isa<UndefValue>(RShuf->getOperand(1)) &&
1212 LShuf->getOperand(0)->getType() == RShuf->getOperand(0)->getType() &&
1213 LShuf->getMask() == RShuf->getMask()) {
1214 Value *NewBO = CreateBinOpAsGiven(Inst, LShuf->getOperand(0),
1215 RShuf->getOperand(0), Builder);
1216 Value *Res = Builder->CreateShuffleVector(NewBO,
1217 UndefValue::get(NewBO->getType()), LShuf->getMask());
1222 // If one argument is a shuffle within one vector, the other is a constant,
1223 // try moving the shuffle after the binary operation.
1224 ShuffleVectorInst *Shuffle = nullptr;
1225 Constant *C1 = nullptr;
1226 if (isa<ShuffleVectorInst>(LHS)) Shuffle = cast<ShuffleVectorInst>(LHS);
1227 if (isa<ShuffleVectorInst>(RHS)) Shuffle = cast<ShuffleVectorInst>(RHS);
1228 if (isa<Constant>(LHS)) C1 = cast<Constant>(LHS);
1229 if (isa<Constant>(RHS)) C1 = cast<Constant>(RHS);
1230 if (Shuffle && C1 && isa<UndefValue>(Shuffle->getOperand(1)) &&
1231 Shuffle->getType() == Shuffle->getOperand(0)->getType()) {
1232 SmallVector<int, 16> ShMask = Shuffle->getShuffleMask();
1233 // Find constant C2 that has property:
1234 // shuffle(C2, ShMask) = C1
1235 // If such constant does not exist (example: ShMask=<0,0> and C1=<1,2>)
1236 // reorder is not possible.
1237 SmallVector<Constant*, 16> C2M(VWidth,
1238 UndefValue::get(C1->getType()->getScalarType()));
1239 bool MayChange = true;
1240 for (unsigned I = 0; I < VWidth; ++I) {
1241 if (ShMask[I] >= 0) {
1242 assert(ShMask[I] < (int)VWidth);
1243 if (!isa<UndefValue>(C2M[ShMask[I]])) {
1247 C2M[ShMask[I]] = C1->getAggregateElement(I);
1251 Constant *C2 = ConstantVector::get(C2M);
1252 Value *NewLHS, *NewRHS;
1253 if (isa<Constant>(LHS)) {
1255 NewRHS = Shuffle->getOperand(0);
1257 NewLHS = Shuffle->getOperand(0);
1260 Value *NewBO = CreateBinOpAsGiven(Inst, NewLHS, NewRHS, Builder);
1261 Value *Res = Builder->CreateShuffleVector(NewBO,
1262 UndefValue::get(Inst.getType()), Shuffle->getMask());
1270 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
1271 SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end());
1273 if (Value *V = SimplifyGEPInst(Ops, DL))
1274 return ReplaceInstUsesWith(GEP, V);
1276 Value *PtrOp = GEP.getOperand(0);
1278 // Eliminate unneeded casts for indices, and replace indices which displace
1279 // by multiples of a zero size type with zero.
1281 bool MadeChange = false;
1282 Type *IntPtrTy = DL->getIntPtrType(GEP.getPointerOperandType());
1284 gep_type_iterator GTI = gep_type_begin(GEP);
1285 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end();
1286 I != E; ++I, ++GTI) {
1287 // Skip indices into struct types.
1288 SequentialType *SeqTy = dyn_cast<SequentialType>(*GTI);
1289 if (!SeqTy) continue;
1291 // If the element type has zero size then any index over it is equivalent
1292 // to an index of zero, so replace it with zero if it is not zero already.
1293 if (SeqTy->getElementType()->isSized() &&
1294 DL->getTypeAllocSize(SeqTy->getElementType()) == 0)
1295 if (!isa<Constant>(*I) || !cast<Constant>(*I)->isNullValue()) {
1296 *I = Constant::getNullValue(IntPtrTy);
1300 Type *IndexTy = (*I)->getType();
1301 if (IndexTy != IntPtrTy) {
1302 // If we are using a wider index than needed for this platform, shrink
1303 // it to what we need. If narrower, sign-extend it to what we need.
1304 // This explicit cast can make subsequent optimizations more obvious.
1305 *I = Builder->CreateIntCast(*I, IntPtrTy, true);
1309 if (MadeChange) return &GEP;
1312 // Check to see if the inputs to the PHI node are getelementptr instructions.
1313 if (PHINode *PN = dyn_cast<PHINode>(PtrOp)) {
1314 GetElementPtrInst *Op1 = dyn_cast<GetElementPtrInst>(PN->getOperand(0));
1320 for (auto I = PN->op_begin()+1, E = PN->op_end(); I !=E; ++I) {
1321 GetElementPtrInst *Op2 = dyn_cast<GetElementPtrInst>(*I);
1322 if (!Op2 || Op1->getNumOperands() != Op2->getNumOperands())
1325 // Keep track of the type as we walk the GEP.
1326 Type *CurTy = Op1->getOperand(0)->getType()->getScalarType();
1328 for (unsigned J = 0, F = Op1->getNumOperands(); J != F; ++J) {
1329 if (Op1->getOperand(J)->getType() != Op2->getOperand(J)->getType())
1332 if (Op1->getOperand(J) != Op2->getOperand(J)) {
1334 // We have not seen any differences yet in the GEPs feeding the
1335 // PHI yet, so we record this one if it is allowed to be a
1338 // The first two arguments can vary for any GEP, the rest have to be
1339 // static for struct slots
1340 if (J > 1 && CurTy->isStructTy())
1345 // The GEP is different by more than one input. While this could be
1346 // extended to support GEPs that vary by more than one variable it
1347 // doesn't make sense since it greatly increases the complexity and
1348 // would result in an R+R+R addressing mode which no backend
1349 // directly supports and would need to be broken into several
1350 // simpler instructions anyway.
1355 // Sink down a layer of the type for the next iteration.
1357 if (CompositeType *CT = dyn_cast<CompositeType>(CurTy)) {
1358 CurTy = CT->getTypeAtIndex(Op1->getOperand(J));
1366 GetElementPtrInst *NewGEP = cast<GetElementPtrInst>(Op1->clone());
1369 // All the GEPs feeding the PHI are identical. Clone one down into our
1370 // BB so that it can be merged with the current GEP.
1371 GEP.getParent()->getInstList().insert(GEP.getParent()->getFirstNonPHI(),
1374 // All the GEPs feeding the PHI differ at a single offset. Clone a GEP
1375 // into the current block so it can be merged, and create a new PHI to
1377 Instruction *InsertPt = Builder->GetInsertPoint();
1378 Builder->SetInsertPoint(PN);
1379 PHINode *NewPN = Builder->CreatePHI(Op1->getOperand(DI)->getType(),
1380 PN->getNumOperands());
1381 Builder->SetInsertPoint(InsertPt);
1383 for (auto &I : PN->operands())
1384 NewPN->addIncoming(cast<GEPOperator>(I)->getOperand(DI),
1385 PN->getIncomingBlock(I));
1387 NewGEP->setOperand(DI, NewPN);
1388 GEP.getParent()->getInstList().insert(GEP.getParent()->getFirstNonPHI(),
1390 NewGEP->setOperand(DI, NewPN);
1393 GEP.setOperand(0, NewGEP);
1397 // Combine Indices - If the source pointer to this getelementptr instruction
1398 // is a getelementptr instruction, combine the indices of the two
1399 // getelementptr instructions into a single instruction.
1401 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) {
1402 if (!shouldMergeGEPs(*cast<GEPOperator>(&GEP), *Src))
1405 // Note that if our source is a gep chain itself then we wait for that
1406 // chain to be resolved before we perform this transformation. This
1407 // avoids us creating a TON of code in some cases.
1408 if (GEPOperator *SrcGEP =
1409 dyn_cast<GEPOperator>(Src->getOperand(0)))
1410 if (SrcGEP->getNumOperands() == 2 && shouldMergeGEPs(*Src, *SrcGEP))
1411 return nullptr; // Wait until our source is folded to completion.
1413 SmallVector<Value*, 8> Indices;
1415 // Find out whether the last index in the source GEP is a sequential idx.
1416 bool EndsWithSequential = false;
1417 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src);
1419 EndsWithSequential = !(*I)->isStructTy();
1421 // Can we combine the two pointer arithmetics offsets?
1422 if (EndsWithSequential) {
1423 // Replace: gep (gep %P, long B), long A, ...
1424 // With: T = long A+B; gep %P, T, ...
1427 Value *SO1 = Src->getOperand(Src->getNumOperands()-1);
1428 Value *GO1 = GEP.getOperand(1);
1429 if (SO1 == Constant::getNullValue(SO1->getType())) {
1431 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
1434 // If they aren't the same type, then the input hasn't been processed
1435 // by the loop above yet (which canonicalizes sequential index types to
1436 // intptr_t). Just avoid transforming this until the input has been
1438 if (SO1->getType() != GO1->getType())
1440 Sum = Builder->CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
1443 // Update the GEP in place if possible.
1444 if (Src->getNumOperands() == 2) {
1445 GEP.setOperand(0, Src->getOperand(0));
1446 GEP.setOperand(1, Sum);
1449 Indices.append(Src->op_begin()+1, Src->op_end()-1);
1450 Indices.push_back(Sum);
1451 Indices.append(GEP.op_begin()+2, GEP.op_end());
1452 } else if (isa<Constant>(*GEP.idx_begin()) &&
1453 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
1454 Src->getNumOperands() != 1) {
1455 // Otherwise we can do the fold if the first index of the GEP is a zero
1456 Indices.append(Src->op_begin()+1, Src->op_end());
1457 Indices.append(GEP.idx_begin()+1, GEP.idx_end());
1460 if (!Indices.empty())
1461 return (GEP.isInBounds() && Src->isInBounds()) ?
1462 GetElementPtrInst::CreateInBounds(Src->getOperand(0), Indices,
1464 GetElementPtrInst::Create(Src->getOperand(0), Indices, GEP.getName());
1467 // Canonicalize (gep i8* X, -(ptrtoint Y)) to (sub (ptrtoint X), (ptrtoint Y))
1468 // The GEP pattern is emitted by the SCEV expander for certain kinds of
1469 // pointer arithmetic.
1470 if (DL && GEP.getNumIndices() == 1 &&
1471 match(GEP.getOperand(1), m_Neg(m_PtrToInt(m_Value())))) {
1472 unsigned AS = GEP.getPointerAddressSpace();
1473 if (GEP.getType() == Builder->getInt8PtrTy(AS) &&
1474 GEP.getOperand(1)->getType()->getScalarSizeInBits() ==
1475 DL->getPointerSizeInBits(AS)) {
1476 Operator *Index = cast<Operator>(GEP.getOperand(1));
1477 Value *PtrToInt = Builder->CreatePtrToInt(PtrOp, Index->getType());
1478 Value *NewSub = Builder->CreateSub(PtrToInt, Index->getOperand(1));
1479 return CastInst::Create(Instruction::IntToPtr, NewSub, GEP.getType());
1483 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0).
1484 Value *StrippedPtr = PtrOp->stripPointerCasts();
1485 PointerType *StrippedPtrTy = dyn_cast<PointerType>(StrippedPtr->getType());
1487 // We do not handle pointer-vector geps here.
1491 if (StrippedPtr != PtrOp) {
1492 bool HasZeroPointerIndex = false;
1493 if (ConstantInt *C = dyn_cast<ConstantInt>(GEP.getOperand(1)))
1494 HasZeroPointerIndex = C->isZero();
1496 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
1497 // into : GEP [10 x i8]* X, i32 0, ...
1499 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
1500 // into : GEP i8* X, ...
1502 // This occurs when the program declares an array extern like "int X[];"
1503 if (HasZeroPointerIndex) {
1504 PointerType *CPTy = cast<PointerType>(PtrOp->getType());
1505 if (ArrayType *CATy =
1506 dyn_cast<ArrayType>(CPTy->getElementType())) {
1507 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
1508 if (CATy->getElementType() == StrippedPtrTy->getElementType()) {
1509 // -> GEP i8* X, ...
1510 SmallVector<Value*, 8> Idx(GEP.idx_begin()+1, GEP.idx_end());
1511 GetElementPtrInst *Res =
1512 GetElementPtrInst::Create(StrippedPtr, Idx, GEP.getName());
1513 Res->setIsInBounds(GEP.isInBounds());
1514 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace())
1516 // Insert Res, and create an addrspacecast.
1518 // GEP (addrspacecast i8 addrspace(1)* X to [0 x i8]*), i32 0, ...
1520 // %0 = GEP i8 addrspace(1)* X, ...
1521 // addrspacecast i8 addrspace(1)* %0 to i8*
1522 return new AddrSpaceCastInst(Builder->Insert(Res), GEP.getType());
1525 if (ArrayType *XATy =
1526 dyn_cast<ArrayType>(StrippedPtrTy->getElementType())){
1527 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
1528 if (CATy->getElementType() == XATy->getElementType()) {
1529 // -> GEP [10 x i8]* X, i32 0, ...
1530 // At this point, we know that the cast source type is a pointer
1531 // to an array of the same type as the destination pointer
1532 // array. Because the array type is never stepped over (there
1533 // is a leading zero) we can fold the cast into this GEP.
1534 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace()) {
1535 GEP.setOperand(0, StrippedPtr);
1538 // Cannot replace the base pointer directly because StrippedPtr's
1539 // address space is different. Instead, create a new GEP followed by
1540 // an addrspacecast.
1542 // GEP (addrspacecast [10 x i8] addrspace(1)* X to [0 x i8]*),
1545 // %0 = GEP [10 x i8] addrspace(1)* X, ...
1546 // addrspacecast i8 addrspace(1)* %0 to i8*
1547 SmallVector<Value*, 8> Idx(GEP.idx_begin(), GEP.idx_end());
1548 Value *NewGEP = GEP.isInBounds() ?
1549 Builder->CreateInBoundsGEP(StrippedPtr, Idx, GEP.getName()) :
1550 Builder->CreateGEP(StrippedPtr, Idx, GEP.getName());
1551 return new AddrSpaceCastInst(NewGEP, GEP.getType());
1555 } else if (GEP.getNumOperands() == 2) {
1556 // Transform things like:
1557 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
1558 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
1559 Type *SrcElTy = StrippedPtrTy->getElementType();
1560 Type *ResElTy = PtrOp->getType()->getPointerElementType();
1561 if (DL && SrcElTy->isArrayTy() &&
1562 DL->getTypeAllocSize(SrcElTy->getArrayElementType()) ==
1563 DL->getTypeAllocSize(ResElTy)) {
1564 Type *IdxType = DL->getIntPtrType(GEP.getType());
1565 Value *Idx[2] = { Constant::getNullValue(IdxType), GEP.getOperand(1) };
1566 Value *NewGEP = GEP.isInBounds() ?
1567 Builder->CreateInBoundsGEP(StrippedPtr, Idx, GEP.getName()) :
1568 Builder->CreateGEP(StrippedPtr, Idx, GEP.getName());
1570 // V and GEP are both pointer types --> BitCast
1571 if (StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace())
1572 return new BitCastInst(NewGEP, GEP.getType());
1573 return new AddrSpaceCastInst(NewGEP, GEP.getType());
1576 // Transform things like:
1577 // %V = mul i64 %N, 4
1578 // %t = getelementptr i8* bitcast (i32* %arr to i8*), i32 %V
1579 // into: %t1 = getelementptr i32* %arr, i32 %N; bitcast
1580 if (DL && ResElTy->isSized() && SrcElTy->isSized()) {
1581 // Check that changing the type amounts to dividing the index by a scale
1583 uint64_t ResSize = DL->getTypeAllocSize(ResElTy);
1584 uint64_t SrcSize = DL->getTypeAllocSize(SrcElTy);
1585 if (ResSize && SrcSize % ResSize == 0) {
1586 Value *Idx = GEP.getOperand(1);
1587 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
1588 uint64_t Scale = SrcSize / ResSize;
1590 // Earlier transforms ensure that the index has type IntPtrType, which
1591 // considerably simplifies the logic by eliminating implicit casts.
1592 assert(Idx->getType() == DL->getIntPtrType(GEP.getType()) &&
1593 "Index not cast to pointer width?");
1596 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
1597 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
1598 // If the multiplication NewIdx * Scale may overflow then the new
1599 // GEP may not be "inbounds".
1600 Value *NewGEP = GEP.isInBounds() && NSW ?
1601 Builder->CreateInBoundsGEP(StrippedPtr, NewIdx, GEP.getName()) :
1602 Builder->CreateGEP(StrippedPtr, NewIdx, GEP.getName());
1604 // The NewGEP must be pointer typed, so must the old one -> BitCast
1605 if (StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace())
1606 return new BitCastInst(NewGEP, GEP.getType());
1607 return new AddrSpaceCastInst(NewGEP, GEP.getType());
1612 // Similarly, transform things like:
1613 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
1614 // (where tmp = 8*tmp2) into:
1615 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
1616 if (DL && ResElTy->isSized() && SrcElTy->isSized() &&
1617 SrcElTy->isArrayTy()) {
1618 // Check that changing to the array element type amounts to dividing the
1619 // index by a scale factor.
1620 uint64_t ResSize = DL->getTypeAllocSize(ResElTy);
1621 uint64_t ArrayEltSize
1622 = DL->getTypeAllocSize(SrcElTy->getArrayElementType());
1623 if (ResSize && ArrayEltSize % ResSize == 0) {
1624 Value *Idx = GEP.getOperand(1);
1625 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
1626 uint64_t Scale = ArrayEltSize / ResSize;
1628 // Earlier transforms ensure that the index has type IntPtrType, which
1629 // considerably simplifies the logic by eliminating implicit casts.
1630 assert(Idx->getType() == DL->getIntPtrType(GEP.getType()) &&
1631 "Index not cast to pointer width?");
1634 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
1635 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
1636 // If the multiplication NewIdx * Scale may overflow then the new
1637 // GEP may not be "inbounds".
1639 Constant::getNullValue(DL->getIntPtrType(GEP.getType())),
1643 Value *NewGEP = GEP.isInBounds() && NSW ?
1644 Builder->CreateInBoundsGEP(StrippedPtr, Off, GEP.getName()) :
1645 Builder->CreateGEP(StrippedPtr, Off, GEP.getName());
1646 // The NewGEP must be pointer typed, so must the old one -> BitCast
1647 if (StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace())
1648 return new BitCastInst(NewGEP, GEP.getType());
1649 return new AddrSpaceCastInst(NewGEP, GEP.getType());
1659 /// See if we can simplify:
1660 /// X = bitcast A* to B*
1661 /// Y = gep X, <...constant indices...>
1662 /// into a gep of the original struct. This is important for SROA and alias
1663 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
1664 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
1665 Value *Operand = BCI->getOperand(0);
1666 PointerType *OpType = cast<PointerType>(Operand->getType());
1667 unsigned OffsetBits = DL->getPointerTypeSizeInBits(OpType);
1668 APInt Offset(OffsetBits, 0);
1669 if (!isa<BitCastInst>(Operand) &&
1670 GEP.accumulateConstantOffset(*DL, Offset) &&
1671 StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace()) {
1673 // If this GEP instruction doesn't move the pointer, just replace the GEP
1674 // with a bitcast of the real input to the dest type.
1676 // If the bitcast is of an allocation, and the allocation will be
1677 // converted to match the type of the cast, don't touch this.
1678 if (isa<AllocaInst>(Operand) || isAllocationFn(Operand, TLI)) {
1679 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
1680 if (Instruction *I = visitBitCast(*BCI)) {
1683 BCI->getParent()->getInstList().insert(BCI, I);
1684 ReplaceInstUsesWith(*BCI, I);
1689 return new BitCastInst(Operand, GEP.getType());
1692 // Otherwise, if the offset is non-zero, we need to find out if there is a
1693 // field at Offset in 'A's type. If so, we can pull the cast through the
1695 SmallVector<Value*, 8> NewIndices;
1696 if (FindElementAtOffset(OpType, Offset.getSExtValue(), NewIndices)) {
1697 Value *NGEP = GEP.isInBounds() ?
1698 Builder->CreateInBoundsGEP(Operand, NewIndices) :
1699 Builder->CreateGEP(Operand, NewIndices);
1701 if (NGEP->getType() == GEP.getType())
1702 return ReplaceInstUsesWith(GEP, NGEP);
1703 NGEP->takeName(&GEP);
1704 return new BitCastInst(NGEP, GEP.getType());
1713 isAllocSiteRemovable(Instruction *AI, SmallVectorImpl<WeakVH> &Users,
1714 const TargetLibraryInfo *TLI) {
1715 SmallVector<Instruction*, 4> Worklist;
1716 Worklist.push_back(AI);
1719 Instruction *PI = Worklist.pop_back_val();
1720 for (User *U : PI->users()) {
1721 Instruction *I = cast<Instruction>(U);
1722 switch (I->getOpcode()) {
1724 // Give up the moment we see something we can't handle.
1727 case Instruction::BitCast:
1728 case Instruction::GetElementPtr:
1730 Worklist.push_back(I);
1733 case Instruction::ICmp: {
1734 ICmpInst *ICI = cast<ICmpInst>(I);
1735 // We can fold eq/ne comparisons with null to false/true, respectively.
1736 if (!ICI->isEquality() || !isa<ConstantPointerNull>(ICI->getOperand(1)))
1742 case Instruction::Call:
1743 // Ignore no-op and store intrinsics.
1744 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1745 switch (II->getIntrinsicID()) {
1749 case Intrinsic::memmove:
1750 case Intrinsic::memcpy:
1751 case Intrinsic::memset: {
1752 MemIntrinsic *MI = cast<MemIntrinsic>(II);
1753 if (MI->isVolatile() || MI->getRawDest() != PI)
1757 case Intrinsic::dbg_declare:
1758 case Intrinsic::dbg_value:
1759 case Intrinsic::invariant_start:
1760 case Intrinsic::invariant_end:
1761 case Intrinsic::lifetime_start:
1762 case Intrinsic::lifetime_end:
1763 case Intrinsic::objectsize:
1769 if (isFreeCall(I, TLI)) {
1775 case Instruction::Store: {
1776 StoreInst *SI = cast<StoreInst>(I);
1777 if (SI->isVolatile() || SI->getPointerOperand() != PI)
1783 llvm_unreachable("missing a return?");
1785 } while (!Worklist.empty());
1789 Instruction *InstCombiner::visitAllocSite(Instruction &MI) {
1790 // If we have a malloc call which is only used in any amount of comparisons
1791 // to null and free calls, delete the calls and replace the comparisons with
1792 // true or false as appropriate.
1793 SmallVector<WeakVH, 64> Users;
1794 if (isAllocSiteRemovable(&MI, Users, TLI)) {
1795 for (unsigned i = 0, e = Users.size(); i != e; ++i) {
1796 Instruction *I = cast_or_null<Instruction>(&*Users[i]);
1799 if (ICmpInst *C = dyn_cast<ICmpInst>(I)) {
1800 ReplaceInstUsesWith(*C,
1801 ConstantInt::get(Type::getInt1Ty(C->getContext()),
1802 C->isFalseWhenEqual()));
1803 } else if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) {
1804 ReplaceInstUsesWith(*I, UndefValue::get(I->getType()));
1805 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1806 if (II->getIntrinsicID() == Intrinsic::objectsize) {
1807 ConstantInt *CI = cast<ConstantInt>(II->getArgOperand(1));
1808 uint64_t DontKnow = CI->isZero() ? -1ULL : 0;
1809 ReplaceInstUsesWith(*I, ConstantInt::get(I->getType(), DontKnow));
1812 EraseInstFromFunction(*I);
1815 if (InvokeInst *II = dyn_cast<InvokeInst>(&MI)) {
1816 // Replace invoke with a NOP intrinsic to maintain the original CFG
1817 Module *M = II->getParent()->getParent()->getParent();
1818 Function *F = Intrinsic::getDeclaration(M, Intrinsic::donothing);
1819 InvokeInst::Create(F, II->getNormalDest(), II->getUnwindDest(),
1820 None, "", II->getParent());
1822 return EraseInstFromFunction(MI);
1827 /// \brief Move the call to free before a NULL test.
1829 /// Check if this free is accessed after its argument has been test
1830 /// against NULL (property 0).
1831 /// If yes, it is legal to move this call in its predecessor block.
1833 /// The move is performed only if the block containing the call to free
1834 /// will be removed, i.e.:
1835 /// 1. it has only one predecessor P, and P has two successors
1836 /// 2. it contains the call and an unconditional branch
1837 /// 3. its successor is the same as its predecessor's successor
1839 /// The profitability is out-of concern here and this function should
1840 /// be called only if the caller knows this transformation would be
1841 /// profitable (e.g., for code size).
1842 static Instruction *
1843 tryToMoveFreeBeforeNullTest(CallInst &FI) {
1844 Value *Op = FI.getArgOperand(0);
1845 BasicBlock *FreeInstrBB = FI.getParent();
1846 BasicBlock *PredBB = FreeInstrBB->getSinglePredecessor();
1848 // Validate part of constraint #1: Only one predecessor
1849 // FIXME: We can extend the number of predecessor, but in that case, we
1850 // would duplicate the call to free in each predecessor and it may
1851 // not be profitable even for code size.
1855 // Validate constraint #2: Does this block contains only the call to
1856 // free and an unconditional branch?
1857 // FIXME: We could check if we can speculate everything in the
1858 // predecessor block
1859 if (FreeInstrBB->size() != 2)
1862 if (!match(FreeInstrBB->getTerminator(), m_UnconditionalBr(SuccBB)))
1865 // Validate the rest of constraint #1 by matching on the pred branch.
1866 TerminatorInst *TI = PredBB->getTerminator();
1867 BasicBlock *TrueBB, *FalseBB;
1868 ICmpInst::Predicate Pred;
1869 if (!match(TI, m_Br(m_ICmp(Pred, m_Specific(Op), m_Zero()), TrueBB, FalseBB)))
1871 if (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE)
1874 // Validate constraint #3: Ensure the null case just falls through.
1875 if (SuccBB != (Pred == ICmpInst::ICMP_EQ ? TrueBB : FalseBB))
1877 assert(FreeInstrBB == (Pred == ICmpInst::ICMP_EQ ? FalseBB : TrueBB) &&
1878 "Broken CFG: missing edge from predecessor to successor");
1885 Instruction *InstCombiner::visitFree(CallInst &FI) {
1886 Value *Op = FI.getArgOperand(0);
1888 // free undef -> unreachable.
1889 if (isa<UndefValue>(Op)) {
1890 // Insert a new store to null because we cannot modify the CFG here.
1891 Builder->CreateStore(ConstantInt::getTrue(FI.getContext()),
1892 UndefValue::get(Type::getInt1PtrTy(FI.getContext())));
1893 return EraseInstFromFunction(FI);
1896 // If we have 'free null' delete the instruction. This can happen in stl code
1897 // when lots of inlining happens.
1898 if (isa<ConstantPointerNull>(Op))
1899 return EraseInstFromFunction(FI);
1901 // If we optimize for code size, try to move the call to free before the null
1902 // test so that simplify cfg can remove the empty block and dead code
1903 // elimination the branch. I.e., helps to turn something like:
1904 // if (foo) free(foo);
1908 if (Instruction *I = tryToMoveFreeBeforeNullTest(FI))
1916 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
1917 // Change br (not X), label True, label False to: br X, label False, True
1919 BasicBlock *TrueDest;
1920 BasicBlock *FalseDest;
1921 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
1922 !isa<Constant>(X)) {
1923 // Swap Destinations and condition...
1925 BI.swapSuccessors();
1929 // Canonicalize fcmp_one -> fcmp_oeq
1930 FCmpInst::Predicate FPred; Value *Y;
1931 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
1932 TrueDest, FalseDest)) &&
1933 BI.getCondition()->hasOneUse())
1934 if (FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
1935 FPred == FCmpInst::FCMP_OGE) {
1936 FCmpInst *Cond = cast<FCmpInst>(BI.getCondition());
1937 Cond->setPredicate(FCmpInst::getInversePredicate(FPred));
1939 // Swap Destinations and condition.
1940 BI.swapSuccessors();
1945 // Canonicalize icmp_ne -> icmp_eq
1946 ICmpInst::Predicate IPred;
1947 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
1948 TrueDest, FalseDest)) &&
1949 BI.getCondition()->hasOneUse())
1950 if (IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
1951 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
1952 IPred == ICmpInst::ICMP_SGE) {
1953 ICmpInst *Cond = cast<ICmpInst>(BI.getCondition());
1954 Cond->setPredicate(ICmpInst::getInversePredicate(IPred));
1955 // Swap Destinations and condition.
1956 BI.swapSuccessors();
1964 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
1965 Value *Cond = SI.getCondition();
1966 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
1967 if (I->getOpcode() == Instruction::Add)
1968 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1969 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
1970 // Skip the first item since that's the default case.
1971 for (SwitchInst::CaseIt i = SI.case_begin(), e = SI.case_end();
1973 ConstantInt* CaseVal = i.getCaseValue();
1974 Constant* NewCaseVal = ConstantExpr::getSub(cast<Constant>(CaseVal),
1976 assert(isa<ConstantInt>(NewCaseVal) &&
1977 "Result of expression should be constant");
1978 i.setValue(cast<ConstantInt>(NewCaseVal));
1980 SI.setCondition(I->getOperand(0));
1988 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
1989 Value *Agg = EV.getAggregateOperand();
1991 if (!EV.hasIndices())
1992 return ReplaceInstUsesWith(EV, Agg);
1994 if (Constant *C = dyn_cast<Constant>(Agg)) {
1995 if (Constant *C2 = C->getAggregateElement(*EV.idx_begin())) {
1996 if (EV.getNumIndices() == 0)
1997 return ReplaceInstUsesWith(EV, C2);
1998 // Extract the remaining indices out of the constant indexed by the
2000 return ExtractValueInst::Create(C2, EV.getIndices().slice(1));
2002 return nullptr; // Can't handle other constants
2005 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
2006 // We're extracting from an insertvalue instruction, compare the indices
2007 const unsigned *exti, *exte, *insi, *inse;
2008 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
2009 exte = EV.idx_end(), inse = IV->idx_end();
2010 exti != exte && insi != inse;
2013 // The insert and extract both reference distinctly different elements.
2014 // This means the extract is not influenced by the insert, and we can
2015 // replace the aggregate operand of the extract with the aggregate
2016 // operand of the insert. i.e., replace
2017 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
2018 // %E = extractvalue { i32, { i32 } } %I, 0
2020 // %E = extractvalue { i32, { i32 } } %A, 0
2021 return ExtractValueInst::Create(IV->getAggregateOperand(),
2024 if (exti == exte && insi == inse)
2025 // Both iterators are at the end: Index lists are identical. Replace
2026 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
2027 // %C = extractvalue { i32, { i32 } } %B, 1, 0
2029 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
2031 // The extract list is a prefix of the insert list. i.e. replace
2032 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
2033 // %E = extractvalue { i32, { i32 } } %I, 1
2035 // %X = extractvalue { i32, { i32 } } %A, 1
2036 // %E = insertvalue { i32 } %X, i32 42, 0
2037 // by switching the order of the insert and extract (though the
2038 // insertvalue should be left in, since it may have other uses).
2039 Value *NewEV = Builder->CreateExtractValue(IV->getAggregateOperand(),
2041 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
2042 makeArrayRef(insi, inse));
2045 // The insert list is a prefix of the extract list
2046 // We can simply remove the common indices from the extract and make it
2047 // operate on the inserted value instead of the insertvalue result.
2049 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
2050 // %E = extractvalue { i32, { i32 } } %I, 1, 0
2052 // %E extractvalue { i32 } { i32 42 }, 0
2053 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
2054 makeArrayRef(exti, exte));
2056 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) {
2057 // We're extracting from an intrinsic, see if we're the only user, which
2058 // allows us to simplify multiple result intrinsics to simpler things that
2059 // just get one value.
2060 if (II->hasOneUse()) {
2061 // Check if we're grabbing the overflow bit or the result of a 'with
2062 // overflow' intrinsic. If it's the latter we can remove the intrinsic
2063 // and replace it with a traditional binary instruction.
2064 switch (II->getIntrinsicID()) {
2065 case Intrinsic::uadd_with_overflow:
2066 case Intrinsic::sadd_with_overflow:
2067 if (*EV.idx_begin() == 0) { // Normal result.
2068 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2069 ReplaceInstUsesWith(*II, UndefValue::get(II->getType()));
2070 EraseInstFromFunction(*II);
2071 return BinaryOperator::CreateAdd(LHS, RHS);
2074 // If the normal result of the add is dead, and the RHS is a constant,
2075 // we can transform this into a range comparison.
2076 // overflow = uadd a, -4 --> overflow = icmp ugt a, 3
2077 if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow)
2078 if (ConstantInt *CI = dyn_cast<ConstantInt>(II->getArgOperand(1)))
2079 return new ICmpInst(ICmpInst::ICMP_UGT, II->getArgOperand(0),
2080 ConstantExpr::getNot(CI));
2082 case Intrinsic::usub_with_overflow:
2083 case Intrinsic::ssub_with_overflow:
2084 if (*EV.idx_begin() == 0) { // Normal result.
2085 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2086 ReplaceInstUsesWith(*II, UndefValue::get(II->getType()));
2087 EraseInstFromFunction(*II);
2088 return BinaryOperator::CreateSub(LHS, RHS);
2091 case Intrinsic::umul_with_overflow:
2092 case Intrinsic::smul_with_overflow:
2093 if (*EV.idx_begin() == 0) { // Normal result.
2094 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2095 ReplaceInstUsesWith(*II, UndefValue::get(II->getType()));
2096 EraseInstFromFunction(*II);
2097 return BinaryOperator::CreateMul(LHS, RHS);
2105 if (LoadInst *L = dyn_cast<LoadInst>(Agg))
2106 // If the (non-volatile) load only has one use, we can rewrite this to a
2107 // load from a GEP. This reduces the size of the load.
2108 // FIXME: If a load is used only by extractvalue instructions then this
2109 // could be done regardless of having multiple uses.
2110 if (L->isSimple() && L->hasOneUse()) {
2111 // extractvalue has integer indices, getelementptr has Value*s. Convert.
2112 SmallVector<Value*, 4> Indices;
2113 // Prefix an i32 0 since we need the first element.
2114 Indices.push_back(Builder->getInt32(0));
2115 for (ExtractValueInst::idx_iterator I = EV.idx_begin(), E = EV.idx_end();
2117 Indices.push_back(Builder->getInt32(*I));
2119 // We need to insert these at the location of the old load, not at that of
2120 // the extractvalue.
2121 Builder->SetInsertPoint(L->getParent(), L);
2122 Value *GEP = Builder->CreateInBoundsGEP(L->getPointerOperand(), Indices);
2123 // Returning the load directly will cause the main loop to insert it in
2124 // the wrong spot, so use ReplaceInstUsesWith().
2125 return ReplaceInstUsesWith(EV, Builder->CreateLoad(GEP));
2127 // We could simplify extracts from other values. Note that nested extracts may
2128 // already be simplified implicitly by the above: extract (extract (insert) )
2129 // will be translated into extract ( insert ( extract ) ) first and then just
2130 // the value inserted, if appropriate. Similarly for extracts from single-use
2131 // loads: extract (extract (load)) will be translated to extract (load (gep))
2132 // and if again single-use then via load (gep (gep)) to load (gep).
2133 // However, double extracts from e.g. function arguments or return values
2134 // aren't handled yet.
2138 enum Personality_Type {
2139 Unknown_Personality,
2140 GNU_Ada_Personality,
2141 GNU_CXX_Personality,
2142 GNU_ObjC_Personality
2145 /// RecognizePersonality - See if the given exception handling personality
2146 /// function is one that we understand. If so, return a description of it;
2147 /// otherwise return Unknown_Personality.
2148 static Personality_Type RecognizePersonality(Value *Pers) {
2149 Function *F = dyn_cast<Function>(Pers->stripPointerCasts());
2151 return Unknown_Personality;
2152 return StringSwitch<Personality_Type>(F->getName())
2153 .Case("__gnat_eh_personality", GNU_Ada_Personality)
2154 .Case("__gxx_personality_v0", GNU_CXX_Personality)
2155 .Case("__objc_personality_v0", GNU_ObjC_Personality)
2156 .Default(Unknown_Personality);
2159 /// isCatchAll - Return 'true' if the given typeinfo will match anything.
2160 static bool isCatchAll(Personality_Type Personality, Constant *TypeInfo) {
2161 switch (Personality) {
2162 case Unknown_Personality:
2164 case GNU_Ada_Personality:
2165 // While __gnat_all_others_value will match any Ada exception, it doesn't
2166 // match foreign exceptions (or didn't, before gcc-4.7).
2168 case GNU_CXX_Personality:
2169 case GNU_ObjC_Personality:
2170 return TypeInfo->isNullValue();
2172 llvm_unreachable("Unknown personality!");
2175 static bool shorter_filter(const Value *LHS, const Value *RHS) {
2177 cast<ArrayType>(LHS->getType())->getNumElements()
2179 cast<ArrayType>(RHS->getType())->getNumElements();
2182 Instruction *InstCombiner::visitLandingPadInst(LandingPadInst &LI) {
2183 // The logic here should be correct for any real-world personality function.
2184 // However if that turns out not to be true, the offending logic can always
2185 // be conditioned on the personality function, like the catch-all logic is.
2186 Personality_Type Personality = RecognizePersonality(LI.getPersonalityFn());
2188 // Simplify the list of clauses, eg by removing repeated catch clauses
2189 // (these are often created by inlining).
2190 bool MakeNewInstruction = false; // If true, recreate using the following:
2191 SmallVector<Constant *, 16> NewClauses; // - Clauses for the new instruction;
2192 bool CleanupFlag = LI.isCleanup(); // - The new instruction is a cleanup.
2194 SmallPtrSet<Value *, 16> AlreadyCaught; // Typeinfos known caught already.
2195 for (unsigned i = 0, e = LI.getNumClauses(); i != e; ++i) {
2196 bool isLastClause = i + 1 == e;
2197 if (LI.isCatch(i)) {
2199 Constant *CatchClause = LI.getClause(i);
2200 Constant *TypeInfo = CatchClause->stripPointerCasts();
2202 // If we already saw this clause, there is no point in having a second
2204 if (AlreadyCaught.insert(TypeInfo)) {
2205 // This catch clause was not already seen.
2206 NewClauses.push_back(CatchClause);
2208 // Repeated catch clause - drop the redundant copy.
2209 MakeNewInstruction = true;
2212 // If this is a catch-all then there is no point in keeping any following
2213 // clauses or marking the landingpad as having a cleanup.
2214 if (isCatchAll(Personality, TypeInfo)) {
2216 MakeNewInstruction = true;
2217 CleanupFlag = false;
2221 // A filter clause. If any of the filter elements were already caught
2222 // then they can be dropped from the filter. It is tempting to try to
2223 // exploit the filter further by saying that any typeinfo that does not
2224 // occur in the filter can't be caught later (and thus can be dropped).
2225 // However this would be wrong, since typeinfos can match without being
2226 // equal (for example if one represents a C++ class, and the other some
2227 // class derived from it).
2228 assert(LI.isFilter(i) && "Unsupported landingpad clause!");
2229 Constant *FilterClause = LI.getClause(i);
2230 ArrayType *FilterType = cast<ArrayType>(FilterClause->getType());
2231 unsigned NumTypeInfos = FilterType->getNumElements();
2233 // An empty filter catches everything, so there is no point in keeping any
2234 // following clauses or marking the landingpad as having a cleanup. By
2235 // dealing with this case here the following code is made a bit simpler.
2236 if (!NumTypeInfos) {
2237 NewClauses.push_back(FilterClause);
2239 MakeNewInstruction = true;
2240 CleanupFlag = false;
2244 bool MakeNewFilter = false; // If true, make a new filter.
2245 SmallVector<Constant *, 16> NewFilterElts; // New elements.
2246 if (isa<ConstantAggregateZero>(FilterClause)) {
2247 // Not an empty filter - it contains at least one null typeinfo.
2248 assert(NumTypeInfos > 0 && "Should have handled empty filter already!");
2249 Constant *TypeInfo =
2250 Constant::getNullValue(FilterType->getElementType());
2251 // If this typeinfo is a catch-all then the filter can never match.
2252 if (isCatchAll(Personality, TypeInfo)) {
2253 // Throw the filter away.
2254 MakeNewInstruction = true;
2258 // There is no point in having multiple copies of this typeinfo, so
2259 // discard all but the first copy if there is more than one.
2260 NewFilterElts.push_back(TypeInfo);
2261 if (NumTypeInfos > 1)
2262 MakeNewFilter = true;
2264 ConstantArray *Filter = cast<ConstantArray>(FilterClause);
2265 SmallPtrSet<Value *, 16> SeenInFilter; // For uniquing the elements.
2266 NewFilterElts.reserve(NumTypeInfos);
2268 // Remove any filter elements that were already caught or that already
2269 // occurred in the filter. While there, see if any of the elements are
2270 // catch-alls. If so, the filter can be discarded.
2271 bool SawCatchAll = false;
2272 for (unsigned j = 0; j != NumTypeInfos; ++j) {
2273 Constant *Elt = Filter->getOperand(j);
2274 Constant *TypeInfo = Elt->stripPointerCasts();
2275 if (isCatchAll(Personality, TypeInfo)) {
2276 // This element is a catch-all. Bail out, noting this fact.
2280 if (AlreadyCaught.count(TypeInfo))
2281 // Already caught by an earlier clause, so having it in the filter
2284 // There is no point in having multiple copies of the same typeinfo in
2285 // a filter, so only add it if we didn't already.
2286 if (SeenInFilter.insert(TypeInfo))
2287 NewFilterElts.push_back(cast<Constant>(Elt));
2289 // A filter containing a catch-all cannot match anything by definition.
2291 // Throw the filter away.
2292 MakeNewInstruction = true;
2296 // If we dropped something from the filter, make a new one.
2297 if (NewFilterElts.size() < NumTypeInfos)
2298 MakeNewFilter = true;
2300 if (MakeNewFilter) {
2301 FilterType = ArrayType::get(FilterType->getElementType(),
2302 NewFilterElts.size());
2303 FilterClause = ConstantArray::get(FilterType, NewFilterElts);
2304 MakeNewInstruction = true;
2307 NewClauses.push_back(FilterClause);
2309 // If the new filter is empty then it will catch everything so there is
2310 // no point in keeping any following clauses or marking the landingpad
2311 // as having a cleanup. The case of the original filter being empty was
2312 // already handled above.
2313 if (MakeNewFilter && !NewFilterElts.size()) {
2314 assert(MakeNewInstruction && "New filter but not a new instruction!");
2315 CleanupFlag = false;
2321 // If several filters occur in a row then reorder them so that the shortest
2322 // filters come first (those with the smallest number of elements). This is
2323 // advantageous because shorter filters are more likely to match, speeding up
2324 // unwinding, but mostly because it increases the effectiveness of the other
2325 // filter optimizations below.
2326 for (unsigned i = 0, e = NewClauses.size(); i + 1 < e; ) {
2328 // Find the maximal 'j' s.t. the range [i, j) consists entirely of filters.
2329 for (j = i; j != e; ++j)
2330 if (!isa<ArrayType>(NewClauses[j]->getType()))
2333 // Check whether the filters are already sorted by length. We need to know
2334 // if sorting them is actually going to do anything so that we only make a
2335 // new landingpad instruction if it does.
2336 for (unsigned k = i; k + 1 < j; ++k)
2337 if (shorter_filter(NewClauses[k+1], NewClauses[k])) {
2338 // Not sorted, so sort the filters now. Doing an unstable sort would be
2339 // correct too but reordering filters pointlessly might confuse users.
2340 std::stable_sort(NewClauses.begin() + i, NewClauses.begin() + j,
2342 MakeNewInstruction = true;
2346 // Look for the next batch of filters.
2350 // If typeinfos matched if and only if equal, then the elements of a filter L
2351 // that occurs later than a filter F could be replaced by the intersection of
2352 // the elements of F and L. In reality two typeinfos can match without being
2353 // equal (for example if one represents a C++ class, and the other some class
2354 // derived from it) so it would be wrong to perform this transform in general.
2355 // However the transform is correct and useful if F is a subset of L. In that
2356 // case L can be replaced by F, and thus removed altogether since repeating a
2357 // filter is pointless. So here we look at all pairs of filters F and L where
2358 // L follows F in the list of clauses, and remove L if every element of F is
2359 // an element of L. This can occur when inlining C++ functions with exception
2361 for (unsigned i = 0; i + 1 < NewClauses.size(); ++i) {
2362 // Examine each filter in turn.
2363 Value *Filter = NewClauses[i];
2364 ArrayType *FTy = dyn_cast<ArrayType>(Filter->getType());
2366 // Not a filter - skip it.
2368 unsigned FElts = FTy->getNumElements();
2369 // Examine each filter following this one. Doing this backwards means that
2370 // we don't have to worry about filters disappearing under us when removed.
2371 for (unsigned j = NewClauses.size() - 1; j != i; --j) {
2372 Value *LFilter = NewClauses[j];
2373 ArrayType *LTy = dyn_cast<ArrayType>(LFilter->getType());
2375 // Not a filter - skip it.
2377 // If Filter is a subset of LFilter, i.e. every element of Filter is also
2378 // an element of LFilter, then discard LFilter.
2379 SmallVectorImpl<Constant *>::iterator J = NewClauses.begin() + j;
2380 // If Filter is empty then it is a subset of LFilter.
2383 NewClauses.erase(J);
2384 MakeNewInstruction = true;
2385 // Move on to the next filter.
2388 unsigned LElts = LTy->getNumElements();
2389 // If Filter is longer than LFilter then it cannot be a subset of it.
2391 // Move on to the next filter.
2393 // At this point we know that LFilter has at least one element.
2394 if (isa<ConstantAggregateZero>(LFilter)) { // LFilter only contains zeros.
2395 // Filter is a subset of LFilter iff Filter contains only zeros (as we
2396 // already know that Filter is not longer than LFilter).
2397 if (isa<ConstantAggregateZero>(Filter)) {
2398 assert(FElts <= LElts && "Should have handled this case earlier!");
2400 NewClauses.erase(J);
2401 MakeNewInstruction = true;
2403 // Move on to the next filter.
2406 ConstantArray *LArray = cast<ConstantArray>(LFilter);
2407 if (isa<ConstantAggregateZero>(Filter)) { // Filter only contains zeros.
2408 // Since Filter is non-empty and contains only zeros, it is a subset of
2409 // LFilter iff LFilter contains a zero.
2410 assert(FElts > 0 && "Should have eliminated the empty filter earlier!");
2411 for (unsigned l = 0; l != LElts; ++l)
2412 if (LArray->getOperand(l)->isNullValue()) {
2413 // LFilter contains a zero - discard it.
2414 NewClauses.erase(J);
2415 MakeNewInstruction = true;
2418 // Move on to the next filter.
2421 // At this point we know that both filters are ConstantArrays. Loop over
2422 // operands to see whether every element of Filter is also an element of
2423 // LFilter. Since filters tend to be short this is probably faster than
2424 // using a method that scales nicely.
2425 ConstantArray *FArray = cast<ConstantArray>(Filter);
2426 bool AllFound = true;
2427 for (unsigned f = 0; f != FElts; ++f) {
2428 Value *FTypeInfo = FArray->getOperand(f)->stripPointerCasts();
2430 for (unsigned l = 0; l != LElts; ++l) {
2431 Value *LTypeInfo = LArray->getOperand(l)->stripPointerCasts();
2432 if (LTypeInfo == FTypeInfo) {
2442 NewClauses.erase(J);
2443 MakeNewInstruction = true;
2445 // Move on to the next filter.
2449 // If we changed any of the clauses, replace the old landingpad instruction
2451 if (MakeNewInstruction) {
2452 LandingPadInst *NLI = LandingPadInst::Create(LI.getType(),
2453 LI.getPersonalityFn(),
2455 for (unsigned i = 0, e = NewClauses.size(); i != e; ++i)
2456 NLI->addClause(NewClauses[i]);
2457 // A landing pad with no clauses must have the cleanup flag set. It is
2458 // theoretically possible, though highly unlikely, that we eliminated all
2459 // clauses. If so, force the cleanup flag to true.
2460 if (NewClauses.empty())
2462 NLI->setCleanup(CleanupFlag);
2466 // Even if none of the clauses changed, we may nonetheless have understood
2467 // that the cleanup flag is pointless. Clear it if so.
2468 if (LI.isCleanup() != CleanupFlag) {
2469 assert(!CleanupFlag && "Adding a cleanup, not removing one?!");
2470 LI.setCleanup(CleanupFlag);
2480 /// TryToSinkInstruction - Try to move the specified instruction from its
2481 /// current block into the beginning of DestBlock, which can only happen if it's
2482 /// safe to move the instruction past all of the instructions between it and the
2483 /// end of its block.
2484 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
2485 assert(I->hasOneUse() && "Invariants didn't hold!");
2487 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
2488 if (isa<PHINode>(I) || isa<LandingPadInst>(I) || I->mayHaveSideEffects() ||
2489 isa<TerminatorInst>(I))
2492 // Do not sink alloca instructions out of the entry block.
2493 if (isa<AllocaInst>(I) && I->getParent() ==
2494 &DestBlock->getParent()->getEntryBlock())
2497 // We can only sink load instructions if there is nothing between the load and
2498 // the end of block that could change the value.
2499 if (I->mayReadFromMemory()) {
2500 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
2502 if (Scan->mayWriteToMemory())
2506 BasicBlock::iterator InsertPos = DestBlock->getFirstInsertionPt();
2507 I->moveBefore(InsertPos);
2513 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
2514 /// all reachable code to the worklist.
2516 /// This has a couple of tricks to make the code faster and more powerful. In
2517 /// particular, we constant fold and DCE instructions as we go, to avoid adding
2518 /// them to the worklist (this significantly speeds up instcombine on code where
2519 /// many instructions are dead or constant). Additionally, if we find a branch
2520 /// whose condition is a known constant, we only visit the reachable successors.
2522 static bool AddReachableCodeToWorklist(BasicBlock *BB,
2523 SmallPtrSet<BasicBlock*, 64> &Visited,
2525 const DataLayout *DL,
2526 const TargetLibraryInfo *TLI) {
2527 bool MadeIRChange = false;
2528 SmallVector<BasicBlock*, 256> Worklist;
2529 Worklist.push_back(BB);
2531 SmallVector<Instruction*, 128> InstrsForInstCombineWorklist;
2532 DenseMap<ConstantExpr*, Constant*> FoldedConstants;
2535 BB = Worklist.pop_back_val();
2537 // We have now visited this block! If we've already been here, ignore it.
2538 if (!Visited.insert(BB)) continue;
2540 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
2541 Instruction *Inst = BBI++;
2543 // DCE instruction if trivially dead.
2544 if (isInstructionTriviallyDead(Inst, TLI)) {
2546 DEBUG(dbgs() << "IC: DCE: " << *Inst << '\n');
2547 Inst->eraseFromParent();
2551 // ConstantProp instruction if trivially constant.
2552 if (!Inst->use_empty() && isa<Constant>(Inst->getOperand(0)))
2553 if (Constant *C = ConstantFoldInstruction(Inst, DL, TLI)) {
2554 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: "
2556 Inst->replaceAllUsesWith(C);
2558 Inst->eraseFromParent();
2563 // See if we can constant fold its operands.
2564 for (User::op_iterator i = Inst->op_begin(), e = Inst->op_end();
2566 ConstantExpr *CE = dyn_cast<ConstantExpr>(i);
2567 if (CE == nullptr) continue;
2569 Constant*& FoldRes = FoldedConstants[CE];
2571 FoldRes = ConstantFoldConstantExpression(CE, DL, TLI);
2575 if (FoldRes != CE) {
2577 MadeIRChange = true;
2582 InstrsForInstCombineWorklist.push_back(Inst);
2585 // Recursively visit successors. If this is a branch or switch on a
2586 // constant, only visit the reachable successor.
2587 TerminatorInst *TI = BB->getTerminator();
2588 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
2589 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
2590 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
2591 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
2592 Worklist.push_back(ReachableBB);
2595 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
2596 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
2597 // See if this is an explicit destination.
2598 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2600 if (i.getCaseValue() == Cond) {
2601 BasicBlock *ReachableBB = i.getCaseSuccessor();
2602 Worklist.push_back(ReachableBB);
2606 // Otherwise it is the default destination.
2607 Worklist.push_back(SI->getDefaultDest());
2612 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
2613 Worklist.push_back(TI->getSuccessor(i));
2614 } while (!Worklist.empty());
2616 // Once we've found all of the instructions to add to instcombine's worklist,
2617 // add them in reverse order. This way instcombine will visit from the top
2618 // of the function down. This jives well with the way that it adds all uses
2619 // of instructions to the worklist after doing a transformation, thus avoiding
2620 // some N^2 behavior in pathological cases.
2621 IC.Worklist.AddInitialGroup(&InstrsForInstCombineWorklist[0],
2622 InstrsForInstCombineWorklist.size());
2624 return MadeIRChange;
2627 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
2628 MadeIRChange = false;
2630 DEBUG(dbgs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
2631 << F.getName() << "\n");
2634 // Do a depth-first traversal of the function, populate the worklist with
2635 // the reachable instructions. Ignore blocks that are not reachable. Keep
2636 // track of which blocks we visit.
2637 SmallPtrSet<BasicBlock*, 64> Visited;
2638 MadeIRChange |= AddReachableCodeToWorklist(F.begin(), Visited, *this, DL,
2641 // Do a quick scan over the function. If we find any blocks that are
2642 // unreachable, remove any instructions inside of them. This prevents
2643 // the instcombine code from having to deal with some bad special cases.
2644 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
2645 if (Visited.count(BB)) continue;
2647 // Delete the instructions backwards, as it has a reduced likelihood of
2648 // having to update as many def-use and use-def chains.
2649 Instruction *EndInst = BB->getTerminator(); // Last not to be deleted.
2650 while (EndInst != BB->begin()) {
2651 // Delete the next to last instruction.
2652 BasicBlock::iterator I = EndInst;
2653 Instruction *Inst = --I;
2654 if (!Inst->use_empty())
2655 Inst->replaceAllUsesWith(UndefValue::get(Inst->getType()));
2656 if (isa<LandingPadInst>(Inst)) {
2660 if (!isa<DbgInfoIntrinsic>(Inst)) {
2662 MadeIRChange = true;
2664 Inst->eraseFromParent();
2669 while (!Worklist.isEmpty()) {
2670 Instruction *I = Worklist.RemoveOne();
2671 if (I == nullptr) continue; // skip null values.
2673 // Check to see if we can DCE the instruction.
2674 if (isInstructionTriviallyDead(I, TLI)) {
2675 DEBUG(dbgs() << "IC: DCE: " << *I << '\n');
2676 EraseInstFromFunction(*I);
2678 MadeIRChange = true;
2682 // Instruction isn't dead, see if we can constant propagate it.
2683 if (!I->use_empty() && isa<Constant>(I->getOperand(0)))
2684 if (Constant *C = ConstantFoldInstruction(I, DL, TLI)) {
2685 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n');
2687 // Add operands to the worklist.
2688 ReplaceInstUsesWith(*I, C);
2690 EraseInstFromFunction(*I);
2691 MadeIRChange = true;
2695 // See if we can trivially sink this instruction to a successor basic block.
2696 if (I->hasOneUse()) {
2697 BasicBlock *BB = I->getParent();
2698 Instruction *UserInst = cast<Instruction>(*I->user_begin());
2699 BasicBlock *UserParent;
2701 // Get the block the use occurs in.
2702 if (PHINode *PN = dyn_cast<PHINode>(UserInst))
2703 UserParent = PN->getIncomingBlock(*I->use_begin());
2705 UserParent = UserInst->getParent();
2707 if (UserParent != BB) {
2708 bool UserIsSuccessor = false;
2709 // See if the user is one of our successors.
2710 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
2711 if (*SI == UserParent) {
2712 UserIsSuccessor = true;
2716 // If the user is one of our immediate successors, and if that successor
2717 // only has us as a predecessors (we'd have to split the critical edge
2718 // otherwise), we can keep going.
2719 if (UserIsSuccessor && UserParent->getSinglePredecessor())
2720 // Okay, the CFG is simple enough, try to sink this instruction.
2721 MadeIRChange |= TryToSinkInstruction(I, UserParent);
2725 // Now that we have an instruction, try combining it to simplify it.
2726 Builder->SetInsertPoint(I->getParent(), I);
2727 Builder->SetCurrentDebugLocation(I->getDebugLoc());
2732 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str(););
2733 DEBUG(dbgs() << "IC: Visiting: " << OrigI << '\n');
2735 if (Instruction *Result = visit(*I)) {
2737 // Should we replace the old instruction with a new one?
2739 DEBUG(dbgs() << "IC: Old = " << *I << '\n'
2740 << " New = " << *Result << '\n');
2742 if (!I->getDebugLoc().isUnknown())
2743 Result->setDebugLoc(I->getDebugLoc());
2744 // Everything uses the new instruction now.
2745 I->replaceAllUsesWith(Result);
2747 // Move the name to the new instruction first.
2748 Result->takeName(I);
2750 // Push the new instruction and any users onto the worklist.
2751 Worklist.Add(Result);
2752 Worklist.AddUsersToWorkList(*Result);
2754 // Insert the new instruction into the basic block...
2755 BasicBlock *InstParent = I->getParent();
2756 BasicBlock::iterator InsertPos = I;
2758 // If we replace a PHI with something that isn't a PHI, fix up the
2760 if (!isa<PHINode>(Result) && isa<PHINode>(InsertPos))
2761 InsertPos = InstParent->getFirstInsertionPt();
2763 InstParent->getInstList().insert(InsertPos, Result);
2765 EraseInstFromFunction(*I);
2768 DEBUG(dbgs() << "IC: Mod = " << OrigI << '\n'
2769 << " New = " << *I << '\n');
2772 // If the instruction was modified, it's possible that it is now dead.
2773 // if so, remove it.
2774 if (isInstructionTriviallyDead(I, TLI)) {
2775 EraseInstFromFunction(*I);
2778 Worklist.AddUsersToWorkList(*I);
2781 MadeIRChange = true;
2786 return MadeIRChange;
2790 class InstCombinerLibCallSimplifier : public LibCallSimplifier {
2793 InstCombinerLibCallSimplifier(const DataLayout *DL,
2794 const TargetLibraryInfo *TLI,
2796 : LibCallSimplifier(DL, TLI, UnsafeFPShrink) {
2800 /// replaceAllUsesWith - override so that instruction replacement
2801 /// can be defined in terms of the instruction combiner framework.
2802 void replaceAllUsesWith(Instruction *I, Value *With) const override {
2803 IC->ReplaceInstUsesWith(*I, With);
2808 bool InstCombiner::runOnFunction(Function &F) {
2809 if (skipOptnoneFunction(F))
2812 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
2813 DL = DLP ? &DLP->getDataLayout() : nullptr;
2814 TLI = &getAnalysis<TargetLibraryInfo>();
2816 MinimizeSize = F.getAttributes().hasAttribute(AttributeSet::FunctionIndex,
2817 Attribute::MinSize);
2819 /// Builder - This is an IRBuilder that automatically inserts new
2820 /// instructions into the worklist when they are created.
2821 IRBuilder<true, TargetFolder, InstCombineIRInserter>
2822 TheBuilder(F.getContext(), TargetFolder(DL),
2823 InstCombineIRInserter(Worklist));
2824 Builder = &TheBuilder;
2826 InstCombinerLibCallSimplifier TheSimplifier(DL, TLI, this);
2827 Simplifier = &TheSimplifier;
2829 bool EverMadeChange = false;
2831 // Lower dbg.declare intrinsics otherwise their value may be clobbered
2833 EverMadeChange = LowerDbgDeclare(F);
2835 // Iterate while there is work to do.
2836 unsigned Iteration = 0;
2837 while (DoOneIteration(F, Iteration++))
2838 EverMadeChange = true;
2841 return EverMadeChange;
2844 FunctionPass *llvm::createInstructionCombiningPass() {
2845 return new InstCombiner();