1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // InstructionCombining - Combine instructions to form fewer, simple
11 // instructions. This pass does not modify the CFG. This pass is where
12 // algebraic simplification happens.
14 // This pass combines things like:
20 // This is a simple worklist driven algorithm.
22 // This pass guarantees that the following canonicalizations are performed on
24 // 1. If a binary operator has a constant operand, it is moved to the RHS
25 // 2. Bitwise operators with constant operands are always grouped so that
26 // shifts are performed first, then or's, then and's, then xor's.
27 // 3. Compare instructions are converted from <,>,<=,>= to ==,!= if possible
28 // 4. All cmp instructions on boolean values are replaced with logical ops
29 // 5. add X, X is represented as (X*2) => (X << 1)
30 // 6. Multiplies with a power-of-two constant argument are transformed into
34 //===----------------------------------------------------------------------===//
36 #include "llvm/Transforms/InstCombine/InstCombine.h"
37 #include "InstCombineInternal.h"
38 #include "llvm-c/Initialization.h"
39 #include "llvm/ADT/SmallPtrSet.h"
40 #include "llvm/ADT/Statistic.h"
41 #include "llvm/ADT/StringSwitch.h"
42 #include "llvm/Analysis/AssumptionCache.h"
43 #include "llvm/Analysis/CFG.h"
44 #include "llvm/Analysis/ConstantFolding.h"
45 #include "llvm/Analysis/GlobalsModRef.h"
46 #include "llvm/Analysis/InstructionSimplify.h"
47 #include "llvm/Analysis/LibCallSemantics.h"
48 #include "llvm/Analysis/LoopInfo.h"
49 #include "llvm/Analysis/MemoryBuiltins.h"
50 #include "llvm/Analysis/TargetLibraryInfo.h"
51 #include "llvm/Analysis/ValueTracking.h"
52 #include "llvm/IR/CFG.h"
53 #include "llvm/IR/DataLayout.h"
54 #include "llvm/IR/Dominators.h"
55 #include "llvm/IR/GetElementPtrTypeIterator.h"
56 #include "llvm/IR/IntrinsicInst.h"
57 #include "llvm/IR/PatternMatch.h"
58 #include "llvm/IR/ValueHandle.h"
59 #include "llvm/Support/CommandLine.h"
60 #include "llvm/Support/Debug.h"
61 #include "llvm/Support/raw_ostream.h"
62 #include "llvm/Transforms/Scalar.h"
63 #include "llvm/Transforms/Utils/Local.h"
67 using namespace llvm::PatternMatch;
69 #define DEBUG_TYPE "instcombine"
71 STATISTIC(NumCombined , "Number of insts combined");
72 STATISTIC(NumConstProp, "Number of constant folds");
73 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
74 STATISTIC(NumSunkInst , "Number of instructions sunk");
75 STATISTIC(NumExpand, "Number of expansions");
76 STATISTIC(NumFactor , "Number of factorizations");
77 STATISTIC(NumReassoc , "Number of reassociations");
79 Value *InstCombiner::EmitGEPOffset(User *GEP) {
80 return llvm::EmitGEPOffset(Builder, DL, GEP);
83 /// ShouldChangeType - Return true if it is desirable to convert a computation
84 /// from 'From' to 'To'. We don't want to convert from a legal to an illegal
85 /// type for example, or from a smaller to a larger illegal type.
86 bool InstCombiner::ShouldChangeType(Type *From, Type *To) const {
87 assert(From->isIntegerTy() && To->isIntegerTy());
89 unsigned FromWidth = From->getPrimitiveSizeInBits();
90 unsigned ToWidth = To->getPrimitiveSizeInBits();
91 bool FromLegal = DL.isLegalInteger(FromWidth);
92 bool ToLegal = DL.isLegalInteger(ToWidth);
94 // If this is a legal integer from type, and the result would be an illegal
95 // type, don't do the transformation.
96 if (FromLegal && !ToLegal)
99 // Otherwise, if both are illegal, do not increase the size of the result. We
100 // do allow things like i160 -> i64, but not i64 -> i160.
101 if (!FromLegal && !ToLegal && ToWidth > FromWidth)
107 // Return true, if No Signed Wrap should be maintained for I.
108 // The No Signed Wrap flag can be kept if the operation "B (I.getOpcode) C",
109 // where both B and C should be ConstantInts, results in a constant that does
110 // not overflow. This function only handles the Add and Sub opcodes. For
111 // all other opcodes, the function conservatively returns false.
112 static bool MaintainNoSignedWrap(BinaryOperator &I, Value *B, Value *C) {
113 OverflowingBinaryOperator *OBO = dyn_cast<OverflowingBinaryOperator>(&I);
114 if (!OBO || !OBO->hasNoSignedWrap()) {
118 // We reason about Add and Sub Only.
119 Instruction::BinaryOps Opcode = I.getOpcode();
120 if (Opcode != Instruction::Add &&
121 Opcode != Instruction::Sub) {
125 ConstantInt *CB = dyn_cast<ConstantInt>(B);
126 ConstantInt *CC = dyn_cast<ConstantInt>(C);
132 const APInt &BVal = CB->getValue();
133 const APInt &CVal = CC->getValue();
134 bool Overflow = false;
136 if (Opcode == Instruction::Add) {
137 BVal.sadd_ov(CVal, Overflow);
139 BVal.ssub_ov(CVal, Overflow);
145 /// Conservatively clears subclassOptionalData after a reassociation or
146 /// commutation. We preserve fast-math flags when applicable as they can be
148 static void ClearSubclassDataAfterReassociation(BinaryOperator &I) {
149 FPMathOperator *FPMO = dyn_cast<FPMathOperator>(&I);
151 I.clearSubclassOptionalData();
155 FastMathFlags FMF = I.getFastMathFlags();
156 I.clearSubclassOptionalData();
157 I.setFastMathFlags(FMF);
160 /// SimplifyAssociativeOrCommutative - This performs a few simplifications for
161 /// operators which are associative or commutative:
163 // Commutative operators:
165 // 1. Order operands such that they are listed from right (least complex) to
166 // left (most complex). This puts constants before unary operators before
169 // Associative operators:
171 // 2. Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
172 // 3. Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
174 // Associative and commutative operators:
176 // 4. Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
177 // 5. Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
178 // 6. Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
179 // if C1 and C2 are constants.
181 bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) {
182 Instruction::BinaryOps Opcode = I.getOpcode();
183 bool Changed = false;
186 // Order operands such that they are listed from right (least complex) to
187 // left (most complex). This puts constants before unary operators before
189 if (I.isCommutative() && getComplexity(I.getOperand(0)) <
190 getComplexity(I.getOperand(1)))
191 Changed = !I.swapOperands();
193 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
194 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
196 if (I.isAssociative()) {
197 // Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
198 if (Op0 && Op0->getOpcode() == Opcode) {
199 Value *A = Op0->getOperand(0);
200 Value *B = Op0->getOperand(1);
201 Value *C = I.getOperand(1);
203 // Does "B op C" simplify?
204 if (Value *V = SimplifyBinOp(Opcode, B, C, DL)) {
205 // It simplifies to V. Form "A op V".
208 // Conservatively clear the optional flags, since they may not be
209 // preserved by the reassociation.
210 if (MaintainNoSignedWrap(I, B, C) &&
211 (!Op0 || (isa<BinaryOperator>(Op0) && Op0->hasNoSignedWrap()))) {
212 // Note: this is only valid because SimplifyBinOp doesn't look at
213 // the operands to Op0.
214 I.clearSubclassOptionalData();
215 I.setHasNoSignedWrap(true);
217 ClearSubclassDataAfterReassociation(I);
226 // Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
227 if (Op1 && Op1->getOpcode() == Opcode) {
228 Value *A = I.getOperand(0);
229 Value *B = Op1->getOperand(0);
230 Value *C = Op1->getOperand(1);
232 // Does "A op B" simplify?
233 if (Value *V = SimplifyBinOp(Opcode, A, B, DL)) {
234 // It simplifies to V. Form "V op C".
237 // Conservatively clear the optional flags, since they may not be
238 // preserved by the reassociation.
239 ClearSubclassDataAfterReassociation(I);
247 if (I.isAssociative() && I.isCommutative()) {
248 // Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
249 if (Op0 && Op0->getOpcode() == Opcode) {
250 Value *A = Op0->getOperand(0);
251 Value *B = Op0->getOperand(1);
252 Value *C = I.getOperand(1);
254 // Does "C op A" simplify?
255 if (Value *V = SimplifyBinOp(Opcode, C, A, DL)) {
256 // It simplifies to V. Form "V op B".
259 // Conservatively clear the optional flags, since they may not be
260 // preserved by the reassociation.
261 ClearSubclassDataAfterReassociation(I);
268 // Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
269 if (Op1 && Op1->getOpcode() == Opcode) {
270 Value *A = I.getOperand(0);
271 Value *B = Op1->getOperand(0);
272 Value *C = Op1->getOperand(1);
274 // Does "C op A" simplify?
275 if (Value *V = SimplifyBinOp(Opcode, C, A, DL)) {
276 // It simplifies to V. Form "B op V".
279 // Conservatively clear the optional flags, since they may not be
280 // preserved by the reassociation.
281 ClearSubclassDataAfterReassociation(I);
288 // Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
289 // if C1 and C2 are constants.
291 Op0->getOpcode() == Opcode && Op1->getOpcode() == Opcode &&
292 isa<Constant>(Op0->getOperand(1)) &&
293 isa<Constant>(Op1->getOperand(1)) &&
294 Op0->hasOneUse() && Op1->hasOneUse()) {
295 Value *A = Op0->getOperand(0);
296 Constant *C1 = cast<Constant>(Op0->getOperand(1));
297 Value *B = Op1->getOperand(0);
298 Constant *C2 = cast<Constant>(Op1->getOperand(1));
300 Constant *Folded = ConstantExpr::get(Opcode, C1, C2);
301 BinaryOperator *New = BinaryOperator::Create(Opcode, A, B);
302 if (isa<FPMathOperator>(New)) {
303 FastMathFlags Flags = I.getFastMathFlags();
304 Flags &= Op0->getFastMathFlags();
305 Flags &= Op1->getFastMathFlags();
306 New->setFastMathFlags(Flags);
308 InsertNewInstWith(New, I);
310 I.setOperand(0, New);
311 I.setOperand(1, Folded);
312 // Conservatively clear the optional flags, since they may not be
313 // preserved by the reassociation.
314 ClearSubclassDataAfterReassociation(I);
321 // No further simplifications.
326 /// LeftDistributesOverRight - Whether "X LOp (Y ROp Z)" is always equal to
327 /// "(X LOp Y) ROp (X LOp Z)".
328 static bool LeftDistributesOverRight(Instruction::BinaryOps LOp,
329 Instruction::BinaryOps ROp) {
334 case Instruction::And:
335 // And distributes over Or and Xor.
339 case Instruction::Or:
340 case Instruction::Xor:
344 case Instruction::Mul:
345 // Multiplication distributes over addition and subtraction.
349 case Instruction::Add:
350 case Instruction::Sub:
354 case Instruction::Or:
355 // Or distributes over And.
359 case Instruction::And:
365 /// RightDistributesOverLeft - Whether "(X LOp Y) ROp Z" is always equal to
366 /// "(X ROp Z) LOp (Y ROp Z)".
367 static bool RightDistributesOverLeft(Instruction::BinaryOps LOp,
368 Instruction::BinaryOps ROp) {
369 if (Instruction::isCommutative(ROp))
370 return LeftDistributesOverRight(ROp, LOp);
375 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
376 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
377 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
378 case Instruction::And:
379 case Instruction::Or:
380 case Instruction::Xor:
384 case Instruction::Shl:
385 case Instruction::LShr:
386 case Instruction::AShr:
390 // TODO: It would be nice to handle division, aka "(X + Y)/Z = X/Z + Y/Z",
391 // but this requires knowing that the addition does not overflow and other
396 /// This function returns identity value for given opcode, which can be used to
397 /// factor patterns like (X * 2) + X ==> (X * 2) + (X * 1) ==> X * (2 + 1).
398 static Value *getIdentityValue(Instruction::BinaryOps OpCode, Value *V) {
399 if (isa<Constant>(V))
402 if (OpCode == Instruction::Mul)
403 return ConstantInt::get(V->getType(), 1);
405 // TODO: We can handle other cases e.g. Instruction::And, Instruction::Or etc.
410 /// This function factors binary ops which can be combined using distributive
411 /// laws. This function tries to transform 'Op' based TopLevelOpcode to enable
412 /// factorization e.g for ADD(SHL(X , 2), MUL(X, 5)), When this function called
413 /// with TopLevelOpcode == Instruction::Add and Op = SHL(X, 2), transforms
414 /// SHL(X, 2) to MUL(X, 4) i.e. returns Instruction::Mul with LHS set to 'X' and
416 static Instruction::BinaryOps
417 getBinOpsForFactorization(Instruction::BinaryOps TopLevelOpcode,
418 BinaryOperator *Op, Value *&LHS, Value *&RHS) {
420 return Instruction::BinaryOpsEnd;
422 LHS = Op->getOperand(0);
423 RHS = Op->getOperand(1);
425 switch (TopLevelOpcode) {
427 return Op->getOpcode();
429 case Instruction::Add:
430 case Instruction::Sub:
431 if (Op->getOpcode() == Instruction::Shl) {
432 if (Constant *CST = dyn_cast<Constant>(Op->getOperand(1))) {
433 // The multiplier is really 1 << CST.
434 RHS = ConstantExpr::getShl(ConstantInt::get(Op->getType(), 1), CST);
435 return Instruction::Mul;
438 return Op->getOpcode();
441 // TODO: We can add other conversions e.g. shr => div etc.
444 /// This tries to simplify binary operations by factorizing out common terms
445 /// (e. g. "(A*B)+(A*C)" -> "A*(B+C)").
446 static Value *tryFactorization(InstCombiner::BuilderTy *Builder,
447 const DataLayout &DL, BinaryOperator &I,
448 Instruction::BinaryOps InnerOpcode, Value *A,
449 Value *B, Value *C, Value *D) {
451 // If any of A, B, C, D are null, we can not factor I, return early.
452 // Checking A and C should be enough.
453 if (!A || !C || !B || !D)
457 Value *SimplifiedInst = nullptr;
458 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
459 Instruction::BinaryOps TopLevelOpcode = I.getOpcode();
461 // Does "X op' Y" always equal "Y op' X"?
462 bool InnerCommutative = Instruction::isCommutative(InnerOpcode);
464 // Does "X op' (Y op Z)" always equal "(X op' Y) op (X op' Z)"?
465 if (LeftDistributesOverRight(InnerOpcode, TopLevelOpcode))
466 // Does the instruction have the form "(A op' B) op (A op' D)" or, in the
467 // commutative case, "(A op' B) op (C op' A)"?
468 if (A == C || (InnerCommutative && A == D)) {
471 // Consider forming "A op' (B op D)".
472 // If "B op D" simplifies then it can be formed with no cost.
473 V = SimplifyBinOp(TopLevelOpcode, B, D, DL);
474 // If "B op D" doesn't simplify then only go on if both of the existing
475 // operations "A op' B" and "C op' D" will be zapped as no longer used.
476 if (!V && LHS->hasOneUse() && RHS->hasOneUse())
477 V = Builder->CreateBinOp(TopLevelOpcode, B, D, RHS->getName());
479 SimplifiedInst = Builder->CreateBinOp(InnerOpcode, A, V);
483 // Does "(X op Y) op' Z" always equal "(X op' Z) op (Y op' Z)"?
484 if (!SimplifiedInst && RightDistributesOverLeft(TopLevelOpcode, InnerOpcode))
485 // Does the instruction have the form "(A op' B) op (C op' B)" or, in the
486 // commutative case, "(A op' B) op (B op' D)"?
487 if (B == D || (InnerCommutative && B == C)) {
490 // Consider forming "(A op C) op' B".
491 // If "A op C" simplifies then it can be formed with no cost.
492 V = SimplifyBinOp(TopLevelOpcode, A, C, DL);
494 // If "A op C" doesn't simplify then only go on if both of the existing
495 // operations "A op' B" and "C op' D" will be zapped as no longer used.
496 if (!V && LHS->hasOneUse() && RHS->hasOneUse())
497 V = Builder->CreateBinOp(TopLevelOpcode, A, C, LHS->getName());
499 SimplifiedInst = Builder->CreateBinOp(InnerOpcode, V, B);
503 if (SimplifiedInst) {
505 SimplifiedInst->takeName(&I);
507 // Check if we can add NSW flag to SimplifiedInst. If so, set NSW flag.
508 // TODO: Check for NUW.
509 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(SimplifiedInst)) {
510 if (isa<OverflowingBinaryOperator>(SimplifiedInst)) {
512 if (isa<OverflowingBinaryOperator>(&I))
513 HasNSW = I.hasNoSignedWrap();
515 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS))
516 if (isa<OverflowingBinaryOperator>(Op0))
517 HasNSW &= Op0->hasNoSignedWrap();
519 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS))
520 if (isa<OverflowingBinaryOperator>(Op1))
521 HasNSW &= Op1->hasNoSignedWrap();
523 // We can propagate 'nsw' if we know that
524 // %Y = mul nsw i16 %X, C
525 // %Z = add nsw i16 %Y, %X
527 // %Z = mul nsw i16 %X, C+1
529 // iff C+1 isn't INT_MIN
531 if (TopLevelOpcode == Instruction::Add &&
532 InnerOpcode == Instruction::Mul)
533 if (match(V, m_APInt(CInt)) && !CInt->isMinSignedValue())
534 BO->setHasNoSignedWrap(HasNSW);
538 return SimplifiedInst;
541 /// SimplifyUsingDistributiveLaws - This tries to simplify binary operations
542 /// which some other binary operation distributes over either by factorizing
543 /// out common terms (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this
544 /// results in simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is
545 /// a win). Returns the simplified value, or null if it didn't simplify.
546 Value *InstCombiner::SimplifyUsingDistributiveLaws(BinaryOperator &I) {
547 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
548 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
549 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
552 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
553 auto TopLevelOpcode = I.getOpcode();
554 auto LHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op0, A, B);
555 auto RHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op1, C, D);
557 // The instruction has the form "(A op' B) op (C op' D)". Try to factorize
559 if (LHSOpcode == RHSOpcode) {
560 if (Value *V = tryFactorization(Builder, DL, I, LHSOpcode, A, B, C, D))
564 // The instruction has the form "(A op' B) op (C)". Try to factorize common
566 if (Value *V = tryFactorization(Builder, DL, I, LHSOpcode, A, B, RHS,
567 getIdentityValue(LHSOpcode, RHS)))
570 // The instruction has the form "(B) op (C op' D)". Try to factorize common
572 if (Value *V = tryFactorization(Builder, DL, I, RHSOpcode, LHS,
573 getIdentityValue(RHSOpcode, LHS), C, D))
577 if (Op0 && RightDistributesOverLeft(Op0->getOpcode(), TopLevelOpcode)) {
578 // The instruction has the form "(A op' B) op C". See if expanding it out
579 // to "(A op C) op' (B op C)" results in simplifications.
580 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
581 Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op'
583 // Do "A op C" and "B op C" both simplify?
584 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, C, DL))
585 if (Value *R = SimplifyBinOp(TopLevelOpcode, B, C, DL)) {
586 // They do! Return "L op' R".
588 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
589 if ((L == A && R == B) ||
590 (Instruction::isCommutative(InnerOpcode) && L == B && R == A))
592 // Otherwise return "L op' R" if it simplifies.
593 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, DL))
595 // Otherwise, create a new instruction.
596 C = Builder->CreateBinOp(InnerOpcode, L, R);
602 if (Op1 && LeftDistributesOverRight(TopLevelOpcode, Op1->getOpcode())) {
603 // The instruction has the form "A op (B op' C)". See if expanding it out
604 // to "(A op B) op' (A op C)" results in simplifications.
605 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
606 Instruction::BinaryOps InnerOpcode = Op1->getOpcode(); // op'
608 // Do "A op B" and "A op C" both simplify?
609 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, B, DL))
610 if (Value *R = SimplifyBinOp(TopLevelOpcode, A, C, DL)) {
611 // They do! Return "L op' R".
613 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
614 if ((L == B && R == C) ||
615 (Instruction::isCommutative(InnerOpcode) && L == C && R == B))
617 // Otherwise return "L op' R" if it simplifies.
618 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, DL))
620 // Otherwise, create a new instruction.
621 A = Builder->CreateBinOp(InnerOpcode, L, R);
627 // (op (select (a, c, b)), (select (a, d, b))) -> (select (a, (op c, d), 0))
628 // (op (select (a, b, c)), (select (a, b, d))) -> (select (a, 0, (op c, d)))
629 if (auto *SI0 = dyn_cast<SelectInst>(LHS)) {
630 if (auto *SI1 = dyn_cast<SelectInst>(RHS)) {
631 if (SI0->getCondition() == SI1->getCondition()) {
633 if (Value *V = SimplifyBinOp(TopLevelOpcode, SI0->getFalseValue(),
634 SI1->getFalseValue(), DL, TLI, DT, AC))
635 SI = Builder->CreateSelect(SI0->getCondition(),
636 Builder->CreateBinOp(TopLevelOpcode,
638 SI1->getTrueValue()),
640 if (Value *V = SimplifyBinOp(TopLevelOpcode, SI0->getTrueValue(),
641 SI1->getTrueValue(), DL, TLI, DT, AC))
642 SI = Builder->CreateSelect(
643 SI0->getCondition(), V,
644 Builder->CreateBinOp(TopLevelOpcode, SI0->getFalseValue(),
645 SI1->getFalseValue()));
657 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
658 // if the LHS is a constant zero (which is the 'negate' form).
660 Value *InstCombiner::dyn_castNegVal(Value *V) const {
661 if (BinaryOperator::isNeg(V))
662 return BinaryOperator::getNegArgument(V);
664 // Constants can be considered to be negated values if they can be folded.
665 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
666 return ConstantExpr::getNeg(C);
668 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
669 if (C->getType()->getElementType()->isIntegerTy())
670 return ConstantExpr::getNeg(C);
675 // dyn_castFNegVal - Given a 'fsub' instruction, return the RHS of the
676 // instruction if the LHS is a constant negative zero (which is the 'negate'
679 Value *InstCombiner::dyn_castFNegVal(Value *V, bool IgnoreZeroSign) const {
680 if (BinaryOperator::isFNeg(V, IgnoreZeroSign))
681 return BinaryOperator::getFNegArgument(V);
683 // Constants can be considered to be negated values if they can be folded.
684 if (ConstantFP *C = dyn_cast<ConstantFP>(V))
685 return ConstantExpr::getFNeg(C);
687 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
688 if (C->getType()->getElementType()->isFloatingPointTy())
689 return ConstantExpr::getFNeg(C);
694 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
696 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
697 return IC->Builder->CreateCast(CI->getOpcode(), SO, I.getType());
700 // Figure out if the constant is the left or the right argument.
701 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
702 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
704 if (Constant *SOC = dyn_cast<Constant>(SO)) {
706 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
707 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
710 Value *Op0 = SO, *Op1 = ConstOperand;
714 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I)) {
715 Value *RI = IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1,
716 SO->getName()+".op");
717 Instruction *FPInst = dyn_cast<Instruction>(RI);
718 if (FPInst && isa<FPMathOperator>(FPInst))
719 FPInst->copyFastMathFlags(BO);
722 if (ICmpInst *CI = dyn_cast<ICmpInst>(&I))
723 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
724 SO->getName()+".cmp");
725 if (FCmpInst *CI = dyn_cast<FCmpInst>(&I))
726 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
727 SO->getName()+".cmp");
728 llvm_unreachable("Unknown binary instruction type!");
731 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
732 // constant as the other operand, try to fold the binary operator into the
733 // select arguments. This also works for Cast instructions, which obviously do
734 // not have a second operand.
735 Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) {
736 // Don't modify shared select instructions
737 if (!SI->hasOneUse()) return nullptr;
738 Value *TV = SI->getOperand(1);
739 Value *FV = SI->getOperand(2);
741 if (isa<Constant>(TV) || isa<Constant>(FV)) {
742 // Bool selects with constant operands can be folded to logical ops.
743 if (SI->getType()->isIntegerTy(1)) return nullptr;
745 // If it's a bitcast involving vectors, make sure it has the same number of
746 // elements on both sides.
747 if (BitCastInst *BC = dyn_cast<BitCastInst>(&Op)) {
748 VectorType *DestTy = dyn_cast<VectorType>(BC->getDestTy());
749 VectorType *SrcTy = dyn_cast<VectorType>(BC->getSrcTy());
751 // Verify that either both or neither are vectors.
752 if ((SrcTy == nullptr) != (DestTy == nullptr)) return nullptr;
753 // If vectors, verify that they have the same number of elements.
754 if (SrcTy && SrcTy->getNumElements() != DestTy->getNumElements())
758 // Test if a CmpInst instruction is used exclusively by a select as
759 // part of a minimum or maximum operation. If so, refrain from doing
760 // any other folding. This helps out other analyses which understand
761 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
762 // and CodeGen. And in this case, at least one of the comparison
763 // operands has at least one user besides the compare (the select),
764 // which would often largely negate the benefit of folding anyway.
765 if (auto *CI = dyn_cast<CmpInst>(SI->getCondition())) {
766 if (CI->hasOneUse()) {
767 Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1);
768 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
769 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
774 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, this);
775 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, this);
777 return SelectInst::Create(SI->getCondition(),
778 SelectTrueVal, SelectFalseVal);
783 /// FoldOpIntoPhi - Given a binary operator, cast instruction, or select which
784 /// has a PHI node as operand #0, see if we can fold the instruction into the
785 /// PHI (which is only possible if all operands to the PHI are constants).
787 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
788 PHINode *PN = cast<PHINode>(I.getOperand(0));
789 unsigned NumPHIValues = PN->getNumIncomingValues();
790 if (NumPHIValues == 0)
793 // We normally only transform phis with a single use. However, if a PHI has
794 // multiple uses and they are all the same operation, we can fold *all* of the
795 // uses into the PHI.
796 if (!PN->hasOneUse()) {
797 // Walk the use list for the instruction, comparing them to I.
798 for (User *U : PN->users()) {
799 Instruction *UI = cast<Instruction>(U);
800 if (UI != &I && !I.isIdenticalTo(UI))
803 // Otherwise, we can replace *all* users with the new PHI we form.
806 // Check to see if all of the operands of the PHI are simple constants
807 // (constantint/constantfp/undef). If there is one non-constant value,
808 // remember the BB it is in. If there is more than one or if *it* is a PHI,
809 // bail out. We don't do arbitrary constant expressions here because moving
810 // their computation can be expensive without a cost model.
811 BasicBlock *NonConstBB = nullptr;
812 for (unsigned i = 0; i != NumPHIValues; ++i) {
813 Value *InVal = PN->getIncomingValue(i);
814 if (isa<Constant>(InVal) && !isa<ConstantExpr>(InVal))
817 if (isa<PHINode>(InVal)) return nullptr; // Itself a phi.
818 if (NonConstBB) return nullptr; // More than one non-const value.
820 NonConstBB = PN->getIncomingBlock(i);
822 // If the InVal is an invoke at the end of the pred block, then we can't
823 // insert a computation after it without breaking the edge.
824 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal))
825 if (II->getParent() == NonConstBB)
828 // If the incoming non-constant value is in I's block, we will remove one
829 // instruction, but insert another equivalent one, leading to infinite
831 if (isPotentiallyReachable(I.getParent(), NonConstBB, DT, LI))
835 // If there is exactly one non-constant value, we can insert a copy of the
836 // operation in that block. However, if this is a critical edge, we would be
837 // inserting the computation on some other paths (e.g. inside a loop). Only
838 // do this if the pred block is unconditionally branching into the phi block.
839 if (NonConstBB != nullptr) {
840 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
841 if (!BI || !BI->isUnconditional()) return nullptr;
844 // Okay, we can do the transformation: create the new PHI node.
845 PHINode *NewPN = PHINode::Create(I.getType(), PN->getNumIncomingValues());
846 InsertNewInstBefore(NewPN, *PN);
849 // If we are going to have to insert a new computation, do so right before the
850 // predecessor's terminator.
852 Builder->SetInsertPoint(NonConstBB->getTerminator());
854 // Next, add all of the operands to the PHI.
855 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) {
856 // We only currently try to fold the condition of a select when it is a phi,
857 // not the true/false values.
858 Value *TrueV = SI->getTrueValue();
859 Value *FalseV = SI->getFalseValue();
860 BasicBlock *PhiTransBB = PN->getParent();
861 for (unsigned i = 0; i != NumPHIValues; ++i) {
862 BasicBlock *ThisBB = PN->getIncomingBlock(i);
863 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB);
864 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB);
865 Value *InV = nullptr;
866 // Beware of ConstantExpr: it may eventually evaluate to getNullValue,
867 // even if currently isNullValue gives false.
868 Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i));
869 if (InC && !isa<ConstantExpr>(InC))
870 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred;
872 InV = Builder->CreateSelect(PN->getIncomingValue(i),
873 TrueVInPred, FalseVInPred, "phitmp");
874 NewPN->addIncoming(InV, ThisBB);
876 } else if (CmpInst *CI = dyn_cast<CmpInst>(&I)) {
877 Constant *C = cast<Constant>(I.getOperand(1));
878 for (unsigned i = 0; i != NumPHIValues; ++i) {
879 Value *InV = nullptr;
880 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
881 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
882 else if (isa<ICmpInst>(CI))
883 InV = Builder->CreateICmp(CI->getPredicate(), PN->getIncomingValue(i),
886 InV = Builder->CreateFCmp(CI->getPredicate(), PN->getIncomingValue(i),
888 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
890 } else if (I.getNumOperands() == 2) {
891 Constant *C = cast<Constant>(I.getOperand(1));
892 for (unsigned i = 0; i != NumPHIValues; ++i) {
893 Value *InV = nullptr;
894 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
895 InV = ConstantExpr::get(I.getOpcode(), InC, C);
897 InV = Builder->CreateBinOp(cast<BinaryOperator>(I).getOpcode(),
898 PN->getIncomingValue(i), C, "phitmp");
899 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
902 CastInst *CI = cast<CastInst>(&I);
903 Type *RetTy = CI->getType();
904 for (unsigned i = 0; i != NumPHIValues; ++i) {
906 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
907 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
909 InV = Builder->CreateCast(CI->getOpcode(),
910 PN->getIncomingValue(i), I.getType(), "phitmp");
911 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
915 for (auto UI = PN->user_begin(), E = PN->user_end(); UI != E;) {
916 Instruction *User = cast<Instruction>(*UI++);
917 if (User == &I) continue;
918 ReplaceInstUsesWith(*User, NewPN);
919 EraseInstFromFunction(*User);
921 return ReplaceInstUsesWith(I, NewPN);
924 /// FindElementAtOffset - Given a pointer type and a constant offset, determine
925 /// whether or not there is a sequence of GEP indices into the pointed type that
926 /// will land us at the specified offset. If so, fill them into NewIndices and
927 /// return the resultant element type, otherwise return null.
928 Type *InstCombiner::FindElementAtOffset(PointerType *PtrTy, int64_t Offset,
929 SmallVectorImpl<Value *> &NewIndices) {
930 Type *Ty = PtrTy->getElementType();
934 // Start with the index over the outer type. Note that the type size
935 // might be zero (even if the offset isn't zero) if the indexed type
936 // is something like [0 x {int, int}]
937 Type *IntPtrTy = DL.getIntPtrType(PtrTy);
938 int64_t FirstIdx = 0;
939 if (int64_t TySize = DL.getTypeAllocSize(Ty)) {
940 FirstIdx = Offset/TySize;
941 Offset -= FirstIdx*TySize;
943 // Handle hosts where % returns negative instead of values [0..TySize).
949 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
952 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
954 // Index into the types. If we fail, set OrigBase to null.
956 // Indexing into tail padding between struct/array elements.
957 if (uint64_t(Offset * 8) >= DL.getTypeSizeInBits(Ty))
960 if (StructType *STy = dyn_cast<StructType>(Ty)) {
961 const StructLayout *SL = DL.getStructLayout(STy);
962 assert(Offset < (int64_t)SL->getSizeInBytes() &&
963 "Offset must stay within the indexed type");
965 unsigned Elt = SL->getElementContainingOffset(Offset);
966 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
969 Offset -= SL->getElementOffset(Elt);
970 Ty = STy->getElementType(Elt);
971 } else if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
972 uint64_t EltSize = DL.getTypeAllocSize(AT->getElementType());
973 assert(EltSize && "Cannot index into a zero-sized array");
974 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
976 Ty = AT->getElementType();
978 // Otherwise, we can't index into the middle of this atomic type, bail.
986 static bool shouldMergeGEPs(GEPOperator &GEP, GEPOperator &Src) {
987 // If this GEP has only 0 indices, it is the same pointer as
988 // Src. If Src is not a trivial GEP too, don't combine
990 if (GEP.hasAllZeroIndices() && !Src.hasAllZeroIndices() &&
996 /// Descale - Return a value X such that Val = X * Scale, or null if none. If
997 /// the multiplication is known not to overflow then NoSignedWrap is set.
998 Value *InstCombiner::Descale(Value *Val, APInt Scale, bool &NoSignedWrap) {
999 assert(isa<IntegerType>(Val->getType()) && "Can only descale integers!");
1000 assert(cast<IntegerType>(Val->getType())->getBitWidth() ==
1001 Scale.getBitWidth() && "Scale not compatible with value!");
1003 // If Val is zero or Scale is one then Val = Val * Scale.
1004 if (match(Val, m_Zero()) || Scale == 1) {
1005 NoSignedWrap = true;
1009 // If Scale is zero then it does not divide Val.
1010 if (Scale.isMinValue())
1013 // Look through chains of multiplications, searching for a constant that is
1014 // divisible by Scale. For example, descaling X*(Y*(Z*4)) by a factor of 4
1015 // will find the constant factor 4 and produce X*(Y*Z). Descaling X*(Y*8) by
1016 // a factor of 4 will produce X*(Y*2). The principle of operation is to bore
1019 // Val = M1 * X || Analysis starts here and works down
1020 // M1 = M2 * Y || Doesn't descend into terms with more
1021 // M2 = Z * 4 \/ than one use
1023 // Then to modify a term at the bottom:
1026 // M1 = Z * Y || Replaced M2 with Z
1028 // Then to work back up correcting nsw flags.
1030 // Op - the term we are currently analyzing. Starts at Val then drills down.
1031 // Replaced with its descaled value before exiting from the drill down loop.
1034 // Parent - initially null, but after drilling down notes where Op came from.
1035 // In the example above, Parent is (Val, 0) when Op is M1, because M1 is the
1036 // 0'th operand of Val.
1037 std::pair<Instruction*, unsigned> Parent;
1039 // RequireNoSignedWrap - Set if the transform requires a descaling at deeper
1040 // levels that doesn't overflow.
1041 bool RequireNoSignedWrap = false;
1043 // logScale - log base 2 of the scale. Negative if not a power of 2.
1044 int32_t logScale = Scale.exactLogBase2();
1046 for (;; Op = Parent.first->getOperand(Parent.second)) { // Drill down
1048 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
1049 // If Op is a constant divisible by Scale then descale to the quotient.
1050 APInt Quotient(Scale), Remainder(Scale); // Init ensures right bitwidth.
1051 APInt::sdivrem(CI->getValue(), Scale, Quotient, Remainder);
1052 if (!Remainder.isMinValue())
1053 // Not divisible by Scale.
1055 // Replace with the quotient in the parent.
1056 Op = ConstantInt::get(CI->getType(), Quotient);
1057 NoSignedWrap = true;
1061 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op)) {
1063 if (BO->getOpcode() == Instruction::Mul) {
1065 NoSignedWrap = BO->hasNoSignedWrap();
1066 if (RequireNoSignedWrap && !NoSignedWrap)
1069 // There are three cases for multiplication: multiplication by exactly
1070 // the scale, multiplication by a constant different to the scale, and
1071 // multiplication by something else.
1072 Value *LHS = BO->getOperand(0);
1073 Value *RHS = BO->getOperand(1);
1075 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
1076 // Multiplication by a constant.
1077 if (CI->getValue() == Scale) {
1078 // Multiplication by exactly the scale, replace the multiplication
1079 // by its left-hand side in the parent.
1084 // Otherwise drill down into the constant.
1085 if (!Op->hasOneUse())
1088 Parent = std::make_pair(BO, 1);
1092 // Multiplication by something else. Drill down into the left-hand side
1093 // since that's where the reassociate pass puts the good stuff.
1094 if (!Op->hasOneUse())
1097 Parent = std::make_pair(BO, 0);
1101 if (logScale > 0 && BO->getOpcode() == Instruction::Shl &&
1102 isa<ConstantInt>(BO->getOperand(1))) {
1103 // Multiplication by a power of 2.
1104 NoSignedWrap = BO->hasNoSignedWrap();
1105 if (RequireNoSignedWrap && !NoSignedWrap)
1108 Value *LHS = BO->getOperand(0);
1109 int32_t Amt = cast<ConstantInt>(BO->getOperand(1))->
1110 getLimitedValue(Scale.getBitWidth());
1113 if (Amt == logScale) {
1114 // Multiplication by exactly the scale, replace the multiplication
1115 // by its left-hand side in the parent.
1119 if (Amt < logScale || !Op->hasOneUse())
1122 // Multiplication by more than the scale. Reduce the multiplying amount
1123 // by the scale in the parent.
1124 Parent = std::make_pair(BO, 1);
1125 Op = ConstantInt::get(BO->getType(), Amt - logScale);
1130 if (!Op->hasOneUse())
1133 if (CastInst *Cast = dyn_cast<CastInst>(Op)) {
1134 if (Cast->getOpcode() == Instruction::SExt) {
1135 // Op is sign-extended from a smaller type, descale in the smaller type.
1136 unsigned SmallSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1137 APInt SmallScale = Scale.trunc(SmallSize);
1138 // Suppose Op = sext X, and we descale X as Y * SmallScale. We want to
1139 // descale Op as (sext Y) * Scale. In order to have
1140 // sext (Y * SmallScale) = (sext Y) * Scale
1141 // some conditions need to hold however: SmallScale must sign-extend to
1142 // Scale and the multiplication Y * SmallScale should not overflow.
1143 if (SmallScale.sext(Scale.getBitWidth()) != Scale)
1144 // SmallScale does not sign-extend to Scale.
1146 assert(SmallScale.exactLogBase2() == logScale);
1147 // Require that Y * SmallScale must not overflow.
1148 RequireNoSignedWrap = true;
1150 // Drill down through the cast.
1151 Parent = std::make_pair(Cast, 0);
1156 if (Cast->getOpcode() == Instruction::Trunc) {
1157 // Op is truncated from a larger type, descale in the larger type.
1158 // Suppose Op = trunc X, and we descale X as Y * sext Scale. Then
1159 // trunc (Y * sext Scale) = (trunc Y) * Scale
1160 // always holds. However (trunc Y) * Scale may overflow even if
1161 // trunc (Y * sext Scale) does not, so nsw flags need to be cleared
1162 // from this point up in the expression (see later).
1163 if (RequireNoSignedWrap)
1166 // Drill down through the cast.
1167 unsigned LargeSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1168 Parent = std::make_pair(Cast, 0);
1169 Scale = Scale.sext(LargeSize);
1170 if (logScale + 1 == (int32_t)Cast->getType()->getPrimitiveSizeInBits())
1172 assert(Scale.exactLogBase2() == logScale);
1177 // Unsupported expression, bail out.
1181 // If Op is zero then Val = Op * Scale.
1182 if (match(Op, m_Zero())) {
1183 NoSignedWrap = true;
1187 // We know that we can successfully descale, so from here on we can safely
1188 // modify the IR. Op holds the descaled version of the deepest term in the
1189 // expression. NoSignedWrap is 'true' if multiplying Op by Scale is known
1193 // The expression only had one term.
1196 // Rewrite the parent using the descaled version of its operand.
1197 assert(Parent.first->hasOneUse() && "Drilled down when more than one use!");
1198 assert(Op != Parent.first->getOperand(Parent.second) &&
1199 "Descaling was a no-op?");
1200 Parent.first->setOperand(Parent.second, Op);
1201 Worklist.Add(Parent.first);
1203 // Now work back up the expression correcting nsw flags. The logic is based
1204 // on the following observation: if X * Y is known not to overflow as a signed
1205 // multiplication, and Y is replaced by a value Z with smaller absolute value,
1206 // then X * Z will not overflow as a signed multiplication either. As we work
1207 // our way up, having NoSignedWrap 'true' means that the descaled value at the
1208 // current level has strictly smaller absolute value than the original.
1209 Instruction *Ancestor = Parent.first;
1211 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Ancestor)) {
1212 // If the multiplication wasn't nsw then we can't say anything about the
1213 // value of the descaled multiplication, and we have to clear nsw flags
1214 // from this point on up.
1215 bool OpNoSignedWrap = BO->hasNoSignedWrap();
1216 NoSignedWrap &= OpNoSignedWrap;
1217 if (NoSignedWrap != OpNoSignedWrap) {
1218 BO->setHasNoSignedWrap(NoSignedWrap);
1219 Worklist.Add(Ancestor);
1221 } else if (Ancestor->getOpcode() == Instruction::Trunc) {
1222 // The fact that the descaled input to the trunc has smaller absolute
1223 // value than the original input doesn't tell us anything useful about
1224 // the absolute values of the truncations.
1225 NoSignedWrap = false;
1227 assert((Ancestor->getOpcode() != Instruction::SExt || NoSignedWrap) &&
1228 "Failed to keep proper track of nsw flags while drilling down?");
1230 if (Ancestor == Val)
1231 // Got to the top, all done!
1234 // Move up one level in the expression.
1235 assert(Ancestor->hasOneUse() && "Drilled down when more than one use!");
1236 Ancestor = Ancestor->user_back();
1240 /// \brief Creates node of binary operation with the same attributes as the
1241 /// specified one but with other operands.
1242 static Value *CreateBinOpAsGiven(BinaryOperator &Inst, Value *LHS, Value *RHS,
1243 InstCombiner::BuilderTy *B) {
1244 Value *BORes = B->CreateBinOp(Inst.getOpcode(), LHS, RHS);
1245 if (BinaryOperator *NewBO = dyn_cast<BinaryOperator>(BORes)) {
1246 if (isa<OverflowingBinaryOperator>(NewBO)) {
1247 NewBO->setHasNoSignedWrap(Inst.hasNoSignedWrap());
1248 NewBO->setHasNoUnsignedWrap(Inst.hasNoUnsignedWrap());
1250 if (isa<PossiblyExactOperator>(NewBO))
1251 NewBO->setIsExact(Inst.isExact());
1256 /// \brief Makes transformation of binary operation specific for vector types.
1257 /// \param Inst Binary operator to transform.
1258 /// \return Pointer to node that must replace the original binary operator, or
1259 /// null pointer if no transformation was made.
1260 Value *InstCombiner::SimplifyVectorOp(BinaryOperator &Inst) {
1261 if (!Inst.getType()->isVectorTy()) return nullptr;
1263 // It may not be safe to reorder shuffles and things like div, urem, etc.
1264 // because we may trap when executing those ops on unknown vector elements.
1266 if (!isSafeToSpeculativelyExecute(&Inst))
1269 unsigned VWidth = cast<VectorType>(Inst.getType())->getNumElements();
1270 Value *LHS = Inst.getOperand(0), *RHS = Inst.getOperand(1);
1271 assert(cast<VectorType>(LHS->getType())->getNumElements() == VWidth);
1272 assert(cast<VectorType>(RHS->getType())->getNumElements() == VWidth);
1274 // If both arguments of binary operation are shuffles, which use the same
1275 // mask and shuffle within a single vector, it is worthwhile to move the
1276 // shuffle after binary operation:
1277 // Op(shuffle(v1, m), shuffle(v2, m)) -> shuffle(Op(v1, v2), m)
1278 if (isa<ShuffleVectorInst>(LHS) && isa<ShuffleVectorInst>(RHS)) {
1279 ShuffleVectorInst *LShuf = cast<ShuffleVectorInst>(LHS);
1280 ShuffleVectorInst *RShuf = cast<ShuffleVectorInst>(RHS);
1281 if (isa<UndefValue>(LShuf->getOperand(1)) &&
1282 isa<UndefValue>(RShuf->getOperand(1)) &&
1283 LShuf->getOperand(0)->getType() == RShuf->getOperand(0)->getType() &&
1284 LShuf->getMask() == RShuf->getMask()) {
1285 Value *NewBO = CreateBinOpAsGiven(Inst, LShuf->getOperand(0),
1286 RShuf->getOperand(0), Builder);
1287 Value *Res = Builder->CreateShuffleVector(NewBO,
1288 UndefValue::get(NewBO->getType()), LShuf->getMask());
1293 // If one argument is a shuffle within one vector, the other is a constant,
1294 // try moving the shuffle after the binary operation.
1295 ShuffleVectorInst *Shuffle = nullptr;
1296 Constant *C1 = nullptr;
1297 if (isa<ShuffleVectorInst>(LHS)) Shuffle = cast<ShuffleVectorInst>(LHS);
1298 if (isa<ShuffleVectorInst>(RHS)) Shuffle = cast<ShuffleVectorInst>(RHS);
1299 if (isa<Constant>(LHS)) C1 = cast<Constant>(LHS);
1300 if (isa<Constant>(RHS)) C1 = cast<Constant>(RHS);
1301 if (Shuffle && C1 &&
1302 (isa<ConstantVector>(C1) || isa<ConstantDataVector>(C1)) &&
1303 isa<UndefValue>(Shuffle->getOperand(1)) &&
1304 Shuffle->getType() == Shuffle->getOperand(0)->getType()) {
1305 SmallVector<int, 16> ShMask = Shuffle->getShuffleMask();
1306 // Find constant C2 that has property:
1307 // shuffle(C2, ShMask) = C1
1308 // If such constant does not exist (example: ShMask=<0,0> and C1=<1,2>)
1309 // reorder is not possible.
1310 SmallVector<Constant*, 16> C2M(VWidth,
1311 UndefValue::get(C1->getType()->getScalarType()));
1312 bool MayChange = true;
1313 for (unsigned I = 0; I < VWidth; ++I) {
1314 if (ShMask[I] >= 0) {
1315 assert(ShMask[I] < (int)VWidth);
1316 if (!isa<UndefValue>(C2M[ShMask[I]])) {
1320 C2M[ShMask[I]] = C1->getAggregateElement(I);
1324 Constant *C2 = ConstantVector::get(C2M);
1325 Value *NewLHS, *NewRHS;
1326 if (isa<Constant>(LHS)) {
1328 NewRHS = Shuffle->getOperand(0);
1330 NewLHS = Shuffle->getOperand(0);
1333 Value *NewBO = CreateBinOpAsGiven(Inst, NewLHS, NewRHS, Builder);
1334 Value *Res = Builder->CreateShuffleVector(NewBO,
1335 UndefValue::get(Inst.getType()), Shuffle->getMask());
1343 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
1344 SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end());
1346 if (Value *V = SimplifyGEPInst(Ops, DL, TLI, DT, AC))
1347 return ReplaceInstUsesWith(GEP, V);
1349 Value *PtrOp = GEP.getOperand(0);
1351 // Eliminate unneeded casts for indices, and replace indices which displace
1352 // by multiples of a zero size type with zero.
1353 bool MadeChange = false;
1354 Type *IntPtrTy = DL.getIntPtrType(GEP.getPointerOperandType());
1356 gep_type_iterator GTI = gep_type_begin(GEP);
1357 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end(); I != E;
1359 // Skip indices into struct types.
1360 SequentialType *SeqTy = dyn_cast<SequentialType>(*GTI);
1364 // If the element type has zero size then any index over it is equivalent
1365 // to an index of zero, so replace it with zero if it is not zero already.
1366 if (SeqTy->getElementType()->isSized() &&
1367 DL.getTypeAllocSize(SeqTy->getElementType()) == 0)
1368 if (!isa<Constant>(*I) || !cast<Constant>(*I)->isNullValue()) {
1369 *I = Constant::getNullValue(IntPtrTy);
1373 Type *IndexTy = (*I)->getType();
1374 if (IndexTy != IntPtrTy) {
1375 // If we are using a wider index than needed for this platform, shrink
1376 // it to what we need. If narrower, sign-extend it to what we need.
1377 // This explicit cast can make subsequent optimizations more obvious.
1378 *I = Builder->CreateIntCast(*I, IntPtrTy, true);
1385 // Check to see if the inputs to the PHI node are getelementptr instructions.
1386 if (PHINode *PN = dyn_cast<PHINode>(PtrOp)) {
1387 GetElementPtrInst *Op1 = dyn_cast<GetElementPtrInst>(PN->getOperand(0));
1391 // Don't fold a GEP into itself through a PHI node. This can only happen
1392 // through the back-edge of a loop. Folding a GEP into itself means that
1393 // the value of the previous iteration needs to be stored in the meantime,
1394 // thus requiring an additional register variable to be live, but not
1395 // actually achieving anything (the GEP still needs to be executed once per
1402 for (auto I = PN->op_begin()+1, E = PN->op_end(); I !=E; ++I) {
1403 GetElementPtrInst *Op2 = dyn_cast<GetElementPtrInst>(*I);
1404 if (!Op2 || Op1->getNumOperands() != Op2->getNumOperands())
1407 // As for Op1 above, don't try to fold a GEP into itself.
1411 // Keep track of the type as we walk the GEP.
1412 Type *CurTy = Op1->getOperand(0)->getType()->getScalarType();
1414 for (unsigned J = 0, F = Op1->getNumOperands(); J != F; ++J) {
1415 if (Op1->getOperand(J)->getType() != Op2->getOperand(J)->getType())
1418 if (Op1->getOperand(J) != Op2->getOperand(J)) {
1420 // We have not seen any differences yet in the GEPs feeding the
1421 // PHI yet, so we record this one if it is allowed to be a
1424 // The first two arguments can vary for any GEP, the rest have to be
1425 // static for struct slots
1426 if (J > 1 && CurTy->isStructTy())
1431 // The GEP is different by more than one input. While this could be
1432 // extended to support GEPs that vary by more than one variable it
1433 // doesn't make sense since it greatly increases the complexity and
1434 // would result in an R+R+R addressing mode which no backend
1435 // directly supports and would need to be broken into several
1436 // simpler instructions anyway.
1441 // Sink down a layer of the type for the next iteration.
1443 if (CompositeType *CT = dyn_cast<CompositeType>(CurTy)) {
1444 CurTy = CT->getTypeAtIndex(Op1->getOperand(J));
1452 GetElementPtrInst *NewGEP = cast<GetElementPtrInst>(Op1->clone());
1455 // All the GEPs feeding the PHI are identical. Clone one down into our
1456 // BB so that it can be merged with the current GEP.
1457 GEP.getParent()->getInstList().insert(
1458 GEP.getParent()->getFirstInsertionPt(), NewGEP);
1460 // All the GEPs feeding the PHI differ at a single offset. Clone a GEP
1461 // into the current block so it can be merged, and create a new PHI to
1463 Instruction *InsertPt = Builder->GetInsertPoint();
1464 Builder->SetInsertPoint(PN);
1465 PHINode *NewPN = Builder->CreatePHI(Op1->getOperand(DI)->getType(),
1466 PN->getNumOperands());
1467 Builder->SetInsertPoint(InsertPt);
1469 for (auto &I : PN->operands())
1470 NewPN->addIncoming(cast<GEPOperator>(I)->getOperand(DI),
1471 PN->getIncomingBlock(I));
1473 NewGEP->setOperand(DI, NewPN);
1474 GEP.getParent()->getInstList().insert(
1475 GEP.getParent()->getFirstInsertionPt(), NewGEP);
1476 NewGEP->setOperand(DI, NewPN);
1479 GEP.setOperand(0, NewGEP);
1483 // Combine Indices - If the source pointer to this getelementptr instruction
1484 // is a getelementptr instruction, combine the indices of the two
1485 // getelementptr instructions into a single instruction.
1487 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) {
1488 if (!shouldMergeGEPs(*cast<GEPOperator>(&GEP), *Src))
1491 // Note that if our source is a gep chain itself then we wait for that
1492 // chain to be resolved before we perform this transformation. This
1493 // avoids us creating a TON of code in some cases.
1494 if (GEPOperator *SrcGEP =
1495 dyn_cast<GEPOperator>(Src->getOperand(0)))
1496 if (SrcGEP->getNumOperands() == 2 && shouldMergeGEPs(*Src, *SrcGEP))
1497 return nullptr; // Wait until our source is folded to completion.
1499 SmallVector<Value*, 8> Indices;
1501 // Find out whether the last index in the source GEP is a sequential idx.
1502 bool EndsWithSequential = false;
1503 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src);
1505 EndsWithSequential = !(*I)->isStructTy();
1507 // Can we combine the two pointer arithmetics offsets?
1508 if (EndsWithSequential) {
1509 // Replace: gep (gep %P, long B), long A, ...
1510 // With: T = long A+B; gep %P, T, ...
1513 Value *SO1 = Src->getOperand(Src->getNumOperands()-1);
1514 Value *GO1 = GEP.getOperand(1);
1515 if (SO1 == Constant::getNullValue(SO1->getType())) {
1517 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
1520 // If they aren't the same type, then the input hasn't been processed
1521 // by the loop above yet (which canonicalizes sequential index types to
1522 // intptr_t). Just avoid transforming this until the input has been
1524 if (SO1->getType() != GO1->getType())
1526 // Only do the combine when GO1 and SO1 are both constants. Only in
1527 // this case, we are sure the cost after the merge is never more than
1528 // that before the merge.
1529 if (!isa<Constant>(GO1) || !isa<Constant>(SO1))
1531 Sum = Builder->CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
1534 // Update the GEP in place if possible.
1535 if (Src->getNumOperands() == 2) {
1536 GEP.setOperand(0, Src->getOperand(0));
1537 GEP.setOperand(1, Sum);
1540 Indices.append(Src->op_begin()+1, Src->op_end()-1);
1541 Indices.push_back(Sum);
1542 Indices.append(GEP.op_begin()+2, GEP.op_end());
1543 } else if (isa<Constant>(*GEP.idx_begin()) &&
1544 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
1545 Src->getNumOperands() != 1) {
1546 // Otherwise we can do the fold if the first index of the GEP is a zero
1547 Indices.append(Src->op_begin()+1, Src->op_end());
1548 Indices.append(GEP.idx_begin()+1, GEP.idx_end());
1551 if (!Indices.empty())
1552 return GEP.isInBounds() && Src->isInBounds()
1553 ? GetElementPtrInst::CreateInBounds(
1554 Src->getSourceElementType(), Src->getOperand(0), Indices,
1556 : GetElementPtrInst::Create(Src->getSourceElementType(),
1557 Src->getOperand(0), Indices,
1561 if (GEP.getNumIndices() == 1) {
1562 unsigned AS = GEP.getPointerAddressSpace();
1563 if (GEP.getOperand(1)->getType()->getScalarSizeInBits() ==
1564 DL.getPointerSizeInBits(AS)) {
1565 Type *PtrTy = GEP.getPointerOperandType();
1566 Type *Ty = PtrTy->getPointerElementType();
1567 uint64_t TyAllocSize = DL.getTypeAllocSize(Ty);
1569 bool Matched = false;
1572 if (TyAllocSize == 1) {
1573 V = GEP.getOperand(1);
1575 } else if (match(GEP.getOperand(1),
1576 m_AShr(m_Value(V), m_ConstantInt(C)))) {
1577 if (TyAllocSize == 1ULL << C)
1579 } else if (match(GEP.getOperand(1),
1580 m_SDiv(m_Value(V), m_ConstantInt(C)))) {
1581 if (TyAllocSize == C)
1586 // Canonicalize (gep i8* X, -(ptrtoint Y))
1587 // to (inttoptr (sub (ptrtoint X), (ptrtoint Y)))
1588 // The GEP pattern is emitted by the SCEV expander for certain kinds of
1589 // pointer arithmetic.
1590 if (match(V, m_Neg(m_PtrToInt(m_Value())))) {
1591 Operator *Index = cast<Operator>(V);
1592 Value *PtrToInt = Builder->CreatePtrToInt(PtrOp, Index->getType());
1593 Value *NewSub = Builder->CreateSub(PtrToInt, Index->getOperand(1));
1594 return CastInst::Create(Instruction::IntToPtr, NewSub, GEP.getType());
1596 // Canonicalize (gep i8* X, (ptrtoint Y)-(ptrtoint X))
1599 if (match(V, m_Sub(m_PtrToInt(m_Value(Y)),
1600 m_PtrToInt(m_Specific(GEP.getOperand(0)))))) {
1601 return CastInst::CreatePointerBitCastOrAddrSpaceCast(Y,
1608 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0).
1609 Value *StrippedPtr = PtrOp->stripPointerCasts();
1610 PointerType *StrippedPtrTy = dyn_cast<PointerType>(StrippedPtr->getType());
1612 // We do not handle pointer-vector geps here.
1616 if (StrippedPtr != PtrOp) {
1617 bool HasZeroPointerIndex = false;
1618 if (ConstantInt *C = dyn_cast<ConstantInt>(GEP.getOperand(1)))
1619 HasZeroPointerIndex = C->isZero();
1621 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
1622 // into : GEP [10 x i8]* X, i32 0, ...
1624 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
1625 // into : GEP i8* X, ...
1627 // This occurs when the program declares an array extern like "int X[];"
1628 if (HasZeroPointerIndex) {
1629 PointerType *CPTy = cast<PointerType>(PtrOp->getType());
1630 if (ArrayType *CATy =
1631 dyn_cast<ArrayType>(CPTy->getElementType())) {
1632 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
1633 if (CATy->getElementType() == StrippedPtrTy->getElementType()) {
1634 // -> GEP i8* X, ...
1635 SmallVector<Value*, 8> Idx(GEP.idx_begin()+1, GEP.idx_end());
1636 GetElementPtrInst *Res = GetElementPtrInst::Create(
1637 StrippedPtrTy->getElementType(), StrippedPtr, Idx, GEP.getName());
1638 Res->setIsInBounds(GEP.isInBounds());
1639 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace())
1641 // Insert Res, and create an addrspacecast.
1643 // GEP (addrspacecast i8 addrspace(1)* X to [0 x i8]*), i32 0, ...
1645 // %0 = GEP i8 addrspace(1)* X, ...
1646 // addrspacecast i8 addrspace(1)* %0 to i8*
1647 return new AddrSpaceCastInst(Builder->Insert(Res), GEP.getType());
1650 if (ArrayType *XATy =
1651 dyn_cast<ArrayType>(StrippedPtrTy->getElementType())){
1652 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
1653 if (CATy->getElementType() == XATy->getElementType()) {
1654 // -> GEP [10 x i8]* X, i32 0, ...
1655 // At this point, we know that the cast source type is a pointer
1656 // to an array of the same type as the destination pointer
1657 // array. Because the array type is never stepped over (there
1658 // is a leading zero) we can fold the cast into this GEP.
1659 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace()) {
1660 GEP.setOperand(0, StrippedPtr);
1661 GEP.setSourceElementType(XATy);
1664 // Cannot replace the base pointer directly because StrippedPtr's
1665 // address space is different. Instead, create a new GEP followed by
1666 // an addrspacecast.
1668 // GEP (addrspacecast [10 x i8] addrspace(1)* X to [0 x i8]*),
1671 // %0 = GEP [10 x i8] addrspace(1)* X, ...
1672 // addrspacecast i8 addrspace(1)* %0 to i8*
1673 SmallVector<Value*, 8> Idx(GEP.idx_begin(), GEP.idx_end());
1674 Value *NewGEP = GEP.isInBounds()
1675 ? Builder->CreateInBoundsGEP(
1676 nullptr, StrippedPtr, Idx, GEP.getName())
1677 : Builder->CreateGEP(nullptr, StrippedPtr, Idx,
1679 return new AddrSpaceCastInst(NewGEP, GEP.getType());
1683 } else if (GEP.getNumOperands() == 2) {
1684 // Transform things like:
1685 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
1686 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
1687 Type *SrcElTy = StrippedPtrTy->getElementType();
1688 Type *ResElTy = PtrOp->getType()->getPointerElementType();
1689 if (SrcElTy->isArrayTy() &&
1690 DL.getTypeAllocSize(SrcElTy->getArrayElementType()) ==
1691 DL.getTypeAllocSize(ResElTy)) {
1692 Type *IdxType = DL.getIntPtrType(GEP.getType());
1693 Value *Idx[2] = { Constant::getNullValue(IdxType), GEP.getOperand(1) };
1696 ? Builder->CreateInBoundsGEP(nullptr, StrippedPtr, Idx,
1698 : Builder->CreateGEP(nullptr, StrippedPtr, Idx, GEP.getName());
1700 // V and GEP are both pointer types --> BitCast
1701 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
1705 // Transform things like:
1706 // %V = mul i64 %N, 4
1707 // %t = getelementptr i8* bitcast (i32* %arr to i8*), i32 %V
1708 // into: %t1 = getelementptr i32* %arr, i32 %N; bitcast
1709 if (ResElTy->isSized() && SrcElTy->isSized()) {
1710 // Check that changing the type amounts to dividing the index by a scale
1712 uint64_t ResSize = DL.getTypeAllocSize(ResElTy);
1713 uint64_t SrcSize = DL.getTypeAllocSize(SrcElTy);
1714 if (ResSize && SrcSize % ResSize == 0) {
1715 Value *Idx = GEP.getOperand(1);
1716 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
1717 uint64_t Scale = SrcSize / ResSize;
1719 // Earlier transforms ensure that the index has type IntPtrType, which
1720 // considerably simplifies the logic by eliminating implicit casts.
1721 assert(Idx->getType() == DL.getIntPtrType(GEP.getType()) &&
1722 "Index not cast to pointer width?");
1725 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
1726 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
1727 // If the multiplication NewIdx * Scale may overflow then the new
1728 // GEP may not be "inbounds".
1730 GEP.isInBounds() && NSW
1731 ? Builder->CreateInBoundsGEP(nullptr, StrippedPtr, NewIdx,
1733 : Builder->CreateGEP(nullptr, StrippedPtr, NewIdx,
1736 // The NewGEP must be pointer typed, so must the old one -> BitCast
1737 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
1743 // Similarly, transform things like:
1744 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
1745 // (where tmp = 8*tmp2) into:
1746 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
1747 if (ResElTy->isSized() && SrcElTy->isSized() && SrcElTy->isArrayTy()) {
1748 // Check that changing to the array element type amounts to dividing the
1749 // index by a scale factor.
1750 uint64_t ResSize = DL.getTypeAllocSize(ResElTy);
1751 uint64_t ArrayEltSize =
1752 DL.getTypeAllocSize(SrcElTy->getArrayElementType());
1753 if (ResSize && ArrayEltSize % ResSize == 0) {
1754 Value *Idx = GEP.getOperand(1);
1755 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
1756 uint64_t Scale = ArrayEltSize / ResSize;
1758 // Earlier transforms ensure that the index has type IntPtrType, which
1759 // considerably simplifies the logic by eliminating implicit casts.
1760 assert(Idx->getType() == DL.getIntPtrType(GEP.getType()) &&
1761 "Index not cast to pointer width?");
1764 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
1765 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
1766 // If the multiplication NewIdx * Scale may overflow then the new
1767 // GEP may not be "inbounds".
1769 Constant::getNullValue(DL.getIntPtrType(GEP.getType())),
1772 Value *NewGEP = GEP.isInBounds() && NSW
1773 ? Builder->CreateInBoundsGEP(
1774 SrcElTy, StrippedPtr, Off, GEP.getName())
1775 : Builder->CreateGEP(SrcElTy, StrippedPtr, Off,
1777 // The NewGEP must be pointer typed, so must the old one -> BitCast
1778 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
1786 // addrspacecast between types is canonicalized as a bitcast, then an
1787 // addrspacecast. To take advantage of the below bitcast + struct GEP, look
1788 // through the addrspacecast.
1789 if (AddrSpaceCastInst *ASC = dyn_cast<AddrSpaceCastInst>(PtrOp)) {
1790 // X = bitcast A addrspace(1)* to B addrspace(1)*
1791 // Y = addrspacecast A addrspace(1)* to B addrspace(2)*
1792 // Z = gep Y, <...constant indices...>
1793 // Into an addrspacecasted GEP of the struct.
1794 if (BitCastInst *BC = dyn_cast<BitCastInst>(ASC->getOperand(0)))
1798 /// See if we can simplify:
1799 /// X = bitcast A* to B*
1800 /// Y = gep X, <...constant indices...>
1801 /// into a gep of the original struct. This is important for SROA and alias
1802 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
1803 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
1804 Value *Operand = BCI->getOperand(0);
1805 PointerType *OpType = cast<PointerType>(Operand->getType());
1806 unsigned OffsetBits = DL.getPointerTypeSizeInBits(GEP.getType());
1807 APInt Offset(OffsetBits, 0);
1808 if (!isa<BitCastInst>(Operand) &&
1809 GEP.accumulateConstantOffset(DL, Offset)) {
1811 // If this GEP instruction doesn't move the pointer, just replace the GEP
1812 // with a bitcast of the real input to the dest type.
1814 // If the bitcast is of an allocation, and the allocation will be
1815 // converted to match the type of the cast, don't touch this.
1816 if (isa<AllocaInst>(Operand) || isAllocationFn(Operand, TLI)) {
1817 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
1818 if (Instruction *I = visitBitCast(*BCI)) {
1821 BCI->getParent()->getInstList().insert(BCI, I);
1822 ReplaceInstUsesWith(*BCI, I);
1828 if (Operand->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
1829 return new AddrSpaceCastInst(Operand, GEP.getType());
1830 return new BitCastInst(Operand, GEP.getType());
1833 // Otherwise, if the offset is non-zero, we need to find out if there is a
1834 // field at Offset in 'A's type. If so, we can pull the cast through the
1836 SmallVector<Value*, 8> NewIndices;
1837 if (FindElementAtOffset(OpType, Offset.getSExtValue(), NewIndices)) {
1840 ? Builder->CreateInBoundsGEP(nullptr, Operand, NewIndices)
1841 : Builder->CreateGEP(nullptr, Operand, NewIndices);
1843 if (NGEP->getType() == GEP.getType())
1844 return ReplaceInstUsesWith(GEP, NGEP);
1845 NGEP->takeName(&GEP);
1847 if (NGEP->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
1848 return new AddrSpaceCastInst(NGEP, GEP.getType());
1849 return new BitCastInst(NGEP, GEP.getType());
1858 isAllocSiteRemovable(Instruction *AI, SmallVectorImpl<WeakVH> &Users,
1859 const TargetLibraryInfo *TLI) {
1860 SmallVector<Instruction*, 4> Worklist;
1861 Worklist.push_back(AI);
1864 Instruction *PI = Worklist.pop_back_val();
1865 for (User *U : PI->users()) {
1866 Instruction *I = cast<Instruction>(U);
1867 switch (I->getOpcode()) {
1869 // Give up the moment we see something we can't handle.
1872 case Instruction::BitCast:
1873 case Instruction::GetElementPtr:
1874 Users.emplace_back(I);
1875 Worklist.push_back(I);
1878 case Instruction::ICmp: {
1879 ICmpInst *ICI = cast<ICmpInst>(I);
1880 // We can fold eq/ne comparisons with null to false/true, respectively.
1881 if (!ICI->isEquality() || !isa<ConstantPointerNull>(ICI->getOperand(1)))
1883 Users.emplace_back(I);
1887 case Instruction::Call:
1888 // Ignore no-op and store intrinsics.
1889 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1890 switch (II->getIntrinsicID()) {
1894 case Intrinsic::memmove:
1895 case Intrinsic::memcpy:
1896 case Intrinsic::memset: {
1897 MemIntrinsic *MI = cast<MemIntrinsic>(II);
1898 if (MI->isVolatile() || MI->getRawDest() != PI)
1902 case Intrinsic::dbg_declare:
1903 case Intrinsic::dbg_value:
1904 case Intrinsic::invariant_start:
1905 case Intrinsic::invariant_end:
1906 case Intrinsic::lifetime_start:
1907 case Intrinsic::lifetime_end:
1908 case Intrinsic::objectsize:
1909 Users.emplace_back(I);
1914 if (isFreeCall(I, TLI)) {
1915 Users.emplace_back(I);
1920 case Instruction::Store: {
1921 StoreInst *SI = cast<StoreInst>(I);
1922 if (SI->isVolatile() || SI->getPointerOperand() != PI)
1924 Users.emplace_back(I);
1928 llvm_unreachable("missing a return?");
1930 } while (!Worklist.empty());
1934 Instruction *InstCombiner::visitAllocSite(Instruction &MI) {
1935 // If we have a malloc call which is only used in any amount of comparisons
1936 // to null and free calls, delete the calls and replace the comparisons with
1937 // true or false as appropriate.
1938 SmallVector<WeakVH, 64> Users;
1939 if (isAllocSiteRemovable(&MI, Users, TLI)) {
1940 for (unsigned i = 0, e = Users.size(); i != e; ++i) {
1941 Instruction *I = cast_or_null<Instruction>(&*Users[i]);
1944 if (ICmpInst *C = dyn_cast<ICmpInst>(I)) {
1945 ReplaceInstUsesWith(*C,
1946 ConstantInt::get(Type::getInt1Ty(C->getContext()),
1947 C->isFalseWhenEqual()));
1948 } else if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) {
1949 ReplaceInstUsesWith(*I, UndefValue::get(I->getType()));
1950 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1951 if (II->getIntrinsicID() == Intrinsic::objectsize) {
1952 ConstantInt *CI = cast<ConstantInt>(II->getArgOperand(1));
1953 uint64_t DontKnow = CI->isZero() ? -1ULL : 0;
1954 ReplaceInstUsesWith(*I, ConstantInt::get(I->getType(), DontKnow));
1957 EraseInstFromFunction(*I);
1960 if (InvokeInst *II = dyn_cast<InvokeInst>(&MI)) {
1961 // Replace invoke with a NOP intrinsic to maintain the original CFG
1962 Module *M = II->getParent()->getParent()->getParent();
1963 Function *F = Intrinsic::getDeclaration(M, Intrinsic::donothing);
1964 InvokeInst::Create(F, II->getNormalDest(), II->getUnwindDest(),
1965 None, "", II->getParent());
1967 return EraseInstFromFunction(MI);
1972 /// \brief Move the call to free before a NULL test.
1974 /// Check if this free is accessed after its argument has been test
1975 /// against NULL (property 0).
1976 /// If yes, it is legal to move this call in its predecessor block.
1978 /// The move is performed only if the block containing the call to free
1979 /// will be removed, i.e.:
1980 /// 1. it has only one predecessor P, and P has two successors
1981 /// 2. it contains the call and an unconditional branch
1982 /// 3. its successor is the same as its predecessor's successor
1984 /// The profitability is out-of concern here and this function should
1985 /// be called only if the caller knows this transformation would be
1986 /// profitable (e.g., for code size).
1987 static Instruction *
1988 tryToMoveFreeBeforeNullTest(CallInst &FI) {
1989 Value *Op = FI.getArgOperand(0);
1990 BasicBlock *FreeInstrBB = FI.getParent();
1991 BasicBlock *PredBB = FreeInstrBB->getSinglePredecessor();
1993 // Validate part of constraint #1: Only one predecessor
1994 // FIXME: We can extend the number of predecessor, but in that case, we
1995 // would duplicate the call to free in each predecessor and it may
1996 // not be profitable even for code size.
2000 // Validate constraint #2: Does this block contains only the call to
2001 // free and an unconditional branch?
2002 // FIXME: We could check if we can speculate everything in the
2003 // predecessor block
2004 if (FreeInstrBB->size() != 2)
2007 if (!match(FreeInstrBB->getTerminator(), m_UnconditionalBr(SuccBB)))
2010 // Validate the rest of constraint #1 by matching on the pred branch.
2011 TerminatorInst *TI = PredBB->getTerminator();
2012 BasicBlock *TrueBB, *FalseBB;
2013 ICmpInst::Predicate Pred;
2014 if (!match(TI, m_Br(m_ICmp(Pred, m_Specific(Op), m_Zero()), TrueBB, FalseBB)))
2016 if (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE)
2019 // Validate constraint #3: Ensure the null case just falls through.
2020 if (SuccBB != (Pred == ICmpInst::ICMP_EQ ? TrueBB : FalseBB))
2022 assert(FreeInstrBB == (Pred == ICmpInst::ICMP_EQ ? FalseBB : TrueBB) &&
2023 "Broken CFG: missing edge from predecessor to successor");
2030 Instruction *InstCombiner::visitFree(CallInst &FI) {
2031 Value *Op = FI.getArgOperand(0);
2033 // free undef -> unreachable.
2034 if (isa<UndefValue>(Op)) {
2035 // Insert a new store to null because we cannot modify the CFG here.
2036 Builder->CreateStore(ConstantInt::getTrue(FI.getContext()),
2037 UndefValue::get(Type::getInt1PtrTy(FI.getContext())));
2038 return EraseInstFromFunction(FI);
2041 // If we have 'free null' delete the instruction. This can happen in stl code
2042 // when lots of inlining happens.
2043 if (isa<ConstantPointerNull>(Op))
2044 return EraseInstFromFunction(FI);
2046 // If we optimize for code size, try to move the call to free before the null
2047 // test so that simplify cfg can remove the empty block and dead code
2048 // elimination the branch. I.e., helps to turn something like:
2049 // if (foo) free(foo);
2053 if (Instruction *I = tryToMoveFreeBeforeNullTest(FI))
2059 Instruction *InstCombiner::visitReturnInst(ReturnInst &RI) {
2060 if (RI.getNumOperands() == 0) // ret void
2063 Value *ResultOp = RI.getOperand(0);
2064 Type *VTy = ResultOp->getType();
2065 if (!VTy->isIntegerTy())
2068 // There might be assume intrinsics dominating this return that completely
2069 // determine the value. If so, constant fold it.
2070 unsigned BitWidth = VTy->getPrimitiveSizeInBits();
2071 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2072 computeKnownBits(ResultOp, KnownZero, KnownOne, 0, &RI);
2073 if ((KnownZero|KnownOne).isAllOnesValue())
2074 RI.setOperand(0, Constant::getIntegerValue(VTy, KnownOne));
2079 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
2080 // Change br (not X), label True, label False to: br X, label False, True
2082 BasicBlock *TrueDest;
2083 BasicBlock *FalseDest;
2084 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
2085 !isa<Constant>(X)) {
2086 // Swap Destinations and condition...
2088 BI.swapSuccessors();
2092 // If the condition is irrelevant, remove the use so that other
2093 // transforms on the condition become more effective.
2094 if (BI.isConditional() &&
2095 BI.getSuccessor(0) == BI.getSuccessor(1) &&
2096 !isa<UndefValue>(BI.getCondition())) {
2097 BI.setCondition(UndefValue::get(BI.getCondition()->getType()));
2101 // Canonicalize fcmp_one -> fcmp_oeq
2102 FCmpInst::Predicate FPred; Value *Y;
2103 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
2104 TrueDest, FalseDest)) &&
2105 BI.getCondition()->hasOneUse())
2106 if (FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
2107 FPred == FCmpInst::FCMP_OGE) {
2108 FCmpInst *Cond = cast<FCmpInst>(BI.getCondition());
2109 Cond->setPredicate(FCmpInst::getInversePredicate(FPred));
2111 // Swap Destinations and condition.
2112 BI.swapSuccessors();
2117 // Canonicalize icmp_ne -> icmp_eq
2118 ICmpInst::Predicate IPred;
2119 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
2120 TrueDest, FalseDest)) &&
2121 BI.getCondition()->hasOneUse())
2122 if (IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
2123 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
2124 IPred == ICmpInst::ICMP_SGE) {
2125 ICmpInst *Cond = cast<ICmpInst>(BI.getCondition());
2126 Cond->setPredicate(ICmpInst::getInversePredicate(IPred));
2127 // Swap Destinations and condition.
2128 BI.swapSuccessors();
2136 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
2137 Value *Cond = SI.getCondition();
2138 unsigned BitWidth = cast<IntegerType>(Cond->getType())->getBitWidth();
2139 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2140 computeKnownBits(Cond, KnownZero, KnownOne, 0, &SI);
2141 unsigned LeadingKnownZeros = KnownZero.countLeadingOnes();
2142 unsigned LeadingKnownOnes = KnownOne.countLeadingOnes();
2144 // Compute the number of leading bits we can ignore.
2145 for (auto &C : SI.cases()) {
2146 LeadingKnownZeros = std::min(
2147 LeadingKnownZeros, C.getCaseValue()->getValue().countLeadingZeros());
2148 LeadingKnownOnes = std::min(
2149 LeadingKnownOnes, C.getCaseValue()->getValue().countLeadingOnes());
2152 unsigned NewWidth = BitWidth - std::max(LeadingKnownZeros, LeadingKnownOnes);
2154 // Truncate the condition operand if the new type is equal to or larger than
2155 // the largest legal integer type. We need to be conservative here since
2156 // x86 generates redundant zero-extension instructions if the operand is
2157 // truncated to i8 or i16.
2158 bool TruncCond = false;
2159 if (NewWidth > 0 && BitWidth > NewWidth &&
2160 NewWidth >= DL.getLargestLegalIntTypeSize()) {
2162 IntegerType *Ty = IntegerType::get(SI.getContext(), NewWidth);
2163 Builder->SetInsertPoint(&SI);
2164 Value *NewCond = Builder->CreateTrunc(SI.getCondition(), Ty, "trunc");
2165 SI.setCondition(NewCond);
2167 for (auto &C : SI.cases())
2168 static_cast<SwitchInst::CaseIt *>(&C)->setValue(ConstantInt::get(
2169 SI.getContext(), C.getCaseValue()->getValue().trunc(NewWidth)));
2172 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
2173 if (I->getOpcode() == Instruction::Add)
2174 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2175 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
2176 // Skip the first item since that's the default case.
2177 for (SwitchInst::CaseIt i = SI.case_begin(), e = SI.case_end();
2179 ConstantInt* CaseVal = i.getCaseValue();
2180 Constant *LHS = CaseVal;
2182 LHS = LeadingKnownZeros
2183 ? ConstantExpr::getZExt(CaseVal, Cond->getType())
2184 : ConstantExpr::getSExt(CaseVal, Cond->getType());
2185 Constant* NewCaseVal = ConstantExpr::getSub(LHS, AddRHS);
2186 assert(isa<ConstantInt>(NewCaseVal) &&
2187 "Result of expression should be constant");
2188 i.setValue(cast<ConstantInt>(NewCaseVal));
2190 SI.setCondition(I->getOperand(0));
2196 return TruncCond ? &SI : nullptr;
2199 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
2200 Value *Agg = EV.getAggregateOperand();
2202 if (!EV.hasIndices())
2203 return ReplaceInstUsesWith(EV, Agg);
2206 SimplifyExtractValueInst(Agg, EV.getIndices(), DL, TLI, DT, AC))
2207 return ReplaceInstUsesWith(EV, V);
2209 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
2210 // We're extracting from an insertvalue instruction, compare the indices
2211 const unsigned *exti, *exte, *insi, *inse;
2212 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
2213 exte = EV.idx_end(), inse = IV->idx_end();
2214 exti != exte && insi != inse;
2217 // The insert and extract both reference distinctly different elements.
2218 // This means the extract is not influenced by the insert, and we can
2219 // replace the aggregate operand of the extract with the aggregate
2220 // operand of the insert. i.e., replace
2221 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
2222 // %E = extractvalue { i32, { i32 } } %I, 0
2224 // %E = extractvalue { i32, { i32 } } %A, 0
2225 return ExtractValueInst::Create(IV->getAggregateOperand(),
2228 if (exti == exte && insi == inse)
2229 // Both iterators are at the end: Index lists are identical. Replace
2230 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
2231 // %C = extractvalue { i32, { i32 } } %B, 1, 0
2233 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
2235 // The extract list is a prefix of the insert list. i.e. replace
2236 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
2237 // %E = extractvalue { i32, { i32 } } %I, 1
2239 // %X = extractvalue { i32, { i32 } } %A, 1
2240 // %E = insertvalue { i32 } %X, i32 42, 0
2241 // by switching the order of the insert and extract (though the
2242 // insertvalue should be left in, since it may have other uses).
2243 Value *NewEV = Builder->CreateExtractValue(IV->getAggregateOperand(),
2245 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
2246 makeArrayRef(insi, inse));
2249 // The insert list is a prefix of the extract list
2250 // We can simply remove the common indices from the extract and make it
2251 // operate on the inserted value instead of the insertvalue result.
2253 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
2254 // %E = extractvalue { i32, { i32 } } %I, 1, 0
2256 // %E extractvalue { i32 } { i32 42 }, 0
2257 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
2258 makeArrayRef(exti, exte));
2260 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) {
2261 // We're extracting from an intrinsic, see if we're the only user, which
2262 // allows us to simplify multiple result intrinsics to simpler things that
2263 // just get one value.
2264 if (II->hasOneUse()) {
2265 // Check if we're grabbing the overflow bit or the result of a 'with
2266 // overflow' intrinsic. If it's the latter we can remove the intrinsic
2267 // and replace it with a traditional binary instruction.
2268 switch (II->getIntrinsicID()) {
2269 case Intrinsic::uadd_with_overflow:
2270 case Intrinsic::sadd_with_overflow:
2271 if (*EV.idx_begin() == 0) { // Normal result.
2272 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2273 ReplaceInstUsesWith(*II, UndefValue::get(II->getType()));
2274 EraseInstFromFunction(*II);
2275 return BinaryOperator::CreateAdd(LHS, RHS);
2278 // If the normal result of the add is dead, and the RHS is a constant,
2279 // we can transform this into a range comparison.
2280 // overflow = uadd a, -4 --> overflow = icmp ugt a, 3
2281 if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow)
2282 if (ConstantInt *CI = dyn_cast<ConstantInt>(II->getArgOperand(1)))
2283 return new ICmpInst(ICmpInst::ICMP_UGT, II->getArgOperand(0),
2284 ConstantExpr::getNot(CI));
2286 case Intrinsic::usub_with_overflow:
2287 case Intrinsic::ssub_with_overflow:
2288 if (*EV.idx_begin() == 0) { // Normal result.
2289 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2290 ReplaceInstUsesWith(*II, UndefValue::get(II->getType()));
2291 EraseInstFromFunction(*II);
2292 return BinaryOperator::CreateSub(LHS, RHS);
2295 case Intrinsic::umul_with_overflow:
2296 case Intrinsic::smul_with_overflow:
2297 if (*EV.idx_begin() == 0) { // Normal result.
2298 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2299 ReplaceInstUsesWith(*II, UndefValue::get(II->getType()));
2300 EraseInstFromFunction(*II);
2301 return BinaryOperator::CreateMul(LHS, RHS);
2309 if (LoadInst *L = dyn_cast<LoadInst>(Agg))
2310 // If the (non-volatile) load only has one use, we can rewrite this to a
2311 // load from a GEP. This reduces the size of the load.
2312 // FIXME: If a load is used only by extractvalue instructions then this
2313 // could be done regardless of having multiple uses.
2314 if (L->isSimple() && L->hasOneUse()) {
2315 // extractvalue has integer indices, getelementptr has Value*s. Convert.
2316 SmallVector<Value*, 4> Indices;
2317 // Prefix an i32 0 since we need the first element.
2318 Indices.push_back(Builder->getInt32(0));
2319 for (ExtractValueInst::idx_iterator I = EV.idx_begin(), E = EV.idx_end();
2321 Indices.push_back(Builder->getInt32(*I));
2323 // We need to insert these at the location of the old load, not at that of
2324 // the extractvalue.
2325 Builder->SetInsertPoint(L->getParent(), L);
2326 Value *GEP = Builder->CreateInBoundsGEP(L->getType(),
2327 L->getPointerOperand(), Indices);
2328 // Returning the load directly will cause the main loop to insert it in
2329 // the wrong spot, so use ReplaceInstUsesWith().
2330 return ReplaceInstUsesWith(EV, Builder->CreateLoad(GEP));
2332 // We could simplify extracts from other values. Note that nested extracts may
2333 // already be simplified implicitly by the above: extract (extract (insert) )
2334 // will be translated into extract ( insert ( extract ) ) first and then just
2335 // the value inserted, if appropriate. Similarly for extracts from single-use
2336 // loads: extract (extract (load)) will be translated to extract (load (gep))
2337 // and if again single-use then via load (gep (gep)) to load (gep).
2338 // However, double extracts from e.g. function arguments or return values
2339 // aren't handled yet.
2343 /// isCatchAll - Return 'true' if the given typeinfo will match anything.
2344 static bool isCatchAll(EHPersonality Personality, Constant *TypeInfo) {
2345 switch (Personality) {
2346 case EHPersonality::GNU_C:
2347 // The GCC C EH personality only exists to support cleanups, so it's not
2348 // clear what the semantics of catch clauses are.
2350 case EHPersonality::Unknown:
2352 case EHPersonality::GNU_Ada:
2353 // While __gnat_all_others_value will match any Ada exception, it doesn't
2354 // match foreign exceptions (or didn't, before gcc-4.7).
2356 case EHPersonality::GNU_CXX:
2357 case EHPersonality::GNU_ObjC:
2358 case EHPersonality::MSVC_X86SEH:
2359 case EHPersonality::MSVC_Win64SEH:
2360 case EHPersonality::MSVC_CXX:
2361 return TypeInfo->isNullValue();
2363 llvm_unreachable("invalid enum");
2366 static bool shorter_filter(const Value *LHS, const Value *RHS) {
2368 cast<ArrayType>(LHS->getType())->getNumElements()
2370 cast<ArrayType>(RHS->getType())->getNumElements();
2373 Instruction *InstCombiner::visitLandingPadInst(LandingPadInst &LI) {
2374 // The logic here should be correct for any real-world personality function.
2375 // However if that turns out not to be true, the offending logic can always
2376 // be conditioned on the personality function, like the catch-all logic is.
2377 EHPersonality Personality =
2378 classifyEHPersonality(LI.getParent()->getParent()->getPersonalityFn());
2380 // Simplify the list of clauses, eg by removing repeated catch clauses
2381 // (these are often created by inlining).
2382 bool MakeNewInstruction = false; // If true, recreate using the following:
2383 SmallVector<Constant *, 16> NewClauses; // - Clauses for the new instruction;
2384 bool CleanupFlag = LI.isCleanup(); // - The new instruction is a cleanup.
2386 SmallPtrSet<Value *, 16> AlreadyCaught; // Typeinfos known caught already.
2387 for (unsigned i = 0, e = LI.getNumClauses(); i != e; ++i) {
2388 bool isLastClause = i + 1 == e;
2389 if (LI.isCatch(i)) {
2391 Constant *CatchClause = LI.getClause(i);
2392 Constant *TypeInfo = CatchClause->stripPointerCasts();
2394 // If we already saw this clause, there is no point in having a second
2396 if (AlreadyCaught.insert(TypeInfo).second) {
2397 // This catch clause was not already seen.
2398 NewClauses.push_back(CatchClause);
2400 // Repeated catch clause - drop the redundant copy.
2401 MakeNewInstruction = true;
2404 // If this is a catch-all then there is no point in keeping any following
2405 // clauses or marking the landingpad as having a cleanup.
2406 if (isCatchAll(Personality, TypeInfo)) {
2408 MakeNewInstruction = true;
2409 CleanupFlag = false;
2413 // A filter clause. If any of the filter elements were already caught
2414 // then they can be dropped from the filter. It is tempting to try to
2415 // exploit the filter further by saying that any typeinfo that does not
2416 // occur in the filter can't be caught later (and thus can be dropped).
2417 // However this would be wrong, since typeinfos can match without being
2418 // equal (for example if one represents a C++ class, and the other some
2419 // class derived from it).
2420 assert(LI.isFilter(i) && "Unsupported landingpad clause!");
2421 Constant *FilterClause = LI.getClause(i);
2422 ArrayType *FilterType = cast<ArrayType>(FilterClause->getType());
2423 unsigned NumTypeInfos = FilterType->getNumElements();
2425 // An empty filter catches everything, so there is no point in keeping any
2426 // following clauses or marking the landingpad as having a cleanup. By
2427 // dealing with this case here the following code is made a bit simpler.
2428 if (!NumTypeInfos) {
2429 NewClauses.push_back(FilterClause);
2431 MakeNewInstruction = true;
2432 CleanupFlag = false;
2436 bool MakeNewFilter = false; // If true, make a new filter.
2437 SmallVector<Constant *, 16> NewFilterElts; // New elements.
2438 if (isa<ConstantAggregateZero>(FilterClause)) {
2439 // Not an empty filter - it contains at least one null typeinfo.
2440 assert(NumTypeInfos > 0 && "Should have handled empty filter already!");
2441 Constant *TypeInfo =
2442 Constant::getNullValue(FilterType->getElementType());
2443 // If this typeinfo is a catch-all then the filter can never match.
2444 if (isCatchAll(Personality, TypeInfo)) {
2445 // Throw the filter away.
2446 MakeNewInstruction = true;
2450 // There is no point in having multiple copies of this typeinfo, so
2451 // discard all but the first copy if there is more than one.
2452 NewFilterElts.push_back(TypeInfo);
2453 if (NumTypeInfos > 1)
2454 MakeNewFilter = true;
2456 ConstantArray *Filter = cast<ConstantArray>(FilterClause);
2457 SmallPtrSet<Value *, 16> SeenInFilter; // For uniquing the elements.
2458 NewFilterElts.reserve(NumTypeInfos);
2460 // Remove any filter elements that were already caught or that already
2461 // occurred in the filter. While there, see if any of the elements are
2462 // catch-alls. If so, the filter can be discarded.
2463 bool SawCatchAll = false;
2464 for (unsigned j = 0; j != NumTypeInfos; ++j) {
2465 Constant *Elt = Filter->getOperand(j);
2466 Constant *TypeInfo = Elt->stripPointerCasts();
2467 if (isCatchAll(Personality, TypeInfo)) {
2468 // This element is a catch-all. Bail out, noting this fact.
2472 if (AlreadyCaught.count(TypeInfo))
2473 // Already caught by an earlier clause, so having it in the filter
2476 // There is no point in having multiple copies of the same typeinfo in
2477 // a filter, so only add it if we didn't already.
2478 if (SeenInFilter.insert(TypeInfo).second)
2479 NewFilterElts.push_back(cast<Constant>(Elt));
2481 // A filter containing a catch-all cannot match anything by definition.
2483 // Throw the filter away.
2484 MakeNewInstruction = true;
2488 // If we dropped something from the filter, make a new one.
2489 if (NewFilterElts.size() < NumTypeInfos)
2490 MakeNewFilter = true;
2492 if (MakeNewFilter) {
2493 FilterType = ArrayType::get(FilterType->getElementType(),
2494 NewFilterElts.size());
2495 FilterClause = ConstantArray::get(FilterType, NewFilterElts);
2496 MakeNewInstruction = true;
2499 NewClauses.push_back(FilterClause);
2501 // If the new filter is empty then it will catch everything so there is
2502 // no point in keeping any following clauses or marking the landingpad
2503 // as having a cleanup. The case of the original filter being empty was
2504 // already handled above.
2505 if (MakeNewFilter && !NewFilterElts.size()) {
2506 assert(MakeNewInstruction && "New filter but not a new instruction!");
2507 CleanupFlag = false;
2513 // If several filters occur in a row then reorder them so that the shortest
2514 // filters come first (those with the smallest number of elements). This is
2515 // advantageous because shorter filters are more likely to match, speeding up
2516 // unwinding, but mostly because it increases the effectiveness of the other
2517 // filter optimizations below.
2518 for (unsigned i = 0, e = NewClauses.size(); i + 1 < e; ) {
2520 // Find the maximal 'j' s.t. the range [i, j) consists entirely of filters.
2521 for (j = i; j != e; ++j)
2522 if (!isa<ArrayType>(NewClauses[j]->getType()))
2525 // Check whether the filters are already sorted by length. We need to know
2526 // if sorting them is actually going to do anything so that we only make a
2527 // new landingpad instruction if it does.
2528 for (unsigned k = i; k + 1 < j; ++k)
2529 if (shorter_filter(NewClauses[k+1], NewClauses[k])) {
2530 // Not sorted, so sort the filters now. Doing an unstable sort would be
2531 // correct too but reordering filters pointlessly might confuse users.
2532 std::stable_sort(NewClauses.begin() + i, NewClauses.begin() + j,
2534 MakeNewInstruction = true;
2538 // Look for the next batch of filters.
2542 // If typeinfos matched if and only if equal, then the elements of a filter L
2543 // that occurs later than a filter F could be replaced by the intersection of
2544 // the elements of F and L. In reality two typeinfos can match without being
2545 // equal (for example if one represents a C++ class, and the other some class
2546 // derived from it) so it would be wrong to perform this transform in general.
2547 // However the transform is correct and useful if F is a subset of L. In that
2548 // case L can be replaced by F, and thus removed altogether since repeating a
2549 // filter is pointless. So here we look at all pairs of filters F and L where
2550 // L follows F in the list of clauses, and remove L if every element of F is
2551 // an element of L. This can occur when inlining C++ functions with exception
2553 for (unsigned i = 0; i + 1 < NewClauses.size(); ++i) {
2554 // Examine each filter in turn.
2555 Value *Filter = NewClauses[i];
2556 ArrayType *FTy = dyn_cast<ArrayType>(Filter->getType());
2558 // Not a filter - skip it.
2560 unsigned FElts = FTy->getNumElements();
2561 // Examine each filter following this one. Doing this backwards means that
2562 // we don't have to worry about filters disappearing under us when removed.
2563 for (unsigned j = NewClauses.size() - 1; j != i; --j) {
2564 Value *LFilter = NewClauses[j];
2565 ArrayType *LTy = dyn_cast<ArrayType>(LFilter->getType());
2567 // Not a filter - skip it.
2569 // If Filter is a subset of LFilter, i.e. every element of Filter is also
2570 // an element of LFilter, then discard LFilter.
2571 SmallVectorImpl<Constant *>::iterator J = NewClauses.begin() + j;
2572 // If Filter is empty then it is a subset of LFilter.
2575 NewClauses.erase(J);
2576 MakeNewInstruction = true;
2577 // Move on to the next filter.
2580 unsigned LElts = LTy->getNumElements();
2581 // If Filter is longer than LFilter then it cannot be a subset of it.
2583 // Move on to the next filter.
2585 // At this point we know that LFilter has at least one element.
2586 if (isa<ConstantAggregateZero>(LFilter)) { // LFilter only contains zeros.
2587 // Filter is a subset of LFilter iff Filter contains only zeros (as we
2588 // already know that Filter is not longer than LFilter).
2589 if (isa<ConstantAggregateZero>(Filter)) {
2590 assert(FElts <= LElts && "Should have handled this case earlier!");
2592 NewClauses.erase(J);
2593 MakeNewInstruction = true;
2595 // Move on to the next filter.
2598 ConstantArray *LArray = cast<ConstantArray>(LFilter);
2599 if (isa<ConstantAggregateZero>(Filter)) { // Filter only contains zeros.
2600 // Since Filter is non-empty and contains only zeros, it is a subset of
2601 // LFilter iff LFilter contains a zero.
2602 assert(FElts > 0 && "Should have eliminated the empty filter earlier!");
2603 for (unsigned l = 0; l != LElts; ++l)
2604 if (LArray->getOperand(l)->isNullValue()) {
2605 // LFilter contains a zero - discard it.
2606 NewClauses.erase(J);
2607 MakeNewInstruction = true;
2610 // Move on to the next filter.
2613 // At this point we know that both filters are ConstantArrays. Loop over
2614 // operands to see whether every element of Filter is also an element of
2615 // LFilter. Since filters tend to be short this is probably faster than
2616 // using a method that scales nicely.
2617 ConstantArray *FArray = cast<ConstantArray>(Filter);
2618 bool AllFound = true;
2619 for (unsigned f = 0; f != FElts; ++f) {
2620 Value *FTypeInfo = FArray->getOperand(f)->stripPointerCasts();
2622 for (unsigned l = 0; l != LElts; ++l) {
2623 Value *LTypeInfo = LArray->getOperand(l)->stripPointerCasts();
2624 if (LTypeInfo == FTypeInfo) {
2634 NewClauses.erase(J);
2635 MakeNewInstruction = true;
2637 // Move on to the next filter.
2641 // If we changed any of the clauses, replace the old landingpad instruction
2643 if (MakeNewInstruction) {
2644 LandingPadInst *NLI = LandingPadInst::Create(LI.getType(),
2646 for (unsigned i = 0, e = NewClauses.size(); i != e; ++i)
2647 NLI->addClause(NewClauses[i]);
2648 // A landing pad with no clauses must have the cleanup flag set. It is
2649 // theoretically possible, though highly unlikely, that we eliminated all
2650 // clauses. If so, force the cleanup flag to true.
2651 if (NewClauses.empty())
2653 NLI->setCleanup(CleanupFlag);
2657 // Even if none of the clauses changed, we may nonetheless have understood
2658 // that the cleanup flag is pointless. Clear it if so.
2659 if (LI.isCleanup() != CleanupFlag) {
2660 assert(!CleanupFlag && "Adding a cleanup, not removing one?!");
2661 LI.setCleanup(CleanupFlag);
2668 /// TryToSinkInstruction - Try to move the specified instruction from its
2669 /// current block into the beginning of DestBlock, which can only happen if it's
2670 /// safe to move the instruction past all of the instructions between it and the
2671 /// end of its block.
2672 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
2673 assert(I->hasOneUse() && "Invariants didn't hold!");
2675 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
2676 if (isa<PHINode>(I) || I->isEHPad() || I->mayHaveSideEffects() ||
2677 isa<TerminatorInst>(I))
2680 // Do not sink alloca instructions out of the entry block.
2681 if (isa<AllocaInst>(I) && I->getParent() ==
2682 &DestBlock->getParent()->getEntryBlock())
2685 // We can only sink load instructions if there is nothing between the load and
2686 // the end of block that could change the value.
2687 if (I->mayReadFromMemory()) {
2688 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
2690 if (Scan->mayWriteToMemory())
2694 BasicBlock::iterator InsertPos = DestBlock->getFirstInsertionPt();
2695 I->moveBefore(InsertPos);
2700 bool InstCombiner::run() {
2701 while (!Worklist.isEmpty()) {
2702 Instruction *I = Worklist.RemoveOne();
2703 if (I == nullptr) continue; // skip null values.
2705 // Check to see if we can DCE the instruction.
2706 if (isInstructionTriviallyDead(I, TLI)) {
2707 DEBUG(dbgs() << "IC: DCE: " << *I << '\n');
2708 EraseInstFromFunction(*I);
2710 MadeIRChange = true;
2714 // Instruction isn't dead, see if we can constant propagate it.
2715 if (!I->use_empty() &&
2716 (I->getNumOperands() == 0 || isa<Constant>(I->getOperand(0)))) {
2717 if (Constant *C = ConstantFoldInstruction(I, DL, TLI)) {
2718 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n');
2720 // Add operands to the worklist.
2721 ReplaceInstUsesWith(*I, C);
2723 EraseInstFromFunction(*I);
2724 MadeIRChange = true;
2729 // See if we can trivially sink this instruction to a successor basic block.
2730 if (I->hasOneUse()) {
2731 BasicBlock *BB = I->getParent();
2732 Instruction *UserInst = cast<Instruction>(*I->user_begin());
2733 BasicBlock *UserParent;
2735 // Get the block the use occurs in.
2736 if (PHINode *PN = dyn_cast<PHINode>(UserInst))
2737 UserParent = PN->getIncomingBlock(*I->use_begin());
2739 UserParent = UserInst->getParent();
2741 if (UserParent != BB) {
2742 bool UserIsSuccessor = false;
2743 // See if the user is one of our successors.
2744 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
2745 if (*SI == UserParent) {
2746 UserIsSuccessor = true;
2750 // If the user is one of our immediate successors, and if that successor
2751 // only has us as a predecessors (we'd have to split the critical edge
2752 // otherwise), we can keep going.
2753 if (UserIsSuccessor && UserParent->getSinglePredecessor()) {
2754 // Okay, the CFG is simple enough, try to sink this instruction.
2755 if (TryToSinkInstruction(I, UserParent)) {
2756 MadeIRChange = true;
2757 // We'll add uses of the sunk instruction below, but since sinking
2758 // can expose opportunities for it's *operands* add them to the
2760 for (Use &U : I->operands())
2761 if (Instruction *OpI = dyn_cast<Instruction>(U.get()))
2768 // Now that we have an instruction, try combining it to simplify it.
2769 Builder->SetInsertPoint(I->getParent(), I);
2770 Builder->SetCurrentDebugLocation(I->getDebugLoc());
2775 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str(););
2776 DEBUG(dbgs() << "IC: Visiting: " << OrigI << '\n');
2778 if (Instruction *Result = visit(*I)) {
2780 // Should we replace the old instruction with a new one?
2782 DEBUG(dbgs() << "IC: Old = " << *I << '\n'
2783 << " New = " << *Result << '\n');
2785 if (I->getDebugLoc())
2786 Result->setDebugLoc(I->getDebugLoc());
2787 // Everything uses the new instruction now.
2788 I->replaceAllUsesWith(Result);
2790 // Move the name to the new instruction first.
2791 Result->takeName(I);
2793 // Push the new instruction and any users onto the worklist.
2794 Worklist.Add(Result);
2795 Worklist.AddUsersToWorkList(*Result);
2797 // Insert the new instruction into the basic block...
2798 BasicBlock *InstParent = I->getParent();
2799 BasicBlock::iterator InsertPos = I;
2801 // If we replace a PHI with something that isn't a PHI, fix up the
2803 if (!isa<PHINode>(Result) && isa<PHINode>(InsertPos))
2804 InsertPos = InstParent->getFirstInsertionPt();
2806 InstParent->getInstList().insert(InsertPos, Result);
2808 EraseInstFromFunction(*I);
2811 DEBUG(dbgs() << "IC: Mod = " << OrigI << '\n'
2812 << " New = " << *I << '\n');
2815 // If the instruction was modified, it's possible that it is now dead.
2816 // if so, remove it.
2817 if (isInstructionTriviallyDead(I, TLI)) {
2818 EraseInstFromFunction(*I);
2821 Worklist.AddUsersToWorkList(*I);
2824 MadeIRChange = true;
2829 return MadeIRChange;
2832 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
2833 /// all reachable code to the worklist.
2835 /// This has a couple of tricks to make the code faster and more powerful. In
2836 /// particular, we constant fold and DCE instructions as we go, to avoid adding
2837 /// them to the worklist (this significantly speeds up instcombine on code where
2838 /// many instructions are dead or constant). Additionally, if we find a branch
2839 /// whose condition is a known constant, we only visit the reachable successors.
2841 static bool AddReachableCodeToWorklist(BasicBlock *BB, const DataLayout &DL,
2842 SmallPtrSetImpl<BasicBlock *> &Visited,
2843 InstCombineWorklist &ICWorklist,
2844 const TargetLibraryInfo *TLI) {
2845 bool MadeIRChange = false;
2846 SmallVector<BasicBlock*, 256> Worklist;
2847 Worklist.push_back(BB);
2849 SmallVector<Instruction*, 128> InstrsForInstCombineWorklist;
2850 DenseMap<ConstantExpr*, Constant*> FoldedConstants;
2853 BB = Worklist.pop_back_val();
2855 // We have now visited this block! If we've already been here, ignore it.
2856 if (!Visited.insert(BB).second)
2859 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
2860 Instruction *Inst = BBI++;
2862 // DCE instruction if trivially dead.
2863 if (isInstructionTriviallyDead(Inst, TLI)) {
2865 DEBUG(dbgs() << "IC: DCE: " << *Inst << '\n');
2866 Inst->eraseFromParent();
2870 // ConstantProp instruction if trivially constant.
2871 if (!Inst->use_empty() &&
2872 (Inst->getNumOperands() == 0 || isa<Constant>(Inst->getOperand(0))))
2873 if (Constant *C = ConstantFoldInstruction(Inst, DL, TLI)) {
2874 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: "
2876 Inst->replaceAllUsesWith(C);
2878 Inst->eraseFromParent();
2882 // See if we can constant fold its operands.
2883 for (User::op_iterator i = Inst->op_begin(), e = Inst->op_end(); i != e;
2885 ConstantExpr *CE = dyn_cast<ConstantExpr>(i);
2889 Constant *&FoldRes = FoldedConstants[CE];
2891 FoldRes = ConstantFoldConstantExpression(CE, DL, TLI);
2895 if (FoldRes != CE) {
2897 MadeIRChange = true;
2901 InstrsForInstCombineWorklist.push_back(Inst);
2904 // Recursively visit successors. If this is a branch or switch on a
2905 // constant, only visit the reachable successor.
2906 TerminatorInst *TI = BB->getTerminator();
2907 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
2908 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
2909 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
2910 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
2911 Worklist.push_back(ReachableBB);
2914 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
2915 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
2916 // See if this is an explicit destination.
2917 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2919 if (i.getCaseValue() == Cond) {
2920 BasicBlock *ReachableBB = i.getCaseSuccessor();
2921 Worklist.push_back(ReachableBB);
2925 // Otherwise it is the default destination.
2926 Worklist.push_back(SI->getDefaultDest());
2931 for (BasicBlock *SuccBB : TI->successors())
2932 Worklist.push_back(SuccBB);
2933 } while (!Worklist.empty());
2935 // Once we've found all of the instructions to add to instcombine's worklist,
2936 // add them in reverse order. This way instcombine will visit from the top
2937 // of the function down. This jives well with the way that it adds all uses
2938 // of instructions to the worklist after doing a transformation, thus avoiding
2939 // some N^2 behavior in pathological cases.
2940 ICWorklist.AddInitialGroup(&InstrsForInstCombineWorklist[0],
2941 InstrsForInstCombineWorklist.size());
2943 return MadeIRChange;
2946 /// \brief Populate the IC worklist from a function, and prune any dead basic
2947 /// blocks discovered in the process.
2949 /// This also does basic constant propagation and other forward fixing to make
2950 /// the combiner itself run much faster.
2951 static bool prepareICWorklistFromFunction(Function &F, const DataLayout &DL,
2952 TargetLibraryInfo *TLI,
2953 InstCombineWorklist &ICWorklist) {
2954 bool MadeIRChange = false;
2956 // Do a depth-first traversal of the function, populate the worklist with
2957 // the reachable instructions. Ignore blocks that are not reachable. Keep
2958 // track of which blocks we visit.
2959 SmallPtrSet<BasicBlock *, 64> Visited;
2961 AddReachableCodeToWorklist(F.begin(), DL, Visited, ICWorklist, TLI);
2963 // Do a quick scan over the function. If we find any blocks that are
2964 // unreachable, remove any instructions inside of them. This prevents
2965 // the instcombine code from having to deal with some bad special cases.
2966 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
2967 if (Visited.count(BB))
2970 // Delete the instructions backwards, as it has a reduced likelihood of
2971 // having to update as many def-use and use-def chains.
2972 Instruction *EndInst = BB->getTerminator(); // Last not to be deleted.
2973 while (EndInst != BB->begin()) {
2974 // Delete the next to last instruction.
2975 BasicBlock::iterator I = EndInst;
2976 Instruction *Inst = --I;
2977 if (!Inst->use_empty())
2978 Inst->replaceAllUsesWith(UndefValue::get(Inst->getType()));
2979 if (Inst->isEHPad()) {
2983 if (!isa<DbgInfoIntrinsic>(Inst)) {
2985 MadeIRChange = true;
2987 Inst->eraseFromParent();
2991 return MadeIRChange;
2995 combineInstructionsOverFunction(Function &F, InstCombineWorklist &Worklist,
2996 AliasAnalysis *AA, AssumptionCache &AC,
2997 TargetLibraryInfo &TLI, DominatorTree &DT,
2998 LoopInfo *LI = nullptr) {
2999 auto &DL = F.getParent()->getDataLayout();
3001 /// Builder - This is an IRBuilder that automatically inserts new
3002 /// instructions into the worklist when they are created.
3003 IRBuilder<true, TargetFolder, InstCombineIRInserter> Builder(
3004 F.getContext(), TargetFolder(DL), InstCombineIRInserter(Worklist, &AC));
3006 // Lower dbg.declare intrinsics otherwise their value may be clobbered
3008 bool DbgDeclaresChanged = LowerDbgDeclare(F);
3010 // Iterate while there is work to do.
3014 DEBUG(dbgs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
3015 << F.getName() << "\n");
3017 bool Changed = false;
3018 if (prepareICWorklistFromFunction(F, DL, &TLI, Worklist))
3021 InstCombiner IC(Worklist, &Builder, F.optForMinSize(),
3022 AA, &AC, &TLI, &DT, DL, LI);
3030 return DbgDeclaresChanged || Iteration > 1;
3033 PreservedAnalyses InstCombinePass::run(Function &F,
3034 AnalysisManager<Function> *AM) {
3035 auto &AC = AM->getResult<AssumptionAnalysis>(F);
3036 auto &DT = AM->getResult<DominatorTreeAnalysis>(F);
3037 auto &TLI = AM->getResult<TargetLibraryAnalysis>(F);
3039 auto *LI = AM->getCachedResult<LoopAnalysis>(F);
3041 // FIXME: The AliasAnalysis is not yet supported in the new pass manager
3042 if (!combineInstructionsOverFunction(F, Worklist, nullptr, AC, TLI, DT, LI))
3043 // No changes, all analyses are preserved.
3044 return PreservedAnalyses::all();
3046 // Mark all the analyses that instcombine updates as preserved.
3047 // FIXME: Need a way to preserve CFG analyses here!
3048 PreservedAnalyses PA;
3049 PA.preserve<DominatorTreeAnalysis>();
3054 /// \brief The legacy pass manager's instcombine pass.
3056 /// This is a basic whole-function wrapper around the instcombine utility. It
3057 /// will try to combine all instructions in the function.
3058 class InstructionCombiningPass : public FunctionPass {
3059 InstCombineWorklist Worklist;
3062 static char ID; // Pass identification, replacement for typeid
3064 InstructionCombiningPass() : FunctionPass(ID) {
3065 initializeInstructionCombiningPassPass(*PassRegistry::getPassRegistry());
3068 void getAnalysisUsage(AnalysisUsage &AU) const override;
3069 bool runOnFunction(Function &F) override;
3073 void InstructionCombiningPass::getAnalysisUsage(AnalysisUsage &AU) const {
3074 AU.setPreservesCFG();
3075 AU.addRequired<AAResultsWrapperPass>();
3076 AU.addRequired<AssumptionCacheTracker>();
3077 AU.addRequired<TargetLibraryInfoWrapperPass>();
3078 AU.addRequired<DominatorTreeWrapperPass>();
3079 AU.addPreserved<DominatorTreeWrapperPass>();
3080 AU.addPreserved<GlobalsAAWrapperPass>();
3083 bool InstructionCombiningPass::runOnFunction(Function &F) {
3084 if (skipOptnoneFunction(F))
3087 // Required analyses.
3088 auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
3089 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
3090 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
3091 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
3093 // Optional analyses.
3094 auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>();
3095 auto *LI = LIWP ? &LIWP->getLoopInfo() : nullptr;
3097 return combineInstructionsOverFunction(F, Worklist, AA, AC, TLI, DT, LI);
3100 char InstructionCombiningPass::ID = 0;
3101 INITIALIZE_PASS_BEGIN(InstructionCombiningPass, "instcombine",
3102 "Combine redundant instructions", false, false)
3103 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
3104 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
3105 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
3106 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
3107 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
3108 INITIALIZE_PASS_END(InstructionCombiningPass, "instcombine",
3109 "Combine redundant instructions", false, false)
3111 // Initialization Routines
3112 void llvm::initializeInstCombine(PassRegistry &Registry) {
3113 initializeInstructionCombiningPassPass(Registry);
3116 void LLVMInitializeInstCombine(LLVMPassRegistryRef R) {
3117 initializeInstructionCombiningPassPass(*unwrap(R));
3120 FunctionPass *llvm::createInstructionCombiningPass() {
3121 return new InstructionCombiningPass();