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/AliasAnalysis.h"
43 #include "llvm/Analysis/AssumptionCache.h"
44 #include "llvm/Analysis/BasicAliasAnalysis.h"
45 #include "llvm/Analysis/CFG.h"
46 #include "llvm/Analysis/ConstantFolding.h"
47 #include "llvm/Analysis/EHPersonalities.h"
48 #include "llvm/Analysis/GlobalsModRef.h"
49 #include "llvm/Analysis/InstructionSimplify.h"
50 #include "llvm/Analysis/LoopInfo.h"
51 #include "llvm/Analysis/MemoryBuiltins.h"
52 #include "llvm/Analysis/TargetLibraryInfo.h"
53 #include "llvm/Analysis/ValueTracking.h"
54 #include "llvm/IR/CFG.h"
55 #include "llvm/IR/DataLayout.h"
56 #include "llvm/IR/Dominators.h"
57 #include "llvm/IR/GetElementPtrTypeIterator.h"
58 #include "llvm/IR/IntrinsicInst.h"
59 #include "llvm/IR/PatternMatch.h"
60 #include "llvm/IR/ValueHandle.h"
61 #include "llvm/Support/CommandLine.h"
62 #include "llvm/Support/Debug.h"
63 #include "llvm/Support/raw_ostream.h"
64 #include "llvm/Transforms/Scalar.h"
65 #include "llvm/Transforms/Utils/Local.h"
69 using namespace llvm::PatternMatch;
71 #define DEBUG_TYPE "instcombine"
73 STATISTIC(NumCombined , "Number of insts combined");
74 STATISTIC(NumConstProp, "Number of constant folds");
75 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
76 STATISTIC(NumSunkInst , "Number of instructions sunk");
77 STATISTIC(NumExpand, "Number of expansions");
78 STATISTIC(NumFactor , "Number of factorizations");
79 STATISTIC(NumReassoc , "Number of reassociations");
82 EnableExpensiveCombines("expensive-combines",
83 cl::desc("Enable expensive instruction combines"));
85 Value *InstCombiner::EmitGEPOffset(User *GEP) {
86 return llvm::EmitGEPOffset(Builder, DL, GEP);
89 /// Return true if it is desirable to convert an integer computation from a
90 /// given bit width to a new bit width.
91 /// We don't want to convert from a legal to an illegal type for example or from
92 /// a smaller to a larger illegal type.
93 bool InstCombiner::ShouldChangeType(unsigned FromWidth,
94 unsigned ToWidth) const {
95 bool FromLegal = DL.isLegalInteger(FromWidth);
96 bool ToLegal = DL.isLegalInteger(ToWidth);
98 // If this is a legal integer from type, and the result would be an illegal
99 // type, don't do the transformation.
100 if (FromLegal && !ToLegal)
103 // Otherwise, if both are illegal, do not increase the size of the result. We
104 // do allow things like i160 -> i64, but not i64 -> i160.
105 if (!FromLegal && !ToLegal && ToWidth > FromWidth)
111 /// Return true if it is desirable to convert a computation from 'From' to 'To'.
112 /// We don't want to convert from a legal to an illegal type for example or from
113 /// a smaller to a larger illegal type.
114 bool InstCombiner::ShouldChangeType(Type *From, Type *To) const {
115 assert(From->isIntegerTy() && To->isIntegerTy());
117 unsigned FromWidth = From->getPrimitiveSizeInBits();
118 unsigned ToWidth = To->getPrimitiveSizeInBits();
119 return ShouldChangeType(FromWidth, ToWidth);
122 // Return true, if No Signed Wrap should be maintained for I.
123 // The No Signed Wrap flag can be kept if the operation "B (I.getOpcode) C",
124 // where both B and C should be ConstantInts, results in a constant that does
125 // not overflow. This function only handles the Add and Sub opcodes. For
126 // all other opcodes, the function conservatively returns false.
127 static bool MaintainNoSignedWrap(BinaryOperator &I, Value *B, Value *C) {
128 OverflowingBinaryOperator *OBO = dyn_cast<OverflowingBinaryOperator>(&I);
129 if (!OBO || !OBO->hasNoSignedWrap())
132 // We reason about Add and Sub Only.
133 Instruction::BinaryOps Opcode = I.getOpcode();
134 if (Opcode != Instruction::Add && Opcode != Instruction::Sub)
137 const APInt *BVal, *CVal;
138 if (!match(B, m_APInt(BVal)) || !match(C, m_APInt(CVal)))
141 bool Overflow = false;
142 if (Opcode == Instruction::Add)
143 BVal->sadd_ov(*CVal, Overflow);
145 BVal->ssub_ov(*CVal, Overflow);
150 /// Conservatively clears subclassOptionalData after a reassociation or
151 /// commutation. We preserve fast-math flags when applicable as they can be
153 static void ClearSubclassDataAfterReassociation(BinaryOperator &I) {
154 FPMathOperator *FPMO = dyn_cast<FPMathOperator>(&I);
156 I.clearSubclassOptionalData();
160 FastMathFlags FMF = I.getFastMathFlags();
161 I.clearSubclassOptionalData();
162 I.setFastMathFlags(FMF);
165 /// Combine constant operands of associative operations either before or after a
166 /// cast to eliminate one of the associative operations:
167 /// (op (cast (op X, C2)), C1) --> (cast (op X, op (C1, C2)))
168 /// (op (cast (op X, C2)), C1) --> (op (cast X), op (C1, C2))
169 static bool simplifyAssocCastAssoc(BinaryOperator *BinOp1) {
170 auto *Cast = dyn_cast<CastInst>(BinOp1->getOperand(0));
171 if (!Cast || !Cast->hasOneUse())
174 // TODO: Enhance logic for other casts and remove this check.
175 auto CastOpcode = Cast->getOpcode();
176 if (CastOpcode != Instruction::ZExt)
179 // TODO: Enhance logic for other BinOps and remove this check.
180 auto AssocOpcode = BinOp1->getOpcode();
181 if (AssocOpcode != Instruction::Xor && AssocOpcode != Instruction::And &&
182 AssocOpcode != Instruction::Or)
185 auto *BinOp2 = dyn_cast<BinaryOperator>(Cast->getOperand(0));
186 if (!BinOp2 || !BinOp2->hasOneUse() || BinOp2->getOpcode() != AssocOpcode)
190 if (!match(BinOp1->getOperand(1), m_Constant(C1)) ||
191 !match(BinOp2->getOperand(1), m_Constant(C2)))
194 // TODO: This assumes a zext cast.
195 // Eg, if it was a trunc, we'd cast C1 to the source type because casting C2
196 // to the destination type might lose bits.
198 // Fold the constants together in the destination type:
199 // (op (cast (op X, C2)), C1) --> (op (cast X), FoldedC)
200 Type *DestTy = C1->getType();
201 Constant *CastC2 = ConstantExpr::getCast(CastOpcode, C2, DestTy);
202 Constant *FoldedC = ConstantExpr::get(AssocOpcode, C1, CastC2);
203 Cast->setOperand(0, BinOp2->getOperand(0));
204 BinOp1->setOperand(1, FoldedC);
208 /// This performs a few simplifications for operators that are associative or
211 /// Commutative operators:
213 /// 1. Order operands such that they are listed from right (least complex) to
214 /// left (most complex). This puts constants before unary operators before
215 /// binary operators.
217 /// Associative operators:
219 /// 2. Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
220 /// 3. Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
222 /// Associative and commutative operators:
224 /// 4. Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
225 /// 5. Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
226 /// 6. Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
227 /// if C1 and C2 are constants.
228 bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) {
229 Instruction::BinaryOps Opcode = I.getOpcode();
230 bool Changed = false;
233 // Order operands such that they are listed from right (least complex) to
234 // left (most complex). This puts constants before unary operators before
236 if (I.isCommutative() && getComplexity(I.getOperand(0)) <
237 getComplexity(I.getOperand(1)))
238 Changed = !I.swapOperands();
240 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
241 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
243 if (I.isAssociative()) {
244 // Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
245 if (Op0 && Op0->getOpcode() == Opcode) {
246 Value *A = Op0->getOperand(0);
247 Value *B = Op0->getOperand(1);
248 Value *C = I.getOperand(1);
250 // Does "B op C" simplify?
251 if (Value *V = SimplifyBinOp(Opcode, B, C, DL)) {
252 // It simplifies to V. Form "A op V".
255 // Conservatively clear the optional flags, since they may not be
256 // preserved by the reassociation.
257 if (MaintainNoSignedWrap(I, B, C) &&
258 (!Op0 || (isa<BinaryOperator>(Op0) && Op0->hasNoSignedWrap()))) {
259 // Note: this is only valid because SimplifyBinOp doesn't look at
260 // the operands to Op0.
261 I.clearSubclassOptionalData();
262 I.setHasNoSignedWrap(true);
264 ClearSubclassDataAfterReassociation(I);
273 // Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
274 if (Op1 && Op1->getOpcode() == Opcode) {
275 Value *A = I.getOperand(0);
276 Value *B = Op1->getOperand(0);
277 Value *C = Op1->getOperand(1);
279 // Does "A op B" simplify?
280 if (Value *V = SimplifyBinOp(Opcode, A, B, DL)) {
281 // It simplifies to V. Form "V op C".
284 // Conservatively clear the optional flags, since they may not be
285 // preserved by the reassociation.
286 ClearSubclassDataAfterReassociation(I);
294 if (I.isAssociative() && I.isCommutative()) {
295 if (simplifyAssocCastAssoc(&I)) {
301 // Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
302 if (Op0 && Op0->getOpcode() == Opcode) {
303 Value *A = Op0->getOperand(0);
304 Value *B = Op0->getOperand(1);
305 Value *C = I.getOperand(1);
307 // Does "C op A" simplify?
308 if (Value *V = SimplifyBinOp(Opcode, C, A, DL)) {
309 // It simplifies to V. Form "V op B".
312 // Conservatively clear the optional flags, since they may not be
313 // preserved by the reassociation.
314 ClearSubclassDataAfterReassociation(I);
321 // Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
322 if (Op1 && Op1->getOpcode() == Opcode) {
323 Value *A = I.getOperand(0);
324 Value *B = Op1->getOperand(0);
325 Value *C = Op1->getOperand(1);
327 // Does "C op A" simplify?
328 if (Value *V = SimplifyBinOp(Opcode, C, A, DL)) {
329 // It simplifies to V. Form "B op V".
332 // Conservatively clear the optional flags, since they may not be
333 // preserved by the reassociation.
334 ClearSubclassDataAfterReassociation(I);
341 // Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
342 // if C1 and C2 are constants.
344 Op0->getOpcode() == Opcode && Op1->getOpcode() == Opcode &&
345 isa<Constant>(Op0->getOperand(1)) &&
346 isa<Constant>(Op1->getOperand(1)) &&
347 Op0->hasOneUse() && Op1->hasOneUse()) {
348 Value *A = Op0->getOperand(0);
349 Constant *C1 = cast<Constant>(Op0->getOperand(1));
350 Value *B = Op1->getOperand(0);
351 Constant *C2 = cast<Constant>(Op1->getOperand(1));
353 Constant *Folded = ConstantExpr::get(Opcode, C1, C2);
354 BinaryOperator *New = BinaryOperator::Create(Opcode, A, B);
355 if (isa<FPMathOperator>(New)) {
356 FastMathFlags Flags = I.getFastMathFlags();
357 Flags &= Op0->getFastMathFlags();
358 Flags &= Op1->getFastMathFlags();
359 New->setFastMathFlags(Flags);
361 InsertNewInstWith(New, I);
363 I.setOperand(0, New);
364 I.setOperand(1, Folded);
365 // Conservatively clear the optional flags, since they may not be
366 // preserved by the reassociation.
367 ClearSubclassDataAfterReassociation(I);
374 // No further simplifications.
379 /// Return whether "X LOp (Y ROp Z)" is always equal to
380 /// "(X LOp Y) ROp (X LOp Z)".
381 static bool LeftDistributesOverRight(Instruction::BinaryOps LOp,
382 Instruction::BinaryOps ROp) {
387 case Instruction::And:
388 // And distributes over Or and Xor.
392 case Instruction::Or:
393 case Instruction::Xor:
397 case Instruction::Mul:
398 // Multiplication distributes over addition and subtraction.
402 case Instruction::Add:
403 case Instruction::Sub:
407 case Instruction::Or:
408 // Or distributes over And.
412 case Instruction::And:
418 /// Return whether "(X LOp Y) ROp Z" is always equal to
419 /// "(X ROp Z) LOp (Y ROp Z)".
420 static bool RightDistributesOverLeft(Instruction::BinaryOps LOp,
421 Instruction::BinaryOps ROp) {
422 if (Instruction::isCommutative(ROp))
423 return LeftDistributesOverRight(ROp, LOp);
428 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
429 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
430 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
431 case Instruction::And:
432 case Instruction::Or:
433 case Instruction::Xor:
437 case Instruction::Shl:
438 case Instruction::LShr:
439 case Instruction::AShr:
443 // TODO: It would be nice to handle division, aka "(X + Y)/Z = X/Z + Y/Z",
444 // but this requires knowing that the addition does not overflow and other
449 /// This function returns identity value for given opcode, which can be used to
450 /// factor patterns like (X * 2) + X ==> (X * 2) + (X * 1) ==> X * (2 + 1).
451 static Value *getIdentityValue(Instruction::BinaryOps OpCode, Value *V) {
452 if (isa<Constant>(V))
455 if (OpCode == Instruction::Mul)
456 return ConstantInt::get(V->getType(), 1);
458 // TODO: We can handle other cases e.g. Instruction::And, Instruction::Or etc.
463 /// This function factors binary ops which can be combined using distributive
464 /// laws. This function tries to transform 'Op' based TopLevelOpcode to enable
465 /// factorization e.g for ADD(SHL(X , 2), MUL(X, 5)), When this function called
466 /// with TopLevelOpcode == Instruction::Add and Op = SHL(X, 2), transforms
467 /// SHL(X, 2) to MUL(X, 4) i.e. returns Instruction::Mul with LHS set to 'X' and
469 static Instruction::BinaryOps
470 getBinOpsForFactorization(Instruction::BinaryOps TopLevelOpcode,
471 BinaryOperator *Op, Value *&LHS, Value *&RHS) {
473 return Instruction::BinaryOpsEnd;
475 LHS = Op->getOperand(0);
476 RHS = Op->getOperand(1);
478 switch (TopLevelOpcode) {
480 return Op->getOpcode();
482 case Instruction::Add:
483 case Instruction::Sub:
484 if (Op->getOpcode() == Instruction::Shl) {
485 if (Constant *CST = dyn_cast<Constant>(Op->getOperand(1))) {
486 // The multiplier is really 1 << CST.
487 RHS = ConstantExpr::getShl(ConstantInt::get(Op->getType(), 1), CST);
488 return Instruction::Mul;
491 return Op->getOpcode();
494 // TODO: We can add other conversions e.g. shr => div etc.
497 /// This tries to simplify binary operations by factorizing out common terms
498 /// (e. g. "(A*B)+(A*C)" -> "A*(B+C)").
499 static Value *tryFactorization(InstCombiner::BuilderTy *Builder,
500 const DataLayout &DL, BinaryOperator &I,
501 Instruction::BinaryOps InnerOpcode, Value *A,
502 Value *B, Value *C, Value *D) {
504 // If any of A, B, C, D are null, we can not factor I, return early.
505 // Checking A and C should be enough.
506 if (!A || !C || !B || !D)
510 Value *SimplifiedInst = nullptr;
511 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
512 Instruction::BinaryOps TopLevelOpcode = I.getOpcode();
514 // Does "X op' Y" always equal "Y op' X"?
515 bool InnerCommutative = Instruction::isCommutative(InnerOpcode);
517 // Does "X op' (Y op Z)" always equal "(X op' Y) op (X op' Z)"?
518 if (LeftDistributesOverRight(InnerOpcode, TopLevelOpcode))
519 // Does the instruction have the form "(A op' B) op (A op' D)" or, in the
520 // commutative case, "(A op' B) op (C op' A)"?
521 if (A == C || (InnerCommutative && A == D)) {
524 // Consider forming "A op' (B op D)".
525 // If "B op D" simplifies then it can be formed with no cost.
526 V = SimplifyBinOp(TopLevelOpcode, B, D, DL);
527 // If "B op D" doesn't simplify then only go on if both of the existing
528 // operations "A op' B" and "C op' D" will be zapped as no longer used.
529 if (!V && LHS->hasOneUse() && RHS->hasOneUse())
530 V = Builder->CreateBinOp(TopLevelOpcode, B, D, RHS->getName());
532 SimplifiedInst = Builder->CreateBinOp(InnerOpcode, A, V);
536 // Does "(X op Y) op' Z" always equal "(X op' Z) op (Y op' Z)"?
537 if (!SimplifiedInst && RightDistributesOverLeft(TopLevelOpcode, InnerOpcode))
538 // Does the instruction have the form "(A op' B) op (C op' B)" or, in the
539 // commutative case, "(A op' B) op (B op' D)"?
540 if (B == D || (InnerCommutative && B == C)) {
543 // Consider forming "(A op C) op' B".
544 // If "A op C" simplifies then it can be formed with no cost.
545 V = SimplifyBinOp(TopLevelOpcode, A, C, DL);
547 // If "A op C" doesn't simplify then only go on if both of the existing
548 // operations "A op' B" and "C op' D" will be zapped as no longer used.
549 if (!V && LHS->hasOneUse() && RHS->hasOneUse())
550 V = Builder->CreateBinOp(TopLevelOpcode, A, C, LHS->getName());
552 SimplifiedInst = Builder->CreateBinOp(InnerOpcode, V, B);
556 if (SimplifiedInst) {
558 SimplifiedInst->takeName(&I);
560 // Check if we can add NSW flag to SimplifiedInst. If so, set NSW flag.
561 // TODO: Check for NUW.
562 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(SimplifiedInst)) {
563 if (isa<OverflowingBinaryOperator>(SimplifiedInst)) {
565 if (isa<OverflowingBinaryOperator>(&I))
566 HasNSW = I.hasNoSignedWrap();
568 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS))
569 if (isa<OverflowingBinaryOperator>(Op0))
570 HasNSW &= Op0->hasNoSignedWrap();
572 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS))
573 if (isa<OverflowingBinaryOperator>(Op1))
574 HasNSW &= Op1->hasNoSignedWrap();
576 // We can propagate 'nsw' if we know that
577 // %Y = mul nsw i16 %X, C
578 // %Z = add nsw i16 %Y, %X
580 // %Z = mul nsw i16 %X, C+1
582 // iff C+1 isn't INT_MIN
584 if (TopLevelOpcode == Instruction::Add &&
585 InnerOpcode == Instruction::Mul)
586 if (match(V, m_APInt(CInt)) && !CInt->isMinSignedValue())
587 BO->setHasNoSignedWrap(HasNSW);
591 return SimplifiedInst;
594 /// This tries to simplify binary operations which some other binary operation
595 /// distributes over either by factorizing out common terms
596 /// (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this results in
597 /// simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is a win).
598 /// Returns the simplified value, or null if it didn't simplify.
599 Value *InstCombiner::SimplifyUsingDistributiveLaws(BinaryOperator &I) {
600 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
601 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
602 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
605 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
606 auto TopLevelOpcode = I.getOpcode();
607 auto LHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op0, A, B);
608 auto RHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op1, C, D);
610 // The instruction has the form "(A op' B) op (C op' D)". Try to factorize
612 if (LHSOpcode == RHSOpcode) {
613 if (Value *V = tryFactorization(Builder, DL, I, LHSOpcode, A, B, C, D))
617 // The instruction has the form "(A op' B) op (C)". Try to factorize common
619 if (Value *V = tryFactorization(Builder, DL, I, LHSOpcode, A, B, RHS,
620 getIdentityValue(LHSOpcode, RHS)))
623 // The instruction has the form "(B) op (C op' D)". Try to factorize common
625 if (Value *V = tryFactorization(Builder, DL, I, RHSOpcode, LHS,
626 getIdentityValue(RHSOpcode, LHS), C, D))
630 if (Op0 && RightDistributesOverLeft(Op0->getOpcode(), TopLevelOpcode)) {
631 // The instruction has the form "(A op' B) op C". See if expanding it out
632 // to "(A op C) op' (B op C)" results in simplifications.
633 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
634 Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op'
636 // Do "A op C" and "B op C" both simplify?
637 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, C, DL))
638 if (Value *R = SimplifyBinOp(TopLevelOpcode, B, C, DL)) {
639 // They do! Return "L op' R".
641 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
642 if ((L == A && R == B) ||
643 (Instruction::isCommutative(InnerOpcode) && L == B && R == A))
645 // Otherwise return "L op' R" if it simplifies.
646 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, DL))
648 // Otherwise, create a new instruction.
649 C = Builder->CreateBinOp(InnerOpcode, L, R);
655 if (Op1 && LeftDistributesOverRight(TopLevelOpcode, Op1->getOpcode())) {
656 // The instruction has the form "A op (B op' C)". See if expanding it out
657 // to "(A op B) op' (A op C)" results in simplifications.
658 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
659 Instruction::BinaryOps InnerOpcode = Op1->getOpcode(); // op'
661 // Do "A op B" and "A op C" both simplify?
662 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, B, DL))
663 if (Value *R = SimplifyBinOp(TopLevelOpcode, A, C, DL)) {
664 // They do! Return "L op' R".
666 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
667 if ((L == B && R == C) ||
668 (Instruction::isCommutative(InnerOpcode) && L == C && R == B))
670 // Otherwise return "L op' R" if it simplifies.
671 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, DL))
673 // Otherwise, create a new instruction.
674 A = Builder->CreateBinOp(InnerOpcode, L, R);
680 // (op (select (a, c, b)), (select (a, d, b))) -> (select (a, (op c, d), 0))
681 // (op (select (a, b, c)), (select (a, b, d))) -> (select (a, 0, (op c, d)))
682 if (auto *SI0 = dyn_cast<SelectInst>(LHS)) {
683 if (auto *SI1 = dyn_cast<SelectInst>(RHS)) {
684 if (SI0->getCondition() == SI1->getCondition()) {
686 if (Value *V = SimplifyBinOp(TopLevelOpcode, SI0->getFalseValue(),
687 SI1->getFalseValue(), DL, TLI, DT, AC))
688 SI = Builder->CreateSelect(SI0->getCondition(),
689 Builder->CreateBinOp(TopLevelOpcode,
691 SI1->getTrueValue()),
693 if (Value *V = SimplifyBinOp(TopLevelOpcode, SI0->getTrueValue(),
694 SI1->getTrueValue(), DL, TLI, DT, AC))
695 SI = Builder->CreateSelect(
696 SI0->getCondition(), V,
697 Builder->CreateBinOp(TopLevelOpcode, SI0->getFalseValue(),
698 SI1->getFalseValue()));
710 /// Given a 'sub' instruction, return the RHS of the instruction if the LHS is a
711 /// constant zero (which is the 'negate' form).
712 Value *InstCombiner::dyn_castNegVal(Value *V) const {
713 if (BinaryOperator::isNeg(V))
714 return BinaryOperator::getNegArgument(V);
716 // Constants can be considered to be negated values if they can be folded.
717 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
718 return ConstantExpr::getNeg(C);
720 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
721 if (C->getType()->getElementType()->isIntegerTy())
722 return ConstantExpr::getNeg(C);
727 /// Given a 'fsub' instruction, return the RHS of the instruction if the LHS is
728 /// a constant negative zero (which is the 'negate' form).
729 Value *InstCombiner::dyn_castFNegVal(Value *V, bool IgnoreZeroSign) const {
730 if (BinaryOperator::isFNeg(V, IgnoreZeroSign))
731 return BinaryOperator::getFNegArgument(V);
733 // Constants can be considered to be negated values if they can be folded.
734 if (ConstantFP *C = dyn_cast<ConstantFP>(V))
735 return ConstantExpr::getFNeg(C);
737 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
738 if (C->getType()->getElementType()->isFloatingPointTy())
739 return ConstantExpr::getFNeg(C);
744 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
746 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
747 return IC->Builder->CreateCast(CI->getOpcode(), SO, I.getType());
750 // Figure out if the constant is the left or the right argument.
751 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
752 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
754 if (Constant *SOC = dyn_cast<Constant>(SO)) {
756 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
757 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
760 Value *Op0 = SO, *Op1 = ConstOperand;
764 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I)) {
765 Value *RI = IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1,
766 SO->getName()+".op");
767 Instruction *FPInst = dyn_cast<Instruction>(RI);
768 if (FPInst && isa<FPMathOperator>(FPInst))
769 FPInst->copyFastMathFlags(BO);
772 if (ICmpInst *CI = dyn_cast<ICmpInst>(&I))
773 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
774 SO->getName()+".cmp");
775 if (FCmpInst *CI = dyn_cast<FCmpInst>(&I))
776 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
777 SO->getName()+".cmp");
778 llvm_unreachable("Unknown binary instruction type!");
781 /// Given an instruction with a select as one operand and a constant as the
782 /// other operand, try to fold the binary operator into the select arguments.
783 /// This also works for Cast instructions, which obviously do not have a second
785 Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) {
786 // Don't modify shared select instructions
787 if (!SI->hasOneUse()) return nullptr;
788 Value *TV = SI->getOperand(1);
789 Value *FV = SI->getOperand(2);
791 if (isa<Constant>(TV) || isa<Constant>(FV)) {
792 // Bool selects with constant operands can be folded to logical ops.
793 if (SI->getType()->isIntegerTy(1)) return nullptr;
795 // If it's a bitcast involving vectors, make sure it has the same number of
796 // elements on both sides.
797 if (BitCastInst *BC = dyn_cast<BitCastInst>(&Op)) {
798 VectorType *DestTy = dyn_cast<VectorType>(BC->getDestTy());
799 VectorType *SrcTy = dyn_cast<VectorType>(BC->getSrcTy());
801 // Verify that either both or neither are vectors.
802 if ((SrcTy == nullptr) != (DestTy == nullptr)) return nullptr;
803 // If vectors, verify that they have the same number of elements.
804 if (SrcTy && SrcTy->getNumElements() != DestTy->getNumElements())
808 // Test if a CmpInst instruction is used exclusively by a select as
809 // part of a minimum or maximum operation. If so, refrain from doing
810 // any other folding. This helps out other analyses which understand
811 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
812 // and CodeGen. And in this case, at least one of the comparison
813 // operands has at least one user besides the compare (the select),
814 // which would often largely negate the benefit of folding anyway.
815 if (auto *CI = dyn_cast<CmpInst>(SI->getCondition())) {
816 if (CI->hasOneUse()) {
817 Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1);
818 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
819 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
824 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, this);
825 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, this);
827 return SelectInst::Create(SI->getCondition(),
828 SelectTrueVal, SelectFalseVal);
833 /// Given a binary operator, cast instruction, or select which has a PHI node as
834 /// operand #0, see if we can fold the instruction into the PHI (which is only
835 /// possible if all operands to the PHI are constants).
836 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
837 PHINode *PN = cast<PHINode>(I.getOperand(0));
838 unsigned NumPHIValues = PN->getNumIncomingValues();
839 if (NumPHIValues == 0)
842 // We normally only transform phis with a single use. However, if a PHI has
843 // multiple uses and they are all the same operation, we can fold *all* of the
844 // uses into the PHI.
845 if (!PN->hasOneUse()) {
846 // Walk the use list for the instruction, comparing them to I.
847 for (User *U : PN->users()) {
848 Instruction *UI = cast<Instruction>(U);
849 if (UI != &I && !I.isIdenticalTo(UI))
852 // Otherwise, we can replace *all* users with the new PHI we form.
855 // Check to see if all of the operands of the PHI are simple constants
856 // (constantint/constantfp/undef). If there is one non-constant value,
857 // remember the BB it is in. If there is more than one or if *it* is a PHI,
858 // bail out. We don't do arbitrary constant expressions here because moving
859 // their computation can be expensive without a cost model.
860 BasicBlock *NonConstBB = nullptr;
861 for (unsigned i = 0; i != NumPHIValues; ++i) {
862 Value *InVal = PN->getIncomingValue(i);
863 if (isa<Constant>(InVal) && !isa<ConstantExpr>(InVal))
866 if (isa<PHINode>(InVal)) return nullptr; // Itself a phi.
867 if (NonConstBB) return nullptr; // More than one non-const value.
869 NonConstBB = PN->getIncomingBlock(i);
871 // If the InVal is an invoke at the end of the pred block, then we can't
872 // insert a computation after it without breaking the edge.
873 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal))
874 if (II->getParent() == NonConstBB)
877 // If the incoming non-constant value is in I's block, we will remove one
878 // instruction, but insert another equivalent one, leading to infinite
880 if (isPotentiallyReachable(I.getParent(), NonConstBB, DT, LI))
884 // If there is exactly one non-constant value, we can insert a copy of the
885 // operation in that block. However, if this is a critical edge, we would be
886 // inserting the computation on some other paths (e.g. inside a loop). Only
887 // do this if the pred block is unconditionally branching into the phi block.
888 if (NonConstBB != nullptr) {
889 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
890 if (!BI || !BI->isUnconditional()) return nullptr;
893 // Okay, we can do the transformation: create the new PHI node.
894 PHINode *NewPN = PHINode::Create(I.getType(), PN->getNumIncomingValues());
895 InsertNewInstBefore(NewPN, *PN);
898 // If we are going to have to insert a new computation, do so right before the
899 // predecessor's terminator.
901 Builder->SetInsertPoint(NonConstBB->getTerminator());
903 // Next, add all of the operands to the PHI.
904 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) {
905 // We only currently try to fold the condition of a select when it is a phi,
906 // not the true/false values.
907 Value *TrueV = SI->getTrueValue();
908 Value *FalseV = SI->getFalseValue();
909 BasicBlock *PhiTransBB = PN->getParent();
910 for (unsigned i = 0; i != NumPHIValues; ++i) {
911 BasicBlock *ThisBB = PN->getIncomingBlock(i);
912 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB);
913 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB);
914 Value *InV = nullptr;
915 // Beware of ConstantExpr: it may eventually evaluate to getNullValue,
916 // even if currently isNullValue gives false.
917 Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i));
918 if (InC && !isa<ConstantExpr>(InC))
919 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred;
921 InV = Builder->CreateSelect(PN->getIncomingValue(i),
922 TrueVInPred, FalseVInPred, "phitmp");
923 NewPN->addIncoming(InV, ThisBB);
925 } else if (CmpInst *CI = dyn_cast<CmpInst>(&I)) {
926 Constant *C = cast<Constant>(I.getOperand(1));
927 for (unsigned i = 0; i != NumPHIValues; ++i) {
928 Value *InV = nullptr;
929 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
930 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
931 else if (isa<ICmpInst>(CI))
932 InV = Builder->CreateICmp(CI->getPredicate(), PN->getIncomingValue(i),
935 InV = Builder->CreateFCmp(CI->getPredicate(), PN->getIncomingValue(i),
937 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
939 } else if (I.getNumOperands() == 2) {
940 Constant *C = cast<Constant>(I.getOperand(1));
941 for (unsigned i = 0; i != NumPHIValues; ++i) {
942 Value *InV = nullptr;
943 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
944 InV = ConstantExpr::get(I.getOpcode(), InC, C);
946 InV = Builder->CreateBinOp(cast<BinaryOperator>(I).getOpcode(),
947 PN->getIncomingValue(i), C, "phitmp");
948 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
951 CastInst *CI = cast<CastInst>(&I);
952 Type *RetTy = CI->getType();
953 for (unsigned i = 0; i != NumPHIValues; ++i) {
955 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
956 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
958 InV = Builder->CreateCast(CI->getOpcode(),
959 PN->getIncomingValue(i), I.getType(), "phitmp");
960 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
964 for (auto UI = PN->user_begin(), E = PN->user_end(); UI != E;) {
965 Instruction *User = cast<Instruction>(*UI++);
966 if (User == &I) continue;
967 replaceInstUsesWith(*User, NewPN);
968 eraseInstFromFunction(*User);
970 return replaceInstUsesWith(I, NewPN);
973 /// Given a pointer type and a constant offset, determine whether or not there
974 /// is a sequence of GEP indices into the pointed type that will land us at the
975 /// specified offset. If so, fill them into NewIndices and return the resultant
976 /// element type, otherwise return null.
977 Type *InstCombiner::FindElementAtOffset(PointerType *PtrTy, int64_t Offset,
978 SmallVectorImpl<Value *> &NewIndices) {
979 Type *Ty = PtrTy->getElementType();
983 // Start with the index over the outer type. Note that the type size
984 // might be zero (even if the offset isn't zero) if the indexed type
985 // is something like [0 x {int, int}]
986 Type *IntPtrTy = DL.getIntPtrType(PtrTy);
987 int64_t FirstIdx = 0;
988 if (int64_t TySize = DL.getTypeAllocSize(Ty)) {
989 FirstIdx = Offset/TySize;
990 Offset -= FirstIdx*TySize;
992 // Handle hosts where % returns negative instead of values [0..TySize).
998 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
1001 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
1003 // Index into the types. If we fail, set OrigBase to null.
1005 // Indexing into tail padding between struct/array elements.
1006 if (uint64_t(Offset * 8) >= DL.getTypeSizeInBits(Ty))
1009 if (StructType *STy = dyn_cast<StructType>(Ty)) {
1010 const StructLayout *SL = DL.getStructLayout(STy);
1011 assert(Offset < (int64_t)SL->getSizeInBytes() &&
1012 "Offset must stay within the indexed type");
1014 unsigned Elt = SL->getElementContainingOffset(Offset);
1015 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
1018 Offset -= SL->getElementOffset(Elt);
1019 Ty = STy->getElementType(Elt);
1020 } else if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
1021 uint64_t EltSize = DL.getTypeAllocSize(AT->getElementType());
1022 assert(EltSize && "Cannot index into a zero-sized array");
1023 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
1025 Ty = AT->getElementType();
1027 // Otherwise, we can't index into the middle of this atomic type, bail.
1035 static bool shouldMergeGEPs(GEPOperator &GEP, GEPOperator &Src) {
1036 // If this GEP has only 0 indices, it is the same pointer as
1037 // Src. If Src is not a trivial GEP too, don't combine
1039 if (GEP.hasAllZeroIndices() && !Src.hasAllZeroIndices() &&
1045 /// Return a value X such that Val = X * Scale, or null if none.
1046 /// If the multiplication is known not to overflow, then NoSignedWrap is set.
1047 Value *InstCombiner::Descale(Value *Val, APInt Scale, bool &NoSignedWrap) {
1048 assert(isa<IntegerType>(Val->getType()) && "Can only descale integers!");
1049 assert(cast<IntegerType>(Val->getType())->getBitWidth() ==
1050 Scale.getBitWidth() && "Scale not compatible with value!");
1052 // If Val is zero or Scale is one then Val = Val * Scale.
1053 if (match(Val, m_Zero()) || Scale == 1) {
1054 NoSignedWrap = true;
1058 // If Scale is zero then it does not divide Val.
1059 if (Scale.isMinValue())
1062 // Look through chains of multiplications, searching for a constant that is
1063 // divisible by Scale. For example, descaling X*(Y*(Z*4)) by a factor of 4
1064 // will find the constant factor 4 and produce X*(Y*Z). Descaling X*(Y*8) by
1065 // a factor of 4 will produce X*(Y*2). The principle of operation is to bore
1068 // Val = M1 * X || Analysis starts here and works down
1069 // M1 = M2 * Y || Doesn't descend into terms with more
1070 // M2 = Z * 4 \/ than one use
1072 // Then to modify a term at the bottom:
1075 // M1 = Z * Y || Replaced M2 with Z
1077 // Then to work back up correcting nsw flags.
1079 // Op - the term we are currently analyzing. Starts at Val then drills down.
1080 // Replaced with its descaled value before exiting from the drill down loop.
1083 // Parent - initially null, but after drilling down notes where Op came from.
1084 // In the example above, Parent is (Val, 0) when Op is M1, because M1 is the
1085 // 0'th operand of Val.
1086 std::pair<Instruction*, unsigned> Parent;
1088 // Set if the transform requires a descaling at deeper levels that doesn't
1090 bool RequireNoSignedWrap = false;
1092 // Log base 2 of the scale. Negative if not a power of 2.
1093 int32_t logScale = Scale.exactLogBase2();
1095 for (;; Op = Parent.first->getOperand(Parent.second)) { // Drill down
1097 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
1098 // If Op is a constant divisible by Scale then descale to the quotient.
1099 APInt Quotient(Scale), Remainder(Scale); // Init ensures right bitwidth.
1100 APInt::sdivrem(CI->getValue(), Scale, Quotient, Remainder);
1101 if (!Remainder.isMinValue())
1102 // Not divisible by Scale.
1104 // Replace with the quotient in the parent.
1105 Op = ConstantInt::get(CI->getType(), Quotient);
1106 NoSignedWrap = true;
1110 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op)) {
1112 if (BO->getOpcode() == Instruction::Mul) {
1114 NoSignedWrap = BO->hasNoSignedWrap();
1115 if (RequireNoSignedWrap && !NoSignedWrap)
1118 // There are three cases for multiplication: multiplication by exactly
1119 // the scale, multiplication by a constant different to the scale, and
1120 // multiplication by something else.
1121 Value *LHS = BO->getOperand(0);
1122 Value *RHS = BO->getOperand(1);
1124 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
1125 // Multiplication by a constant.
1126 if (CI->getValue() == Scale) {
1127 // Multiplication by exactly the scale, replace the multiplication
1128 // by its left-hand side in the parent.
1133 // Otherwise drill down into the constant.
1134 if (!Op->hasOneUse())
1137 Parent = std::make_pair(BO, 1);
1141 // Multiplication by something else. Drill down into the left-hand side
1142 // since that's where the reassociate pass puts the good stuff.
1143 if (!Op->hasOneUse())
1146 Parent = std::make_pair(BO, 0);
1150 if (logScale > 0 && BO->getOpcode() == Instruction::Shl &&
1151 isa<ConstantInt>(BO->getOperand(1))) {
1152 // Multiplication by a power of 2.
1153 NoSignedWrap = BO->hasNoSignedWrap();
1154 if (RequireNoSignedWrap && !NoSignedWrap)
1157 Value *LHS = BO->getOperand(0);
1158 int32_t Amt = cast<ConstantInt>(BO->getOperand(1))->
1159 getLimitedValue(Scale.getBitWidth());
1162 if (Amt == logScale) {
1163 // Multiplication by exactly the scale, replace the multiplication
1164 // by its left-hand side in the parent.
1168 if (Amt < logScale || !Op->hasOneUse())
1171 // Multiplication by more than the scale. Reduce the multiplying amount
1172 // by the scale in the parent.
1173 Parent = std::make_pair(BO, 1);
1174 Op = ConstantInt::get(BO->getType(), Amt - logScale);
1179 if (!Op->hasOneUse())
1182 if (CastInst *Cast = dyn_cast<CastInst>(Op)) {
1183 if (Cast->getOpcode() == Instruction::SExt) {
1184 // Op is sign-extended from a smaller type, descale in the smaller type.
1185 unsigned SmallSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1186 APInt SmallScale = Scale.trunc(SmallSize);
1187 // Suppose Op = sext X, and we descale X as Y * SmallScale. We want to
1188 // descale Op as (sext Y) * Scale. In order to have
1189 // sext (Y * SmallScale) = (sext Y) * Scale
1190 // some conditions need to hold however: SmallScale must sign-extend to
1191 // Scale and the multiplication Y * SmallScale should not overflow.
1192 if (SmallScale.sext(Scale.getBitWidth()) != Scale)
1193 // SmallScale does not sign-extend to Scale.
1195 assert(SmallScale.exactLogBase2() == logScale);
1196 // Require that Y * SmallScale must not overflow.
1197 RequireNoSignedWrap = true;
1199 // Drill down through the cast.
1200 Parent = std::make_pair(Cast, 0);
1205 if (Cast->getOpcode() == Instruction::Trunc) {
1206 // Op is truncated from a larger type, descale in the larger type.
1207 // Suppose Op = trunc X, and we descale X as Y * sext Scale. Then
1208 // trunc (Y * sext Scale) = (trunc Y) * Scale
1209 // always holds. However (trunc Y) * Scale may overflow even if
1210 // trunc (Y * sext Scale) does not, so nsw flags need to be cleared
1211 // from this point up in the expression (see later).
1212 if (RequireNoSignedWrap)
1215 // Drill down through the cast.
1216 unsigned LargeSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1217 Parent = std::make_pair(Cast, 0);
1218 Scale = Scale.sext(LargeSize);
1219 if (logScale + 1 == (int32_t)Cast->getType()->getPrimitiveSizeInBits())
1221 assert(Scale.exactLogBase2() == logScale);
1226 // Unsupported expression, bail out.
1230 // If Op is zero then Val = Op * Scale.
1231 if (match(Op, m_Zero())) {
1232 NoSignedWrap = true;
1236 // We know that we can successfully descale, so from here on we can safely
1237 // modify the IR. Op holds the descaled version of the deepest term in the
1238 // expression. NoSignedWrap is 'true' if multiplying Op by Scale is known
1242 // The expression only had one term.
1245 // Rewrite the parent using the descaled version of its operand.
1246 assert(Parent.first->hasOneUse() && "Drilled down when more than one use!");
1247 assert(Op != Parent.first->getOperand(Parent.second) &&
1248 "Descaling was a no-op?");
1249 Parent.first->setOperand(Parent.second, Op);
1250 Worklist.Add(Parent.first);
1252 // Now work back up the expression correcting nsw flags. The logic is based
1253 // on the following observation: if X * Y is known not to overflow as a signed
1254 // multiplication, and Y is replaced by a value Z with smaller absolute value,
1255 // then X * Z will not overflow as a signed multiplication either. As we work
1256 // our way up, having NoSignedWrap 'true' means that the descaled value at the
1257 // current level has strictly smaller absolute value than the original.
1258 Instruction *Ancestor = Parent.first;
1260 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Ancestor)) {
1261 // If the multiplication wasn't nsw then we can't say anything about the
1262 // value of the descaled multiplication, and we have to clear nsw flags
1263 // from this point on up.
1264 bool OpNoSignedWrap = BO->hasNoSignedWrap();
1265 NoSignedWrap &= OpNoSignedWrap;
1266 if (NoSignedWrap != OpNoSignedWrap) {
1267 BO->setHasNoSignedWrap(NoSignedWrap);
1268 Worklist.Add(Ancestor);
1270 } else if (Ancestor->getOpcode() == Instruction::Trunc) {
1271 // The fact that the descaled input to the trunc has smaller absolute
1272 // value than the original input doesn't tell us anything useful about
1273 // the absolute values of the truncations.
1274 NoSignedWrap = false;
1276 assert((Ancestor->getOpcode() != Instruction::SExt || NoSignedWrap) &&
1277 "Failed to keep proper track of nsw flags while drilling down?");
1279 if (Ancestor == Val)
1280 // Got to the top, all done!
1283 // Move up one level in the expression.
1284 assert(Ancestor->hasOneUse() && "Drilled down when more than one use!");
1285 Ancestor = Ancestor->user_back();
1289 /// \brief Creates node of binary operation with the same attributes as the
1290 /// specified one but with other operands.
1291 static Value *CreateBinOpAsGiven(BinaryOperator &Inst, Value *LHS, Value *RHS,
1292 InstCombiner::BuilderTy *B) {
1293 Value *BO = B->CreateBinOp(Inst.getOpcode(), LHS, RHS);
1294 // If LHS and RHS are constant, BO won't be a binary operator.
1295 if (BinaryOperator *NewBO = dyn_cast<BinaryOperator>(BO))
1296 NewBO->copyIRFlags(&Inst);
1300 /// \brief Makes transformation of binary operation specific for vector types.
1301 /// \param Inst Binary operator to transform.
1302 /// \return Pointer to node that must replace the original binary operator, or
1303 /// null pointer if no transformation was made.
1304 Value *InstCombiner::SimplifyVectorOp(BinaryOperator &Inst) {
1305 if (!Inst.getType()->isVectorTy()) return nullptr;
1307 // It may not be safe to reorder shuffles and things like div, urem, etc.
1308 // because we may trap when executing those ops on unknown vector elements.
1310 if (!isSafeToSpeculativelyExecute(&Inst))
1313 unsigned VWidth = cast<VectorType>(Inst.getType())->getNumElements();
1314 Value *LHS = Inst.getOperand(0), *RHS = Inst.getOperand(1);
1315 assert(cast<VectorType>(LHS->getType())->getNumElements() == VWidth);
1316 assert(cast<VectorType>(RHS->getType())->getNumElements() == VWidth);
1318 // If both arguments of binary operation are shuffles, which use the same
1319 // mask and shuffle within a single vector, it is worthwhile to move the
1320 // shuffle after binary operation:
1321 // Op(shuffle(v1, m), shuffle(v2, m)) -> shuffle(Op(v1, v2), m)
1322 if (isa<ShuffleVectorInst>(LHS) && isa<ShuffleVectorInst>(RHS)) {
1323 ShuffleVectorInst *LShuf = cast<ShuffleVectorInst>(LHS);
1324 ShuffleVectorInst *RShuf = cast<ShuffleVectorInst>(RHS);
1325 if (isa<UndefValue>(LShuf->getOperand(1)) &&
1326 isa<UndefValue>(RShuf->getOperand(1)) &&
1327 LShuf->getOperand(0)->getType() == RShuf->getOperand(0)->getType() &&
1328 LShuf->getMask() == RShuf->getMask()) {
1329 Value *NewBO = CreateBinOpAsGiven(Inst, LShuf->getOperand(0),
1330 RShuf->getOperand(0), Builder);
1331 return Builder->CreateShuffleVector(NewBO,
1332 UndefValue::get(NewBO->getType()), LShuf->getMask());
1336 // If one argument is a shuffle within one vector, the other is a constant,
1337 // try moving the shuffle after the binary operation.
1338 ShuffleVectorInst *Shuffle = nullptr;
1339 Constant *C1 = nullptr;
1340 if (isa<ShuffleVectorInst>(LHS)) Shuffle = cast<ShuffleVectorInst>(LHS);
1341 if (isa<ShuffleVectorInst>(RHS)) Shuffle = cast<ShuffleVectorInst>(RHS);
1342 if (isa<Constant>(LHS)) C1 = cast<Constant>(LHS);
1343 if (isa<Constant>(RHS)) C1 = cast<Constant>(RHS);
1344 if (Shuffle && C1 &&
1345 (isa<ConstantVector>(C1) || isa<ConstantDataVector>(C1)) &&
1346 isa<UndefValue>(Shuffle->getOperand(1)) &&
1347 Shuffle->getType() == Shuffle->getOperand(0)->getType()) {
1348 SmallVector<int, 16> ShMask = Shuffle->getShuffleMask();
1349 // Find constant C2 that has property:
1350 // shuffle(C2, ShMask) = C1
1351 // If such constant does not exist (example: ShMask=<0,0> and C1=<1,2>)
1352 // reorder is not possible.
1353 SmallVector<Constant*, 16> C2M(VWidth,
1354 UndefValue::get(C1->getType()->getScalarType()));
1355 bool MayChange = true;
1356 for (unsigned I = 0; I < VWidth; ++I) {
1357 if (ShMask[I] >= 0) {
1358 assert(ShMask[I] < (int)VWidth);
1359 if (!isa<UndefValue>(C2M[ShMask[I]])) {
1363 C2M[ShMask[I]] = C1->getAggregateElement(I);
1367 Constant *C2 = ConstantVector::get(C2M);
1368 Value *NewLHS = isa<Constant>(LHS) ? C2 : Shuffle->getOperand(0);
1369 Value *NewRHS = isa<Constant>(LHS) ? Shuffle->getOperand(0) : C2;
1370 Value *NewBO = CreateBinOpAsGiven(Inst, NewLHS, NewRHS, Builder);
1371 return Builder->CreateShuffleVector(NewBO,
1372 UndefValue::get(Inst.getType()), Shuffle->getMask());
1379 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
1380 SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end());
1382 if (Value *V = SimplifyGEPInst(GEP.getSourceElementType(), Ops, DL, TLI, DT, AC))
1383 return replaceInstUsesWith(GEP, V);
1385 Value *PtrOp = GEP.getOperand(0);
1387 // Eliminate unneeded casts for indices, and replace indices which displace
1388 // by multiples of a zero size type with zero.
1389 bool MadeChange = false;
1391 DL.getIntPtrType(GEP.getPointerOperandType()->getScalarType());
1393 gep_type_iterator GTI = gep_type_begin(GEP);
1394 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end(); I != E;
1396 // Skip indices into struct types.
1397 if (isa<StructType>(*GTI))
1400 // Index type should have the same width as IntPtr
1401 Type *IndexTy = (*I)->getType();
1402 Type *NewIndexType = IndexTy->isVectorTy() ?
1403 VectorType::get(IntPtrTy, IndexTy->getVectorNumElements()) : IntPtrTy;
1405 // If the element type has zero size then any index over it is equivalent
1406 // to an index of zero, so replace it with zero if it is not zero already.
1407 Type *EltTy = GTI.getIndexedType();
1408 if (EltTy->isSized() && DL.getTypeAllocSize(EltTy) == 0)
1409 if (!isa<Constant>(*I) || !cast<Constant>(*I)->isNullValue()) {
1410 *I = Constant::getNullValue(NewIndexType);
1414 if (IndexTy != NewIndexType) {
1415 // If we are using a wider index than needed for this platform, shrink
1416 // it to what we need. If narrower, sign-extend it to what we need.
1417 // This explicit cast can make subsequent optimizations more obvious.
1418 *I = Builder->CreateIntCast(*I, NewIndexType, true);
1425 // Check to see if the inputs to the PHI node are getelementptr instructions.
1426 if (PHINode *PN = dyn_cast<PHINode>(PtrOp)) {
1427 GetElementPtrInst *Op1 = dyn_cast<GetElementPtrInst>(PN->getOperand(0));
1431 // Don't fold a GEP into itself through a PHI node. This can only happen
1432 // through the back-edge of a loop. Folding a GEP into itself means that
1433 // the value of the previous iteration needs to be stored in the meantime,
1434 // thus requiring an additional register variable to be live, but not
1435 // actually achieving anything (the GEP still needs to be executed once per
1442 for (auto I = PN->op_begin()+1, E = PN->op_end(); I !=E; ++I) {
1443 GetElementPtrInst *Op2 = dyn_cast<GetElementPtrInst>(*I);
1444 if (!Op2 || Op1->getNumOperands() != Op2->getNumOperands())
1447 // As for Op1 above, don't try to fold a GEP into itself.
1451 // Keep track of the type as we walk the GEP.
1452 Type *CurTy = nullptr;
1454 for (unsigned J = 0, F = Op1->getNumOperands(); J != F; ++J) {
1455 if (Op1->getOperand(J)->getType() != Op2->getOperand(J)->getType())
1458 if (Op1->getOperand(J) != Op2->getOperand(J)) {
1460 // We have not seen any differences yet in the GEPs feeding the
1461 // PHI yet, so we record this one if it is allowed to be a
1464 // The first two arguments can vary for any GEP, the rest have to be
1465 // static for struct slots
1466 if (J > 1 && CurTy->isStructTy())
1471 // The GEP is different by more than one input. While this could be
1472 // extended to support GEPs that vary by more than one variable it
1473 // doesn't make sense since it greatly increases the complexity and
1474 // would result in an R+R+R addressing mode which no backend
1475 // directly supports and would need to be broken into several
1476 // simpler instructions anyway.
1481 // Sink down a layer of the type for the next iteration.
1484 CurTy = Op1->getSourceElementType();
1485 } else if (CompositeType *CT = dyn_cast<CompositeType>(CurTy)) {
1486 CurTy = CT->getTypeAtIndex(Op1->getOperand(J));
1494 // If not all GEPs are identical we'll have to create a new PHI node.
1495 // Check that the old PHI node has only one use so that it will get
1497 if (DI != -1 && !PN->hasOneUse())
1500 GetElementPtrInst *NewGEP = cast<GetElementPtrInst>(Op1->clone());
1502 // All the GEPs feeding the PHI are identical. Clone one down into our
1503 // BB so that it can be merged with the current GEP.
1504 GEP.getParent()->getInstList().insert(
1505 GEP.getParent()->getFirstInsertionPt(), NewGEP);
1507 // All the GEPs feeding the PHI differ at a single offset. Clone a GEP
1508 // into the current block so it can be merged, and create a new PHI to
1512 IRBuilderBase::InsertPointGuard Guard(*Builder);
1513 Builder->SetInsertPoint(PN);
1514 NewPN = Builder->CreatePHI(Op1->getOperand(DI)->getType(),
1515 PN->getNumOperands());
1518 for (auto &I : PN->operands())
1519 NewPN->addIncoming(cast<GEPOperator>(I)->getOperand(DI),
1520 PN->getIncomingBlock(I));
1522 NewGEP->setOperand(DI, NewPN);
1523 GEP.getParent()->getInstList().insert(
1524 GEP.getParent()->getFirstInsertionPt(), NewGEP);
1525 NewGEP->setOperand(DI, NewPN);
1528 GEP.setOperand(0, NewGEP);
1532 // Combine Indices - If the source pointer to this getelementptr instruction
1533 // is a getelementptr instruction, combine the indices of the two
1534 // getelementptr instructions into a single instruction.
1536 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) {
1537 if (!shouldMergeGEPs(*cast<GEPOperator>(&GEP), *Src))
1540 // Note that if our source is a gep chain itself then we wait for that
1541 // chain to be resolved before we perform this transformation. This
1542 // avoids us creating a TON of code in some cases.
1543 if (GEPOperator *SrcGEP =
1544 dyn_cast<GEPOperator>(Src->getOperand(0)))
1545 if (SrcGEP->getNumOperands() == 2 && shouldMergeGEPs(*Src, *SrcGEP))
1546 return nullptr; // Wait until our source is folded to completion.
1548 SmallVector<Value*, 8> Indices;
1550 // Find out whether the last index in the source GEP is a sequential idx.
1551 bool EndsWithSequential = false;
1552 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src);
1554 EndsWithSequential = !(*I)->isStructTy();
1556 // Can we combine the two pointer arithmetics offsets?
1557 if (EndsWithSequential) {
1558 // Replace: gep (gep %P, long B), long A, ...
1559 // With: T = long A+B; gep %P, T, ...
1562 Value *SO1 = Src->getOperand(Src->getNumOperands()-1);
1563 Value *GO1 = GEP.getOperand(1);
1564 if (SO1 == Constant::getNullValue(SO1->getType())) {
1566 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
1569 // If they aren't the same type, then the input hasn't been processed
1570 // by the loop above yet (which canonicalizes sequential index types to
1571 // intptr_t). Just avoid transforming this until the input has been
1573 if (SO1->getType() != GO1->getType())
1575 // Only do the combine when GO1 and SO1 are both constants. Only in
1576 // this case, we are sure the cost after the merge is never more than
1577 // that before the merge.
1578 if (!isa<Constant>(GO1) || !isa<Constant>(SO1))
1580 Sum = Builder->CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
1583 // Update the GEP in place if possible.
1584 if (Src->getNumOperands() == 2) {
1585 GEP.setOperand(0, Src->getOperand(0));
1586 GEP.setOperand(1, Sum);
1589 Indices.append(Src->op_begin()+1, Src->op_end()-1);
1590 Indices.push_back(Sum);
1591 Indices.append(GEP.op_begin()+2, GEP.op_end());
1592 } else if (isa<Constant>(*GEP.idx_begin()) &&
1593 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
1594 Src->getNumOperands() != 1) {
1595 // Otherwise we can do the fold if the first index of the GEP is a zero
1596 Indices.append(Src->op_begin()+1, Src->op_end());
1597 Indices.append(GEP.idx_begin()+1, GEP.idx_end());
1600 if (!Indices.empty())
1601 return GEP.isInBounds() && Src->isInBounds()
1602 ? GetElementPtrInst::CreateInBounds(
1603 Src->getSourceElementType(), Src->getOperand(0), Indices,
1605 : GetElementPtrInst::Create(Src->getSourceElementType(),
1606 Src->getOperand(0), Indices,
1610 if (GEP.getNumIndices() == 1) {
1611 unsigned AS = GEP.getPointerAddressSpace();
1612 if (GEP.getOperand(1)->getType()->getScalarSizeInBits() ==
1613 DL.getPointerSizeInBits(AS)) {
1614 Type *Ty = GEP.getSourceElementType();
1615 uint64_t TyAllocSize = DL.getTypeAllocSize(Ty);
1617 bool Matched = false;
1620 if (TyAllocSize == 1) {
1621 V = GEP.getOperand(1);
1623 } else if (match(GEP.getOperand(1),
1624 m_AShr(m_Value(V), m_ConstantInt(C)))) {
1625 if (TyAllocSize == 1ULL << C)
1627 } else if (match(GEP.getOperand(1),
1628 m_SDiv(m_Value(V), m_ConstantInt(C)))) {
1629 if (TyAllocSize == C)
1634 // Canonicalize (gep i8* X, -(ptrtoint Y))
1635 // to (inttoptr (sub (ptrtoint X), (ptrtoint Y)))
1636 // The GEP pattern is emitted by the SCEV expander for certain kinds of
1637 // pointer arithmetic.
1638 if (match(V, m_Neg(m_PtrToInt(m_Value())))) {
1639 Operator *Index = cast<Operator>(V);
1640 Value *PtrToInt = Builder->CreatePtrToInt(PtrOp, Index->getType());
1641 Value *NewSub = Builder->CreateSub(PtrToInt, Index->getOperand(1));
1642 return CastInst::Create(Instruction::IntToPtr, NewSub, GEP.getType());
1644 // Canonicalize (gep i8* X, (ptrtoint Y)-(ptrtoint X))
1647 if (match(V, m_Sub(m_PtrToInt(m_Value(Y)),
1648 m_PtrToInt(m_Specific(GEP.getOperand(0)))))) {
1649 return CastInst::CreatePointerBitCastOrAddrSpaceCast(Y,
1656 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0).
1657 Value *StrippedPtr = PtrOp->stripPointerCasts();
1658 PointerType *StrippedPtrTy = dyn_cast<PointerType>(StrippedPtr->getType());
1660 // We do not handle pointer-vector geps here.
1664 if (StrippedPtr != PtrOp) {
1665 bool HasZeroPointerIndex = false;
1666 if (ConstantInt *C = dyn_cast<ConstantInt>(GEP.getOperand(1)))
1667 HasZeroPointerIndex = C->isZero();
1669 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
1670 // into : GEP [10 x i8]* X, i32 0, ...
1672 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
1673 // into : GEP i8* X, ...
1675 // This occurs when the program declares an array extern like "int X[];"
1676 if (HasZeroPointerIndex) {
1677 if (ArrayType *CATy =
1678 dyn_cast<ArrayType>(GEP.getSourceElementType())) {
1679 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
1680 if (CATy->getElementType() == StrippedPtrTy->getElementType()) {
1681 // -> GEP i8* X, ...
1682 SmallVector<Value*, 8> Idx(GEP.idx_begin()+1, GEP.idx_end());
1683 GetElementPtrInst *Res = GetElementPtrInst::Create(
1684 StrippedPtrTy->getElementType(), StrippedPtr, Idx, GEP.getName());
1685 Res->setIsInBounds(GEP.isInBounds());
1686 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace())
1688 // Insert Res, and create an addrspacecast.
1690 // GEP (addrspacecast i8 addrspace(1)* X to [0 x i8]*), i32 0, ...
1692 // %0 = GEP i8 addrspace(1)* X, ...
1693 // addrspacecast i8 addrspace(1)* %0 to i8*
1694 return new AddrSpaceCastInst(Builder->Insert(Res), GEP.getType());
1697 if (ArrayType *XATy =
1698 dyn_cast<ArrayType>(StrippedPtrTy->getElementType())){
1699 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
1700 if (CATy->getElementType() == XATy->getElementType()) {
1701 // -> GEP [10 x i8]* X, i32 0, ...
1702 // At this point, we know that the cast source type is a pointer
1703 // to an array of the same type as the destination pointer
1704 // array. Because the array type is never stepped over (there
1705 // is a leading zero) we can fold the cast into this GEP.
1706 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace()) {
1707 GEP.setOperand(0, StrippedPtr);
1708 GEP.setSourceElementType(XATy);
1711 // Cannot replace the base pointer directly because StrippedPtr's
1712 // address space is different. Instead, create a new GEP followed by
1713 // an addrspacecast.
1715 // GEP (addrspacecast [10 x i8] addrspace(1)* X to [0 x i8]*),
1718 // %0 = GEP [10 x i8] addrspace(1)* X, ...
1719 // addrspacecast i8 addrspace(1)* %0 to i8*
1720 SmallVector<Value*, 8> Idx(GEP.idx_begin(), GEP.idx_end());
1721 Value *NewGEP = GEP.isInBounds()
1722 ? Builder->CreateInBoundsGEP(
1723 nullptr, StrippedPtr, Idx, GEP.getName())
1724 : Builder->CreateGEP(nullptr, StrippedPtr, Idx,
1726 return new AddrSpaceCastInst(NewGEP, GEP.getType());
1730 } else if (GEP.getNumOperands() == 2) {
1731 // Transform things like:
1732 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
1733 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
1734 Type *SrcElTy = StrippedPtrTy->getElementType();
1735 Type *ResElTy = GEP.getSourceElementType();
1736 if (SrcElTy->isArrayTy() &&
1737 DL.getTypeAllocSize(SrcElTy->getArrayElementType()) ==
1738 DL.getTypeAllocSize(ResElTy)) {
1739 Type *IdxType = DL.getIntPtrType(GEP.getType());
1740 Value *Idx[2] = { Constant::getNullValue(IdxType), GEP.getOperand(1) };
1743 ? Builder->CreateInBoundsGEP(nullptr, StrippedPtr, Idx,
1745 : Builder->CreateGEP(nullptr, StrippedPtr, Idx, GEP.getName());
1747 // V and GEP are both pointer types --> BitCast
1748 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
1752 // Transform things like:
1753 // %V = mul i64 %N, 4
1754 // %t = getelementptr i8* bitcast (i32* %arr to i8*), i32 %V
1755 // into: %t1 = getelementptr i32* %arr, i32 %N; bitcast
1756 if (ResElTy->isSized() && SrcElTy->isSized()) {
1757 // Check that changing the type amounts to dividing the index by a scale
1759 uint64_t ResSize = DL.getTypeAllocSize(ResElTy);
1760 uint64_t SrcSize = DL.getTypeAllocSize(SrcElTy);
1761 if (ResSize && SrcSize % ResSize == 0) {
1762 Value *Idx = GEP.getOperand(1);
1763 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
1764 uint64_t Scale = SrcSize / ResSize;
1766 // Earlier transforms ensure that the index has type IntPtrType, which
1767 // considerably simplifies the logic by eliminating implicit casts.
1768 assert(Idx->getType() == DL.getIntPtrType(GEP.getType()) &&
1769 "Index not cast to pointer width?");
1772 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
1773 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
1774 // If the multiplication NewIdx * Scale may overflow then the new
1775 // GEP may not be "inbounds".
1777 GEP.isInBounds() && NSW
1778 ? Builder->CreateInBoundsGEP(nullptr, StrippedPtr, NewIdx,
1780 : Builder->CreateGEP(nullptr, StrippedPtr, NewIdx,
1783 // The NewGEP must be pointer typed, so must the old one -> BitCast
1784 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
1790 // Similarly, transform things like:
1791 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
1792 // (where tmp = 8*tmp2) into:
1793 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
1794 if (ResElTy->isSized() && SrcElTy->isSized() && SrcElTy->isArrayTy()) {
1795 // Check that changing to the array element type amounts to dividing the
1796 // index by a scale factor.
1797 uint64_t ResSize = DL.getTypeAllocSize(ResElTy);
1798 uint64_t ArrayEltSize =
1799 DL.getTypeAllocSize(SrcElTy->getArrayElementType());
1800 if (ResSize && ArrayEltSize % ResSize == 0) {
1801 Value *Idx = GEP.getOperand(1);
1802 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
1803 uint64_t Scale = ArrayEltSize / ResSize;
1805 // Earlier transforms ensure that the index has type IntPtrType, which
1806 // considerably simplifies the logic by eliminating implicit casts.
1807 assert(Idx->getType() == DL.getIntPtrType(GEP.getType()) &&
1808 "Index not cast to pointer width?");
1811 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
1812 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
1813 // If the multiplication NewIdx * Scale may overflow then the new
1814 // GEP may not be "inbounds".
1816 Constant::getNullValue(DL.getIntPtrType(GEP.getType())),
1819 Value *NewGEP = GEP.isInBounds() && NSW
1820 ? Builder->CreateInBoundsGEP(
1821 SrcElTy, StrippedPtr, Off, GEP.getName())
1822 : Builder->CreateGEP(SrcElTy, StrippedPtr, Off,
1824 // The NewGEP must be pointer typed, so must the old one -> BitCast
1825 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
1833 // addrspacecast between types is canonicalized as a bitcast, then an
1834 // addrspacecast. To take advantage of the below bitcast + struct GEP, look
1835 // through the addrspacecast.
1836 if (AddrSpaceCastInst *ASC = dyn_cast<AddrSpaceCastInst>(PtrOp)) {
1837 // X = bitcast A addrspace(1)* to B addrspace(1)*
1838 // Y = addrspacecast A addrspace(1)* to B addrspace(2)*
1839 // Z = gep Y, <...constant indices...>
1840 // Into an addrspacecasted GEP of the struct.
1841 if (BitCastInst *BC = dyn_cast<BitCastInst>(ASC->getOperand(0)))
1845 /// See if we can simplify:
1846 /// X = bitcast A* to B*
1847 /// Y = gep X, <...constant indices...>
1848 /// into a gep of the original struct. This is important for SROA and alias
1849 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
1850 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
1851 Value *Operand = BCI->getOperand(0);
1852 PointerType *OpType = cast<PointerType>(Operand->getType());
1853 unsigned OffsetBits = DL.getPointerTypeSizeInBits(GEP.getType());
1854 APInt Offset(OffsetBits, 0);
1855 if (!isa<BitCastInst>(Operand) &&
1856 GEP.accumulateConstantOffset(DL, Offset)) {
1858 // If this GEP instruction doesn't move the pointer, just replace the GEP
1859 // with a bitcast of the real input to the dest type.
1861 // If the bitcast is of an allocation, and the allocation will be
1862 // converted to match the type of the cast, don't touch this.
1863 if (isa<AllocaInst>(Operand) || isAllocationFn(Operand, TLI)) {
1864 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
1865 if (Instruction *I = visitBitCast(*BCI)) {
1868 BCI->getParent()->getInstList().insert(BCI->getIterator(), I);
1869 replaceInstUsesWith(*BCI, I);
1875 if (Operand->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
1876 return new AddrSpaceCastInst(Operand, GEP.getType());
1877 return new BitCastInst(Operand, GEP.getType());
1880 // Otherwise, if the offset is non-zero, we need to find out if there is a
1881 // field at Offset in 'A's type. If so, we can pull the cast through the
1883 SmallVector<Value*, 8> NewIndices;
1884 if (FindElementAtOffset(OpType, Offset.getSExtValue(), NewIndices)) {
1887 ? Builder->CreateInBoundsGEP(nullptr, Operand, NewIndices)
1888 : Builder->CreateGEP(nullptr, Operand, NewIndices);
1890 if (NGEP->getType() == GEP.getType())
1891 return replaceInstUsesWith(GEP, NGEP);
1892 NGEP->takeName(&GEP);
1894 if (NGEP->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
1895 return new AddrSpaceCastInst(NGEP, GEP.getType());
1896 return new BitCastInst(NGEP, GEP.getType());
1904 static bool isNeverEqualToUnescapedAlloc(Value *V, const TargetLibraryInfo *TLI,
1906 if (isa<ConstantPointerNull>(V))
1908 if (auto *LI = dyn_cast<LoadInst>(V))
1909 return isa<GlobalVariable>(LI->getPointerOperand());
1910 // Two distinct allocations will never be equal.
1911 // We rely on LookThroughBitCast in isAllocLikeFn being false, since looking
1912 // through bitcasts of V can cause
1913 // the result statement below to be true, even when AI and V (ex:
1914 // i8* ->i32* ->i8* of AI) are the same allocations.
1915 return isAllocLikeFn(V, TLI) && V != AI;
1919 isAllocSiteRemovable(Instruction *AI, SmallVectorImpl<WeakVH> &Users,
1920 const TargetLibraryInfo *TLI) {
1921 SmallVector<Instruction*, 4> Worklist;
1922 Worklist.push_back(AI);
1925 Instruction *PI = Worklist.pop_back_val();
1926 for (User *U : PI->users()) {
1927 Instruction *I = cast<Instruction>(U);
1928 switch (I->getOpcode()) {
1930 // Give up the moment we see something we can't handle.
1933 case Instruction::BitCast:
1934 case Instruction::GetElementPtr:
1935 Users.emplace_back(I);
1936 Worklist.push_back(I);
1939 case Instruction::ICmp: {
1940 ICmpInst *ICI = cast<ICmpInst>(I);
1941 // We can fold eq/ne comparisons with null to false/true, respectively.
1942 // We also fold comparisons in some conditions provided the alloc has
1943 // not escaped (see isNeverEqualToUnescapedAlloc).
1944 if (!ICI->isEquality())
1946 unsigned OtherIndex = (ICI->getOperand(0) == PI) ? 1 : 0;
1947 if (!isNeverEqualToUnescapedAlloc(ICI->getOperand(OtherIndex), TLI, AI))
1949 Users.emplace_back(I);
1953 case Instruction::Call:
1954 // Ignore no-op and store intrinsics.
1955 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1956 switch (II->getIntrinsicID()) {
1960 case Intrinsic::memmove:
1961 case Intrinsic::memcpy:
1962 case Intrinsic::memset: {
1963 MemIntrinsic *MI = cast<MemIntrinsic>(II);
1964 if (MI->isVolatile() || MI->getRawDest() != PI)
1968 case Intrinsic::dbg_declare:
1969 case Intrinsic::dbg_value:
1970 case Intrinsic::invariant_start:
1971 case Intrinsic::invariant_end:
1972 case Intrinsic::lifetime_start:
1973 case Intrinsic::lifetime_end:
1974 case Intrinsic::objectsize:
1975 Users.emplace_back(I);
1980 if (isFreeCall(I, TLI)) {
1981 Users.emplace_back(I);
1986 case Instruction::Store: {
1987 StoreInst *SI = cast<StoreInst>(I);
1988 if (SI->isVolatile() || SI->getPointerOperand() != PI)
1990 Users.emplace_back(I);
1994 llvm_unreachable("missing a return?");
1996 } while (!Worklist.empty());
2000 Instruction *InstCombiner::visitAllocSite(Instruction &MI) {
2001 // If we have a malloc call which is only used in any amount of comparisons
2002 // to null and free calls, delete the calls and replace the comparisons with
2003 // true or false as appropriate.
2004 SmallVector<WeakVH, 64> Users;
2005 if (isAllocSiteRemovable(&MI, Users, TLI)) {
2006 for (unsigned i = 0, e = Users.size(); i != e; ++i) {
2007 // Lowering all @llvm.objectsize calls first because they may
2008 // use a bitcast/GEP of the alloca we are removing.
2012 Instruction *I = cast<Instruction>(&*Users[i]);
2014 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
2015 if (II->getIntrinsicID() == Intrinsic::objectsize) {
2017 if (!getObjectSize(II->getArgOperand(0), Size, DL, TLI)) {
2018 ConstantInt *CI = cast<ConstantInt>(II->getArgOperand(1));
2019 Size = CI->isZero() ? -1ULL : 0;
2021 replaceInstUsesWith(*I, ConstantInt::get(I->getType(), Size));
2022 eraseInstFromFunction(*I);
2023 Users[i] = nullptr; // Skip examining in the next loop.
2027 for (unsigned i = 0, e = Users.size(); i != e; ++i) {
2031 Instruction *I = cast<Instruction>(&*Users[i]);
2033 if (ICmpInst *C = dyn_cast<ICmpInst>(I)) {
2034 replaceInstUsesWith(*C,
2035 ConstantInt::get(Type::getInt1Ty(C->getContext()),
2036 C->isFalseWhenEqual()));
2037 } else if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) {
2038 replaceInstUsesWith(*I, UndefValue::get(I->getType()));
2040 eraseInstFromFunction(*I);
2043 if (InvokeInst *II = dyn_cast<InvokeInst>(&MI)) {
2044 // Replace invoke with a NOP intrinsic to maintain the original CFG
2045 Module *M = II->getModule();
2046 Function *F = Intrinsic::getDeclaration(M, Intrinsic::donothing);
2047 InvokeInst::Create(F, II->getNormalDest(), II->getUnwindDest(),
2048 None, "", II->getParent());
2050 return eraseInstFromFunction(MI);
2055 /// \brief Move the call to free before a NULL test.
2057 /// Check if this free is accessed after its argument has been test
2058 /// against NULL (property 0).
2059 /// If yes, it is legal to move this call in its predecessor block.
2061 /// The move is performed only if the block containing the call to free
2062 /// will be removed, i.e.:
2063 /// 1. it has only one predecessor P, and P has two successors
2064 /// 2. it contains the call and an unconditional branch
2065 /// 3. its successor is the same as its predecessor's successor
2067 /// The profitability is out-of concern here and this function should
2068 /// be called only if the caller knows this transformation would be
2069 /// profitable (e.g., for code size).
2070 static Instruction *
2071 tryToMoveFreeBeforeNullTest(CallInst &FI) {
2072 Value *Op = FI.getArgOperand(0);
2073 BasicBlock *FreeInstrBB = FI.getParent();
2074 BasicBlock *PredBB = FreeInstrBB->getSinglePredecessor();
2076 // Validate part of constraint #1: Only one predecessor
2077 // FIXME: We can extend the number of predecessor, but in that case, we
2078 // would duplicate the call to free in each predecessor and it may
2079 // not be profitable even for code size.
2083 // Validate constraint #2: Does this block contains only the call to
2084 // free and an unconditional branch?
2085 // FIXME: We could check if we can speculate everything in the
2086 // predecessor block
2087 if (FreeInstrBB->size() != 2)
2090 if (!match(FreeInstrBB->getTerminator(), m_UnconditionalBr(SuccBB)))
2093 // Validate the rest of constraint #1 by matching on the pred branch.
2094 TerminatorInst *TI = PredBB->getTerminator();
2095 BasicBlock *TrueBB, *FalseBB;
2096 ICmpInst::Predicate Pred;
2097 if (!match(TI, m_Br(m_ICmp(Pred, m_Specific(Op), m_Zero()), TrueBB, FalseBB)))
2099 if (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE)
2102 // Validate constraint #3: Ensure the null case just falls through.
2103 if (SuccBB != (Pred == ICmpInst::ICMP_EQ ? TrueBB : FalseBB))
2105 assert(FreeInstrBB == (Pred == ICmpInst::ICMP_EQ ? FalseBB : TrueBB) &&
2106 "Broken CFG: missing edge from predecessor to successor");
2113 Instruction *InstCombiner::visitFree(CallInst &FI) {
2114 Value *Op = FI.getArgOperand(0);
2116 // free undef -> unreachable.
2117 if (isa<UndefValue>(Op)) {
2118 // Insert a new store to null because we cannot modify the CFG here.
2119 Builder->CreateStore(ConstantInt::getTrue(FI.getContext()),
2120 UndefValue::get(Type::getInt1PtrTy(FI.getContext())));
2121 return eraseInstFromFunction(FI);
2124 // If we have 'free null' delete the instruction. This can happen in stl code
2125 // when lots of inlining happens.
2126 if (isa<ConstantPointerNull>(Op))
2127 return eraseInstFromFunction(FI);
2129 // If we optimize for code size, try to move the call to free before the null
2130 // test so that simplify cfg can remove the empty block and dead code
2131 // elimination the branch. I.e., helps to turn something like:
2132 // if (foo) free(foo);
2136 if (Instruction *I = tryToMoveFreeBeforeNullTest(FI))
2142 Instruction *InstCombiner::visitReturnInst(ReturnInst &RI) {
2143 if (RI.getNumOperands() == 0) // ret void
2146 Value *ResultOp = RI.getOperand(0);
2147 Type *VTy = ResultOp->getType();
2148 if (!VTy->isIntegerTy())
2151 // There might be assume intrinsics dominating this return that completely
2152 // determine the value. If so, constant fold it.
2153 unsigned BitWidth = VTy->getPrimitiveSizeInBits();
2154 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2155 computeKnownBits(ResultOp, KnownZero, KnownOne, 0, &RI);
2156 if ((KnownZero|KnownOne).isAllOnesValue())
2157 RI.setOperand(0, Constant::getIntegerValue(VTy, KnownOne));
2162 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
2163 // Change br (not X), label True, label False to: br X, label False, True
2165 BasicBlock *TrueDest;
2166 BasicBlock *FalseDest;
2167 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
2168 !isa<Constant>(X)) {
2169 // Swap Destinations and condition...
2171 BI.swapSuccessors();
2175 // If the condition is irrelevant, remove the use so that other
2176 // transforms on the condition become more effective.
2177 if (BI.isConditional() &&
2178 BI.getSuccessor(0) == BI.getSuccessor(1) &&
2179 !isa<UndefValue>(BI.getCondition())) {
2180 BI.setCondition(UndefValue::get(BI.getCondition()->getType()));
2184 // Canonicalize fcmp_one -> fcmp_oeq
2185 FCmpInst::Predicate FPred; Value *Y;
2186 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
2187 TrueDest, FalseDest)) &&
2188 BI.getCondition()->hasOneUse())
2189 if (FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
2190 FPred == FCmpInst::FCMP_OGE) {
2191 FCmpInst *Cond = cast<FCmpInst>(BI.getCondition());
2192 Cond->setPredicate(FCmpInst::getInversePredicate(FPred));
2194 // Swap Destinations and condition.
2195 BI.swapSuccessors();
2200 // Canonicalize icmp_ne -> icmp_eq
2201 ICmpInst::Predicate IPred;
2202 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
2203 TrueDest, FalseDest)) &&
2204 BI.getCondition()->hasOneUse())
2205 if (IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
2206 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
2207 IPred == ICmpInst::ICMP_SGE) {
2208 ICmpInst *Cond = cast<ICmpInst>(BI.getCondition());
2209 Cond->setPredicate(ICmpInst::getInversePredicate(IPred));
2210 // Swap Destinations and condition.
2211 BI.swapSuccessors();
2219 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
2220 Value *Cond = SI.getCondition();
2221 unsigned BitWidth = cast<IntegerType>(Cond->getType())->getBitWidth();
2222 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2223 computeKnownBits(Cond, KnownZero, KnownOne, 0, &SI);
2224 unsigned LeadingKnownZeros = KnownZero.countLeadingOnes();
2225 unsigned LeadingKnownOnes = KnownOne.countLeadingOnes();
2227 // Compute the number of leading bits we can ignore.
2228 // TODO: A better way to determine this would use ComputeNumSignBits().
2229 for (auto &C : SI.cases()) {
2230 LeadingKnownZeros = std::min(
2231 LeadingKnownZeros, C.getCaseValue()->getValue().countLeadingZeros());
2232 LeadingKnownOnes = std::min(
2233 LeadingKnownOnes, C.getCaseValue()->getValue().countLeadingOnes());
2236 unsigned NewWidth = BitWidth - std::max(LeadingKnownZeros, LeadingKnownOnes);
2238 // Shrink the condition operand if the new type is smaller than the old type.
2239 // This may produce a non-standard type for the switch, but that's ok because
2240 // the backend should extend back to a legal type for the target.
2241 bool TruncCond = false;
2242 if (NewWidth > 0 && NewWidth < BitWidth) {
2244 IntegerType *Ty = IntegerType::get(SI.getContext(), NewWidth);
2245 Builder->SetInsertPoint(&SI);
2246 Value *NewCond = Builder->CreateTrunc(Cond, Ty, "trunc");
2247 SI.setCondition(NewCond);
2249 for (auto &C : SI.cases())
2250 static_cast<SwitchInst::CaseIt *>(&C)->setValue(ConstantInt::get(
2251 SI.getContext(), C.getCaseValue()->getValue().trunc(NewWidth)));
2254 ConstantInt *AddRHS = nullptr;
2255 if (match(Cond, m_Add(m_Value(), m_ConstantInt(AddRHS)))) {
2256 Instruction *I = cast<Instruction>(Cond);
2257 // Change 'switch (X+4) case 1:' into 'switch (X) case -3'.
2258 for (SwitchInst::CaseIt i = SI.case_begin(), e = SI.case_end(); i != e;
2260 ConstantInt *CaseVal = i.getCaseValue();
2261 Constant *LHS = CaseVal;
2263 LHS = LeadingKnownZeros
2264 ? ConstantExpr::getZExt(CaseVal, Cond->getType())
2265 : ConstantExpr::getSExt(CaseVal, Cond->getType());
2267 Constant *NewCaseVal = ConstantExpr::getSub(LHS, AddRHS);
2268 assert(isa<ConstantInt>(NewCaseVal) &&
2269 "Result of expression should be constant");
2270 i.setValue(cast<ConstantInt>(NewCaseVal));
2272 SI.setCondition(I->getOperand(0));
2277 return TruncCond ? &SI : nullptr;
2280 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
2281 Value *Agg = EV.getAggregateOperand();
2283 if (!EV.hasIndices())
2284 return replaceInstUsesWith(EV, Agg);
2287 SimplifyExtractValueInst(Agg, EV.getIndices(), DL, TLI, DT, AC))
2288 return replaceInstUsesWith(EV, V);
2290 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
2291 // We're extracting from an insertvalue instruction, compare the indices
2292 const unsigned *exti, *exte, *insi, *inse;
2293 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
2294 exte = EV.idx_end(), inse = IV->idx_end();
2295 exti != exte && insi != inse;
2298 // The insert and extract both reference distinctly different elements.
2299 // This means the extract is not influenced by the insert, and we can
2300 // replace the aggregate operand of the extract with the aggregate
2301 // operand of the insert. i.e., replace
2302 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
2303 // %E = extractvalue { i32, { i32 } } %I, 0
2305 // %E = extractvalue { i32, { i32 } } %A, 0
2306 return ExtractValueInst::Create(IV->getAggregateOperand(),
2309 if (exti == exte && insi == inse)
2310 // Both iterators are at the end: Index lists are identical. Replace
2311 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
2312 // %C = extractvalue { i32, { i32 } } %B, 1, 0
2314 return replaceInstUsesWith(EV, IV->getInsertedValueOperand());
2316 // The extract list is a prefix of the insert list. i.e. replace
2317 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
2318 // %E = extractvalue { i32, { i32 } } %I, 1
2320 // %X = extractvalue { i32, { i32 } } %A, 1
2321 // %E = insertvalue { i32 } %X, i32 42, 0
2322 // by switching the order of the insert and extract (though the
2323 // insertvalue should be left in, since it may have other uses).
2324 Value *NewEV = Builder->CreateExtractValue(IV->getAggregateOperand(),
2326 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
2327 makeArrayRef(insi, inse));
2330 // The insert list is a prefix of the extract list
2331 // We can simply remove the common indices from the extract and make it
2332 // operate on the inserted value instead of the insertvalue result.
2334 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
2335 // %E = extractvalue { i32, { i32 } } %I, 1, 0
2337 // %E extractvalue { i32 } { i32 42 }, 0
2338 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
2339 makeArrayRef(exti, exte));
2341 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) {
2342 // We're extracting from an intrinsic, see if we're the only user, which
2343 // allows us to simplify multiple result intrinsics to simpler things that
2344 // just get one value.
2345 if (II->hasOneUse()) {
2346 // Check if we're grabbing the overflow bit or the result of a 'with
2347 // overflow' intrinsic. If it's the latter we can remove the intrinsic
2348 // and replace it with a traditional binary instruction.
2349 switch (II->getIntrinsicID()) {
2350 case Intrinsic::uadd_with_overflow:
2351 case Intrinsic::sadd_with_overflow:
2352 if (*EV.idx_begin() == 0) { // Normal result.
2353 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2354 replaceInstUsesWith(*II, UndefValue::get(II->getType()));
2355 eraseInstFromFunction(*II);
2356 return BinaryOperator::CreateAdd(LHS, RHS);
2359 // If the normal result of the add is dead, and the RHS is a constant,
2360 // we can transform this into a range comparison.
2361 // overflow = uadd a, -4 --> overflow = icmp ugt a, 3
2362 if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow)
2363 if (ConstantInt *CI = dyn_cast<ConstantInt>(II->getArgOperand(1)))
2364 return new ICmpInst(ICmpInst::ICMP_UGT, II->getArgOperand(0),
2365 ConstantExpr::getNot(CI));
2367 case Intrinsic::usub_with_overflow:
2368 case Intrinsic::ssub_with_overflow:
2369 if (*EV.idx_begin() == 0) { // Normal result.
2370 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2371 replaceInstUsesWith(*II, UndefValue::get(II->getType()));
2372 eraseInstFromFunction(*II);
2373 return BinaryOperator::CreateSub(LHS, RHS);
2376 case Intrinsic::umul_with_overflow:
2377 case Intrinsic::smul_with_overflow:
2378 if (*EV.idx_begin() == 0) { // Normal result.
2379 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2380 replaceInstUsesWith(*II, UndefValue::get(II->getType()));
2381 eraseInstFromFunction(*II);
2382 return BinaryOperator::CreateMul(LHS, RHS);
2390 if (LoadInst *L = dyn_cast<LoadInst>(Agg))
2391 // If the (non-volatile) load only has one use, we can rewrite this to a
2392 // load from a GEP. This reduces the size of the load. If a load is used
2393 // only by extractvalue instructions then this either must have been
2394 // optimized before, or it is a struct with padding, in which case we
2395 // don't want to do the transformation as it loses padding knowledge.
2396 if (L->isSimple() && L->hasOneUse()) {
2397 // extractvalue has integer indices, getelementptr has Value*s. Convert.
2398 SmallVector<Value*, 4> Indices;
2399 // Prefix an i32 0 since we need the first element.
2400 Indices.push_back(Builder->getInt32(0));
2401 for (ExtractValueInst::idx_iterator I = EV.idx_begin(), E = EV.idx_end();
2403 Indices.push_back(Builder->getInt32(*I));
2405 // We need to insert these at the location of the old load, not at that of
2406 // the extractvalue.
2407 Builder->SetInsertPoint(L);
2408 Value *GEP = Builder->CreateInBoundsGEP(L->getType(),
2409 L->getPointerOperand(), Indices);
2410 // Returning the load directly will cause the main loop to insert it in
2411 // the wrong spot, so use replaceInstUsesWith().
2412 return replaceInstUsesWith(EV, Builder->CreateLoad(GEP));
2414 // We could simplify extracts from other values. Note that nested extracts may
2415 // already be simplified implicitly by the above: extract (extract (insert) )
2416 // will be translated into extract ( insert ( extract ) ) first and then just
2417 // the value inserted, if appropriate. Similarly for extracts from single-use
2418 // loads: extract (extract (load)) will be translated to extract (load (gep))
2419 // and if again single-use then via load (gep (gep)) to load (gep).
2420 // However, double extracts from e.g. function arguments or return values
2421 // aren't handled yet.
2425 /// Return 'true' if the given typeinfo will match anything.
2426 static bool isCatchAll(EHPersonality Personality, Constant *TypeInfo) {
2427 switch (Personality) {
2428 case EHPersonality::GNU_C:
2429 case EHPersonality::GNU_C_SjLj:
2430 case EHPersonality::Rust:
2431 // The GCC C EH and Rust personality only exists to support cleanups, so
2432 // it's not clear what the semantics of catch clauses are.
2434 case EHPersonality::Unknown:
2436 case EHPersonality::GNU_Ada:
2437 // While __gnat_all_others_value will match any Ada exception, it doesn't
2438 // match foreign exceptions (or didn't, before gcc-4.7).
2440 case EHPersonality::GNU_CXX:
2441 case EHPersonality::GNU_CXX_SjLj:
2442 case EHPersonality::GNU_ObjC:
2443 case EHPersonality::MSVC_X86SEH:
2444 case EHPersonality::MSVC_Win64SEH:
2445 case EHPersonality::MSVC_CXX:
2446 case EHPersonality::CoreCLR:
2447 return TypeInfo->isNullValue();
2449 llvm_unreachable("invalid enum");
2452 static bool shorter_filter(const Value *LHS, const Value *RHS) {
2454 cast<ArrayType>(LHS->getType())->getNumElements()
2456 cast<ArrayType>(RHS->getType())->getNumElements();
2459 Instruction *InstCombiner::visitLandingPadInst(LandingPadInst &LI) {
2460 // The logic here should be correct for any real-world personality function.
2461 // However if that turns out not to be true, the offending logic can always
2462 // be conditioned on the personality function, like the catch-all logic is.
2463 EHPersonality Personality =
2464 classifyEHPersonality(LI.getParent()->getParent()->getPersonalityFn());
2466 // Simplify the list of clauses, eg by removing repeated catch clauses
2467 // (these are often created by inlining).
2468 bool MakeNewInstruction = false; // If true, recreate using the following:
2469 SmallVector<Constant *, 16> NewClauses; // - Clauses for the new instruction;
2470 bool CleanupFlag = LI.isCleanup(); // - The new instruction is a cleanup.
2472 SmallPtrSet<Value *, 16> AlreadyCaught; // Typeinfos known caught already.
2473 for (unsigned i = 0, e = LI.getNumClauses(); i != e; ++i) {
2474 bool isLastClause = i + 1 == e;
2475 if (LI.isCatch(i)) {
2477 Constant *CatchClause = LI.getClause(i);
2478 Constant *TypeInfo = CatchClause->stripPointerCasts();
2480 // If we already saw this clause, there is no point in having a second
2482 if (AlreadyCaught.insert(TypeInfo).second) {
2483 // This catch clause was not already seen.
2484 NewClauses.push_back(CatchClause);
2486 // Repeated catch clause - drop the redundant copy.
2487 MakeNewInstruction = true;
2490 // If this is a catch-all then there is no point in keeping any following
2491 // clauses or marking the landingpad as having a cleanup.
2492 if (isCatchAll(Personality, TypeInfo)) {
2494 MakeNewInstruction = true;
2495 CleanupFlag = false;
2499 // A filter clause. If any of the filter elements were already caught
2500 // then they can be dropped from the filter. It is tempting to try to
2501 // exploit the filter further by saying that any typeinfo that does not
2502 // occur in the filter can't be caught later (and thus can be dropped).
2503 // However this would be wrong, since typeinfos can match without being
2504 // equal (for example if one represents a C++ class, and the other some
2505 // class derived from it).
2506 assert(LI.isFilter(i) && "Unsupported landingpad clause!");
2507 Constant *FilterClause = LI.getClause(i);
2508 ArrayType *FilterType = cast<ArrayType>(FilterClause->getType());
2509 unsigned NumTypeInfos = FilterType->getNumElements();
2511 // An empty filter catches everything, so there is no point in keeping any
2512 // following clauses or marking the landingpad as having a cleanup. By
2513 // dealing with this case here the following code is made a bit simpler.
2514 if (!NumTypeInfos) {
2515 NewClauses.push_back(FilterClause);
2517 MakeNewInstruction = true;
2518 CleanupFlag = false;
2522 bool MakeNewFilter = false; // If true, make a new filter.
2523 SmallVector<Constant *, 16> NewFilterElts; // New elements.
2524 if (isa<ConstantAggregateZero>(FilterClause)) {
2525 // Not an empty filter - it contains at least one null typeinfo.
2526 assert(NumTypeInfos > 0 && "Should have handled empty filter already!");
2527 Constant *TypeInfo =
2528 Constant::getNullValue(FilterType->getElementType());
2529 // If this typeinfo is a catch-all then the filter can never match.
2530 if (isCatchAll(Personality, TypeInfo)) {
2531 // Throw the filter away.
2532 MakeNewInstruction = true;
2536 // There is no point in having multiple copies of this typeinfo, so
2537 // discard all but the first copy if there is more than one.
2538 NewFilterElts.push_back(TypeInfo);
2539 if (NumTypeInfos > 1)
2540 MakeNewFilter = true;
2542 ConstantArray *Filter = cast<ConstantArray>(FilterClause);
2543 SmallPtrSet<Value *, 16> SeenInFilter; // For uniquing the elements.
2544 NewFilterElts.reserve(NumTypeInfos);
2546 // Remove any filter elements that were already caught or that already
2547 // occurred in the filter. While there, see if any of the elements are
2548 // catch-alls. If so, the filter can be discarded.
2549 bool SawCatchAll = false;
2550 for (unsigned j = 0; j != NumTypeInfos; ++j) {
2551 Constant *Elt = Filter->getOperand(j);
2552 Constant *TypeInfo = Elt->stripPointerCasts();
2553 if (isCatchAll(Personality, TypeInfo)) {
2554 // This element is a catch-all. Bail out, noting this fact.
2559 // Even if we've seen a type in a catch clause, we don't want to
2560 // remove it from the filter. An unexpected type handler may be
2561 // set up for a call site which throws an exception of the same
2562 // type caught. In order for the exception thrown by the unexpected
2563 // handler to propogate correctly, the filter must be correctly
2564 // described for the call site.
2568 // void unexpected() { throw 1;}
2569 // void foo() throw (int) {
2570 // std::set_unexpected(unexpected);
2573 // } catch (int i) {}
2576 // There is no point in having multiple copies of the same typeinfo in
2577 // a filter, so only add it if we didn't already.
2578 if (SeenInFilter.insert(TypeInfo).second)
2579 NewFilterElts.push_back(cast<Constant>(Elt));
2581 // A filter containing a catch-all cannot match anything by definition.
2583 // Throw the filter away.
2584 MakeNewInstruction = true;
2588 // If we dropped something from the filter, make a new one.
2589 if (NewFilterElts.size() < NumTypeInfos)
2590 MakeNewFilter = true;
2592 if (MakeNewFilter) {
2593 FilterType = ArrayType::get(FilterType->getElementType(),
2594 NewFilterElts.size());
2595 FilterClause = ConstantArray::get(FilterType, NewFilterElts);
2596 MakeNewInstruction = true;
2599 NewClauses.push_back(FilterClause);
2601 // If the new filter is empty then it will catch everything so there is
2602 // no point in keeping any following clauses or marking the landingpad
2603 // as having a cleanup. The case of the original filter being empty was
2604 // already handled above.
2605 if (MakeNewFilter && !NewFilterElts.size()) {
2606 assert(MakeNewInstruction && "New filter but not a new instruction!");
2607 CleanupFlag = false;
2613 // If several filters occur in a row then reorder them so that the shortest
2614 // filters come first (those with the smallest number of elements). This is
2615 // advantageous because shorter filters are more likely to match, speeding up
2616 // unwinding, but mostly because it increases the effectiveness of the other
2617 // filter optimizations below.
2618 for (unsigned i = 0, e = NewClauses.size(); i + 1 < e; ) {
2620 // Find the maximal 'j' s.t. the range [i, j) consists entirely of filters.
2621 for (j = i; j != e; ++j)
2622 if (!isa<ArrayType>(NewClauses[j]->getType()))
2625 // Check whether the filters are already sorted by length. We need to know
2626 // if sorting them is actually going to do anything so that we only make a
2627 // new landingpad instruction if it does.
2628 for (unsigned k = i; k + 1 < j; ++k)
2629 if (shorter_filter(NewClauses[k+1], NewClauses[k])) {
2630 // Not sorted, so sort the filters now. Doing an unstable sort would be
2631 // correct too but reordering filters pointlessly might confuse users.
2632 std::stable_sort(NewClauses.begin() + i, NewClauses.begin() + j,
2634 MakeNewInstruction = true;
2638 // Look for the next batch of filters.
2642 // If typeinfos matched if and only if equal, then the elements of a filter L
2643 // that occurs later than a filter F could be replaced by the intersection of
2644 // the elements of F and L. In reality two typeinfos can match without being
2645 // equal (for example if one represents a C++ class, and the other some class
2646 // derived from it) so it would be wrong to perform this transform in general.
2647 // However the transform is correct and useful if F is a subset of L. In that
2648 // case L can be replaced by F, and thus removed altogether since repeating a
2649 // filter is pointless. So here we look at all pairs of filters F and L where
2650 // L follows F in the list of clauses, and remove L if every element of F is
2651 // an element of L. This can occur when inlining C++ functions with exception
2653 for (unsigned i = 0; i + 1 < NewClauses.size(); ++i) {
2654 // Examine each filter in turn.
2655 Value *Filter = NewClauses[i];
2656 ArrayType *FTy = dyn_cast<ArrayType>(Filter->getType());
2658 // Not a filter - skip it.
2660 unsigned FElts = FTy->getNumElements();
2661 // Examine each filter following this one. Doing this backwards means that
2662 // we don't have to worry about filters disappearing under us when removed.
2663 for (unsigned j = NewClauses.size() - 1; j != i; --j) {
2664 Value *LFilter = NewClauses[j];
2665 ArrayType *LTy = dyn_cast<ArrayType>(LFilter->getType());
2667 // Not a filter - skip it.
2669 // If Filter is a subset of LFilter, i.e. every element of Filter is also
2670 // an element of LFilter, then discard LFilter.
2671 SmallVectorImpl<Constant *>::iterator J = NewClauses.begin() + j;
2672 // If Filter is empty then it is a subset of LFilter.
2675 NewClauses.erase(J);
2676 MakeNewInstruction = true;
2677 // Move on to the next filter.
2680 unsigned LElts = LTy->getNumElements();
2681 // If Filter is longer than LFilter then it cannot be a subset of it.
2683 // Move on to the next filter.
2685 // At this point we know that LFilter has at least one element.
2686 if (isa<ConstantAggregateZero>(LFilter)) { // LFilter only contains zeros.
2687 // Filter is a subset of LFilter iff Filter contains only zeros (as we
2688 // already know that Filter is not longer than LFilter).
2689 if (isa<ConstantAggregateZero>(Filter)) {
2690 assert(FElts <= LElts && "Should have handled this case earlier!");
2692 NewClauses.erase(J);
2693 MakeNewInstruction = true;
2695 // Move on to the next filter.
2698 ConstantArray *LArray = cast<ConstantArray>(LFilter);
2699 if (isa<ConstantAggregateZero>(Filter)) { // Filter only contains zeros.
2700 // Since Filter is non-empty and contains only zeros, it is a subset of
2701 // LFilter iff LFilter contains a zero.
2702 assert(FElts > 0 && "Should have eliminated the empty filter earlier!");
2703 for (unsigned l = 0; l != LElts; ++l)
2704 if (LArray->getOperand(l)->isNullValue()) {
2705 // LFilter contains a zero - discard it.
2706 NewClauses.erase(J);
2707 MakeNewInstruction = true;
2710 // Move on to the next filter.
2713 // At this point we know that both filters are ConstantArrays. Loop over
2714 // operands to see whether every element of Filter is also an element of
2715 // LFilter. Since filters tend to be short this is probably faster than
2716 // using a method that scales nicely.
2717 ConstantArray *FArray = cast<ConstantArray>(Filter);
2718 bool AllFound = true;
2719 for (unsigned f = 0; f != FElts; ++f) {
2720 Value *FTypeInfo = FArray->getOperand(f)->stripPointerCasts();
2722 for (unsigned l = 0; l != LElts; ++l) {
2723 Value *LTypeInfo = LArray->getOperand(l)->stripPointerCasts();
2724 if (LTypeInfo == FTypeInfo) {
2734 NewClauses.erase(J);
2735 MakeNewInstruction = true;
2737 // Move on to the next filter.
2741 // If we changed any of the clauses, replace the old landingpad instruction
2743 if (MakeNewInstruction) {
2744 LandingPadInst *NLI = LandingPadInst::Create(LI.getType(),
2746 for (unsigned i = 0, e = NewClauses.size(); i != e; ++i)
2747 NLI->addClause(NewClauses[i]);
2748 // A landing pad with no clauses must have the cleanup flag set. It is
2749 // theoretically possible, though highly unlikely, that we eliminated all
2750 // clauses. If so, force the cleanup flag to true.
2751 if (NewClauses.empty())
2753 NLI->setCleanup(CleanupFlag);
2757 // Even if none of the clauses changed, we may nonetheless have understood
2758 // that the cleanup flag is pointless. Clear it if so.
2759 if (LI.isCleanup() != CleanupFlag) {
2760 assert(!CleanupFlag && "Adding a cleanup, not removing one?!");
2761 LI.setCleanup(CleanupFlag);
2768 /// Try to move the specified instruction from its current block into the
2769 /// beginning of DestBlock, which can only happen if it's safe to move the
2770 /// instruction past all of the instructions between it and the end of its
2772 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
2773 assert(I->hasOneUse() && "Invariants didn't hold!");
2775 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
2776 if (isa<PHINode>(I) || I->isEHPad() || I->mayHaveSideEffects() ||
2777 isa<TerminatorInst>(I))
2780 // Do not sink alloca instructions out of the entry block.
2781 if (isa<AllocaInst>(I) && I->getParent() ==
2782 &DestBlock->getParent()->getEntryBlock())
2785 // Do not sink into catchswitch blocks.
2786 if (isa<CatchSwitchInst>(DestBlock->getTerminator()))
2789 // Do not sink convergent call instructions.
2790 if (auto *CI = dyn_cast<CallInst>(I)) {
2791 if (CI->isConvergent())
2794 // We can only sink load instructions if there is nothing between the load and
2795 // the end of block that could change the value.
2796 if (I->mayReadFromMemory()) {
2797 for (BasicBlock::iterator Scan = I->getIterator(),
2798 E = I->getParent()->end();
2800 if (Scan->mayWriteToMemory())
2804 BasicBlock::iterator InsertPos = DestBlock->getFirstInsertionPt();
2805 I->moveBefore(&*InsertPos);
2810 bool InstCombiner::run() {
2811 while (!Worklist.isEmpty()) {
2812 Instruction *I = Worklist.RemoveOne();
2813 if (I == nullptr) continue; // skip null values.
2815 // Check to see if we can DCE the instruction.
2816 if (isInstructionTriviallyDead(I, TLI)) {
2817 DEBUG(dbgs() << "IC: DCE: " << *I << '\n');
2818 eraseInstFromFunction(*I);
2820 MadeIRChange = true;
2824 // Instruction isn't dead, see if we can constant propagate it.
2825 if (!I->use_empty() &&
2826 (I->getNumOperands() == 0 || isa<Constant>(I->getOperand(0)))) {
2827 if (Constant *C = ConstantFoldInstruction(I, DL, TLI)) {
2828 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n');
2830 // Add operands to the worklist.
2831 replaceInstUsesWith(*I, C);
2833 if (isInstructionTriviallyDead(I, TLI))
2834 eraseInstFromFunction(*I);
2835 MadeIRChange = true;
2840 // In general, it is possible for computeKnownBits to determine all bits in
2841 // a value even when the operands are not all constants.
2842 if (ExpensiveCombines && !I->use_empty() && I->getType()->isIntegerTy()) {
2843 unsigned BitWidth = I->getType()->getScalarSizeInBits();
2844 APInt KnownZero(BitWidth, 0);
2845 APInt KnownOne(BitWidth, 0);
2846 computeKnownBits(I, KnownZero, KnownOne, /*Depth*/0, I);
2847 if ((KnownZero | KnownOne).isAllOnesValue()) {
2848 Constant *C = ConstantInt::get(I->getContext(), KnownOne);
2849 DEBUG(dbgs() << "IC: ConstFold (all bits known) to: " << *C <<
2850 " from: " << *I << '\n');
2852 // Add operands to the worklist.
2853 replaceInstUsesWith(*I, C);
2855 if (isInstructionTriviallyDead(I, TLI))
2856 eraseInstFromFunction(*I);
2857 MadeIRChange = true;
2862 // See if we can trivially sink this instruction to a successor basic block.
2863 if (I->hasOneUse()) {
2864 BasicBlock *BB = I->getParent();
2865 Instruction *UserInst = cast<Instruction>(*I->user_begin());
2866 BasicBlock *UserParent;
2868 // Get the block the use occurs in.
2869 if (PHINode *PN = dyn_cast<PHINode>(UserInst))
2870 UserParent = PN->getIncomingBlock(*I->use_begin());
2872 UserParent = UserInst->getParent();
2874 if (UserParent != BB) {
2875 bool UserIsSuccessor = false;
2876 // See if the user is one of our successors.
2877 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
2878 if (*SI == UserParent) {
2879 UserIsSuccessor = true;
2883 // If the user is one of our immediate successors, and if that successor
2884 // only has us as a predecessors (we'd have to split the critical edge
2885 // otherwise), we can keep going.
2886 if (UserIsSuccessor && UserParent->getSinglePredecessor()) {
2887 // Okay, the CFG is simple enough, try to sink this instruction.
2888 if (TryToSinkInstruction(I, UserParent)) {
2889 DEBUG(dbgs() << "IC: Sink: " << *I << '\n');
2890 MadeIRChange = true;
2891 // We'll add uses of the sunk instruction below, but since sinking
2892 // can expose opportunities for it's *operands* add them to the
2894 for (Use &U : I->operands())
2895 if (Instruction *OpI = dyn_cast<Instruction>(U.get()))
2902 // Now that we have an instruction, try combining it to simplify it.
2903 Builder->SetInsertPoint(I);
2904 Builder->SetCurrentDebugLocation(I->getDebugLoc());
2909 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str(););
2910 DEBUG(dbgs() << "IC: Visiting: " << OrigI << '\n');
2912 if (Instruction *Result = visit(*I)) {
2914 // Should we replace the old instruction with a new one?
2916 DEBUG(dbgs() << "IC: Old = " << *I << '\n'
2917 << " New = " << *Result << '\n');
2919 if (I->getDebugLoc())
2920 Result->setDebugLoc(I->getDebugLoc());
2921 // Everything uses the new instruction now.
2922 I->replaceAllUsesWith(Result);
2924 // Move the name to the new instruction first.
2925 Result->takeName(I);
2927 // Push the new instruction and any users onto the worklist.
2928 Worklist.Add(Result);
2929 Worklist.AddUsersToWorkList(*Result);
2931 // Insert the new instruction into the basic block...
2932 BasicBlock *InstParent = I->getParent();
2933 BasicBlock::iterator InsertPos = I->getIterator();
2935 // If we replace a PHI with something that isn't a PHI, fix up the
2937 if (!isa<PHINode>(Result) && isa<PHINode>(InsertPos))
2938 InsertPos = InstParent->getFirstInsertionPt();
2940 InstParent->getInstList().insert(InsertPos, Result);
2942 eraseInstFromFunction(*I);
2945 DEBUG(dbgs() << "IC: Mod = " << OrigI << '\n'
2946 << " New = " << *I << '\n');
2949 // If the instruction was modified, it's possible that it is now dead.
2950 // if so, remove it.
2951 if (isInstructionTriviallyDead(I, TLI)) {
2952 eraseInstFromFunction(*I);
2955 Worklist.AddUsersToWorkList(*I);
2958 MadeIRChange = true;
2963 return MadeIRChange;
2966 /// Walk the function in depth-first order, adding all reachable code to the
2969 /// This has a couple of tricks to make the code faster and more powerful. In
2970 /// particular, we constant fold and DCE instructions as we go, to avoid adding
2971 /// them to the worklist (this significantly speeds up instcombine on code where
2972 /// many instructions are dead or constant). Additionally, if we find a branch
2973 /// whose condition is a known constant, we only visit the reachable successors.
2975 static bool AddReachableCodeToWorklist(BasicBlock *BB, const DataLayout &DL,
2976 SmallPtrSetImpl<BasicBlock *> &Visited,
2977 InstCombineWorklist &ICWorklist,
2978 const TargetLibraryInfo *TLI) {
2979 bool MadeIRChange = false;
2980 SmallVector<BasicBlock*, 256> Worklist;
2981 Worklist.push_back(BB);
2983 SmallVector<Instruction*, 128> InstrsForInstCombineWorklist;
2984 DenseMap<ConstantExpr*, Constant*> FoldedConstants;
2987 BB = Worklist.pop_back_val();
2989 // We have now visited this block! If we've already been here, ignore it.
2990 if (!Visited.insert(BB).second)
2993 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
2994 Instruction *Inst = &*BBI++;
2996 // DCE instruction if trivially dead.
2997 if (isInstructionTriviallyDead(Inst, TLI)) {
2999 DEBUG(dbgs() << "IC: DCE: " << *Inst << '\n');
3000 Inst->eraseFromParent();
3004 // ConstantProp instruction if trivially constant.
3005 if (!Inst->use_empty() &&
3006 (Inst->getNumOperands() == 0 || isa<Constant>(Inst->getOperand(0))))
3007 if (Constant *C = ConstantFoldInstruction(Inst, DL, TLI)) {
3008 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: "
3010 Inst->replaceAllUsesWith(C);
3012 if (isInstructionTriviallyDead(Inst, TLI))
3013 Inst->eraseFromParent();
3017 // See if we can constant fold its operands.
3018 for (User::op_iterator i = Inst->op_begin(), e = Inst->op_end(); i != e;
3020 ConstantExpr *CE = dyn_cast<ConstantExpr>(i);
3024 Constant *&FoldRes = FoldedConstants[CE];
3026 FoldRes = ConstantFoldConstantExpression(CE, DL, TLI);
3030 if (FoldRes != CE) {
3032 MadeIRChange = true;
3036 InstrsForInstCombineWorklist.push_back(Inst);
3039 // Recursively visit successors. If this is a branch or switch on a
3040 // constant, only visit the reachable successor.
3041 TerminatorInst *TI = BB->getTerminator();
3042 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
3043 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
3044 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
3045 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
3046 Worklist.push_back(ReachableBB);
3049 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
3050 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
3051 // See if this is an explicit destination.
3052 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3054 if (i.getCaseValue() == Cond) {
3055 BasicBlock *ReachableBB = i.getCaseSuccessor();
3056 Worklist.push_back(ReachableBB);
3060 // Otherwise it is the default destination.
3061 Worklist.push_back(SI->getDefaultDest());
3066 for (BasicBlock *SuccBB : TI->successors())
3067 Worklist.push_back(SuccBB);
3068 } while (!Worklist.empty());
3070 // Once we've found all of the instructions to add to instcombine's worklist,
3071 // add them in reverse order. This way instcombine will visit from the top
3072 // of the function down. This jives well with the way that it adds all uses
3073 // of instructions to the worklist after doing a transformation, thus avoiding
3074 // some N^2 behavior in pathological cases.
3075 ICWorklist.AddInitialGroup(InstrsForInstCombineWorklist);
3077 return MadeIRChange;
3080 /// \brief Populate the IC worklist from a function, and prune any dead basic
3081 /// blocks discovered in the process.
3083 /// This also does basic constant propagation and other forward fixing to make
3084 /// the combiner itself run much faster.
3085 static bool prepareICWorklistFromFunction(Function &F, const DataLayout &DL,
3086 TargetLibraryInfo *TLI,
3087 InstCombineWorklist &ICWorklist) {
3088 bool MadeIRChange = false;
3090 // Do a depth-first traversal of the function, populate the worklist with
3091 // the reachable instructions. Ignore blocks that are not reachable. Keep
3092 // track of which blocks we visit.
3093 SmallPtrSet<BasicBlock *, 32> Visited;
3095 AddReachableCodeToWorklist(&F.front(), DL, Visited, ICWorklist, TLI);
3097 // Do a quick scan over the function. If we find any blocks that are
3098 // unreachable, remove any instructions inside of them. This prevents
3099 // the instcombine code from having to deal with some bad special cases.
3100 for (BasicBlock &BB : F) {
3101 if (Visited.count(&BB))
3104 unsigned NumDeadInstInBB = removeAllNonTerminatorAndEHPadInstructions(&BB);
3105 MadeIRChange |= NumDeadInstInBB > 0;
3106 NumDeadInst += NumDeadInstInBB;
3109 return MadeIRChange;
3113 combineInstructionsOverFunction(Function &F, InstCombineWorklist &Worklist,
3114 AliasAnalysis *AA, AssumptionCache &AC,
3115 TargetLibraryInfo &TLI, DominatorTree &DT,
3116 bool ExpensiveCombines = true,
3117 LoopInfo *LI = nullptr) {
3118 auto &DL = F.getParent()->getDataLayout();
3119 ExpensiveCombines |= EnableExpensiveCombines;
3121 /// Builder - This is an IRBuilder that automatically inserts new
3122 /// instructions into the worklist when they are created.
3123 IRBuilder<TargetFolder, InstCombineIRInserter> Builder(
3124 F.getContext(), TargetFolder(DL), InstCombineIRInserter(Worklist, &AC));
3126 // Lower dbg.declare intrinsics otherwise their value may be clobbered
3128 bool DbgDeclaresChanged = LowerDbgDeclare(F);
3130 // Iterate while there is work to do.
3134 DEBUG(dbgs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
3135 << F.getName() << "\n");
3137 bool Changed = prepareICWorklistFromFunction(F, DL, &TLI, Worklist);
3139 InstCombiner IC(Worklist, &Builder, F.optForMinSize(), ExpensiveCombines,
3140 AA, &AC, &TLI, &DT, DL, LI);
3141 Changed |= IC.run();
3147 return DbgDeclaresChanged || Iteration > 1;
3150 PreservedAnalyses InstCombinePass::run(Function &F,
3151 AnalysisManager<Function> &AM) {
3152 auto &AC = AM.getResult<AssumptionAnalysis>(F);
3153 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
3154 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
3156 auto *LI = AM.getCachedResult<LoopAnalysis>(F);
3158 // FIXME: The AliasAnalysis is not yet supported in the new pass manager
3159 if (!combineInstructionsOverFunction(F, Worklist, nullptr, AC, TLI, DT,
3160 ExpensiveCombines, LI))
3161 // No changes, all analyses are preserved.
3162 return PreservedAnalyses::all();
3164 // Mark all the analyses that instcombine updates as preserved.
3165 // FIXME: This should also 'preserve the CFG'.
3166 PreservedAnalyses PA;
3167 PA.preserve<DominatorTreeAnalysis>();
3171 void InstructionCombiningPass::getAnalysisUsage(AnalysisUsage &AU) const {
3172 AU.setPreservesCFG();
3173 AU.addRequired<AAResultsWrapperPass>();
3174 AU.addRequired<AssumptionCacheTracker>();
3175 AU.addRequired<TargetLibraryInfoWrapperPass>();
3176 AU.addRequired<DominatorTreeWrapperPass>();
3177 AU.addPreserved<DominatorTreeWrapperPass>();
3178 AU.addPreserved<AAResultsWrapperPass>();
3179 AU.addPreserved<BasicAAWrapperPass>();
3180 AU.addPreserved<GlobalsAAWrapperPass>();
3183 bool InstructionCombiningPass::runOnFunction(Function &F) {
3184 if (skipFunction(F))
3187 // Required analyses.
3188 auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
3189 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
3190 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
3191 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
3193 // Optional analyses.
3194 auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>();
3195 auto *LI = LIWP ? &LIWP->getLoopInfo() : nullptr;
3197 return combineInstructionsOverFunction(F, Worklist, AA, AC, TLI, DT,
3198 ExpensiveCombines, LI);
3201 char InstructionCombiningPass::ID = 0;
3202 INITIALIZE_PASS_BEGIN(InstructionCombiningPass, "instcombine",
3203 "Combine redundant instructions", false, false)
3204 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
3205 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
3206 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
3207 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
3208 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
3209 INITIALIZE_PASS_END(InstructionCombiningPass, "instcombine",
3210 "Combine redundant instructions", false, false)
3212 // Initialization Routines
3213 void llvm::initializeInstCombine(PassRegistry &Registry) {
3214 initializeInstructionCombiningPassPass(Registry);
3217 void LLVMInitializeInstCombine(LLVMPassRegistryRef R) {
3218 initializeInstructionCombiningPassPass(*unwrap(R));
3221 FunctionPass *llvm::createInstructionCombiningPass(bool ExpensiveCombines) {
3222 return new InstructionCombiningPass(ExpensiveCombines);