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 if (!BinOp1->isBitwiseLogicOp())
183 auto AssocOpcode = BinOp1->getOpcode();
184 auto *BinOp2 = dyn_cast<BinaryOperator>(Cast->getOperand(0));
185 if (!BinOp2 || !BinOp2->hasOneUse() || BinOp2->getOpcode() != AssocOpcode)
189 if (!match(BinOp1->getOperand(1), m_Constant(C1)) ||
190 !match(BinOp2->getOperand(1), m_Constant(C2)))
193 // TODO: This assumes a zext cast.
194 // Eg, if it was a trunc, we'd cast C1 to the source type because casting C2
195 // to the destination type might lose bits.
197 // Fold the constants together in the destination type:
198 // (op (cast (op X, C2)), C1) --> (op (cast X), FoldedC)
199 Type *DestTy = C1->getType();
200 Constant *CastC2 = ConstantExpr::getCast(CastOpcode, C2, DestTy);
201 Constant *FoldedC = ConstantExpr::get(AssocOpcode, C1, CastC2);
202 Cast->setOperand(0, BinOp2->getOperand(0));
203 BinOp1->setOperand(1, FoldedC);
207 /// This performs a few simplifications for operators that are associative or
210 /// Commutative operators:
212 /// 1. Order operands such that they are listed from right (least complex) to
213 /// left (most complex). This puts constants before unary operators before
214 /// binary operators.
216 /// Associative operators:
218 /// 2. Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
219 /// 3. Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
221 /// Associative and commutative operators:
223 /// 4. Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
224 /// 5. Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
225 /// 6. Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
226 /// if C1 and C2 are constants.
227 bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) {
228 Instruction::BinaryOps Opcode = I.getOpcode();
229 bool Changed = false;
232 // Order operands such that they are listed from right (least complex) to
233 // left (most complex). This puts constants before unary operators before
235 if (I.isCommutative() && getComplexity(I.getOperand(0)) <
236 getComplexity(I.getOperand(1)))
237 Changed = !I.swapOperands();
239 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
240 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
242 if (I.isAssociative()) {
243 // Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
244 if (Op0 && Op0->getOpcode() == Opcode) {
245 Value *A = Op0->getOperand(0);
246 Value *B = Op0->getOperand(1);
247 Value *C = I.getOperand(1);
249 // Does "B op C" simplify?
250 if (Value *V = SimplifyBinOp(Opcode, B, C, DL)) {
251 // It simplifies to V. Form "A op V".
254 // Conservatively clear the optional flags, since they may not be
255 // preserved by the reassociation.
256 if (MaintainNoSignedWrap(I, B, C) &&
257 (!Op0 || (isa<BinaryOperator>(Op0) && Op0->hasNoSignedWrap()))) {
258 // Note: this is only valid because SimplifyBinOp doesn't look at
259 // the operands to Op0.
260 I.clearSubclassOptionalData();
261 I.setHasNoSignedWrap(true);
263 ClearSubclassDataAfterReassociation(I);
272 // Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
273 if (Op1 && Op1->getOpcode() == Opcode) {
274 Value *A = I.getOperand(0);
275 Value *B = Op1->getOperand(0);
276 Value *C = Op1->getOperand(1);
278 // Does "A op B" simplify?
279 if (Value *V = SimplifyBinOp(Opcode, A, B, DL)) {
280 // It simplifies to V. Form "V op C".
283 // Conservatively clear the optional flags, since they may not be
284 // preserved by the reassociation.
285 ClearSubclassDataAfterReassociation(I);
293 if (I.isAssociative() && I.isCommutative()) {
294 if (simplifyAssocCastAssoc(&I)) {
300 // Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
301 if (Op0 && Op0->getOpcode() == Opcode) {
302 Value *A = Op0->getOperand(0);
303 Value *B = Op0->getOperand(1);
304 Value *C = I.getOperand(1);
306 // Does "C op A" simplify?
307 if (Value *V = SimplifyBinOp(Opcode, C, A, DL)) {
308 // It simplifies to V. Form "V op B".
311 // Conservatively clear the optional flags, since they may not be
312 // preserved by the reassociation.
313 ClearSubclassDataAfterReassociation(I);
320 // Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
321 if (Op1 && Op1->getOpcode() == Opcode) {
322 Value *A = I.getOperand(0);
323 Value *B = Op1->getOperand(0);
324 Value *C = Op1->getOperand(1);
326 // Does "C op A" simplify?
327 if (Value *V = SimplifyBinOp(Opcode, C, A, DL)) {
328 // It simplifies to V. Form "B op V".
331 // Conservatively clear the optional flags, since they may not be
332 // preserved by the reassociation.
333 ClearSubclassDataAfterReassociation(I);
340 // Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
341 // if C1 and C2 are constants.
343 Op0->getOpcode() == Opcode && Op1->getOpcode() == Opcode &&
344 isa<Constant>(Op0->getOperand(1)) &&
345 isa<Constant>(Op1->getOperand(1)) &&
346 Op0->hasOneUse() && Op1->hasOneUse()) {
347 Value *A = Op0->getOperand(0);
348 Constant *C1 = cast<Constant>(Op0->getOperand(1));
349 Value *B = Op1->getOperand(0);
350 Constant *C2 = cast<Constant>(Op1->getOperand(1));
352 Constant *Folded = ConstantExpr::get(Opcode, C1, C2);
353 BinaryOperator *New = BinaryOperator::Create(Opcode, A, B);
354 if (isa<FPMathOperator>(New)) {
355 FastMathFlags Flags = I.getFastMathFlags();
356 Flags &= Op0->getFastMathFlags();
357 Flags &= Op1->getFastMathFlags();
358 New->setFastMathFlags(Flags);
360 InsertNewInstWith(New, I);
362 I.setOperand(0, New);
363 I.setOperand(1, Folded);
364 // Conservatively clear the optional flags, since they may not be
365 // preserved by the reassociation.
366 ClearSubclassDataAfterReassociation(I);
373 // No further simplifications.
378 /// Return whether "X LOp (Y ROp Z)" is always equal to
379 /// "(X LOp Y) ROp (X LOp Z)".
380 static bool LeftDistributesOverRight(Instruction::BinaryOps LOp,
381 Instruction::BinaryOps ROp) {
386 case Instruction::And:
387 // And distributes over Or and Xor.
391 case Instruction::Or:
392 case Instruction::Xor:
396 case Instruction::Mul:
397 // Multiplication distributes over addition and subtraction.
401 case Instruction::Add:
402 case Instruction::Sub:
406 case Instruction::Or:
407 // Or distributes over And.
411 case Instruction::And:
417 /// Return whether "(X LOp Y) ROp Z" is always equal to
418 /// "(X ROp Z) LOp (Y ROp Z)".
419 static bool RightDistributesOverLeft(Instruction::BinaryOps LOp,
420 Instruction::BinaryOps ROp) {
421 if (Instruction::isCommutative(ROp))
422 return LeftDistributesOverRight(ROp, LOp);
427 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
428 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
429 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
430 case Instruction::And:
431 case Instruction::Or:
432 case Instruction::Xor:
436 case Instruction::Shl:
437 case Instruction::LShr:
438 case Instruction::AShr:
442 // TODO: It would be nice to handle division, aka "(X + Y)/Z = X/Z + Y/Z",
443 // but this requires knowing that the addition does not overflow and other
448 /// This function returns identity value for given opcode, which can be used to
449 /// factor patterns like (X * 2) + X ==> (X * 2) + (X * 1) ==> X * (2 + 1).
450 static Value *getIdentityValue(Instruction::BinaryOps OpCode, Value *V) {
451 if (isa<Constant>(V))
454 if (OpCode == Instruction::Mul)
455 return ConstantInt::get(V->getType(), 1);
457 // TODO: We can handle other cases e.g. Instruction::And, Instruction::Or etc.
462 /// This function factors binary ops which can be combined using distributive
463 /// laws. This function tries to transform 'Op' based TopLevelOpcode to enable
464 /// factorization e.g for ADD(SHL(X , 2), MUL(X, 5)), When this function called
465 /// with TopLevelOpcode == Instruction::Add and Op = SHL(X, 2), transforms
466 /// SHL(X, 2) to MUL(X, 4) i.e. returns Instruction::Mul with LHS set to 'X' and
468 static Instruction::BinaryOps
469 getBinOpsForFactorization(Instruction::BinaryOps TopLevelOpcode,
470 BinaryOperator *Op, Value *&LHS, Value *&RHS) {
472 return Instruction::BinaryOpsEnd;
474 LHS = Op->getOperand(0);
475 RHS = Op->getOperand(1);
477 switch (TopLevelOpcode) {
479 return Op->getOpcode();
481 case Instruction::Add:
482 case Instruction::Sub:
483 if (Op->getOpcode() == Instruction::Shl) {
484 if (Constant *CST = dyn_cast<Constant>(Op->getOperand(1))) {
485 // The multiplier is really 1 << CST.
486 RHS = ConstantExpr::getShl(ConstantInt::get(Op->getType(), 1), CST);
487 return Instruction::Mul;
490 return Op->getOpcode();
493 // TODO: We can add other conversions e.g. shr => div etc.
496 /// This tries to simplify binary operations by factorizing out common terms
497 /// (e. g. "(A*B)+(A*C)" -> "A*(B+C)").
498 static Value *tryFactorization(InstCombiner::BuilderTy *Builder,
499 const DataLayout &DL, BinaryOperator &I,
500 Instruction::BinaryOps InnerOpcode, Value *A,
501 Value *B, Value *C, Value *D) {
503 // If any of A, B, C, D are null, we can not factor I, return early.
504 // Checking A and C should be enough.
505 if (!A || !C || !B || !D)
509 Value *SimplifiedInst = nullptr;
510 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
511 Instruction::BinaryOps TopLevelOpcode = I.getOpcode();
513 // Does "X op' Y" always equal "Y op' X"?
514 bool InnerCommutative = Instruction::isCommutative(InnerOpcode);
516 // Does "X op' (Y op Z)" always equal "(X op' Y) op (X op' Z)"?
517 if (LeftDistributesOverRight(InnerOpcode, TopLevelOpcode))
518 // Does the instruction have the form "(A op' B) op (A op' D)" or, in the
519 // commutative case, "(A op' B) op (C op' A)"?
520 if (A == C || (InnerCommutative && A == D)) {
523 // Consider forming "A op' (B op D)".
524 // If "B op D" simplifies then it can be formed with no cost.
525 V = SimplifyBinOp(TopLevelOpcode, B, D, DL);
526 // If "B op D" doesn't simplify then only go on if both of the existing
527 // operations "A op' B" and "C op' D" will be zapped as no longer used.
528 if (!V && LHS->hasOneUse() && RHS->hasOneUse())
529 V = Builder->CreateBinOp(TopLevelOpcode, B, D, RHS->getName());
531 SimplifiedInst = Builder->CreateBinOp(InnerOpcode, A, V);
535 // Does "(X op Y) op' Z" always equal "(X op' Z) op (Y op' Z)"?
536 if (!SimplifiedInst && RightDistributesOverLeft(TopLevelOpcode, InnerOpcode))
537 // Does the instruction have the form "(A op' B) op (C op' B)" or, in the
538 // commutative case, "(A op' B) op (B op' D)"?
539 if (B == D || (InnerCommutative && B == C)) {
542 // Consider forming "(A op C) op' B".
543 // If "A op C" simplifies then it can be formed with no cost.
544 V = SimplifyBinOp(TopLevelOpcode, A, C, DL);
546 // If "A op C" doesn't simplify then only go on if both of the existing
547 // operations "A op' B" and "C op' D" will be zapped as no longer used.
548 if (!V && LHS->hasOneUse() && RHS->hasOneUse())
549 V = Builder->CreateBinOp(TopLevelOpcode, A, C, LHS->getName());
551 SimplifiedInst = Builder->CreateBinOp(InnerOpcode, V, B);
555 if (SimplifiedInst) {
557 SimplifiedInst->takeName(&I);
559 // Check if we can add NSW flag to SimplifiedInst. If so, set NSW flag.
560 // TODO: Check for NUW.
561 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(SimplifiedInst)) {
562 if (isa<OverflowingBinaryOperator>(SimplifiedInst)) {
564 if (isa<OverflowingBinaryOperator>(&I))
565 HasNSW = I.hasNoSignedWrap();
567 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS))
568 if (isa<OverflowingBinaryOperator>(Op0))
569 HasNSW &= Op0->hasNoSignedWrap();
571 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS))
572 if (isa<OverflowingBinaryOperator>(Op1))
573 HasNSW &= Op1->hasNoSignedWrap();
575 // We can propagate 'nsw' if we know that
576 // %Y = mul nsw i16 %X, C
577 // %Z = add nsw i16 %Y, %X
579 // %Z = mul nsw i16 %X, C+1
581 // iff C+1 isn't INT_MIN
583 if (TopLevelOpcode == Instruction::Add &&
584 InnerOpcode == Instruction::Mul)
585 if (match(V, m_APInt(CInt)) && !CInt->isMinSignedValue())
586 BO->setHasNoSignedWrap(HasNSW);
590 return SimplifiedInst;
593 /// This tries to simplify binary operations which some other binary operation
594 /// distributes over either by factorizing out common terms
595 /// (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this results in
596 /// simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is a win).
597 /// Returns the simplified value, or null if it didn't simplify.
598 Value *InstCombiner::SimplifyUsingDistributiveLaws(BinaryOperator &I) {
599 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
600 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
601 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
604 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
605 auto TopLevelOpcode = I.getOpcode();
606 auto LHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op0, A, B);
607 auto RHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op1, C, D);
609 // The instruction has the form "(A op' B) op (C op' D)". Try to factorize
611 if (LHSOpcode == RHSOpcode) {
612 if (Value *V = tryFactorization(Builder, DL, I, LHSOpcode, A, B, C, D))
616 // The instruction has the form "(A op' B) op (C)". Try to factorize common
618 if (Value *V = tryFactorization(Builder, DL, I, LHSOpcode, A, B, RHS,
619 getIdentityValue(LHSOpcode, RHS)))
622 // The instruction has the form "(B) op (C op' D)". Try to factorize common
624 if (Value *V = tryFactorization(Builder, DL, I, RHSOpcode, LHS,
625 getIdentityValue(RHSOpcode, LHS), C, D))
629 if (Op0 && RightDistributesOverLeft(Op0->getOpcode(), TopLevelOpcode)) {
630 // The instruction has the form "(A op' B) op C". See if expanding it out
631 // to "(A op C) op' (B op C)" results in simplifications.
632 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
633 Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op'
635 // Do "A op C" and "B op C" both simplify?
636 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, C, DL))
637 if (Value *R = SimplifyBinOp(TopLevelOpcode, B, C, DL)) {
638 // They do! Return "L op' R".
640 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
641 if ((L == A && R == B) ||
642 (Instruction::isCommutative(InnerOpcode) && L == B && R == A))
644 // Otherwise return "L op' R" if it simplifies.
645 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, DL))
647 // Otherwise, create a new instruction.
648 C = Builder->CreateBinOp(InnerOpcode, L, R);
654 if (Op1 && LeftDistributesOverRight(TopLevelOpcode, Op1->getOpcode())) {
655 // The instruction has the form "A op (B op' C)". See if expanding it out
656 // to "(A op B) op' (A op C)" results in simplifications.
657 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
658 Instruction::BinaryOps InnerOpcode = Op1->getOpcode(); // op'
660 // Do "A op B" and "A op C" both simplify?
661 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, B, DL))
662 if (Value *R = SimplifyBinOp(TopLevelOpcode, A, C, DL)) {
663 // They do! Return "L op' R".
665 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
666 if ((L == B && R == C) ||
667 (Instruction::isCommutative(InnerOpcode) && L == C && R == B))
669 // Otherwise return "L op' R" if it simplifies.
670 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, DL))
672 // Otherwise, create a new instruction.
673 A = Builder->CreateBinOp(InnerOpcode, L, R);
679 // (op (select (a, c, b)), (select (a, d, b))) -> (select (a, (op c, d), 0))
680 // (op (select (a, b, c)), (select (a, b, d))) -> (select (a, 0, (op c, d)))
681 if (auto *SI0 = dyn_cast<SelectInst>(LHS)) {
682 if (auto *SI1 = dyn_cast<SelectInst>(RHS)) {
683 if (SI0->getCondition() == SI1->getCondition()) {
685 if (Value *V = SimplifyBinOp(TopLevelOpcode, SI0->getFalseValue(),
686 SI1->getFalseValue(), DL, &TLI, &DT, &AC))
687 SI = Builder->CreateSelect(SI0->getCondition(),
688 Builder->CreateBinOp(TopLevelOpcode,
690 SI1->getTrueValue()),
692 if (Value *V = SimplifyBinOp(TopLevelOpcode, SI0->getTrueValue(),
693 SI1->getTrueValue(), DL, &TLI, &DT, &AC))
694 SI = Builder->CreateSelect(
695 SI0->getCondition(), V,
696 Builder->CreateBinOp(TopLevelOpcode, SI0->getFalseValue(),
697 SI1->getFalseValue()));
709 /// Given a 'sub' instruction, return the RHS of the instruction if the LHS is a
710 /// constant zero (which is the 'negate' form).
711 Value *InstCombiner::dyn_castNegVal(Value *V) const {
712 if (BinaryOperator::isNeg(V))
713 return BinaryOperator::getNegArgument(V);
715 // Constants can be considered to be negated values if they can be folded.
716 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
717 return ConstantExpr::getNeg(C);
719 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
720 if (C->getType()->getElementType()->isIntegerTy())
721 return ConstantExpr::getNeg(C);
726 /// Given a 'fsub' instruction, return the RHS of the instruction if the LHS is
727 /// a constant negative zero (which is the 'negate' form).
728 Value *InstCombiner::dyn_castFNegVal(Value *V, bool IgnoreZeroSign) const {
729 if (BinaryOperator::isFNeg(V, IgnoreZeroSign))
730 return BinaryOperator::getFNegArgument(V);
732 // Constants can be considered to be negated values if they can be folded.
733 if (ConstantFP *C = dyn_cast<ConstantFP>(V))
734 return ConstantExpr::getFNeg(C);
736 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
737 if (C->getType()->getElementType()->isFloatingPointTy())
738 return ConstantExpr::getFNeg(C);
743 static Value *foldOperationIntoSelectOperand(Instruction &I, Value *SO,
745 if (auto *Cast = dyn_cast<CastInst>(&I))
746 return IC->Builder->CreateCast(Cast->getOpcode(), SO, I.getType());
748 assert(I.isBinaryOp() && "Unexpected opcode for select folding");
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 (auto *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 auto *BO = cast<BinaryOperator>(&I);
765 Value *RI = IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1,
766 SO->getName() + ".op");
767 auto *FPInst = dyn_cast<Instruction>(RI);
768 if (FPInst && isa<FPMathOperator>(FPInst))
769 FPInst->copyFastMathFlags(BO);
773 /// Given an instruction with a select as one operand and a constant as the
774 /// other operand, try to fold the binary operator into the select arguments.
775 /// This also works for Cast instructions, which obviously do not have a second
777 Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) {
778 // Don't modify shared select instructions.
779 if (!SI->hasOneUse())
782 Value *TV = SI->getTrueValue();
783 Value *FV = SI->getFalseValue();
784 if (!(isa<Constant>(TV) || isa<Constant>(FV)))
787 // Bool selects with constant operands can be folded to logical ops.
788 if (SI->getType()->getScalarType()->isIntegerTy(1))
791 // If it's a bitcast involving vectors, make sure it has the same number of
792 // elements on both sides.
793 if (auto *BC = dyn_cast<BitCastInst>(&Op)) {
794 VectorType *DestTy = dyn_cast<VectorType>(BC->getDestTy());
795 VectorType *SrcTy = dyn_cast<VectorType>(BC->getSrcTy());
797 // Verify that either both or neither are vectors.
798 if ((SrcTy == nullptr) != (DestTy == nullptr))
801 // If vectors, verify that they have the same number of elements.
802 if (SrcTy && SrcTy->getNumElements() != DestTy->getNumElements())
806 // Test if a CmpInst instruction is used exclusively by a select as
807 // part of a minimum or maximum operation. If so, refrain from doing
808 // any other folding. This helps out other analyses which understand
809 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
810 // and CodeGen. And in this case, at least one of the comparison
811 // operands has at least one user besides the compare (the select),
812 // which would often largely negate the benefit of folding anyway.
813 if (auto *CI = dyn_cast<CmpInst>(SI->getCondition())) {
814 if (CI->hasOneUse()) {
815 Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1);
816 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
817 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
822 Value *NewTV = foldOperationIntoSelectOperand(Op, TV, this);
823 Value *NewFV = foldOperationIntoSelectOperand(Op, FV, this);
824 return SelectInst::Create(SI->getCondition(), NewTV, NewFV, "", nullptr, SI);
827 /// Given a binary operator, cast instruction, or select which has a PHI node as
828 /// operand #0, see if we can fold the instruction into the PHI (which is only
829 /// possible if all operands to the PHI are constants).
830 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
831 PHINode *PN = cast<PHINode>(I.getOperand(0));
832 unsigned NumPHIValues = PN->getNumIncomingValues();
833 if (NumPHIValues == 0)
836 // We normally only transform phis with a single use. However, if a PHI has
837 // multiple uses and they are all the same operation, we can fold *all* of the
838 // uses into the PHI.
839 if (!PN->hasOneUse()) {
840 // Walk the use list for the instruction, comparing them to I.
841 for (User *U : PN->users()) {
842 Instruction *UI = cast<Instruction>(U);
843 if (UI != &I && !I.isIdenticalTo(UI))
846 // Otherwise, we can replace *all* users with the new PHI we form.
849 // Check to see if all of the operands of the PHI are simple constants
850 // (constantint/constantfp/undef). If there is one non-constant value,
851 // remember the BB it is in. If there is more than one or if *it* is a PHI,
852 // bail out. We don't do arbitrary constant expressions here because moving
853 // their computation can be expensive without a cost model.
854 BasicBlock *NonConstBB = nullptr;
855 for (unsigned i = 0; i != NumPHIValues; ++i) {
856 Value *InVal = PN->getIncomingValue(i);
857 if (isa<Constant>(InVal) && !isa<ConstantExpr>(InVal))
860 if (isa<PHINode>(InVal)) return nullptr; // Itself a phi.
861 if (NonConstBB) return nullptr; // More than one non-const value.
863 NonConstBB = PN->getIncomingBlock(i);
865 // If the InVal is an invoke at the end of the pred block, then we can't
866 // insert a computation after it without breaking the edge.
867 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal))
868 if (II->getParent() == NonConstBB)
871 // If the incoming non-constant value is in I's block, we will remove one
872 // instruction, but insert another equivalent one, leading to infinite
874 if (isPotentiallyReachable(I.getParent(), NonConstBB, &DT, LI))
878 // If there is exactly one non-constant value, we can insert a copy of the
879 // operation in that block. However, if this is a critical edge, we would be
880 // inserting the computation on some other paths (e.g. inside a loop). Only
881 // do this if the pred block is unconditionally branching into the phi block.
882 if (NonConstBB != nullptr) {
883 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
884 if (!BI || !BI->isUnconditional()) return nullptr;
887 // Okay, we can do the transformation: create the new PHI node.
888 PHINode *NewPN = PHINode::Create(I.getType(), PN->getNumIncomingValues());
889 InsertNewInstBefore(NewPN, *PN);
892 // If we are going to have to insert a new computation, do so right before the
893 // predecessor's terminator.
895 Builder->SetInsertPoint(NonConstBB->getTerminator());
897 // Next, add all of the operands to the PHI.
898 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) {
899 // We only currently try to fold the condition of a select when it is a phi,
900 // not the true/false values.
901 Value *TrueV = SI->getTrueValue();
902 Value *FalseV = SI->getFalseValue();
903 BasicBlock *PhiTransBB = PN->getParent();
904 for (unsigned i = 0; i != NumPHIValues; ++i) {
905 BasicBlock *ThisBB = PN->getIncomingBlock(i);
906 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB);
907 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB);
908 Value *InV = nullptr;
909 // Beware of ConstantExpr: it may eventually evaluate to getNullValue,
910 // even if currently isNullValue gives false.
911 Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i));
912 if (InC && !isa<ConstantExpr>(InC))
913 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred;
915 InV = Builder->CreateSelect(PN->getIncomingValue(i),
916 TrueVInPred, FalseVInPred, "phitmp");
917 NewPN->addIncoming(InV, ThisBB);
919 } else if (CmpInst *CI = dyn_cast<CmpInst>(&I)) {
920 Constant *C = cast<Constant>(I.getOperand(1));
921 for (unsigned i = 0; i != NumPHIValues; ++i) {
922 Value *InV = nullptr;
923 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
924 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
925 else if (isa<ICmpInst>(CI))
926 InV = Builder->CreateICmp(CI->getPredicate(), PN->getIncomingValue(i),
929 InV = Builder->CreateFCmp(CI->getPredicate(), PN->getIncomingValue(i),
931 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
933 } else if (I.getNumOperands() == 2) {
934 Constant *C = cast<Constant>(I.getOperand(1));
935 for (unsigned i = 0; i != NumPHIValues; ++i) {
936 Value *InV = nullptr;
937 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
938 InV = ConstantExpr::get(I.getOpcode(), InC, C);
940 InV = Builder->CreateBinOp(cast<BinaryOperator>(I).getOpcode(),
941 PN->getIncomingValue(i), C, "phitmp");
942 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
945 CastInst *CI = cast<CastInst>(&I);
946 Type *RetTy = CI->getType();
947 for (unsigned i = 0; i != NumPHIValues; ++i) {
949 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
950 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
952 InV = Builder->CreateCast(CI->getOpcode(),
953 PN->getIncomingValue(i), I.getType(), "phitmp");
954 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
958 for (auto UI = PN->user_begin(), E = PN->user_end(); UI != E;) {
959 Instruction *User = cast<Instruction>(*UI++);
960 if (User == &I) continue;
961 replaceInstUsesWith(*User, NewPN);
962 eraseInstFromFunction(*User);
964 return replaceInstUsesWith(I, NewPN);
967 /// Given a pointer type and a constant offset, determine whether or not there
968 /// is a sequence of GEP indices into the pointed type that will land us at the
969 /// specified offset. If so, fill them into NewIndices and return the resultant
970 /// element type, otherwise return null.
971 Type *InstCombiner::FindElementAtOffset(PointerType *PtrTy, int64_t Offset,
972 SmallVectorImpl<Value *> &NewIndices) {
973 Type *Ty = PtrTy->getElementType();
977 // Start with the index over the outer type. Note that the type size
978 // might be zero (even if the offset isn't zero) if the indexed type
979 // is something like [0 x {int, int}]
980 Type *IntPtrTy = DL.getIntPtrType(PtrTy);
981 int64_t FirstIdx = 0;
982 if (int64_t TySize = DL.getTypeAllocSize(Ty)) {
983 FirstIdx = Offset/TySize;
984 Offset -= FirstIdx*TySize;
986 // Handle hosts where % returns negative instead of values [0..TySize).
992 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
995 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
997 // Index into the types. If we fail, set OrigBase to null.
999 // Indexing into tail padding between struct/array elements.
1000 if (uint64_t(Offset * 8) >= DL.getTypeSizeInBits(Ty))
1003 if (StructType *STy = dyn_cast<StructType>(Ty)) {
1004 const StructLayout *SL = DL.getStructLayout(STy);
1005 assert(Offset < (int64_t)SL->getSizeInBytes() &&
1006 "Offset must stay within the indexed type");
1008 unsigned Elt = SL->getElementContainingOffset(Offset);
1009 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
1012 Offset -= SL->getElementOffset(Elt);
1013 Ty = STy->getElementType(Elt);
1014 } else if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
1015 uint64_t EltSize = DL.getTypeAllocSize(AT->getElementType());
1016 assert(EltSize && "Cannot index into a zero-sized array");
1017 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
1019 Ty = AT->getElementType();
1021 // Otherwise, we can't index into the middle of this atomic type, bail.
1029 static bool shouldMergeGEPs(GEPOperator &GEP, GEPOperator &Src) {
1030 // If this GEP has only 0 indices, it is the same pointer as
1031 // Src. If Src is not a trivial GEP too, don't combine
1033 if (GEP.hasAllZeroIndices() && !Src.hasAllZeroIndices() &&
1039 /// Return a value X such that Val = X * Scale, or null if none.
1040 /// If the multiplication is known not to overflow, then NoSignedWrap is set.
1041 Value *InstCombiner::Descale(Value *Val, APInt Scale, bool &NoSignedWrap) {
1042 assert(isa<IntegerType>(Val->getType()) && "Can only descale integers!");
1043 assert(cast<IntegerType>(Val->getType())->getBitWidth() ==
1044 Scale.getBitWidth() && "Scale not compatible with value!");
1046 // If Val is zero or Scale is one then Val = Val * Scale.
1047 if (match(Val, m_Zero()) || Scale == 1) {
1048 NoSignedWrap = true;
1052 // If Scale is zero then it does not divide Val.
1053 if (Scale.isMinValue())
1056 // Look through chains of multiplications, searching for a constant that is
1057 // divisible by Scale. For example, descaling X*(Y*(Z*4)) by a factor of 4
1058 // will find the constant factor 4 and produce X*(Y*Z). Descaling X*(Y*8) by
1059 // a factor of 4 will produce X*(Y*2). The principle of operation is to bore
1062 // Val = M1 * X || Analysis starts here and works down
1063 // M1 = M2 * Y || Doesn't descend into terms with more
1064 // M2 = Z * 4 \/ than one use
1066 // Then to modify a term at the bottom:
1069 // M1 = Z * Y || Replaced M2 with Z
1071 // Then to work back up correcting nsw flags.
1073 // Op - the term we are currently analyzing. Starts at Val then drills down.
1074 // Replaced with its descaled value before exiting from the drill down loop.
1077 // Parent - initially null, but after drilling down notes where Op came from.
1078 // In the example above, Parent is (Val, 0) when Op is M1, because M1 is the
1079 // 0'th operand of Val.
1080 std::pair<Instruction*, unsigned> Parent;
1082 // Set if the transform requires a descaling at deeper levels that doesn't
1084 bool RequireNoSignedWrap = false;
1086 // Log base 2 of the scale. Negative if not a power of 2.
1087 int32_t logScale = Scale.exactLogBase2();
1089 for (;; Op = Parent.first->getOperand(Parent.second)) { // Drill down
1091 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
1092 // If Op is a constant divisible by Scale then descale to the quotient.
1093 APInt Quotient(Scale), Remainder(Scale); // Init ensures right bitwidth.
1094 APInt::sdivrem(CI->getValue(), Scale, Quotient, Remainder);
1095 if (!Remainder.isMinValue())
1096 // Not divisible by Scale.
1098 // Replace with the quotient in the parent.
1099 Op = ConstantInt::get(CI->getType(), Quotient);
1100 NoSignedWrap = true;
1104 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op)) {
1106 if (BO->getOpcode() == Instruction::Mul) {
1108 NoSignedWrap = BO->hasNoSignedWrap();
1109 if (RequireNoSignedWrap && !NoSignedWrap)
1112 // There are three cases for multiplication: multiplication by exactly
1113 // the scale, multiplication by a constant different to the scale, and
1114 // multiplication by something else.
1115 Value *LHS = BO->getOperand(0);
1116 Value *RHS = BO->getOperand(1);
1118 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
1119 // Multiplication by a constant.
1120 if (CI->getValue() == Scale) {
1121 // Multiplication by exactly the scale, replace the multiplication
1122 // by its left-hand side in the parent.
1127 // Otherwise drill down into the constant.
1128 if (!Op->hasOneUse())
1131 Parent = std::make_pair(BO, 1);
1135 // Multiplication by something else. Drill down into the left-hand side
1136 // since that's where the reassociate pass puts the good stuff.
1137 if (!Op->hasOneUse())
1140 Parent = std::make_pair(BO, 0);
1144 if (logScale > 0 && BO->getOpcode() == Instruction::Shl &&
1145 isa<ConstantInt>(BO->getOperand(1))) {
1146 // Multiplication by a power of 2.
1147 NoSignedWrap = BO->hasNoSignedWrap();
1148 if (RequireNoSignedWrap && !NoSignedWrap)
1151 Value *LHS = BO->getOperand(0);
1152 int32_t Amt = cast<ConstantInt>(BO->getOperand(1))->
1153 getLimitedValue(Scale.getBitWidth());
1156 if (Amt == logScale) {
1157 // Multiplication by exactly the scale, replace the multiplication
1158 // by its left-hand side in the parent.
1162 if (Amt < logScale || !Op->hasOneUse())
1165 // Multiplication by more than the scale. Reduce the multiplying amount
1166 // by the scale in the parent.
1167 Parent = std::make_pair(BO, 1);
1168 Op = ConstantInt::get(BO->getType(), Amt - logScale);
1173 if (!Op->hasOneUse())
1176 if (CastInst *Cast = dyn_cast<CastInst>(Op)) {
1177 if (Cast->getOpcode() == Instruction::SExt) {
1178 // Op is sign-extended from a smaller type, descale in the smaller type.
1179 unsigned SmallSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1180 APInt SmallScale = Scale.trunc(SmallSize);
1181 // Suppose Op = sext X, and we descale X as Y * SmallScale. We want to
1182 // descale Op as (sext Y) * Scale. In order to have
1183 // sext (Y * SmallScale) = (sext Y) * Scale
1184 // some conditions need to hold however: SmallScale must sign-extend to
1185 // Scale and the multiplication Y * SmallScale should not overflow.
1186 if (SmallScale.sext(Scale.getBitWidth()) != Scale)
1187 // SmallScale does not sign-extend to Scale.
1189 assert(SmallScale.exactLogBase2() == logScale);
1190 // Require that Y * SmallScale must not overflow.
1191 RequireNoSignedWrap = true;
1193 // Drill down through the cast.
1194 Parent = std::make_pair(Cast, 0);
1199 if (Cast->getOpcode() == Instruction::Trunc) {
1200 // Op is truncated from a larger type, descale in the larger type.
1201 // Suppose Op = trunc X, and we descale X as Y * sext Scale. Then
1202 // trunc (Y * sext Scale) = (trunc Y) * Scale
1203 // always holds. However (trunc Y) * Scale may overflow even if
1204 // trunc (Y * sext Scale) does not, so nsw flags need to be cleared
1205 // from this point up in the expression (see later).
1206 if (RequireNoSignedWrap)
1209 // Drill down through the cast.
1210 unsigned LargeSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1211 Parent = std::make_pair(Cast, 0);
1212 Scale = Scale.sext(LargeSize);
1213 if (logScale + 1 == (int32_t)Cast->getType()->getPrimitiveSizeInBits())
1215 assert(Scale.exactLogBase2() == logScale);
1220 // Unsupported expression, bail out.
1224 // If Op is zero then Val = Op * Scale.
1225 if (match(Op, m_Zero())) {
1226 NoSignedWrap = true;
1230 // We know that we can successfully descale, so from here on we can safely
1231 // modify the IR. Op holds the descaled version of the deepest term in the
1232 // expression. NoSignedWrap is 'true' if multiplying Op by Scale is known
1236 // The expression only had one term.
1239 // Rewrite the parent using the descaled version of its operand.
1240 assert(Parent.first->hasOneUse() && "Drilled down when more than one use!");
1241 assert(Op != Parent.first->getOperand(Parent.second) &&
1242 "Descaling was a no-op?");
1243 Parent.first->setOperand(Parent.second, Op);
1244 Worklist.Add(Parent.first);
1246 // Now work back up the expression correcting nsw flags. The logic is based
1247 // on the following observation: if X * Y is known not to overflow as a signed
1248 // multiplication, and Y is replaced by a value Z with smaller absolute value,
1249 // then X * Z will not overflow as a signed multiplication either. As we work
1250 // our way up, having NoSignedWrap 'true' means that the descaled value at the
1251 // current level has strictly smaller absolute value than the original.
1252 Instruction *Ancestor = Parent.first;
1254 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Ancestor)) {
1255 // If the multiplication wasn't nsw then we can't say anything about the
1256 // value of the descaled multiplication, and we have to clear nsw flags
1257 // from this point on up.
1258 bool OpNoSignedWrap = BO->hasNoSignedWrap();
1259 NoSignedWrap &= OpNoSignedWrap;
1260 if (NoSignedWrap != OpNoSignedWrap) {
1261 BO->setHasNoSignedWrap(NoSignedWrap);
1262 Worklist.Add(Ancestor);
1264 } else if (Ancestor->getOpcode() == Instruction::Trunc) {
1265 // The fact that the descaled input to the trunc has smaller absolute
1266 // value than the original input doesn't tell us anything useful about
1267 // the absolute values of the truncations.
1268 NoSignedWrap = false;
1270 assert((Ancestor->getOpcode() != Instruction::SExt || NoSignedWrap) &&
1271 "Failed to keep proper track of nsw flags while drilling down?");
1273 if (Ancestor == Val)
1274 // Got to the top, all done!
1277 // Move up one level in the expression.
1278 assert(Ancestor->hasOneUse() && "Drilled down when more than one use!");
1279 Ancestor = Ancestor->user_back();
1283 /// \brief Creates node of binary operation with the same attributes as the
1284 /// specified one but with other operands.
1285 static Value *CreateBinOpAsGiven(BinaryOperator &Inst, Value *LHS, Value *RHS,
1286 InstCombiner::BuilderTy *B) {
1287 Value *BO = B->CreateBinOp(Inst.getOpcode(), LHS, RHS);
1288 // If LHS and RHS are constant, BO won't be a binary operator.
1289 if (BinaryOperator *NewBO = dyn_cast<BinaryOperator>(BO))
1290 NewBO->copyIRFlags(&Inst);
1294 /// \brief Makes transformation of binary operation specific for vector types.
1295 /// \param Inst Binary operator to transform.
1296 /// \return Pointer to node that must replace the original binary operator, or
1297 /// null pointer if no transformation was made.
1298 Value *InstCombiner::SimplifyVectorOp(BinaryOperator &Inst) {
1299 if (!Inst.getType()->isVectorTy()) return nullptr;
1301 // It may not be safe to reorder shuffles and things like div, urem, etc.
1302 // because we may trap when executing those ops on unknown vector elements.
1304 if (!isSafeToSpeculativelyExecute(&Inst))
1307 unsigned VWidth = cast<VectorType>(Inst.getType())->getNumElements();
1308 Value *LHS = Inst.getOperand(0), *RHS = Inst.getOperand(1);
1309 assert(cast<VectorType>(LHS->getType())->getNumElements() == VWidth);
1310 assert(cast<VectorType>(RHS->getType())->getNumElements() == VWidth);
1312 // If both arguments of binary operation are shuffles, which use the same
1313 // mask and shuffle within a single vector, it is worthwhile to move the
1314 // shuffle after binary operation:
1315 // Op(shuffle(v1, m), shuffle(v2, m)) -> shuffle(Op(v1, v2), m)
1316 if (isa<ShuffleVectorInst>(LHS) && isa<ShuffleVectorInst>(RHS)) {
1317 ShuffleVectorInst *LShuf = cast<ShuffleVectorInst>(LHS);
1318 ShuffleVectorInst *RShuf = cast<ShuffleVectorInst>(RHS);
1319 if (isa<UndefValue>(LShuf->getOperand(1)) &&
1320 isa<UndefValue>(RShuf->getOperand(1)) &&
1321 LShuf->getOperand(0)->getType() == RShuf->getOperand(0)->getType() &&
1322 LShuf->getMask() == RShuf->getMask()) {
1323 Value *NewBO = CreateBinOpAsGiven(Inst, LShuf->getOperand(0),
1324 RShuf->getOperand(0), Builder);
1325 return Builder->CreateShuffleVector(NewBO,
1326 UndefValue::get(NewBO->getType()), LShuf->getMask());
1330 // If one argument is a shuffle within one vector, the other is a constant,
1331 // try moving the shuffle after the binary operation.
1332 ShuffleVectorInst *Shuffle = nullptr;
1333 Constant *C1 = nullptr;
1334 if (isa<ShuffleVectorInst>(LHS)) Shuffle = cast<ShuffleVectorInst>(LHS);
1335 if (isa<ShuffleVectorInst>(RHS)) Shuffle = cast<ShuffleVectorInst>(RHS);
1336 if (isa<Constant>(LHS)) C1 = cast<Constant>(LHS);
1337 if (isa<Constant>(RHS)) C1 = cast<Constant>(RHS);
1338 if (Shuffle && C1 &&
1339 (isa<ConstantVector>(C1) || isa<ConstantDataVector>(C1)) &&
1340 isa<UndefValue>(Shuffle->getOperand(1)) &&
1341 Shuffle->getType() == Shuffle->getOperand(0)->getType()) {
1342 SmallVector<int, 16> ShMask = Shuffle->getShuffleMask();
1343 // Find constant C2 that has property:
1344 // shuffle(C2, ShMask) = C1
1345 // If such constant does not exist (example: ShMask=<0,0> and C1=<1,2>)
1346 // reorder is not possible.
1347 SmallVector<Constant*, 16> C2M(VWidth,
1348 UndefValue::get(C1->getType()->getScalarType()));
1349 bool MayChange = true;
1350 for (unsigned I = 0; I < VWidth; ++I) {
1351 if (ShMask[I] >= 0) {
1352 assert(ShMask[I] < (int)VWidth);
1353 if (!isa<UndefValue>(C2M[ShMask[I]])) {
1357 C2M[ShMask[I]] = C1->getAggregateElement(I);
1361 Constant *C2 = ConstantVector::get(C2M);
1362 Value *NewLHS = isa<Constant>(LHS) ? C2 : Shuffle->getOperand(0);
1363 Value *NewRHS = isa<Constant>(LHS) ? Shuffle->getOperand(0) : C2;
1364 Value *NewBO = CreateBinOpAsGiven(Inst, NewLHS, NewRHS, Builder);
1365 return Builder->CreateShuffleVector(NewBO,
1366 UndefValue::get(Inst.getType()), Shuffle->getMask());
1373 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
1374 SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end());
1377 SimplifyGEPInst(GEP.getSourceElementType(), Ops, DL, &TLI, &DT, &AC))
1378 return replaceInstUsesWith(GEP, V);
1380 Value *PtrOp = GEP.getOperand(0);
1382 // Eliminate unneeded casts for indices, and replace indices which displace
1383 // by multiples of a zero size type with zero.
1384 bool MadeChange = false;
1386 DL.getIntPtrType(GEP.getPointerOperandType()->getScalarType());
1388 gep_type_iterator GTI = gep_type_begin(GEP);
1389 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end(); I != E;
1391 // Skip indices into struct types.
1395 // Index type should have the same width as IntPtr
1396 Type *IndexTy = (*I)->getType();
1397 Type *NewIndexType = IndexTy->isVectorTy() ?
1398 VectorType::get(IntPtrTy, IndexTy->getVectorNumElements()) : IntPtrTy;
1400 // If the element type has zero size then any index over it is equivalent
1401 // to an index of zero, so replace it with zero if it is not zero already.
1402 Type *EltTy = GTI.getIndexedType();
1403 if (EltTy->isSized() && DL.getTypeAllocSize(EltTy) == 0)
1404 if (!isa<Constant>(*I) || !cast<Constant>(*I)->isNullValue()) {
1405 *I = Constant::getNullValue(NewIndexType);
1409 if (IndexTy != NewIndexType) {
1410 // If we are using a wider index than needed for this platform, shrink
1411 // it to what we need. If narrower, sign-extend it to what we need.
1412 // This explicit cast can make subsequent optimizations more obvious.
1413 *I = Builder->CreateIntCast(*I, NewIndexType, true);
1420 // Check to see if the inputs to the PHI node are getelementptr instructions.
1421 if (PHINode *PN = dyn_cast<PHINode>(PtrOp)) {
1422 GetElementPtrInst *Op1 = dyn_cast<GetElementPtrInst>(PN->getOperand(0));
1426 // Don't fold a GEP into itself through a PHI node. This can only happen
1427 // through the back-edge of a loop. Folding a GEP into itself means that
1428 // the value of the previous iteration needs to be stored in the meantime,
1429 // thus requiring an additional register variable to be live, but not
1430 // actually achieving anything (the GEP still needs to be executed once per
1437 for (auto I = PN->op_begin()+1, E = PN->op_end(); I !=E; ++I) {
1438 GetElementPtrInst *Op2 = dyn_cast<GetElementPtrInst>(*I);
1439 if (!Op2 || Op1->getNumOperands() != Op2->getNumOperands())
1442 // As for Op1 above, don't try to fold a GEP into itself.
1446 // Keep track of the type as we walk the GEP.
1447 Type *CurTy = nullptr;
1449 for (unsigned J = 0, F = Op1->getNumOperands(); J != F; ++J) {
1450 if (Op1->getOperand(J)->getType() != Op2->getOperand(J)->getType())
1453 if (Op1->getOperand(J) != Op2->getOperand(J)) {
1455 // We have not seen any differences yet in the GEPs feeding the
1456 // PHI yet, so we record this one if it is allowed to be a
1459 // The first two arguments can vary for any GEP, the rest have to be
1460 // static for struct slots
1461 if (J > 1 && CurTy->isStructTy())
1466 // The GEP is different by more than one input. While this could be
1467 // extended to support GEPs that vary by more than one variable it
1468 // doesn't make sense since it greatly increases the complexity and
1469 // would result in an R+R+R addressing mode which no backend
1470 // directly supports and would need to be broken into several
1471 // simpler instructions anyway.
1476 // Sink down a layer of the type for the next iteration.
1479 CurTy = Op1->getSourceElementType();
1480 } else if (CompositeType *CT = dyn_cast<CompositeType>(CurTy)) {
1481 CurTy = CT->getTypeAtIndex(Op1->getOperand(J));
1489 // If not all GEPs are identical we'll have to create a new PHI node.
1490 // Check that the old PHI node has only one use so that it will get
1492 if (DI != -1 && !PN->hasOneUse())
1495 GetElementPtrInst *NewGEP = cast<GetElementPtrInst>(Op1->clone());
1497 // All the GEPs feeding the PHI are identical. Clone one down into our
1498 // BB so that it can be merged with the current GEP.
1499 GEP.getParent()->getInstList().insert(
1500 GEP.getParent()->getFirstInsertionPt(), NewGEP);
1502 // All the GEPs feeding the PHI differ at a single offset. Clone a GEP
1503 // into the current block so it can be merged, and create a new PHI to
1507 IRBuilderBase::InsertPointGuard Guard(*Builder);
1508 Builder->SetInsertPoint(PN);
1509 NewPN = Builder->CreatePHI(Op1->getOperand(DI)->getType(),
1510 PN->getNumOperands());
1513 for (auto &I : PN->operands())
1514 NewPN->addIncoming(cast<GEPOperator>(I)->getOperand(DI),
1515 PN->getIncomingBlock(I));
1517 NewGEP->setOperand(DI, NewPN);
1518 GEP.getParent()->getInstList().insert(
1519 GEP.getParent()->getFirstInsertionPt(), NewGEP);
1520 NewGEP->setOperand(DI, NewPN);
1523 GEP.setOperand(0, NewGEP);
1527 // Combine Indices - If the source pointer to this getelementptr instruction
1528 // is a getelementptr instruction, combine the indices of the two
1529 // getelementptr instructions into a single instruction.
1531 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) {
1532 if (!shouldMergeGEPs(*cast<GEPOperator>(&GEP), *Src))
1535 // Note that if our source is a gep chain itself then we wait for that
1536 // chain to be resolved before we perform this transformation. This
1537 // avoids us creating a TON of code in some cases.
1538 if (GEPOperator *SrcGEP =
1539 dyn_cast<GEPOperator>(Src->getOperand(0)))
1540 if (SrcGEP->getNumOperands() == 2 && shouldMergeGEPs(*Src, *SrcGEP))
1541 return nullptr; // Wait until our source is folded to completion.
1543 SmallVector<Value*, 8> Indices;
1545 // Find out whether the last index in the source GEP is a sequential idx.
1546 bool EndsWithSequential = false;
1547 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src);
1549 EndsWithSequential = I.isSequential();
1551 // Can we combine the two pointer arithmetics offsets?
1552 if (EndsWithSequential) {
1553 // Replace: gep (gep %P, long B), long A, ...
1554 // With: T = long A+B; gep %P, T, ...
1557 Value *SO1 = Src->getOperand(Src->getNumOperands()-1);
1558 Value *GO1 = GEP.getOperand(1);
1559 if (SO1 == Constant::getNullValue(SO1->getType())) {
1561 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
1564 // If they aren't the same type, then the input hasn't been processed
1565 // by the loop above yet (which canonicalizes sequential index types to
1566 // intptr_t). Just avoid transforming this until the input has been
1568 if (SO1->getType() != GO1->getType())
1570 // Only do the combine when GO1 and SO1 are both constants. Only in
1571 // this case, we are sure the cost after the merge is never more than
1572 // that before the merge.
1573 if (!isa<Constant>(GO1) || !isa<Constant>(SO1))
1575 Sum = Builder->CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
1578 // Update the GEP in place if possible.
1579 if (Src->getNumOperands() == 2) {
1580 GEP.setOperand(0, Src->getOperand(0));
1581 GEP.setOperand(1, Sum);
1584 Indices.append(Src->op_begin()+1, Src->op_end()-1);
1585 Indices.push_back(Sum);
1586 Indices.append(GEP.op_begin()+2, GEP.op_end());
1587 } else if (isa<Constant>(*GEP.idx_begin()) &&
1588 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
1589 Src->getNumOperands() != 1) {
1590 // Otherwise we can do the fold if the first index of the GEP is a zero
1591 Indices.append(Src->op_begin()+1, Src->op_end());
1592 Indices.append(GEP.idx_begin()+1, GEP.idx_end());
1595 if (!Indices.empty())
1596 return GEP.isInBounds() && Src->isInBounds()
1597 ? GetElementPtrInst::CreateInBounds(
1598 Src->getSourceElementType(), Src->getOperand(0), Indices,
1600 : GetElementPtrInst::Create(Src->getSourceElementType(),
1601 Src->getOperand(0), Indices,
1605 if (GEP.getNumIndices() == 1) {
1606 unsigned AS = GEP.getPointerAddressSpace();
1607 if (GEP.getOperand(1)->getType()->getScalarSizeInBits() ==
1608 DL.getPointerSizeInBits(AS)) {
1609 Type *Ty = GEP.getSourceElementType();
1610 uint64_t TyAllocSize = DL.getTypeAllocSize(Ty);
1612 bool Matched = false;
1615 if (TyAllocSize == 1) {
1616 V = GEP.getOperand(1);
1618 } else if (match(GEP.getOperand(1),
1619 m_AShr(m_Value(V), m_ConstantInt(C)))) {
1620 if (TyAllocSize == 1ULL << C)
1622 } else if (match(GEP.getOperand(1),
1623 m_SDiv(m_Value(V), m_ConstantInt(C)))) {
1624 if (TyAllocSize == C)
1629 // Canonicalize (gep i8* X, -(ptrtoint Y))
1630 // to (inttoptr (sub (ptrtoint X), (ptrtoint Y)))
1631 // The GEP pattern is emitted by the SCEV expander for certain kinds of
1632 // pointer arithmetic.
1633 if (match(V, m_Neg(m_PtrToInt(m_Value())))) {
1634 Operator *Index = cast<Operator>(V);
1635 Value *PtrToInt = Builder->CreatePtrToInt(PtrOp, Index->getType());
1636 Value *NewSub = Builder->CreateSub(PtrToInt, Index->getOperand(1));
1637 return CastInst::Create(Instruction::IntToPtr, NewSub, GEP.getType());
1639 // Canonicalize (gep i8* X, (ptrtoint Y)-(ptrtoint X))
1642 if (match(V, m_Sub(m_PtrToInt(m_Value(Y)),
1643 m_PtrToInt(m_Specific(GEP.getOperand(0)))))) {
1644 return CastInst::CreatePointerBitCastOrAddrSpaceCast(Y,
1651 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0).
1652 Value *StrippedPtr = PtrOp->stripPointerCasts();
1653 PointerType *StrippedPtrTy = dyn_cast<PointerType>(StrippedPtr->getType());
1655 // We do not handle pointer-vector geps here.
1659 if (StrippedPtr != PtrOp) {
1660 bool HasZeroPointerIndex = false;
1661 if (ConstantInt *C = dyn_cast<ConstantInt>(GEP.getOperand(1)))
1662 HasZeroPointerIndex = C->isZero();
1664 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
1665 // into : GEP [10 x i8]* X, i32 0, ...
1667 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
1668 // into : GEP i8* X, ...
1670 // This occurs when the program declares an array extern like "int X[];"
1671 if (HasZeroPointerIndex) {
1672 if (ArrayType *CATy =
1673 dyn_cast<ArrayType>(GEP.getSourceElementType())) {
1674 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
1675 if (CATy->getElementType() == StrippedPtrTy->getElementType()) {
1676 // -> GEP i8* X, ...
1677 SmallVector<Value*, 8> Idx(GEP.idx_begin()+1, GEP.idx_end());
1678 GetElementPtrInst *Res = GetElementPtrInst::Create(
1679 StrippedPtrTy->getElementType(), StrippedPtr, Idx, GEP.getName());
1680 Res->setIsInBounds(GEP.isInBounds());
1681 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace())
1683 // Insert Res, and create an addrspacecast.
1685 // GEP (addrspacecast i8 addrspace(1)* X to [0 x i8]*), i32 0, ...
1687 // %0 = GEP i8 addrspace(1)* X, ...
1688 // addrspacecast i8 addrspace(1)* %0 to i8*
1689 return new AddrSpaceCastInst(Builder->Insert(Res), GEP.getType());
1692 if (ArrayType *XATy =
1693 dyn_cast<ArrayType>(StrippedPtrTy->getElementType())){
1694 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
1695 if (CATy->getElementType() == XATy->getElementType()) {
1696 // -> GEP [10 x i8]* X, i32 0, ...
1697 // At this point, we know that the cast source type is a pointer
1698 // to an array of the same type as the destination pointer
1699 // array. Because the array type is never stepped over (there
1700 // is a leading zero) we can fold the cast into this GEP.
1701 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace()) {
1702 GEP.setOperand(0, StrippedPtr);
1703 GEP.setSourceElementType(XATy);
1706 // Cannot replace the base pointer directly because StrippedPtr's
1707 // address space is different. Instead, create a new GEP followed by
1708 // an addrspacecast.
1710 // GEP (addrspacecast [10 x i8] addrspace(1)* X to [0 x i8]*),
1713 // %0 = GEP [10 x i8] addrspace(1)* X, ...
1714 // addrspacecast i8 addrspace(1)* %0 to i8*
1715 SmallVector<Value*, 8> Idx(GEP.idx_begin(), GEP.idx_end());
1716 Value *NewGEP = GEP.isInBounds()
1717 ? Builder->CreateInBoundsGEP(
1718 nullptr, StrippedPtr, Idx, GEP.getName())
1719 : Builder->CreateGEP(nullptr, StrippedPtr, Idx,
1721 return new AddrSpaceCastInst(NewGEP, GEP.getType());
1725 } else if (GEP.getNumOperands() == 2) {
1726 // Transform things like:
1727 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
1728 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
1729 Type *SrcElTy = StrippedPtrTy->getElementType();
1730 Type *ResElTy = GEP.getSourceElementType();
1731 if (SrcElTy->isArrayTy() &&
1732 DL.getTypeAllocSize(SrcElTy->getArrayElementType()) ==
1733 DL.getTypeAllocSize(ResElTy)) {
1734 Type *IdxType = DL.getIntPtrType(GEP.getType());
1735 Value *Idx[2] = { Constant::getNullValue(IdxType), GEP.getOperand(1) };
1738 ? Builder->CreateInBoundsGEP(nullptr, StrippedPtr, Idx,
1740 : Builder->CreateGEP(nullptr, StrippedPtr, Idx, GEP.getName());
1742 // V and GEP are both pointer types --> BitCast
1743 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
1747 // Transform things like:
1748 // %V = mul i64 %N, 4
1749 // %t = getelementptr i8* bitcast (i32* %arr to i8*), i32 %V
1750 // into: %t1 = getelementptr i32* %arr, i32 %N; bitcast
1751 if (ResElTy->isSized() && SrcElTy->isSized()) {
1752 // Check that changing the type amounts to dividing the index by a scale
1754 uint64_t ResSize = DL.getTypeAllocSize(ResElTy);
1755 uint64_t SrcSize = DL.getTypeAllocSize(SrcElTy);
1756 if (ResSize && SrcSize % ResSize == 0) {
1757 Value *Idx = GEP.getOperand(1);
1758 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
1759 uint64_t Scale = SrcSize / ResSize;
1761 // Earlier transforms ensure that the index has type IntPtrType, which
1762 // considerably simplifies the logic by eliminating implicit casts.
1763 assert(Idx->getType() == DL.getIntPtrType(GEP.getType()) &&
1764 "Index not cast to pointer width?");
1767 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
1768 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
1769 // If the multiplication NewIdx * Scale may overflow then the new
1770 // GEP may not be "inbounds".
1772 GEP.isInBounds() && NSW
1773 ? Builder->CreateInBoundsGEP(nullptr, StrippedPtr, NewIdx,
1775 : Builder->CreateGEP(nullptr, StrippedPtr, NewIdx,
1778 // The NewGEP must be pointer typed, so must the old one -> BitCast
1779 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
1785 // Similarly, transform things like:
1786 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
1787 // (where tmp = 8*tmp2) into:
1788 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
1789 if (ResElTy->isSized() && SrcElTy->isSized() && SrcElTy->isArrayTy()) {
1790 // Check that changing to the array element type amounts to dividing the
1791 // index by a scale factor.
1792 uint64_t ResSize = DL.getTypeAllocSize(ResElTy);
1793 uint64_t ArrayEltSize =
1794 DL.getTypeAllocSize(SrcElTy->getArrayElementType());
1795 if (ResSize && ArrayEltSize % ResSize == 0) {
1796 Value *Idx = GEP.getOperand(1);
1797 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
1798 uint64_t Scale = ArrayEltSize / ResSize;
1800 // Earlier transforms ensure that the index has type IntPtrType, which
1801 // considerably simplifies the logic by eliminating implicit casts.
1802 assert(Idx->getType() == DL.getIntPtrType(GEP.getType()) &&
1803 "Index not cast to pointer width?");
1806 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
1807 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
1808 // If the multiplication NewIdx * Scale may overflow then the new
1809 // GEP may not be "inbounds".
1811 Constant::getNullValue(DL.getIntPtrType(GEP.getType())),
1814 Value *NewGEP = GEP.isInBounds() && NSW
1815 ? Builder->CreateInBoundsGEP(
1816 SrcElTy, StrippedPtr, Off, GEP.getName())
1817 : Builder->CreateGEP(SrcElTy, StrippedPtr, Off,
1819 // The NewGEP must be pointer typed, so must the old one -> BitCast
1820 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
1828 // addrspacecast between types is canonicalized as a bitcast, then an
1829 // addrspacecast. To take advantage of the below bitcast + struct GEP, look
1830 // through the addrspacecast.
1831 if (AddrSpaceCastInst *ASC = dyn_cast<AddrSpaceCastInst>(PtrOp)) {
1832 // X = bitcast A addrspace(1)* to B addrspace(1)*
1833 // Y = addrspacecast A addrspace(1)* to B addrspace(2)*
1834 // Z = gep Y, <...constant indices...>
1835 // Into an addrspacecasted GEP of the struct.
1836 if (BitCastInst *BC = dyn_cast<BitCastInst>(ASC->getOperand(0)))
1840 /// See if we can simplify:
1841 /// X = bitcast A* to B*
1842 /// Y = gep X, <...constant indices...>
1843 /// into a gep of the original struct. This is important for SROA and alias
1844 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
1845 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
1846 Value *Operand = BCI->getOperand(0);
1847 PointerType *OpType = cast<PointerType>(Operand->getType());
1848 unsigned OffsetBits = DL.getPointerTypeSizeInBits(GEP.getType());
1849 APInt Offset(OffsetBits, 0);
1850 if (!isa<BitCastInst>(Operand) &&
1851 GEP.accumulateConstantOffset(DL, Offset)) {
1853 // If this GEP instruction doesn't move the pointer, just replace the GEP
1854 // with a bitcast of the real input to the dest type.
1856 // If the bitcast is of an allocation, and the allocation will be
1857 // converted to match the type of the cast, don't touch this.
1858 if (isa<AllocaInst>(Operand) || isAllocationFn(Operand, &TLI)) {
1859 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
1860 if (Instruction *I = visitBitCast(*BCI)) {
1863 BCI->getParent()->getInstList().insert(BCI->getIterator(), I);
1864 replaceInstUsesWith(*BCI, I);
1870 if (Operand->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
1871 return new AddrSpaceCastInst(Operand, GEP.getType());
1872 return new BitCastInst(Operand, GEP.getType());
1875 // Otherwise, if the offset is non-zero, we need to find out if there is a
1876 // field at Offset in 'A's type. If so, we can pull the cast through the
1878 SmallVector<Value*, 8> NewIndices;
1879 if (FindElementAtOffset(OpType, Offset.getSExtValue(), NewIndices)) {
1882 ? Builder->CreateInBoundsGEP(nullptr, Operand, NewIndices)
1883 : Builder->CreateGEP(nullptr, Operand, NewIndices);
1885 if (NGEP->getType() == GEP.getType())
1886 return replaceInstUsesWith(GEP, NGEP);
1887 NGEP->takeName(&GEP);
1889 if (NGEP->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
1890 return new AddrSpaceCastInst(NGEP, GEP.getType());
1891 return new BitCastInst(NGEP, GEP.getType());
1896 if (!GEP.isInBounds()) {
1898 DL.getPointerSizeInBits(PtrOp->getType()->getPointerAddressSpace());
1899 APInt BasePtrOffset(PtrWidth, 0);
1900 Value *UnderlyingPtrOp =
1901 PtrOp->stripAndAccumulateInBoundsConstantOffsets(DL,
1903 if (auto *AI = dyn_cast<AllocaInst>(UnderlyingPtrOp)) {
1904 if (GEP.accumulateConstantOffset(DL, BasePtrOffset) &&
1905 BasePtrOffset.isNonNegative()) {
1906 APInt AllocSize(PtrWidth, DL.getTypeAllocSize(AI->getAllocatedType()));
1907 if (BasePtrOffset.ule(AllocSize)) {
1908 return GetElementPtrInst::CreateInBounds(
1909 PtrOp, makeArrayRef(Ops).slice(1), GEP.getName());
1918 static bool isNeverEqualToUnescapedAlloc(Value *V, const TargetLibraryInfo *TLI,
1920 if (isa<ConstantPointerNull>(V))
1922 if (auto *LI = dyn_cast<LoadInst>(V))
1923 return isa<GlobalVariable>(LI->getPointerOperand());
1924 // Two distinct allocations will never be equal.
1925 // We rely on LookThroughBitCast in isAllocLikeFn being false, since looking
1926 // through bitcasts of V can cause
1927 // the result statement below to be true, even when AI and V (ex:
1928 // i8* ->i32* ->i8* of AI) are the same allocations.
1929 return isAllocLikeFn(V, TLI) && V != AI;
1933 isAllocSiteRemovable(Instruction *AI, SmallVectorImpl<WeakVH> &Users,
1934 const TargetLibraryInfo *TLI) {
1935 SmallVector<Instruction*, 4> Worklist;
1936 Worklist.push_back(AI);
1939 Instruction *PI = Worklist.pop_back_val();
1940 for (User *U : PI->users()) {
1941 Instruction *I = cast<Instruction>(U);
1942 switch (I->getOpcode()) {
1944 // Give up the moment we see something we can't handle.
1947 case Instruction::BitCast:
1948 case Instruction::GetElementPtr:
1949 Users.emplace_back(I);
1950 Worklist.push_back(I);
1953 case Instruction::ICmp: {
1954 ICmpInst *ICI = cast<ICmpInst>(I);
1955 // We can fold eq/ne comparisons with null to false/true, respectively.
1956 // We also fold comparisons in some conditions provided the alloc has
1957 // not escaped (see isNeverEqualToUnescapedAlloc).
1958 if (!ICI->isEquality())
1960 unsigned OtherIndex = (ICI->getOperand(0) == PI) ? 1 : 0;
1961 if (!isNeverEqualToUnescapedAlloc(ICI->getOperand(OtherIndex), TLI, AI))
1963 Users.emplace_back(I);
1967 case Instruction::Call:
1968 // Ignore no-op and store intrinsics.
1969 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1970 switch (II->getIntrinsicID()) {
1974 case Intrinsic::memmove:
1975 case Intrinsic::memcpy:
1976 case Intrinsic::memset: {
1977 MemIntrinsic *MI = cast<MemIntrinsic>(II);
1978 if (MI->isVolatile() || MI->getRawDest() != PI)
1982 case Intrinsic::dbg_declare:
1983 case Intrinsic::dbg_value:
1984 case Intrinsic::invariant_start:
1985 case Intrinsic::invariant_end:
1986 case Intrinsic::lifetime_start:
1987 case Intrinsic::lifetime_end:
1988 case Intrinsic::objectsize:
1989 Users.emplace_back(I);
1994 if (isFreeCall(I, TLI)) {
1995 Users.emplace_back(I);
2000 case Instruction::Store: {
2001 StoreInst *SI = cast<StoreInst>(I);
2002 if (SI->isVolatile() || SI->getPointerOperand() != PI)
2004 Users.emplace_back(I);
2008 llvm_unreachable("missing a return?");
2010 } while (!Worklist.empty());
2014 Instruction *InstCombiner::visitAllocSite(Instruction &MI) {
2015 // If we have a malloc call which is only used in any amount of comparisons
2016 // to null and free calls, delete the calls and replace the comparisons with
2017 // true or false as appropriate.
2018 SmallVector<WeakVH, 64> Users;
2019 if (isAllocSiteRemovable(&MI, Users, &TLI)) {
2020 for (unsigned i = 0, e = Users.size(); i != e; ++i) {
2021 // Lowering all @llvm.objectsize calls first because they may
2022 // use a bitcast/GEP of the alloca we are removing.
2026 Instruction *I = cast<Instruction>(&*Users[i]);
2028 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
2029 if (II->getIntrinsicID() == Intrinsic::objectsize) {
2030 ConstantInt *Result = lowerObjectSizeCall(II, DL, &TLI,
2031 /*MustSucceed=*/true);
2032 replaceInstUsesWith(*I, Result);
2033 eraseInstFromFunction(*I);
2034 Users[i] = nullptr; // Skip examining in the next loop.
2038 for (unsigned i = 0, e = Users.size(); i != e; ++i) {
2042 Instruction *I = cast<Instruction>(&*Users[i]);
2044 if (ICmpInst *C = dyn_cast<ICmpInst>(I)) {
2045 replaceInstUsesWith(*C,
2046 ConstantInt::get(Type::getInt1Ty(C->getContext()),
2047 C->isFalseWhenEqual()));
2048 } else if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) {
2049 replaceInstUsesWith(*I, UndefValue::get(I->getType()));
2051 eraseInstFromFunction(*I);
2054 if (InvokeInst *II = dyn_cast<InvokeInst>(&MI)) {
2055 // Replace invoke with a NOP intrinsic to maintain the original CFG
2056 Module *M = II->getModule();
2057 Function *F = Intrinsic::getDeclaration(M, Intrinsic::donothing);
2058 InvokeInst::Create(F, II->getNormalDest(), II->getUnwindDest(),
2059 None, "", II->getParent());
2061 return eraseInstFromFunction(MI);
2066 /// \brief Move the call to free before a NULL test.
2068 /// Check if this free is accessed after its argument has been test
2069 /// against NULL (property 0).
2070 /// If yes, it is legal to move this call in its predecessor block.
2072 /// The move is performed only if the block containing the call to free
2073 /// will be removed, i.e.:
2074 /// 1. it has only one predecessor P, and P has two successors
2075 /// 2. it contains the call and an unconditional branch
2076 /// 3. its successor is the same as its predecessor's successor
2078 /// The profitability is out-of concern here and this function should
2079 /// be called only if the caller knows this transformation would be
2080 /// profitable (e.g., for code size).
2081 static Instruction *
2082 tryToMoveFreeBeforeNullTest(CallInst &FI) {
2083 Value *Op = FI.getArgOperand(0);
2084 BasicBlock *FreeInstrBB = FI.getParent();
2085 BasicBlock *PredBB = FreeInstrBB->getSinglePredecessor();
2087 // Validate part of constraint #1: Only one predecessor
2088 // FIXME: We can extend the number of predecessor, but in that case, we
2089 // would duplicate the call to free in each predecessor and it may
2090 // not be profitable even for code size.
2094 // Validate constraint #2: Does this block contains only the call to
2095 // free and an unconditional branch?
2096 // FIXME: We could check if we can speculate everything in the
2097 // predecessor block
2098 if (FreeInstrBB->size() != 2)
2101 if (!match(FreeInstrBB->getTerminator(), m_UnconditionalBr(SuccBB)))
2104 // Validate the rest of constraint #1 by matching on the pred branch.
2105 TerminatorInst *TI = PredBB->getTerminator();
2106 BasicBlock *TrueBB, *FalseBB;
2107 ICmpInst::Predicate Pred;
2108 if (!match(TI, m_Br(m_ICmp(Pred, m_Specific(Op), m_Zero()), TrueBB, FalseBB)))
2110 if (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE)
2113 // Validate constraint #3: Ensure the null case just falls through.
2114 if (SuccBB != (Pred == ICmpInst::ICMP_EQ ? TrueBB : FalseBB))
2116 assert(FreeInstrBB == (Pred == ICmpInst::ICMP_EQ ? FalseBB : TrueBB) &&
2117 "Broken CFG: missing edge from predecessor to successor");
2124 Instruction *InstCombiner::visitFree(CallInst &FI) {
2125 Value *Op = FI.getArgOperand(0);
2127 // free undef -> unreachable.
2128 if (isa<UndefValue>(Op)) {
2129 // Insert a new store to null because we cannot modify the CFG here.
2130 Builder->CreateStore(ConstantInt::getTrue(FI.getContext()),
2131 UndefValue::get(Type::getInt1PtrTy(FI.getContext())));
2132 return eraseInstFromFunction(FI);
2135 // If we have 'free null' delete the instruction. This can happen in stl code
2136 // when lots of inlining happens.
2137 if (isa<ConstantPointerNull>(Op))
2138 return eraseInstFromFunction(FI);
2140 // If we optimize for code size, try to move the call to free before the null
2141 // test so that simplify cfg can remove the empty block and dead code
2142 // elimination the branch. I.e., helps to turn something like:
2143 // if (foo) free(foo);
2147 if (Instruction *I = tryToMoveFreeBeforeNullTest(FI))
2153 Instruction *InstCombiner::visitReturnInst(ReturnInst &RI) {
2154 if (RI.getNumOperands() == 0) // ret void
2157 Value *ResultOp = RI.getOperand(0);
2158 Type *VTy = ResultOp->getType();
2159 if (!VTy->isIntegerTy())
2162 // There might be assume intrinsics dominating this return that completely
2163 // determine the value. If so, constant fold it.
2164 unsigned BitWidth = VTy->getPrimitiveSizeInBits();
2165 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2166 computeKnownBits(ResultOp, KnownZero, KnownOne, 0, &RI);
2167 if ((KnownZero|KnownOne).isAllOnesValue())
2168 RI.setOperand(0, Constant::getIntegerValue(VTy, KnownOne));
2173 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
2174 // Change br (not X), label True, label False to: br X, label False, True
2176 BasicBlock *TrueDest;
2177 BasicBlock *FalseDest;
2178 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
2179 !isa<Constant>(X)) {
2180 // Swap Destinations and condition...
2182 BI.swapSuccessors();
2186 // If the condition is irrelevant, remove the use so that other
2187 // transforms on the condition become more effective.
2188 if (BI.isConditional() &&
2189 BI.getSuccessor(0) == BI.getSuccessor(1) &&
2190 !isa<UndefValue>(BI.getCondition())) {
2191 BI.setCondition(UndefValue::get(BI.getCondition()->getType()));
2195 // Canonicalize fcmp_one -> fcmp_oeq
2196 FCmpInst::Predicate FPred; Value *Y;
2197 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
2198 TrueDest, FalseDest)) &&
2199 BI.getCondition()->hasOneUse())
2200 if (FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
2201 FPred == FCmpInst::FCMP_OGE) {
2202 FCmpInst *Cond = cast<FCmpInst>(BI.getCondition());
2203 Cond->setPredicate(FCmpInst::getInversePredicate(FPred));
2205 // Swap Destinations and condition.
2206 BI.swapSuccessors();
2211 // Canonicalize icmp_ne -> icmp_eq
2212 ICmpInst::Predicate IPred;
2213 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
2214 TrueDest, FalseDest)) &&
2215 BI.getCondition()->hasOneUse())
2216 if (IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
2217 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
2218 IPred == ICmpInst::ICMP_SGE) {
2219 ICmpInst *Cond = cast<ICmpInst>(BI.getCondition());
2220 Cond->setPredicate(ICmpInst::getInversePredicate(IPred));
2221 // Swap Destinations and condition.
2222 BI.swapSuccessors();
2230 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
2231 Value *Cond = SI.getCondition();
2233 ConstantInt *AddRHS;
2234 if (match(Cond, m_Add(m_Value(Op0), m_ConstantInt(AddRHS)))) {
2235 // Change 'switch (X+4) case 1:' into 'switch (X) case -3'.
2236 for (SwitchInst::CaseIt CaseIter : SI.cases()) {
2237 Constant *NewCase = ConstantExpr::getSub(CaseIter.getCaseValue(), AddRHS);
2238 assert(isa<ConstantInt>(NewCase) &&
2239 "Result of expression should be constant");
2240 CaseIter.setValue(cast<ConstantInt>(NewCase));
2242 SI.setCondition(Op0);
2246 unsigned BitWidth = cast<IntegerType>(Cond->getType())->getBitWidth();
2247 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2248 computeKnownBits(Cond, KnownZero, KnownOne, 0, &SI);
2249 unsigned LeadingKnownZeros = KnownZero.countLeadingOnes();
2250 unsigned LeadingKnownOnes = KnownOne.countLeadingOnes();
2252 // Compute the number of leading bits we can ignore.
2253 // TODO: A better way to determine this would use ComputeNumSignBits().
2254 for (auto &C : SI.cases()) {
2255 LeadingKnownZeros = std::min(
2256 LeadingKnownZeros, C.getCaseValue()->getValue().countLeadingZeros());
2257 LeadingKnownOnes = std::min(
2258 LeadingKnownOnes, C.getCaseValue()->getValue().countLeadingOnes());
2261 unsigned NewWidth = BitWidth - std::max(LeadingKnownZeros, LeadingKnownOnes);
2263 // Shrink the condition operand if the new type is smaller than the old type.
2264 // This may produce a non-standard type for the switch, but that's ok because
2265 // the backend should extend back to a legal type for the target.
2266 if (NewWidth > 0 && NewWidth < BitWidth) {
2267 IntegerType *Ty = IntegerType::get(SI.getContext(), NewWidth);
2268 Builder->SetInsertPoint(&SI);
2269 Value *NewCond = Builder->CreateTrunc(Cond, Ty, "trunc");
2270 SI.setCondition(NewCond);
2272 for (SwitchInst::CaseIt CaseIter : SI.cases()) {
2273 APInt TruncatedCase = CaseIter.getCaseValue()->getValue().trunc(NewWidth);
2274 CaseIter.setValue(ConstantInt::get(SI.getContext(), TruncatedCase));
2282 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
2283 Value *Agg = EV.getAggregateOperand();
2285 if (!EV.hasIndices())
2286 return replaceInstUsesWith(EV, Agg);
2289 SimplifyExtractValueInst(Agg, EV.getIndices(), DL, &TLI, &DT, &AC))
2290 return replaceInstUsesWith(EV, V);
2292 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
2293 // We're extracting from an insertvalue instruction, compare the indices
2294 const unsigned *exti, *exte, *insi, *inse;
2295 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
2296 exte = EV.idx_end(), inse = IV->idx_end();
2297 exti != exte && insi != inse;
2300 // The insert and extract both reference distinctly different elements.
2301 // This means the extract is not influenced by the insert, and we can
2302 // replace the aggregate operand of the extract with the aggregate
2303 // operand of the insert. i.e., replace
2304 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
2305 // %E = extractvalue { i32, { i32 } } %I, 0
2307 // %E = extractvalue { i32, { i32 } } %A, 0
2308 return ExtractValueInst::Create(IV->getAggregateOperand(),
2311 if (exti == exte && insi == inse)
2312 // Both iterators are at the end: Index lists are identical. Replace
2313 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
2314 // %C = extractvalue { i32, { i32 } } %B, 1, 0
2316 return replaceInstUsesWith(EV, IV->getInsertedValueOperand());
2318 // The extract list is a prefix of the insert list. i.e. replace
2319 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
2320 // %E = extractvalue { i32, { i32 } } %I, 1
2322 // %X = extractvalue { i32, { i32 } } %A, 1
2323 // %E = insertvalue { i32 } %X, i32 42, 0
2324 // by switching the order of the insert and extract (though the
2325 // insertvalue should be left in, since it may have other uses).
2326 Value *NewEV = Builder->CreateExtractValue(IV->getAggregateOperand(),
2328 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
2329 makeArrayRef(insi, inse));
2332 // The insert list is a prefix of the extract list
2333 // We can simply remove the common indices from the extract and make it
2334 // operate on the inserted value instead of the insertvalue result.
2336 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
2337 // %E = extractvalue { i32, { i32 } } %I, 1, 0
2339 // %E extractvalue { i32 } { i32 42 }, 0
2340 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
2341 makeArrayRef(exti, exte));
2343 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) {
2344 // We're extracting from an intrinsic, see if we're the only user, which
2345 // allows us to simplify multiple result intrinsics to simpler things that
2346 // just get one value.
2347 if (II->hasOneUse()) {
2348 // Check if we're grabbing the overflow bit or the result of a 'with
2349 // overflow' intrinsic. If it's the latter we can remove the intrinsic
2350 // and replace it with a traditional binary instruction.
2351 switch (II->getIntrinsicID()) {
2352 case Intrinsic::uadd_with_overflow:
2353 case Intrinsic::sadd_with_overflow:
2354 if (*EV.idx_begin() == 0) { // Normal result.
2355 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2356 replaceInstUsesWith(*II, UndefValue::get(II->getType()));
2357 eraseInstFromFunction(*II);
2358 return BinaryOperator::CreateAdd(LHS, RHS);
2361 // If the normal result of the add is dead, and the RHS is a constant,
2362 // we can transform this into a range comparison.
2363 // overflow = uadd a, -4 --> overflow = icmp ugt a, 3
2364 if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow)
2365 if (ConstantInt *CI = dyn_cast<ConstantInt>(II->getArgOperand(1)))
2366 return new ICmpInst(ICmpInst::ICMP_UGT, II->getArgOperand(0),
2367 ConstantExpr::getNot(CI));
2369 case Intrinsic::usub_with_overflow:
2370 case Intrinsic::ssub_with_overflow:
2371 if (*EV.idx_begin() == 0) { // Normal result.
2372 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2373 replaceInstUsesWith(*II, UndefValue::get(II->getType()));
2374 eraseInstFromFunction(*II);
2375 return BinaryOperator::CreateSub(LHS, RHS);
2378 case Intrinsic::umul_with_overflow:
2379 case Intrinsic::smul_with_overflow:
2380 if (*EV.idx_begin() == 0) { // Normal result.
2381 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2382 replaceInstUsesWith(*II, UndefValue::get(II->getType()));
2383 eraseInstFromFunction(*II);
2384 return BinaryOperator::CreateMul(LHS, RHS);
2392 if (LoadInst *L = dyn_cast<LoadInst>(Agg))
2393 // If the (non-volatile) load only has one use, we can rewrite this to a
2394 // load from a GEP. This reduces the size of the load. If a load is used
2395 // only by extractvalue instructions then this either must have been
2396 // optimized before, or it is a struct with padding, in which case we
2397 // don't want to do the transformation as it loses padding knowledge.
2398 if (L->isSimple() && L->hasOneUse()) {
2399 // extractvalue has integer indices, getelementptr has Value*s. Convert.
2400 SmallVector<Value*, 4> Indices;
2401 // Prefix an i32 0 since we need the first element.
2402 Indices.push_back(Builder->getInt32(0));
2403 for (ExtractValueInst::idx_iterator I = EV.idx_begin(), E = EV.idx_end();
2405 Indices.push_back(Builder->getInt32(*I));
2407 // We need to insert these at the location of the old load, not at that of
2408 // the extractvalue.
2409 Builder->SetInsertPoint(L);
2410 Value *GEP = Builder->CreateInBoundsGEP(L->getType(),
2411 L->getPointerOperand(), Indices);
2412 // Returning the load directly will cause the main loop to insert it in
2413 // the wrong spot, so use replaceInstUsesWith().
2414 return replaceInstUsesWith(EV, Builder->CreateLoad(GEP));
2416 // We could simplify extracts from other values. Note that nested extracts may
2417 // already be simplified implicitly by the above: extract (extract (insert) )
2418 // will be translated into extract ( insert ( extract ) ) first and then just
2419 // the value inserted, if appropriate. Similarly for extracts from single-use
2420 // loads: extract (extract (load)) will be translated to extract (load (gep))
2421 // and if again single-use then via load (gep (gep)) to load (gep).
2422 // However, double extracts from e.g. function arguments or return values
2423 // aren't handled yet.
2427 /// Return 'true' if the given typeinfo will match anything.
2428 static bool isCatchAll(EHPersonality Personality, Constant *TypeInfo) {
2429 switch (Personality) {
2430 case EHPersonality::GNU_C:
2431 case EHPersonality::GNU_C_SjLj:
2432 case EHPersonality::Rust:
2433 // The GCC C EH and Rust personality only exists to support cleanups, so
2434 // it's not clear what the semantics of catch clauses are.
2436 case EHPersonality::Unknown:
2438 case EHPersonality::GNU_Ada:
2439 // While __gnat_all_others_value will match any Ada exception, it doesn't
2440 // match foreign exceptions (or didn't, before gcc-4.7).
2442 case EHPersonality::GNU_CXX:
2443 case EHPersonality::GNU_CXX_SjLj:
2444 case EHPersonality::GNU_ObjC:
2445 case EHPersonality::MSVC_X86SEH:
2446 case EHPersonality::MSVC_Win64SEH:
2447 case EHPersonality::MSVC_CXX:
2448 case EHPersonality::CoreCLR:
2449 return TypeInfo->isNullValue();
2451 llvm_unreachable("invalid enum");
2454 static bool shorter_filter(const Value *LHS, const Value *RHS) {
2456 cast<ArrayType>(LHS->getType())->getNumElements()
2458 cast<ArrayType>(RHS->getType())->getNumElements();
2461 Instruction *InstCombiner::visitLandingPadInst(LandingPadInst &LI) {
2462 // The logic here should be correct for any real-world personality function.
2463 // However if that turns out not to be true, the offending logic can always
2464 // be conditioned on the personality function, like the catch-all logic is.
2465 EHPersonality Personality =
2466 classifyEHPersonality(LI.getParent()->getParent()->getPersonalityFn());
2468 // Simplify the list of clauses, eg by removing repeated catch clauses
2469 // (these are often created by inlining).
2470 bool MakeNewInstruction = false; // If true, recreate using the following:
2471 SmallVector<Constant *, 16> NewClauses; // - Clauses for the new instruction;
2472 bool CleanupFlag = LI.isCleanup(); // - The new instruction is a cleanup.
2474 SmallPtrSet<Value *, 16> AlreadyCaught; // Typeinfos known caught already.
2475 for (unsigned i = 0, e = LI.getNumClauses(); i != e; ++i) {
2476 bool isLastClause = i + 1 == e;
2477 if (LI.isCatch(i)) {
2479 Constant *CatchClause = LI.getClause(i);
2480 Constant *TypeInfo = CatchClause->stripPointerCasts();
2482 // If we already saw this clause, there is no point in having a second
2484 if (AlreadyCaught.insert(TypeInfo).second) {
2485 // This catch clause was not already seen.
2486 NewClauses.push_back(CatchClause);
2488 // Repeated catch clause - drop the redundant copy.
2489 MakeNewInstruction = true;
2492 // If this is a catch-all then there is no point in keeping any following
2493 // clauses or marking the landingpad as having a cleanup.
2494 if (isCatchAll(Personality, TypeInfo)) {
2496 MakeNewInstruction = true;
2497 CleanupFlag = false;
2501 // A filter clause. If any of the filter elements were already caught
2502 // then they can be dropped from the filter. It is tempting to try to
2503 // exploit the filter further by saying that any typeinfo that does not
2504 // occur in the filter can't be caught later (and thus can be dropped).
2505 // However this would be wrong, since typeinfos can match without being
2506 // equal (for example if one represents a C++ class, and the other some
2507 // class derived from it).
2508 assert(LI.isFilter(i) && "Unsupported landingpad clause!");
2509 Constant *FilterClause = LI.getClause(i);
2510 ArrayType *FilterType = cast<ArrayType>(FilterClause->getType());
2511 unsigned NumTypeInfos = FilterType->getNumElements();
2513 // An empty filter catches everything, so there is no point in keeping any
2514 // following clauses or marking the landingpad as having a cleanup. By
2515 // dealing with this case here the following code is made a bit simpler.
2516 if (!NumTypeInfos) {
2517 NewClauses.push_back(FilterClause);
2519 MakeNewInstruction = true;
2520 CleanupFlag = false;
2524 bool MakeNewFilter = false; // If true, make a new filter.
2525 SmallVector<Constant *, 16> NewFilterElts; // New elements.
2526 if (isa<ConstantAggregateZero>(FilterClause)) {
2527 // Not an empty filter - it contains at least one null typeinfo.
2528 assert(NumTypeInfos > 0 && "Should have handled empty filter already!");
2529 Constant *TypeInfo =
2530 Constant::getNullValue(FilterType->getElementType());
2531 // If this typeinfo is a catch-all then the filter can never match.
2532 if (isCatchAll(Personality, TypeInfo)) {
2533 // Throw the filter away.
2534 MakeNewInstruction = true;
2538 // There is no point in having multiple copies of this typeinfo, so
2539 // discard all but the first copy if there is more than one.
2540 NewFilterElts.push_back(TypeInfo);
2541 if (NumTypeInfos > 1)
2542 MakeNewFilter = true;
2544 ConstantArray *Filter = cast<ConstantArray>(FilterClause);
2545 SmallPtrSet<Value *, 16> SeenInFilter; // For uniquing the elements.
2546 NewFilterElts.reserve(NumTypeInfos);
2548 // Remove any filter elements that were already caught or that already
2549 // occurred in the filter. While there, see if any of the elements are
2550 // catch-alls. If so, the filter can be discarded.
2551 bool SawCatchAll = false;
2552 for (unsigned j = 0; j != NumTypeInfos; ++j) {
2553 Constant *Elt = Filter->getOperand(j);
2554 Constant *TypeInfo = Elt->stripPointerCasts();
2555 if (isCatchAll(Personality, TypeInfo)) {
2556 // This element is a catch-all. Bail out, noting this fact.
2561 // Even if we've seen a type in a catch clause, we don't want to
2562 // remove it from the filter. An unexpected type handler may be
2563 // set up for a call site which throws an exception of the same
2564 // type caught. In order for the exception thrown by the unexpected
2565 // handler to propagate correctly, the filter must be correctly
2566 // described for the call site.
2570 // void unexpected() { throw 1;}
2571 // void foo() throw (int) {
2572 // std::set_unexpected(unexpected);
2575 // } catch (int i) {}
2578 // There is no point in having multiple copies of the same typeinfo in
2579 // a filter, so only add it if we didn't already.
2580 if (SeenInFilter.insert(TypeInfo).second)
2581 NewFilterElts.push_back(cast<Constant>(Elt));
2583 // A filter containing a catch-all cannot match anything by definition.
2585 // Throw the filter away.
2586 MakeNewInstruction = true;
2590 // If we dropped something from the filter, make a new one.
2591 if (NewFilterElts.size() < NumTypeInfos)
2592 MakeNewFilter = true;
2594 if (MakeNewFilter) {
2595 FilterType = ArrayType::get(FilterType->getElementType(),
2596 NewFilterElts.size());
2597 FilterClause = ConstantArray::get(FilterType, NewFilterElts);
2598 MakeNewInstruction = true;
2601 NewClauses.push_back(FilterClause);
2603 // If the new filter is empty then it will catch everything so there is
2604 // no point in keeping any following clauses or marking the landingpad
2605 // as having a cleanup. The case of the original filter being empty was
2606 // already handled above.
2607 if (MakeNewFilter && !NewFilterElts.size()) {
2608 assert(MakeNewInstruction && "New filter but not a new instruction!");
2609 CleanupFlag = false;
2615 // If several filters occur in a row then reorder them so that the shortest
2616 // filters come first (those with the smallest number of elements). This is
2617 // advantageous because shorter filters are more likely to match, speeding up
2618 // unwinding, but mostly because it increases the effectiveness of the other
2619 // filter optimizations below.
2620 for (unsigned i = 0, e = NewClauses.size(); i + 1 < e; ) {
2622 // Find the maximal 'j' s.t. the range [i, j) consists entirely of filters.
2623 for (j = i; j != e; ++j)
2624 if (!isa<ArrayType>(NewClauses[j]->getType()))
2627 // Check whether the filters are already sorted by length. We need to know
2628 // if sorting them is actually going to do anything so that we only make a
2629 // new landingpad instruction if it does.
2630 for (unsigned k = i; k + 1 < j; ++k)
2631 if (shorter_filter(NewClauses[k+1], NewClauses[k])) {
2632 // Not sorted, so sort the filters now. Doing an unstable sort would be
2633 // correct too but reordering filters pointlessly might confuse users.
2634 std::stable_sort(NewClauses.begin() + i, NewClauses.begin() + j,
2636 MakeNewInstruction = true;
2640 // Look for the next batch of filters.
2644 // If typeinfos matched if and only if equal, then the elements of a filter L
2645 // that occurs later than a filter F could be replaced by the intersection of
2646 // the elements of F and L. In reality two typeinfos can match without being
2647 // equal (for example if one represents a C++ class, and the other some class
2648 // derived from it) so it would be wrong to perform this transform in general.
2649 // However the transform is correct and useful if F is a subset of L. In that
2650 // case L can be replaced by F, and thus removed altogether since repeating a
2651 // filter is pointless. So here we look at all pairs of filters F and L where
2652 // L follows F in the list of clauses, and remove L if every element of F is
2653 // an element of L. This can occur when inlining C++ functions with exception
2655 for (unsigned i = 0; i + 1 < NewClauses.size(); ++i) {
2656 // Examine each filter in turn.
2657 Value *Filter = NewClauses[i];
2658 ArrayType *FTy = dyn_cast<ArrayType>(Filter->getType());
2660 // Not a filter - skip it.
2662 unsigned FElts = FTy->getNumElements();
2663 // Examine each filter following this one. Doing this backwards means that
2664 // we don't have to worry about filters disappearing under us when removed.
2665 for (unsigned j = NewClauses.size() - 1; j != i; --j) {
2666 Value *LFilter = NewClauses[j];
2667 ArrayType *LTy = dyn_cast<ArrayType>(LFilter->getType());
2669 // Not a filter - skip it.
2671 // If Filter is a subset of LFilter, i.e. every element of Filter is also
2672 // an element of LFilter, then discard LFilter.
2673 SmallVectorImpl<Constant *>::iterator J = NewClauses.begin() + j;
2674 // If Filter is empty then it is a subset of LFilter.
2677 NewClauses.erase(J);
2678 MakeNewInstruction = true;
2679 // Move on to the next filter.
2682 unsigned LElts = LTy->getNumElements();
2683 // If Filter is longer than LFilter then it cannot be a subset of it.
2685 // Move on to the next filter.
2687 // At this point we know that LFilter has at least one element.
2688 if (isa<ConstantAggregateZero>(LFilter)) { // LFilter only contains zeros.
2689 // Filter is a subset of LFilter iff Filter contains only zeros (as we
2690 // already know that Filter is not longer than LFilter).
2691 if (isa<ConstantAggregateZero>(Filter)) {
2692 assert(FElts <= LElts && "Should have handled this case earlier!");
2694 NewClauses.erase(J);
2695 MakeNewInstruction = true;
2697 // Move on to the next filter.
2700 ConstantArray *LArray = cast<ConstantArray>(LFilter);
2701 if (isa<ConstantAggregateZero>(Filter)) { // Filter only contains zeros.
2702 // Since Filter is non-empty and contains only zeros, it is a subset of
2703 // LFilter iff LFilter contains a zero.
2704 assert(FElts > 0 && "Should have eliminated the empty filter earlier!");
2705 for (unsigned l = 0; l != LElts; ++l)
2706 if (LArray->getOperand(l)->isNullValue()) {
2707 // LFilter contains a zero - discard it.
2708 NewClauses.erase(J);
2709 MakeNewInstruction = true;
2712 // Move on to the next filter.
2715 // At this point we know that both filters are ConstantArrays. Loop over
2716 // operands to see whether every element of Filter is also an element of
2717 // LFilter. Since filters tend to be short this is probably faster than
2718 // using a method that scales nicely.
2719 ConstantArray *FArray = cast<ConstantArray>(Filter);
2720 bool AllFound = true;
2721 for (unsigned f = 0; f != FElts; ++f) {
2722 Value *FTypeInfo = FArray->getOperand(f)->stripPointerCasts();
2724 for (unsigned l = 0; l != LElts; ++l) {
2725 Value *LTypeInfo = LArray->getOperand(l)->stripPointerCasts();
2726 if (LTypeInfo == FTypeInfo) {
2736 NewClauses.erase(J);
2737 MakeNewInstruction = true;
2739 // Move on to the next filter.
2743 // If we changed any of the clauses, replace the old landingpad instruction
2745 if (MakeNewInstruction) {
2746 LandingPadInst *NLI = LandingPadInst::Create(LI.getType(),
2748 for (unsigned i = 0, e = NewClauses.size(); i != e; ++i)
2749 NLI->addClause(NewClauses[i]);
2750 // A landing pad with no clauses must have the cleanup flag set. It is
2751 // theoretically possible, though highly unlikely, that we eliminated all
2752 // clauses. If so, force the cleanup flag to true.
2753 if (NewClauses.empty())
2755 NLI->setCleanup(CleanupFlag);
2759 // Even if none of the clauses changed, we may nonetheless have understood
2760 // that the cleanup flag is pointless. Clear it if so.
2761 if (LI.isCleanup() != CleanupFlag) {
2762 assert(!CleanupFlag && "Adding a cleanup, not removing one?!");
2763 LI.setCleanup(CleanupFlag);
2770 /// Try to move the specified instruction from its current block into the
2771 /// beginning of DestBlock, which can only happen if it's safe to move the
2772 /// instruction past all of the instructions between it and the end of its
2774 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
2775 assert(I->hasOneUse() && "Invariants didn't hold!");
2777 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
2778 if (isa<PHINode>(I) || I->isEHPad() || I->mayHaveSideEffects() ||
2779 isa<TerminatorInst>(I))
2782 // Do not sink alloca instructions out of the entry block.
2783 if (isa<AllocaInst>(I) && I->getParent() ==
2784 &DestBlock->getParent()->getEntryBlock())
2787 // Do not sink into catchswitch blocks.
2788 if (isa<CatchSwitchInst>(DestBlock->getTerminator()))
2791 // Do not sink convergent call instructions.
2792 if (auto *CI = dyn_cast<CallInst>(I)) {
2793 if (CI->isConvergent())
2796 // We can only sink load instructions if there is nothing between the load and
2797 // the end of block that could change the value.
2798 if (I->mayReadFromMemory()) {
2799 for (BasicBlock::iterator Scan = I->getIterator(),
2800 E = I->getParent()->end();
2802 if (Scan->mayWriteToMemory())
2806 BasicBlock::iterator InsertPos = DestBlock->getFirstInsertionPt();
2807 I->moveBefore(&*InsertPos);
2812 bool InstCombiner::run() {
2813 while (!Worklist.isEmpty()) {
2814 Instruction *I = Worklist.RemoveOne();
2815 if (I == nullptr) continue; // skip null values.
2817 // Check to see if we can DCE the instruction.
2818 if (isInstructionTriviallyDead(I, &TLI)) {
2819 DEBUG(dbgs() << "IC: DCE: " << *I << '\n');
2820 eraseInstFromFunction(*I);
2822 MadeIRChange = true;
2826 // Instruction isn't dead, see if we can constant propagate it.
2827 if (!I->use_empty() &&
2828 (I->getNumOperands() == 0 || isa<Constant>(I->getOperand(0)))) {
2829 if (Constant *C = ConstantFoldInstruction(I, DL, &TLI)) {
2830 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n');
2832 // Add operands to the worklist.
2833 replaceInstUsesWith(*I, C);
2835 if (isInstructionTriviallyDead(I, &TLI))
2836 eraseInstFromFunction(*I);
2837 MadeIRChange = true;
2842 // In general, it is possible for computeKnownBits to determine all bits in
2843 // a value even when the operands are not all constants.
2844 Type *Ty = I->getType();
2845 if (ExpensiveCombines && !I->use_empty() && Ty->isIntOrIntVectorTy()) {
2846 unsigned BitWidth = Ty->getScalarSizeInBits();
2847 APInt KnownZero(BitWidth, 0);
2848 APInt KnownOne(BitWidth, 0);
2849 computeKnownBits(I, KnownZero, KnownOne, /*Depth*/0, I);
2850 if ((KnownZero | KnownOne).isAllOnesValue()) {
2851 Constant *C = ConstantInt::get(Ty, KnownOne);
2852 DEBUG(dbgs() << "IC: ConstFold (all bits known) to: " << *C <<
2853 " from: " << *I << '\n');
2855 // Add operands to the worklist.
2856 replaceInstUsesWith(*I, C);
2858 if (isInstructionTriviallyDead(I, &TLI))
2859 eraseInstFromFunction(*I);
2860 MadeIRChange = true;
2865 // See if we can trivially sink this instruction to a successor basic block.
2866 if (I->hasOneUse()) {
2867 BasicBlock *BB = I->getParent();
2868 Instruction *UserInst = cast<Instruction>(*I->user_begin());
2869 BasicBlock *UserParent;
2871 // Get the block the use occurs in.
2872 if (PHINode *PN = dyn_cast<PHINode>(UserInst))
2873 UserParent = PN->getIncomingBlock(*I->use_begin());
2875 UserParent = UserInst->getParent();
2877 if (UserParent != BB) {
2878 bool UserIsSuccessor = false;
2879 // See if the user is one of our successors.
2880 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
2881 if (*SI == UserParent) {
2882 UserIsSuccessor = true;
2886 // If the user is one of our immediate successors, and if that successor
2887 // only has us as a predecessors (we'd have to split the critical edge
2888 // otherwise), we can keep going.
2889 if (UserIsSuccessor && UserParent->getUniquePredecessor()) {
2890 // Okay, the CFG is simple enough, try to sink this instruction.
2891 if (TryToSinkInstruction(I, UserParent)) {
2892 DEBUG(dbgs() << "IC: Sink: " << *I << '\n');
2893 MadeIRChange = true;
2894 // We'll add uses of the sunk instruction below, but since sinking
2895 // can expose opportunities for it's *operands* add them to the
2897 for (Use &U : I->operands())
2898 if (Instruction *OpI = dyn_cast<Instruction>(U.get()))
2905 // Now that we have an instruction, try combining it to simplify it.
2906 Builder->SetInsertPoint(I);
2907 Builder->SetCurrentDebugLocation(I->getDebugLoc());
2912 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str(););
2913 DEBUG(dbgs() << "IC: Visiting: " << OrigI << '\n');
2915 if (Instruction *Result = visit(*I)) {
2917 // Should we replace the old instruction with a new one?
2919 DEBUG(dbgs() << "IC: Old = " << *I << '\n'
2920 << " New = " << *Result << '\n');
2922 if (I->getDebugLoc())
2923 Result->setDebugLoc(I->getDebugLoc());
2924 // Everything uses the new instruction now.
2925 I->replaceAllUsesWith(Result);
2927 // Move the name to the new instruction first.
2928 Result->takeName(I);
2930 // Push the new instruction and any users onto the worklist.
2931 Worklist.Add(Result);
2932 Worklist.AddUsersToWorkList(*Result);
2934 // Insert the new instruction into the basic block...
2935 BasicBlock *InstParent = I->getParent();
2936 BasicBlock::iterator InsertPos = I->getIterator();
2938 // If we replace a PHI with something that isn't a PHI, fix up the
2940 if (!isa<PHINode>(Result) && isa<PHINode>(InsertPos))
2941 InsertPos = InstParent->getFirstInsertionPt();
2943 InstParent->getInstList().insert(InsertPos, Result);
2945 eraseInstFromFunction(*I);
2947 DEBUG(dbgs() << "IC: Mod = " << OrigI << '\n'
2948 << " New = " << *I << '\n');
2950 // If the instruction was modified, it's possible that it is now dead.
2951 // if so, remove it.
2952 if (isInstructionTriviallyDead(I, &TLI)) {
2953 eraseInstFromFunction(*I);
2956 Worklist.AddUsersToWorkList(*I);
2959 MadeIRChange = true;
2964 return MadeIRChange;
2967 /// Walk the function in depth-first order, adding all reachable code to the
2970 /// This has a couple of tricks to make the code faster and more powerful. In
2971 /// particular, we constant fold and DCE instructions as we go, to avoid adding
2972 /// them to the worklist (this significantly speeds up instcombine on code where
2973 /// many instructions are dead or constant). Additionally, if we find a branch
2974 /// whose condition is a known constant, we only visit the reachable successors.
2976 static bool AddReachableCodeToWorklist(BasicBlock *BB, const DataLayout &DL,
2977 SmallPtrSetImpl<BasicBlock *> &Visited,
2978 InstCombineWorklist &ICWorklist,
2979 const TargetLibraryInfo *TLI) {
2980 bool MadeIRChange = false;
2981 SmallVector<BasicBlock*, 256> Worklist;
2982 Worklist.push_back(BB);
2984 SmallVector<Instruction*, 128> InstrsForInstCombineWorklist;
2985 DenseMap<Constant *, Constant *> FoldedConstants;
2988 BB = Worklist.pop_back_val();
2990 // We have now visited this block! If we've already been here, ignore it.
2991 if (!Visited.insert(BB).second)
2994 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
2995 Instruction *Inst = &*BBI++;
2997 // DCE instruction if trivially dead.
2998 if (isInstructionTriviallyDead(Inst, TLI)) {
3000 DEBUG(dbgs() << "IC: DCE: " << *Inst << '\n');
3001 Inst->eraseFromParent();
3005 // ConstantProp instruction if trivially constant.
3006 if (!Inst->use_empty() &&
3007 (Inst->getNumOperands() == 0 || isa<Constant>(Inst->getOperand(0))))
3008 if (Constant *C = ConstantFoldInstruction(Inst, DL, TLI)) {
3009 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: "
3011 Inst->replaceAllUsesWith(C);
3013 if (isInstructionTriviallyDead(Inst, TLI))
3014 Inst->eraseFromParent();
3018 // See if we can constant fold its operands.
3019 for (User::op_iterator i = Inst->op_begin(), e = Inst->op_end(); i != e;
3021 if (!isa<ConstantVector>(i) && !isa<ConstantExpr>(i))
3024 auto *C = cast<Constant>(i);
3025 Constant *&FoldRes = FoldedConstants[C];
3027 FoldRes = ConstantFoldConstant(C, DL, TLI);
3033 MadeIRChange = true;
3037 InstrsForInstCombineWorklist.push_back(Inst);
3040 // Recursively visit successors. If this is a branch or switch on a
3041 // constant, only visit the reachable successor.
3042 TerminatorInst *TI = BB->getTerminator();
3043 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
3044 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
3045 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
3046 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
3047 Worklist.push_back(ReachableBB);
3050 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
3051 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
3052 // See if this is an explicit destination.
3053 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3055 if (i.getCaseValue() == Cond) {
3056 BasicBlock *ReachableBB = i.getCaseSuccessor();
3057 Worklist.push_back(ReachableBB);
3061 // Otherwise it is the default destination.
3062 Worklist.push_back(SI->getDefaultDest());
3067 for (BasicBlock *SuccBB : TI->successors())
3068 Worklist.push_back(SuccBB);
3069 } while (!Worklist.empty());
3071 // Once we've found all of the instructions to add to instcombine's worklist,
3072 // add them in reverse order. This way instcombine will visit from the top
3073 // of the function down. This jives well with the way that it adds all uses
3074 // of instructions to the worklist after doing a transformation, thus avoiding
3075 // some N^2 behavior in pathological cases.
3076 ICWorklist.AddInitialGroup(InstrsForInstCombineWorklist);
3078 return MadeIRChange;
3081 /// \brief Populate the IC worklist from a function, and prune any dead basic
3082 /// blocks discovered in the process.
3084 /// This also does basic constant propagation and other forward fixing to make
3085 /// the combiner itself run much faster.
3086 static bool prepareICWorklistFromFunction(Function &F, const DataLayout &DL,
3087 TargetLibraryInfo *TLI,
3088 InstCombineWorklist &ICWorklist) {
3089 bool MadeIRChange = false;
3091 // Do a depth-first traversal of the function, populate the worklist with
3092 // the reachable instructions. Ignore blocks that are not reachable. Keep
3093 // track of which blocks we visit.
3094 SmallPtrSet<BasicBlock *, 32> Visited;
3096 AddReachableCodeToWorklist(&F.front(), DL, Visited, ICWorklist, TLI);
3098 // Do a quick scan over the function. If we find any blocks that are
3099 // unreachable, remove any instructions inside of them. This prevents
3100 // the instcombine code from having to deal with some bad special cases.
3101 for (BasicBlock &BB : F) {
3102 if (Visited.count(&BB))
3105 unsigned NumDeadInstInBB = removeAllNonTerminatorAndEHPadInstructions(&BB);
3106 MadeIRChange |= NumDeadInstInBB > 0;
3107 NumDeadInst += NumDeadInstInBB;
3110 return MadeIRChange;
3114 combineInstructionsOverFunction(Function &F, InstCombineWorklist &Worklist,
3115 AliasAnalysis *AA, AssumptionCache &AC,
3116 TargetLibraryInfo &TLI, DominatorTree &DT,
3117 bool ExpensiveCombines = true,
3118 LoopInfo *LI = nullptr) {
3119 auto &DL = F.getParent()->getDataLayout();
3120 ExpensiveCombines |= EnableExpensiveCombines;
3122 /// Builder - This is an IRBuilder that automatically inserts new
3123 /// instructions into the worklist when they are created.
3124 IRBuilder<TargetFolder, IRBuilderCallbackInserter> Builder(
3125 F.getContext(), TargetFolder(DL),
3126 IRBuilderCallbackInserter([&Worklist, &AC](Instruction *I) {
3129 using namespace llvm::PatternMatch;
3130 if (match(I, m_Intrinsic<Intrinsic::assume>()))
3131 AC.registerAssumption(cast<CallInst>(I));
3134 // Lower dbg.declare intrinsics otherwise their value may be clobbered
3136 bool DbgDeclaresChanged = LowerDbgDeclare(F);
3138 // Iterate while there is work to do.
3142 DEBUG(dbgs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
3143 << F.getName() << "\n");
3145 bool Changed = prepareICWorklistFromFunction(F, DL, &TLI, Worklist);
3147 InstCombiner IC(Worklist, &Builder, F.optForMinSize(), ExpensiveCombines,
3148 AA, AC, TLI, DT, DL, LI);
3149 Changed |= IC.run();
3155 return DbgDeclaresChanged || Iteration > 1;
3158 PreservedAnalyses InstCombinePass::run(Function &F,
3159 FunctionAnalysisManager &AM) {
3160 auto &AC = AM.getResult<AssumptionAnalysis>(F);
3161 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
3162 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
3164 auto *LI = AM.getCachedResult<LoopAnalysis>(F);
3166 // FIXME: The AliasAnalysis is not yet supported in the new pass manager
3167 if (!combineInstructionsOverFunction(F, Worklist, nullptr, AC, TLI, DT,
3168 ExpensiveCombines, LI))
3169 // No changes, all analyses are preserved.
3170 return PreservedAnalyses::all();
3172 // Mark all the analyses that instcombine updates as preserved.
3173 // FIXME: This should also 'preserve the CFG'.
3174 PreservedAnalyses PA;
3175 PA.preserve<DominatorTreeAnalysis>();
3179 void InstructionCombiningPass::getAnalysisUsage(AnalysisUsage &AU) const {
3180 AU.setPreservesCFG();
3181 AU.addRequired<AAResultsWrapperPass>();
3182 AU.addRequired<AssumptionCacheTracker>();
3183 AU.addRequired<TargetLibraryInfoWrapperPass>();
3184 AU.addRequired<DominatorTreeWrapperPass>();
3185 AU.addPreserved<DominatorTreeWrapperPass>();
3186 AU.addPreserved<AAResultsWrapperPass>();
3187 AU.addPreserved<BasicAAWrapperPass>();
3188 AU.addPreserved<GlobalsAAWrapperPass>();
3191 bool InstructionCombiningPass::runOnFunction(Function &F) {
3192 if (skipFunction(F))
3195 // Required analyses.
3196 auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
3197 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
3198 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
3199 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
3201 // Optional analyses.
3202 auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>();
3203 auto *LI = LIWP ? &LIWP->getLoopInfo() : nullptr;
3205 return combineInstructionsOverFunction(F, Worklist, AA, AC, TLI, DT,
3206 ExpensiveCombines, LI);
3209 char InstructionCombiningPass::ID = 0;
3210 INITIALIZE_PASS_BEGIN(InstructionCombiningPass, "instcombine",
3211 "Combine redundant instructions", false, false)
3212 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
3213 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
3214 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
3215 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
3216 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
3217 INITIALIZE_PASS_END(InstructionCombiningPass, "instcombine",
3218 "Combine redundant instructions", false, false)
3220 // Initialization Routines
3221 void llvm::initializeInstCombine(PassRegistry &Registry) {
3222 initializeInstructionCombiningPassPass(Registry);
3225 void LLVMInitializeInstCombine(LLVMPassRegistryRef R) {
3226 initializeInstructionCombiningPassPass(*unwrap(R));
3229 FunctionPass *llvm::createInstructionCombiningPass(bool ExpensiveCombines) {
3230 return new InstructionCombiningPass(ExpensiveCombines);