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 "InstCombineInternal.h"
37 #include "llvm-c/Initialization.h"
38 #include "llvm/ADT/APInt.h"
39 #include "llvm/ADT/ArrayRef.h"
40 #include "llvm/ADT/DenseMap.h"
41 #include "llvm/ADT/None.h"
42 #include "llvm/ADT/SmallPtrSet.h"
43 #include "llvm/ADT/SmallVector.h"
44 #include "llvm/ADT/Statistic.h"
45 #include "llvm/ADT/TinyPtrVector.h"
46 #include "llvm/Analysis/AliasAnalysis.h"
47 #include "llvm/Analysis/AssumptionCache.h"
48 #include "llvm/Analysis/BasicAliasAnalysis.h"
49 #include "llvm/Analysis/CFG.h"
50 #include "llvm/Analysis/ConstantFolding.h"
51 #include "llvm/Analysis/EHPersonalities.h"
52 #include "llvm/Analysis/GlobalsModRef.h"
53 #include "llvm/Analysis/InstructionSimplify.h"
54 #include "llvm/Analysis/LoopInfo.h"
55 #include "llvm/Analysis/MemoryBuiltins.h"
56 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
57 #include "llvm/Analysis/TargetFolder.h"
58 #include "llvm/Analysis/TargetLibraryInfo.h"
59 #include "llvm/Analysis/ValueTracking.h"
60 #include "llvm/IR/BasicBlock.h"
61 #include "llvm/IR/CFG.h"
62 #include "llvm/IR/Constant.h"
63 #include "llvm/IR/Constants.h"
64 #include "llvm/IR/DIBuilder.h"
65 #include "llvm/IR/DataLayout.h"
66 #include "llvm/IR/DerivedTypes.h"
67 #include "llvm/IR/Dominators.h"
68 #include "llvm/IR/Function.h"
69 #include "llvm/IR/GetElementPtrTypeIterator.h"
70 #include "llvm/IR/IRBuilder.h"
71 #include "llvm/IR/InstrTypes.h"
72 #include "llvm/IR/Instruction.h"
73 #include "llvm/IR/Instructions.h"
74 #include "llvm/IR/IntrinsicInst.h"
75 #include "llvm/IR/Intrinsics.h"
76 #include "llvm/IR/Metadata.h"
77 #include "llvm/IR/Operator.h"
78 #include "llvm/IR/PassManager.h"
79 #include "llvm/IR/PatternMatch.h"
80 #include "llvm/IR/Type.h"
81 #include "llvm/IR/Use.h"
82 #include "llvm/IR/User.h"
83 #include "llvm/IR/Value.h"
84 #include "llvm/IR/ValueHandle.h"
85 #include "llvm/Pass.h"
86 #include "llvm/Support/CBindingWrapping.h"
87 #include "llvm/Support/Casting.h"
88 #include "llvm/Support/CommandLine.h"
89 #include "llvm/Support/Compiler.h"
90 #include "llvm/Support/Debug.h"
91 #include "llvm/Support/DebugCounter.h"
92 #include "llvm/Support/ErrorHandling.h"
93 #include "llvm/Support/KnownBits.h"
94 #include "llvm/Support/raw_ostream.h"
95 #include "llvm/Transforms/InstCombine/InstCombine.h"
96 #include "llvm/Transforms/InstCombine/InstCombineWorklist.h"
97 #include "llvm/Transforms/Scalar.h"
98 #include "llvm/Transforms/Utils/Local.h"
106 using namespace llvm;
107 using namespace llvm::PatternMatch;
109 #define DEBUG_TYPE "instcombine"
111 STATISTIC(NumCombined , "Number of insts combined");
112 STATISTIC(NumConstProp, "Number of constant folds");
113 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
114 STATISTIC(NumSunkInst , "Number of instructions sunk");
115 STATISTIC(NumExpand, "Number of expansions");
116 STATISTIC(NumFactor , "Number of factorizations");
117 STATISTIC(NumReassoc , "Number of reassociations");
118 DEBUG_COUNTER(VisitCounter, "instcombine-visit",
119 "Controls which instructions are visited");
122 EnableExpensiveCombines("expensive-combines",
123 cl::desc("Enable expensive instruction combines"));
125 static cl::opt<unsigned>
126 MaxArraySize("instcombine-maxarray-size", cl::init(1024),
127 cl::desc("Maximum array size considered when doing a combine"));
129 // FIXME: Remove this flag when it is no longer necessary to convert
130 // llvm.dbg.declare to avoid inaccurate debug info. Setting this to false
131 // increases variable availability at the cost of accuracy. Variables that
132 // cannot be promoted by mem2reg or SROA will be described as living in memory
133 // for their entire lifetime. However, passes like DSE and instcombine can
134 // delete stores to the alloca, leading to misleading and inaccurate debug
135 // information. This flag can be removed when those passes are fixed.
136 static cl::opt<unsigned> ShouldLowerDbgDeclare("instcombine-lower-dbg-declare",
137 cl::Hidden, cl::init(true));
139 Value *InstCombiner::EmitGEPOffset(User *GEP) {
140 return llvm::EmitGEPOffset(&Builder, DL, GEP);
143 /// Return true if it is desirable to convert an integer computation from a
144 /// given bit width to a new bit width.
145 /// We don't want to convert from a legal to an illegal type or from a smaller
146 /// to a larger illegal type. A width of '1' is always treated as a legal type
147 /// because i1 is a fundamental type in IR, and there are many specialized
148 /// optimizations for i1 types.
149 bool InstCombiner::shouldChangeType(unsigned FromWidth,
150 unsigned ToWidth) const {
151 bool FromLegal = FromWidth == 1 || DL.isLegalInteger(FromWidth);
152 bool ToLegal = ToWidth == 1 || DL.isLegalInteger(ToWidth);
154 // If this is a legal integer from type, and the result would be an illegal
155 // type, don't do the transformation.
156 if (FromLegal && !ToLegal)
159 // Otherwise, if both are illegal, do not increase the size of the result. We
160 // do allow things like i160 -> i64, but not i64 -> i160.
161 if (!FromLegal && !ToLegal && ToWidth > FromWidth)
167 /// Return true if it is desirable to convert a computation from 'From' to 'To'.
168 /// We don't want to convert from a legal to an illegal type or from a smaller
169 /// to a larger illegal type. i1 is always treated as a legal type because it is
170 /// a fundamental type in IR, and there are many specialized optimizations for
172 bool InstCombiner::shouldChangeType(Type *From, Type *To) const {
173 assert(From->isIntegerTy() && To->isIntegerTy());
175 unsigned FromWidth = From->getPrimitiveSizeInBits();
176 unsigned ToWidth = To->getPrimitiveSizeInBits();
177 return shouldChangeType(FromWidth, ToWidth);
180 // Return true, if No Signed Wrap should be maintained for I.
181 // The No Signed Wrap flag can be kept if the operation "B (I.getOpcode) C",
182 // where both B and C should be ConstantInts, results in a constant that does
183 // not overflow. This function only handles the Add and Sub opcodes. For
184 // all other opcodes, the function conservatively returns false.
185 static bool MaintainNoSignedWrap(BinaryOperator &I, Value *B, Value *C) {
186 OverflowingBinaryOperator *OBO = dyn_cast<OverflowingBinaryOperator>(&I);
187 if (!OBO || !OBO->hasNoSignedWrap())
190 // We reason about Add and Sub Only.
191 Instruction::BinaryOps Opcode = I.getOpcode();
192 if (Opcode != Instruction::Add && Opcode != Instruction::Sub)
195 const APInt *BVal, *CVal;
196 if (!match(B, m_APInt(BVal)) || !match(C, m_APInt(CVal)))
199 bool Overflow = false;
200 if (Opcode == Instruction::Add)
201 (void)BVal->sadd_ov(*CVal, Overflow);
203 (void)BVal->ssub_ov(*CVal, Overflow);
208 /// Conservatively clears subclassOptionalData after a reassociation or
209 /// commutation. We preserve fast-math flags when applicable as they can be
211 static void ClearSubclassDataAfterReassociation(BinaryOperator &I) {
212 FPMathOperator *FPMO = dyn_cast<FPMathOperator>(&I);
214 I.clearSubclassOptionalData();
218 FastMathFlags FMF = I.getFastMathFlags();
219 I.clearSubclassOptionalData();
220 I.setFastMathFlags(FMF);
223 /// Combine constant operands of associative operations either before or after a
224 /// cast to eliminate one of the associative operations:
225 /// (op (cast (op X, C2)), C1) --> (cast (op X, op (C1, C2)))
226 /// (op (cast (op X, C2)), C1) --> (op (cast X), op (C1, C2))
227 static bool simplifyAssocCastAssoc(BinaryOperator *BinOp1) {
228 auto *Cast = dyn_cast<CastInst>(BinOp1->getOperand(0));
229 if (!Cast || !Cast->hasOneUse())
232 // TODO: Enhance logic for other casts and remove this check.
233 auto CastOpcode = Cast->getOpcode();
234 if (CastOpcode != Instruction::ZExt)
237 // TODO: Enhance logic for other BinOps and remove this check.
238 if (!BinOp1->isBitwiseLogicOp())
241 auto AssocOpcode = BinOp1->getOpcode();
242 auto *BinOp2 = dyn_cast<BinaryOperator>(Cast->getOperand(0));
243 if (!BinOp2 || !BinOp2->hasOneUse() || BinOp2->getOpcode() != AssocOpcode)
247 if (!match(BinOp1->getOperand(1), m_Constant(C1)) ||
248 !match(BinOp2->getOperand(1), m_Constant(C2)))
251 // TODO: This assumes a zext cast.
252 // Eg, if it was a trunc, we'd cast C1 to the source type because casting C2
253 // to the destination type might lose bits.
255 // Fold the constants together in the destination type:
256 // (op (cast (op X, C2)), C1) --> (op (cast X), FoldedC)
257 Type *DestTy = C1->getType();
258 Constant *CastC2 = ConstantExpr::getCast(CastOpcode, C2, DestTy);
259 Constant *FoldedC = ConstantExpr::get(AssocOpcode, C1, CastC2);
260 Cast->setOperand(0, BinOp2->getOperand(0));
261 BinOp1->setOperand(1, FoldedC);
265 /// This performs a few simplifications for operators that are associative or
268 /// Commutative operators:
270 /// 1. Order operands such that they are listed from right (least complex) to
271 /// left (most complex). This puts constants before unary operators before
272 /// binary operators.
274 /// Associative operators:
276 /// 2. Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
277 /// 3. Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
279 /// Associative and commutative operators:
281 /// 4. Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
282 /// 5. Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
283 /// 6. Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
284 /// if C1 and C2 are constants.
285 bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) {
286 Instruction::BinaryOps Opcode = I.getOpcode();
287 bool Changed = false;
290 // Order operands such that they are listed from right (least complex) to
291 // left (most complex). This puts constants before unary operators before
293 if (I.isCommutative() && getComplexity(I.getOperand(0)) <
294 getComplexity(I.getOperand(1)))
295 Changed = !I.swapOperands();
297 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
298 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
300 if (I.isAssociative()) {
301 // Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
302 if (Op0 && Op0->getOpcode() == Opcode) {
303 Value *A = Op0->getOperand(0);
304 Value *B = Op0->getOperand(1);
305 Value *C = I.getOperand(1);
307 // Does "B op C" simplify?
308 if (Value *V = SimplifyBinOp(Opcode, B, C, SQ.getWithInstruction(&I))) {
309 // It simplifies to V. Form "A op V".
312 // Conservatively clear the optional flags, since they may not be
313 // preserved by the reassociation.
314 if (MaintainNoSignedWrap(I, B, C) &&
315 (!Op0 || (isa<BinaryOperator>(Op0) && Op0->hasNoSignedWrap()))) {
316 // Note: this is only valid because SimplifyBinOp doesn't look at
317 // the operands to Op0.
318 I.clearSubclassOptionalData();
319 I.setHasNoSignedWrap(true);
321 ClearSubclassDataAfterReassociation(I);
330 // Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
331 if (Op1 && Op1->getOpcode() == Opcode) {
332 Value *A = I.getOperand(0);
333 Value *B = Op1->getOperand(0);
334 Value *C = Op1->getOperand(1);
336 // Does "A op B" simplify?
337 if (Value *V = SimplifyBinOp(Opcode, A, B, SQ.getWithInstruction(&I))) {
338 // It simplifies to V. Form "V op C".
341 // Conservatively clear the optional flags, since they may not be
342 // preserved by the reassociation.
343 ClearSubclassDataAfterReassociation(I);
351 if (I.isAssociative() && I.isCommutative()) {
352 if (simplifyAssocCastAssoc(&I)) {
358 // Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
359 if (Op0 && Op0->getOpcode() == Opcode) {
360 Value *A = Op0->getOperand(0);
361 Value *B = Op0->getOperand(1);
362 Value *C = I.getOperand(1);
364 // Does "C op A" simplify?
365 if (Value *V = SimplifyBinOp(Opcode, C, A, SQ.getWithInstruction(&I))) {
366 // It simplifies to V. Form "V op B".
369 // Conservatively clear the optional flags, since they may not be
370 // preserved by the reassociation.
371 ClearSubclassDataAfterReassociation(I);
378 // Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
379 if (Op1 && Op1->getOpcode() == Opcode) {
380 Value *A = I.getOperand(0);
381 Value *B = Op1->getOperand(0);
382 Value *C = Op1->getOperand(1);
384 // Does "C op A" simplify?
385 if (Value *V = SimplifyBinOp(Opcode, C, A, SQ.getWithInstruction(&I))) {
386 // It simplifies to V. Form "B op V".
389 // Conservatively clear the optional flags, since they may not be
390 // preserved by the reassociation.
391 ClearSubclassDataAfterReassociation(I);
398 // Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
399 // if C1 and C2 are constants.
401 Op0->getOpcode() == Opcode && Op1->getOpcode() == Opcode &&
402 isa<Constant>(Op0->getOperand(1)) &&
403 isa<Constant>(Op1->getOperand(1)) &&
404 Op0->hasOneUse() && Op1->hasOneUse()) {
405 Value *A = Op0->getOperand(0);
406 Constant *C1 = cast<Constant>(Op0->getOperand(1));
407 Value *B = Op1->getOperand(0);
408 Constant *C2 = cast<Constant>(Op1->getOperand(1));
410 Constant *Folded = ConstantExpr::get(Opcode, C1, C2);
411 BinaryOperator *New = BinaryOperator::Create(Opcode, A, B);
412 if (isa<FPMathOperator>(New)) {
413 FastMathFlags Flags = I.getFastMathFlags();
414 Flags &= Op0->getFastMathFlags();
415 Flags &= Op1->getFastMathFlags();
416 New->setFastMathFlags(Flags);
418 InsertNewInstWith(New, I);
420 I.setOperand(0, New);
421 I.setOperand(1, Folded);
422 // Conservatively clear the optional flags, since they may not be
423 // preserved by the reassociation.
424 ClearSubclassDataAfterReassociation(I);
431 // No further simplifications.
436 /// Return whether "X LOp (Y ROp Z)" is always equal to
437 /// "(X LOp Y) ROp (X LOp Z)".
438 static bool LeftDistributesOverRight(Instruction::BinaryOps LOp,
439 Instruction::BinaryOps ROp) {
444 case Instruction::And:
445 // And distributes over Or and Xor.
449 case Instruction::Or:
450 case Instruction::Xor:
454 case Instruction::Mul:
455 // Multiplication distributes over addition and subtraction.
459 case Instruction::Add:
460 case Instruction::Sub:
464 case Instruction::Or:
465 // Or distributes over And.
469 case Instruction::And:
475 /// Return whether "(X LOp Y) ROp Z" is always equal to
476 /// "(X ROp Z) LOp (Y ROp Z)".
477 static bool RightDistributesOverLeft(Instruction::BinaryOps LOp,
478 Instruction::BinaryOps ROp) {
479 if (Instruction::isCommutative(ROp))
480 return LeftDistributesOverRight(ROp, LOp);
485 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
486 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
487 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
488 case Instruction::And:
489 case Instruction::Or:
490 case Instruction::Xor:
494 case Instruction::Shl:
495 case Instruction::LShr:
496 case Instruction::AShr:
500 // TODO: It would be nice to handle division, aka "(X + Y)/Z = X/Z + Y/Z",
501 // but this requires knowing that the addition does not overflow and other
506 /// This function returns identity value for given opcode, which can be used to
507 /// factor patterns like (X * 2) + X ==> (X * 2) + (X * 1) ==> X * (2 + 1).
508 static Value *getIdentityValue(Instruction::BinaryOps Opcode, Value *V) {
509 if (isa<Constant>(V))
512 return ConstantExpr::getBinOpIdentity(Opcode, V->getType());
515 /// This function factors binary ops which can be combined using distributive
516 /// laws. This function tries to transform 'Op' based TopLevelOpcode to enable
517 /// factorization e.g for ADD(SHL(X , 2), MUL(X, 5)), When this function called
518 /// with TopLevelOpcode == Instruction::Add and Op = SHL(X, 2), transforms
519 /// SHL(X, 2) to MUL(X, 4) i.e. returns Instruction::Mul with LHS set to 'X' and
521 static Instruction::BinaryOps
522 getBinOpsForFactorization(Instruction::BinaryOps TopLevelOpcode,
523 BinaryOperator *Op, Value *&LHS, Value *&RHS) {
524 assert(Op && "Expected a binary operator");
526 LHS = Op->getOperand(0);
527 RHS = Op->getOperand(1);
529 switch (TopLevelOpcode) {
531 return Op->getOpcode();
533 case Instruction::Add:
534 case Instruction::Sub:
535 if (Op->getOpcode() == Instruction::Shl) {
536 if (Constant *CST = dyn_cast<Constant>(Op->getOperand(1))) {
537 // The multiplier is really 1 << CST.
538 RHS = ConstantExpr::getShl(ConstantInt::get(Op->getType(), 1), CST);
539 return Instruction::Mul;
542 return Op->getOpcode();
545 // TODO: We can add other conversions e.g. shr => div etc.
548 /// This tries to simplify binary operations by factorizing out common terms
549 /// (e. g. "(A*B)+(A*C)" -> "A*(B+C)").
550 Value *InstCombiner::tryFactorization(BinaryOperator &I,
551 Instruction::BinaryOps InnerOpcode,
552 Value *A, Value *B, Value *C, Value *D) {
553 assert(A && B && C && D && "All values must be provided");
556 Value *SimplifiedInst = nullptr;
557 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
558 Instruction::BinaryOps TopLevelOpcode = I.getOpcode();
560 // Does "X op' Y" always equal "Y op' X"?
561 bool InnerCommutative = Instruction::isCommutative(InnerOpcode);
563 // Does "X op' (Y op Z)" always equal "(X op' Y) op (X op' Z)"?
564 if (LeftDistributesOverRight(InnerOpcode, TopLevelOpcode))
565 // Does the instruction have the form "(A op' B) op (A op' D)" or, in the
566 // commutative case, "(A op' B) op (C op' A)"?
567 if (A == C || (InnerCommutative && A == D)) {
570 // Consider forming "A op' (B op D)".
571 // If "B op D" simplifies then it can be formed with no cost.
572 V = SimplifyBinOp(TopLevelOpcode, B, D, SQ.getWithInstruction(&I));
573 // If "B op D" doesn't simplify then only go on if both of the existing
574 // operations "A op' B" and "C op' D" will be zapped as no longer used.
575 if (!V && LHS->hasOneUse() && RHS->hasOneUse())
576 V = Builder.CreateBinOp(TopLevelOpcode, B, D, RHS->getName());
578 SimplifiedInst = Builder.CreateBinOp(InnerOpcode, A, V);
582 // Does "(X op Y) op' Z" always equal "(X op' Z) op (Y op' Z)"?
583 if (!SimplifiedInst && RightDistributesOverLeft(TopLevelOpcode, InnerOpcode))
584 // Does the instruction have the form "(A op' B) op (C op' B)" or, in the
585 // commutative case, "(A op' B) op (B op' D)"?
586 if (B == D || (InnerCommutative && B == C)) {
589 // Consider forming "(A op C) op' B".
590 // If "A op C" simplifies then it can be formed with no cost.
591 V = SimplifyBinOp(TopLevelOpcode, A, C, SQ.getWithInstruction(&I));
593 // If "A op C" doesn't simplify then only go on if both of the existing
594 // operations "A op' B" and "C op' D" will be zapped as no longer used.
595 if (!V && LHS->hasOneUse() && RHS->hasOneUse())
596 V = Builder.CreateBinOp(TopLevelOpcode, A, C, LHS->getName());
598 SimplifiedInst = Builder.CreateBinOp(InnerOpcode, V, B);
602 if (SimplifiedInst) {
604 SimplifiedInst->takeName(&I);
606 // Check if we can add NSW flag to SimplifiedInst. If so, set NSW flag.
607 // TODO: Check for NUW.
608 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(SimplifiedInst)) {
609 if (isa<OverflowingBinaryOperator>(SimplifiedInst)) {
611 if (isa<OverflowingBinaryOperator>(&I))
612 HasNSW = I.hasNoSignedWrap();
614 if (auto *LOBO = dyn_cast<OverflowingBinaryOperator>(LHS))
615 HasNSW &= LOBO->hasNoSignedWrap();
617 if (auto *ROBO = dyn_cast<OverflowingBinaryOperator>(RHS))
618 HasNSW &= ROBO->hasNoSignedWrap();
620 // We can propagate 'nsw' if we know that
621 // %Y = mul nsw i16 %X, C
622 // %Z = add nsw i16 %Y, %X
624 // %Z = mul nsw i16 %X, C+1
626 // iff C+1 isn't INT_MIN
628 if (TopLevelOpcode == Instruction::Add &&
629 InnerOpcode == Instruction::Mul)
630 if (match(V, m_APInt(CInt)) && !CInt->isMinSignedValue())
631 BO->setHasNoSignedWrap(HasNSW);
635 return SimplifiedInst;
638 /// This tries to simplify binary operations which some other binary operation
639 /// distributes over either by factorizing out common terms
640 /// (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this results in
641 /// simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is a win).
642 /// Returns the simplified value, or null if it didn't simplify.
643 Value *InstCombiner::SimplifyUsingDistributiveLaws(BinaryOperator &I) {
644 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
645 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
646 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
647 Instruction::BinaryOps TopLevelOpcode = I.getOpcode();
651 Value *A, *B, *C, *D;
652 Instruction::BinaryOps LHSOpcode, RHSOpcode;
654 LHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op0, A, B);
656 RHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op1, C, D);
658 // The instruction has the form "(A op' B) op (C op' D)". Try to factorize
660 if (Op0 && Op1 && LHSOpcode == RHSOpcode)
661 if (Value *V = tryFactorization(I, LHSOpcode, A, B, C, D))
664 // The instruction has the form "(A op' B) op (C)". Try to factorize common
667 if (Value *Ident = getIdentityValue(LHSOpcode, RHS))
669 tryFactorization(I, LHSOpcode, A, B, RHS, Ident))
672 // The instruction has the form "(B) op (C op' D)". Try to factorize common
675 if (Value *Ident = getIdentityValue(RHSOpcode, LHS))
677 tryFactorization(I, RHSOpcode, LHS, Ident, C, D))
682 if (Op0 && RightDistributesOverLeft(Op0->getOpcode(), TopLevelOpcode)) {
683 // The instruction has the form "(A op' B) op C". See if expanding it out
684 // to "(A op C) op' (B op C)" results in simplifications.
685 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
686 Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op'
688 Value *L = SimplifyBinOp(TopLevelOpcode, A, C, SQ.getWithInstruction(&I));
689 Value *R = SimplifyBinOp(TopLevelOpcode, B, C, SQ.getWithInstruction(&I));
691 // Do "A op C" and "B op C" both simplify?
693 // They do! Return "L op' R".
695 C = Builder.CreateBinOp(InnerOpcode, L, R);
700 // Does "A op C" simplify to the identity value for the inner opcode?
701 if (L && L == ConstantExpr::getBinOpIdentity(InnerOpcode, L->getType())) {
702 // They do! Return "B op C".
704 C = Builder.CreateBinOp(TopLevelOpcode, B, C);
709 // Does "B op C" simplify to the identity value for the inner opcode?
710 if (R && R == ConstantExpr::getBinOpIdentity(InnerOpcode, R->getType())) {
711 // They do! Return "A op C".
713 C = Builder.CreateBinOp(TopLevelOpcode, A, C);
719 if (Op1 && LeftDistributesOverRight(TopLevelOpcode, Op1->getOpcode())) {
720 // The instruction has the form "A op (B op' C)". See if expanding it out
721 // to "(A op B) op' (A op C)" results in simplifications.
722 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
723 Instruction::BinaryOps InnerOpcode = Op1->getOpcode(); // op'
725 Value *L = SimplifyBinOp(TopLevelOpcode, A, B, SQ.getWithInstruction(&I));
726 Value *R = SimplifyBinOp(TopLevelOpcode, A, C, SQ.getWithInstruction(&I));
728 // Do "A op B" and "A op C" both simplify?
730 // They do! Return "L op' R".
732 A = Builder.CreateBinOp(InnerOpcode, L, R);
737 // Does "A op B" simplify to the identity value for the inner opcode?
738 if (L && L == ConstantExpr::getBinOpIdentity(InnerOpcode, L->getType())) {
739 // They do! Return "A op C".
741 A = Builder.CreateBinOp(TopLevelOpcode, A, C);
746 // Does "A op C" simplify to the identity value for the inner opcode?
747 if (R && R == ConstantExpr::getBinOpIdentity(InnerOpcode, R->getType())) {
748 // They do! Return "A op B".
750 A = Builder.CreateBinOp(TopLevelOpcode, A, B);
756 return SimplifySelectsFeedingBinaryOp(I, LHS, RHS);
759 Value *InstCombiner::SimplifySelectsFeedingBinaryOp(BinaryOperator &I,
760 Value *LHS, Value *RHS) {
761 Instruction::BinaryOps Opcode = I.getOpcode();
762 // (op (select (a, b, c)), (select (a, d, e))) -> (select (a, (op b, d), (op
764 Value *A, *B, *C, *D, *E;
766 if (match(LHS, m_Select(m_Value(A), m_Value(B), m_Value(C))) &&
767 match(RHS, m_Select(m_Specific(A), m_Value(D), m_Value(E)))) {
768 bool SelectsHaveOneUse = LHS->hasOneUse() && RHS->hasOneUse();
769 BuilderTy::FastMathFlagGuard Guard(Builder);
770 if (isa<FPMathOperator>(&I))
771 Builder.setFastMathFlags(I.getFastMathFlags());
773 Value *V1 = SimplifyBinOp(Opcode, C, E, SQ.getWithInstruction(&I));
774 Value *V2 = SimplifyBinOp(Opcode, B, D, SQ.getWithInstruction(&I));
776 SI = Builder.CreateSelect(A, V2, V1);
777 else if (V2 && SelectsHaveOneUse)
778 SI = Builder.CreateSelect(A, V2, Builder.CreateBinOp(Opcode, C, E));
779 else if (V1 && SelectsHaveOneUse)
780 SI = Builder.CreateSelect(A, Builder.CreateBinOp(Opcode, B, D), V1);
789 /// Given a 'sub' instruction, return the RHS of the instruction if the LHS is a
790 /// constant zero (which is the 'negate' form).
791 Value *InstCombiner::dyn_castNegVal(Value *V) const {
792 if (BinaryOperator::isNeg(V))
793 return BinaryOperator::getNegArgument(V);
795 // Constants can be considered to be negated values if they can be folded.
796 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
797 return ConstantExpr::getNeg(C);
799 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
800 if (C->getType()->getElementType()->isIntegerTy())
801 return ConstantExpr::getNeg(C);
803 if (ConstantVector *CV = dyn_cast<ConstantVector>(V)) {
804 for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) {
805 Constant *Elt = CV->getAggregateElement(i);
809 if (isa<UndefValue>(Elt))
812 if (!isa<ConstantInt>(Elt))
815 return ConstantExpr::getNeg(CV);
821 /// Given a 'fsub' instruction, return the RHS of the instruction if the LHS is
822 /// a constant negative zero (which is the 'negate' form).
823 Value *InstCombiner::dyn_castFNegVal(Value *V, bool IgnoreZeroSign) const {
824 if (BinaryOperator::isFNeg(V, IgnoreZeroSign))
825 return BinaryOperator::getFNegArgument(V);
827 // Constants can be considered to be negated values if they can be folded.
828 if (ConstantFP *C = dyn_cast<ConstantFP>(V))
829 return ConstantExpr::getFNeg(C);
831 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
832 if (C->getType()->getElementType()->isFloatingPointTy())
833 return ConstantExpr::getFNeg(C);
838 static Value *foldOperationIntoSelectOperand(Instruction &I, Value *SO,
839 InstCombiner::BuilderTy &Builder) {
840 if (auto *Cast = dyn_cast<CastInst>(&I))
841 return Builder.CreateCast(Cast->getOpcode(), SO, I.getType());
843 assert(I.isBinaryOp() && "Unexpected opcode for select folding");
845 // Figure out if the constant is the left or the right argument.
846 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
847 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
849 if (auto *SOC = dyn_cast<Constant>(SO)) {
851 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
852 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
855 Value *Op0 = SO, *Op1 = ConstOperand;
859 auto *BO = cast<BinaryOperator>(&I);
860 Value *RI = Builder.CreateBinOp(BO->getOpcode(), Op0, Op1,
861 SO->getName() + ".op");
862 auto *FPInst = dyn_cast<Instruction>(RI);
863 if (FPInst && isa<FPMathOperator>(FPInst))
864 FPInst->copyFastMathFlags(BO);
868 Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) {
869 // Don't modify shared select instructions.
870 if (!SI->hasOneUse())
873 Value *TV = SI->getTrueValue();
874 Value *FV = SI->getFalseValue();
875 if (!(isa<Constant>(TV) || isa<Constant>(FV)))
878 // Bool selects with constant operands can be folded to logical ops.
879 if (SI->getType()->isIntOrIntVectorTy(1))
882 // If it's a bitcast involving vectors, make sure it has the same number of
883 // elements on both sides.
884 if (auto *BC = dyn_cast<BitCastInst>(&Op)) {
885 VectorType *DestTy = dyn_cast<VectorType>(BC->getDestTy());
886 VectorType *SrcTy = dyn_cast<VectorType>(BC->getSrcTy());
888 // Verify that either both or neither are vectors.
889 if ((SrcTy == nullptr) != (DestTy == nullptr))
892 // If vectors, verify that they have the same number of elements.
893 if (SrcTy && SrcTy->getNumElements() != DestTy->getNumElements())
897 // Test if a CmpInst instruction is used exclusively by a select as
898 // part of a minimum or maximum operation. If so, refrain from doing
899 // any other folding. This helps out other analyses which understand
900 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
901 // and CodeGen. And in this case, at least one of the comparison
902 // operands has at least one user besides the compare (the select),
903 // which would often largely negate the benefit of folding anyway.
904 if (auto *CI = dyn_cast<CmpInst>(SI->getCondition())) {
905 if (CI->hasOneUse()) {
906 Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1);
907 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
908 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
913 Value *NewTV = foldOperationIntoSelectOperand(Op, TV, Builder);
914 Value *NewFV = foldOperationIntoSelectOperand(Op, FV, Builder);
915 return SelectInst::Create(SI->getCondition(), NewTV, NewFV, "", nullptr, SI);
918 static Value *foldOperationIntoPhiValue(BinaryOperator *I, Value *InV,
919 InstCombiner::BuilderTy &Builder) {
920 bool ConstIsRHS = isa<Constant>(I->getOperand(1));
921 Constant *C = cast<Constant>(I->getOperand(ConstIsRHS));
923 if (auto *InC = dyn_cast<Constant>(InV)) {
925 return ConstantExpr::get(I->getOpcode(), InC, C);
926 return ConstantExpr::get(I->getOpcode(), C, InC);
929 Value *Op0 = InV, *Op1 = C;
933 Value *RI = Builder.CreateBinOp(I->getOpcode(), Op0, Op1, "phitmp");
934 auto *FPInst = dyn_cast<Instruction>(RI);
935 if (FPInst && isa<FPMathOperator>(FPInst))
936 FPInst->copyFastMathFlags(I);
940 Instruction *InstCombiner::foldOpIntoPhi(Instruction &I, PHINode *PN) {
941 unsigned NumPHIValues = PN->getNumIncomingValues();
942 if (NumPHIValues == 0)
945 // We normally only transform phis with a single use. However, if a PHI has
946 // multiple uses and they are all the same operation, we can fold *all* of the
947 // uses into the PHI.
948 if (!PN->hasOneUse()) {
949 // Walk the use list for the instruction, comparing them to I.
950 for (User *U : PN->users()) {
951 Instruction *UI = cast<Instruction>(U);
952 if (UI != &I && !I.isIdenticalTo(UI))
955 // Otherwise, we can replace *all* users with the new PHI we form.
958 // Check to see if all of the operands of the PHI are simple constants
959 // (constantint/constantfp/undef). If there is one non-constant value,
960 // remember the BB it is in. If there is more than one or if *it* is a PHI,
961 // bail out. We don't do arbitrary constant expressions here because moving
962 // their computation can be expensive without a cost model.
963 BasicBlock *NonConstBB = nullptr;
964 for (unsigned i = 0; i != NumPHIValues; ++i) {
965 Value *InVal = PN->getIncomingValue(i);
966 if (isa<Constant>(InVal) && !isa<ConstantExpr>(InVal))
969 if (isa<PHINode>(InVal)) return nullptr; // Itself a phi.
970 if (NonConstBB) return nullptr; // More than one non-const value.
972 NonConstBB = PN->getIncomingBlock(i);
974 // If the InVal is an invoke at the end of the pred block, then we can't
975 // insert a computation after it without breaking the edge.
976 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal))
977 if (II->getParent() == NonConstBB)
980 // If the incoming non-constant value is in I's block, we will remove one
981 // instruction, but insert another equivalent one, leading to infinite
983 if (isPotentiallyReachable(I.getParent(), NonConstBB, &DT, LI))
987 // If there is exactly one non-constant value, we can insert a copy of the
988 // operation in that block. However, if this is a critical edge, we would be
989 // inserting the computation on some other paths (e.g. inside a loop). Only
990 // do this if the pred block is unconditionally branching into the phi block.
991 if (NonConstBB != nullptr) {
992 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
993 if (!BI || !BI->isUnconditional()) return nullptr;
996 // Okay, we can do the transformation: create the new PHI node.
997 PHINode *NewPN = PHINode::Create(I.getType(), PN->getNumIncomingValues());
998 InsertNewInstBefore(NewPN, *PN);
1001 // If we are going to have to insert a new computation, do so right before the
1002 // predecessor's terminator.
1004 Builder.SetInsertPoint(NonConstBB->getTerminator());
1006 // Next, add all of the operands to the PHI.
1007 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) {
1008 // We only currently try to fold the condition of a select when it is a phi,
1009 // not the true/false values.
1010 Value *TrueV = SI->getTrueValue();
1011 Value *FalseV = SI->getFalseValue();
1012 BasicBlock *PhiTransBB = PN->getParent();
1013 for (unsigned i = 0; i != NumPHIValues; ++i) {
1014 BasicBlock *ThisBB = PN->getIncomingBlock(i);
1015 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB);
1016 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB);
1017 Value *InV = nullptr;
1018 // Beware of ConstantExpr: it may eventually evaluate to getNullValue,
1019 // even if currently isNullValue gives false.
1020 Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i));
1021 // For vector constants, we cannot use isNullValue to fold into
1022 // FalseVInPred versus TrueVInPred. When we have individual nonzero
1023 // elements in the vector, we will incorrectly fold InC to
1025 if (InC && !isa<ConstantExpr>(InC) && isa<ConstantInt>(InC))
1026 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred;
1028 // Generate the select in the same block as PN's current incoming block.
1029 // Note: ThisBB need not be the NonConstBB because vector constants
1030 // which are constants by definition are handled here.
1031 // FIXME: This can lead to an increase in IR generation because we might
1032 // generate selects for vector constant phi operand, that could not be
1033 // folded to TrueVInPred or FalseVInPred as done for ConstantInt. For
1034 // non-vector phis, this transformation was always profitable because
1035 // the select would be generated exactly once in the NonConstBB.
1036 Builder.SetInsertPoint(ThisBB->getTerminator());
1037 InV = Builder.CreateSelect(PN->getIncomingValue(i), TrueVInPred,
1038 FalseVInPred, "phitmp");
1040 NewPN->addIncoming(InV, ThisBB);
1042 } else if (CmpInst *CI = dyn_cast<CmpInst>(&I)) {
1043 Constant *C = cast<Constant>(I.getOperand(1));
1044 for (unsigned i = 0; i != NumPHIValues; ++i) {
1045 Value *InV = nullptr;
1046 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
1047 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1048 else if (isa<ICmpInst>(CI))
1049 InV = Builder.CreateICmp(CI->getPredicate(), PN->getIncomingValue(i),
1052 InV = Builder.CreateFCmp(CI->getPredicate(), PN->getIncomingValue(i),
1054 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1056 } else if (auto *BO = dyn_cast<BinaryOperator>(&I)) {
1057 for (unsigned i = 0; i != NumPHIValues; ++i) {
1058 Value *InV = foldOperationIntoPhiValue(BO, PN->getIncomingValue(i),
1060 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1063 CastInst *CI = cast<CastInst>(&I);
1064 Type *RetTy = CI->getType();
1065 for (unsigned i = 0; i != NumPHIValues; ++i) {
1067 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
1068 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1070 InV = Builder.CreateCast(CI->getOpcode(), PN->getIncomingValue(i),
1071 I.getType(), "phitmp");
1072 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1076 for (auto UI = PN->user_begin(), E = PN->user_end(); UI != E;) {
1077 Instruction *User = cast<Instruction>(*UI++);
1078 if (User == &I) continue;
1079 replaceInstUsesWith(*User, NewPN);
1080 eraseInstFromFunction(*User);
1082 return replaceInstUsesWith(I, NewPN);
1085 Instruction *InstCombiner::foldOpWithConstantIntoOperand(BinaryOperator &I) {
1086 assert(isa<Constant>(I.getOperand(1)) && "Unexpected operand type");
1088 if (auto *Sel = dyn_cast<SelectInst>(I.getOperand(0))) {
1089 if (Instruction *NewSel = FoldOpIntoSelect(I, Sel))
1091 } else if (auto *PN = dyn_cast<PHINode>(I.getOperand(0))) {
1092 if (Instruction *NewPhi = foldOpIntoPhi(I, PN))
1098 /// Given a pointer type and a constant offset, determine whether or not there
1099 /// is a sequence of GEP indices into the pointed type that will land us at the
1100 /// specified offset. If so, fill them into NewIndices and return the resultant
1101 /// element type, otherwise return null.
1102 Type *InstCombiner::FindElementAtOffset(PointerType *PtrTy, int64_t Offset,
1103 SmallVectorImpl<Value *> &NewIndices) {
1104 Type *Ty = PtrTy->getElementType();
1108 // Start with the index over the outer type. Note that the type size
1109 // might be zero (even if the offset isn't zero) if the indexed type
1110 // is something like [0 x {int, int}]
1111 Type *IntPtrTy = DL.getIntPtrType(PtrTy);
1112 int64_t FirstIdx = 0;
1113 if (int64_t TySize = DL.getTypeAllocSize(Ty)) {
1114 FirstIdx = Offset/TySize;
1115 Offset -= FirstIdx*TySize;
1117 // Handle hosts where % returns negative instead of values [0..TySize).
1121 assert(Offset >= 0);
1123 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
1126 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
1128 // Index into the types. If we fail, set OrigBase to null.
1130 // Indexing into tail padding between struct/array elements.
1131 if (uint64_t(Offset * 8) >= DL.getTypeSizeInBits(Ty))
1134 if (StructType *STy = dyn_cast<StructType>(Ty)) {
1135 const StructLayout *SL = DL.getStructLayout(STy);
1136 assert(Offset < (int64_t)SL->getSizeInBytes() &&
1137 "Offset must stay within the indexed type");
1139 unsigned Elt = SL->getElementContainingOffset(Offset);
1140 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
1143 Offset -= SL->getElementOffset(Elt);
1144 Ty = STy->getElementType(Elt);
1145 } else if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
1146 uint64_t EltSize = DL.getTypeAllocSize(AT->getElementType());
1147 assert(EltSize && "Cannot index into a zero-sized array");
1148 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
1150 Ty = AT->getElementType();
1152 // Otherwise, we can't index into the middle of this atomic type, bail.
1160 static bool shouldMergeGEPs(GEPOperator &GEP, GEPOperator &Src) {
1161 // If this GEP has only 0 indices, it is the same pointer as
1162 // Src. If Src is not a trivial GEP too, don't combine
1164 if (GEP.hasAllZeroIndices() && !Src.hasAllZeroIndices() &&
1170 /// Return a value X such that Val = X * Scale, or null if none.
1171 /// If the multiplication is known not to overflow, then NoSignedWrap is set.
1172 Value *InstCombiner::Descale(Value *Val, APInt Scale, bool &NoSignedWrap) {
1173 assert(isa<IntegerType>(Val->getType()) && "Can only descale integers!");
1174 assert(cast<IntegerType>(Val->getType())->getBitWidth() ==
1175 Scale.getBitWidth() && "Scale not compatible with value!");
1177 // If Val is zero or Scale is one then Val = Val * Scale.
1178 if (match(Val, m_Zero()) || Scale == 1) {
1179 NoSignedWrap = true;
1183 // If Scale is zero then it does not divide Val.
1184 if (Scale.isMinValue())
1187 // Look through chains of multiplications, searching for a constant that is
1188 // divisible by Scale. For example, descaling X*(Y*(Z*4)) by a factor of 4
1189 // will find the constant factor 4 and produce X*(Y*Z). Descaling X*(Y*8) by
1190 // a factor of 4 will produce X*(Y*2). The principle of operation is to bore
1193 // Val = M1 * X || Analysis starts here and works down
1194 // M1 = M2 * Y || Doesn't descend into terms with more
1195 // M2 = Z * 4 \/ than one use
1197 // Then to modify a term at the bottom:
1200 // M1 = Z * Y || Replaced M2 with Z
1202 // Then to work back up correcting nsw flags.
1204 // Op - the term we are currently analyzing. Starts at Val then drills down.
1205 // Replaced with its descaled value before exiting from the drill down loop.
1208 // Parent - initially null, but after drilling down notes where Op came from.
1209 // In the example above, Parent is (Val, 0) when Op is M1, because M1 is the
1210 // 0'th operand of Val.
1211 std::pair<Instruction *, unsigned> Parent;
1213 // Set if the transform requires a descaling at deeper levels that doesn't
1215 bool RequireNoSignedWrap = false;
1217 // Log base 2 of the scale. Negative if not a power of 2.
1218 int32_t logScale = Scale.exactLogBase2();
1220 for (;; Op = Parent.first->getOperand(Parent.second)) { // Drill down
1221 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
1222 // If Op is a constant divisible by Scale then descale to the quotient.
1223 APInt Quotient(Scale), Remainder(Scale); // Init ensures right bitwidth.
1224 APInt::sdivrem(CI->getValue(), Scale, Quotient, Remainder);
1225 if (!Remainder.isMinValue())
1226 // Not divisible by Scale.
1228 // Replace with the quotient in the parent.
1229 Op = ConstantInt::get(CI->getType(), Quotient);
1230 NoSignedWrap = true;
1234 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op)) {
1235 if (BO->getOpcode() == Instruction::Mul) {
1237 NoSignedWrap = BO->hasNoSignedWrap();
1238 if (RequireNoSignedWrap && !NoSignedWrap)
1241 // There are three cases for multiplication: multiplication by exactly
1242 // the scale, multiplication by a constant different to the scale, and
1243 // multiplication by something else.
1244 Value *LHS = BO->getOperand(0);
1245 Value *RHS = BO->getOperand(1);
1247 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
1248 // Multiplication by a constant.
1249 if (CI->getValue() == Scale) {
1250 // Multiplication by exactly the scale, replace the multiplication
1251 // by its left-hand side in the parent.
1256 // Otherwise drill down into the constant.
1257 if (!Op->hasOneUse())
1260 Parent = std::make_pair(BO, 1);
1264 // Multiplication by something else. Drill down into the left-hand side
1265 // since that's where the reassociate pass puts the good stuff.
1266 if (!Op->hasOneUse())
1269 Parent = std::make_pair(BO, 0);
1273 if (logScale > 0 && BO->getOpcode() == Instruction::Shl &&
1274 isa<ConstantInt>(BO->getOperand(1))) {
1275 // Multiplication by a power of 2.
1276 NoSignedWrap = BO->hasNoSignedWrap();
1277 if (RequireNoSignedWrap && !NoSignedWrap)
1280 Value *LHS = BO->getOperand(0);
1281 int32_t Amt = cast<ConstantInt>(BO->getOperand(1))->
1282 getLimitedValue(Scale.getBitWidth());
1285 if (Amt == logScale) {
1286 // Multiplication by exactly the scale, replace the multiplication
1287 // by its left-hand side in the parent.
1291 if (Amt < logScale || !Op->hasOneUse())
1294 // Multiplication by more than the scale. Reduce the multiplying amount
1295 // by the scale in the parent.
1296 Parent = std::make_pair(BO, 1);
1297 Op = ConstantInt::get(BO->getType(), Amt - logScale);
1302 if (!Op->hasOneUse())
1305 if (CastInst *Cast = dyn_cast<CastInst>(Op)) {
1306 if (Cast->getOpcode() == Instruction::SExt) {
1307 // Op is sign-extended from a smaller type, descale in the smaller type.
1308 unsigned SmallSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1309 APInt SmallScale = Scale.trunc(SmallSize);
1310 // Suppose Op = sext X, and we descale X as Y * SmallScale. We want to
1311 // descale Op as (sext Y) * Scale. In order to have
1312 // sext (Y * SmallScale) = (sext Y) * Scale
1313 // some conditions need to hold however: SmallScale must sign-extend to
1314 // Scale and the multiplication Y * SmallScale should not overflow.
1315 if (SmallScale.sext(Scale.getBitWidth()) != Scale)
1316 // SmallScale does not sign-extend to Scale.
1318 assert(SmallScale.exactLogBase2() == logScale);
1319 // Require that Y * SmallScale must not overflow.
1320 RequireNoSignedWrap = true;
1322 // Drill down through the cast.
1323 Parent = std::make_pair(Cast, 0);
1328 if (Cast->getOpcode() == Instruction::Trunc) {
1329 // Op is truncated from a larger type, descale in the larger type.
1330 // Suppose Op = trunc X, and we descale X as Y * sext Scale. Then
1331 // trunc (Y * sext Scale) = (trunc Y) * Scale
1332 // always holds. However (trunc Y) * Scale may overflow even if
1333 // trunc (Y * sext Scale) does not, so nsw flags need to be cleared
1334 // from this point up in the expression (see later).
1335 if (RequireNoSignedWrap)
1338 // Drill down through the cast.
1339 unsigned LargeSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1340 Parent = std::make_pair(Cast, 0);
1341 Scale = Scale.sext(LargeSize);
1342 if (logScale + 1 == (int32_t)Cast->getType()->getPrimitiveSizeInBits())
1344 assert(Scale.exactLogBase2() == logScale);
1349 // Unsupported expression, bail out.
1353 // If Op is zero then Val = Op * Scale.
1354 if (match(Op, m_Zero())) {
1355 NoSignedWrap = true;
1359 // We know that we can successfully descale, so from here on we can safely
1360 // modify the IR. Op holds the descaled version of the deepest term in the
1361 // expression. NoSignedWrap is 'true' if multiplying Op by Scale is known
1365 // The expression only had one term.
1368 // Rewrite the parent using the descaled version of its operand.
1369 assert(Parent.first->hasOneUse() && "Drilled down when more than one use!");
1370 assert(Op != Parent.first->getOperand(Parent.second) &&
1371 "Descaling was a no-op?");
1372 Parent.first->setOperand(Parent.second, Op);
1373 Worklist.Add(Parent.first);
1375 // Now work back up the expression correcting nsw flags. The logic is based
1376 // on the following observation: if X * Y is known not to overflow as a signed
1377 // multiplication, and Y is replaced by a value Z with smaller absolute value,
1378 // then X * Z will not overflow as a signed multiplication either. As we work
1379 // our way up, having NoSignedWrap 'true' means that the descaled value at the
1380 // current level has strictly smaller absolute value than the original.
1381 Instruction *Ancestor = Parent.first;
1383 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Ancestor)) {
1384 // If the multiplication wasn't nsw then we can't say anything about the
1385 // value of the descaled multiplication, and we have to clear nsw flags
1386 // from this point on up.
1387 bool OpNoSignedWrap = BO->hasNoSignedWrap();
1388 NoSignedWrap &= OpNoSignedWrap;
1389 if (NoSignedWrap != OpNoSignedWrap) {
1390 BO->setHasNoSignedWrap(NoSignedWrap);
1391 Worklist.Add(Ancestor);
1393 } else if (Ancestor->getOpcode() == Instruction::Trunc) {
1394 // The fact that the descaled input to the trunc has smaller absolute
1395 // value than the original input doesn't tell us anything useful about
1396 // the absolute values of the truncations.
1397 NoSignedWrap = false;
1399 assert((Ancestor->getOpcode() != Instruction::SExt || NoSignedWrap) &&
1400 "Failed to keep proper track of nsw flags while drilling down?");
1402 if (Ancestor == Val)
1403 // Got to the top, all done!
1406 // Move up one level in the expression.
1407 assert(Ancestor->hasOneUse() && "Drilled down when more than one use!");
1408 Ancestor = Ancestor->user_back();
1412 /// \brief Creates node of binary operation with the same attributes as the
1413 /// specified one but with other operands.
1414 static Value *CreateBinOpAsGiven(BinaryOperator &Inst, Value *LHS, Value *RHS,
1415 InstCombiner::BuilderTy &B) {
1416 Value *BO = B.CreateBinOp(Inst.getOpcode(), LHS, RHS);
1417 // If LHS and RHS are constant, BO won't be a binary operator.
1418 if (BinaryOperator *NewBO = dyn_cast<BinaryOperator>(BO))
1419 NewBO->copyIRFlags(&Inst);
1423 /// \brief Makes transformation of binary operation specific for vector types.
1424 /// \param Inst Binary operator to transform.
1425 /// \return Pointer to node that must replace the original binary operator, or
1426 /// null pointer if no transformation was made.
1427 Value *InstCombiner::SimplifyVectorOp(BinaryOperator &Inst) {
1428 if (!Inst.getType()->isVectorTy()) return nullptr;
1430 // It may not be safe to reorder shuffles and things like div, urem, etc.
1431 // because we may trap when executing those ops on unknown vector elements.
1433 if (!isSafeToSpeculativelyExecute(&Inst))
1436 unsigned VWidth = cast<VectorType>(Inst.getType())->getNumElements();
1437 Value *LHS = Inst.getOperand(0), *RHS = Inst.getOperand(1);
1438 assert(cast<VectorType>(LHS->getType())->getNumElements() == VWidth);
1439 assert(cast<VectorType>(RHS->getType())->getNumElements() == VWidth);
1441 // If both arguments of the binary operation are shuffles that use the same
1442 // mask and shuffle within a single vector, move the shuffle after the binop:
1443 // Op(shuffle(v1, m), shuffle(v2, m)) -> shuffle(Op(v1, v2), m)
1444 auto *LShuf = dyn_cast<ShuffleVectorInst>(LHS);
1445 auto *RShuf = dyn_cast<ShuffleVectorInst>(RHS);
1446 if (LShuf && RShuf && LShuf->getMask() == RShuf->getMask() &&
1447 isa<UndefValue>(LShuf->getOperand(1)) &&
1448 isa<UndefValue>(RShuf->getOperand(1)) &&
1449 LShuf->getOperand(0)->getType() == RShuf->getOperand(0)->getType()) {
1450 Value *NewBO = CreateBinOpAsGiven(Inst, LShuf->getOperand(0),
1451 RShuf->getOperand(0), Builder);
1452 return Builder.CreateShuffleVector(
1453 NewBO, UndefValue::get(NewBO->getType()), LShuf->getMask());
1456 // If one argument is a shuffle within one vector, the other is a constant,
1457 // try moving the shuffle after the binary operation.
1458 ShuffleVectorInst *Shuffle = nullptr;
1459 Constant *C1 = nullptr;
1460 if (isa<ShuffleVectorInst>(LHS)) Shuffle = cast<ShuffleVectorInst>(LHS);
1461 if (isa<ShuffleVectorInst>(RHS)) Shuffle = cast<ShuffleVectorInst>(RHS);
1462 if (isa<Constant>(LHS)) C1 = cast<Constant>(LHS);
1463 if (isa<Constant>(RHS)) C1 = cast<Constant>(RHS);
1464 if (Shuffle && C1 &&
1465 (isa<ConstantVector>(C1) || isa<ConstantDataVector>(C1)) &&
1466 isa<UndefValue>(Shuffle->getOperand(1)) &&
1467 Shuffle->getType() == Shuffle->getOperand(0)->getType()) {
1468 SmallVector<int, 16> ShMask = Shuffle->getShuffleMask();
1469 // Find constant C2 that has property:
1470 // shuffle(C2, ShMask) = C1
1471 // If such constant does not exist (example: ShMask=<0,0> and C1=<1,2>)
1472 // reorder is not possible.
1473 SmallVector<Constant*, 16> C2M(VWidth,
1474 UndefValue::get(C1->getType()->getScalarType()));
1475 bool MayChange = true;
1476 for (unsigned I = 0; I < VWidth; ++I) {
1477 if (ShMask[I] >= 0) {
1478 assert(ShMask[I] < (int)VWidth);
1479 if (!isa<UndefValue>(C2M[ShMask[I]])) {
1483 C2M[ShMask[I]] = C1->getAggregateElement(I);
1487 Constant *C2 = ConstantVector::get(C2M);
1488 Value *NewLHS = isa<Constant>(LHS) ? C2 : Shuffle->getOperand(0);
1489 Value *NewRHS = isa<Constant>(LHS) ? Shuffle->getOperand(0) : C2;
1490 Value *NewBO = CreateBinOpAsGiven(Inst, NewLHS, NewRHS, Builder);
1491 return Builder.CreateShuffleVector(NewBO,
1492 UndefValue::get(Inst.getType()), Shuffle->getMask());
1499 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
1500 SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end());
1502 if (Value *V = SimplifyGEPInst(GEP.getSourceElementType(), Ops,
1503 SQ.getWithInstruction(&GEP)))
1504 return replaceInstUsesWith(GEP, V);
1506 Value *PtrOp = GEP.getOperand(0);
1508 // Eliminate unneeded casts for indices, and replace indices which displace
1509 // by multiples of a zero size type with zero.
1510 bool MadeChange = false;
1512 DL.getIntPtrType(GEP.getPointerOperandType()->getScalarType());
1514 gep_type_iterator GTI = gep_type_begin(GEP);
1515 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end(); I != E;
1517 // Skip indices into struct types.
1521 // Index type should have the same width as IntPtr
1522 Type *IndexTy = (*I)->getType();
1523 Type *NewIndexType = IndexTy->isVectorTy() ?
1524 VectorType::get(IntPtrTy, IndexTy->getVectorNumElements()) : IntPtrTy;
1526 // If the element type has zero size then any index over it is equivalent
1527 // to an index of zero, so replace it with zero if it is not zero already.
1528 Type *EltTy = GTI.getIndexedType();
1529 if (EltTy->isSized() && DL.getTypeAllocSize(EltTy) == 0)
1530 if (!isa<Constant>(*I) || !cast<Constant>(*I)->isNullValue()) {
1531 *I = Constant::getNullValue(NewIndexType);
1535 if (IndexTy != NewIndexType) {
1536 // If we are using a wider index than needed for this platform, shrink
1537 // it to what we need. If narrower, sign-extend it to what we need.
1538 // This explicit cast can make subsequent optimizations more obvious.
1539 *I = Builder.CreateIntCast(*I, NewIndexType, true);
1546 // Check to see if the inputs to the PHI node are getelementptr instructions.
1547 if (PHINode *PN = dyn_cast<PHINode>(PtrOp)) {
1548 GetElementPtrInst *Op1 = dyn_cast<GetElementPtrInst>(PN->getOperand(0));
1552 // Don't fold a GEP into itself through a PHI node. This can only happen
1553 // through the back-edge of a loop. Folding a GEP into itself means that
1554 // the value of the previous iteration needs to be stored in the meantime,
1555 // thus requiring an additional register variable to be live, but not
1556 // actually achieving anything (the GEP still needs to be executed once per
1563 for (auto I = PN->op_begin()+1, E = PN->op_end(); I !=E; ++I) {
1564 GetElementPtrInst *Op2 = dyn_cast<GetElementPtrInst>(*I);
1565 if (!Op2 || Op1->getNumOperands() != Op2->getNumOperands())
1568 // As for Op1 above, don't try to fold a GEP into itself.
1572 // Keep track of the type as we walk the GEP.
1573 Type *CurTy = nullptr;
1575 for (unsigned J = 0, F = Op1->getNumOperands(); J != F; ++J) {
1576 if (Op1->getOperand(J)->getType() != Op2->getOperand(J)->getType())
1579 if (Op1->getOperand(J) != Op2->getOperand(J)) {
1581 // We have not seen any differences yet in the GEPs feeding the
1582 // PHI yet, so we record this one if it is allowed to be a
1585 // The first two arguments can vary for any GEP, the rest have to be
1586 // static for struct slots
1587 if (J > 1 && CurTy->isStructTy())
1592 // The GEP is different by more than one input. While this could be
1593 // extended to support GEPs that vary by more than one variable it
1594 // doesn't make sense since it greatly increases the complexity and
1595 // would result in an R+R+R addressing mode which no backend
1596 // directly supports and would need to be broken into several
1597 // simpler instructions anyway.
1602 // Sink down a layer of the type for the next iteration.
1605 CurTy = Op1->getSourceElementType();
1606 } else if (CompositeType *CT = dyn_cast<CompositeType>(CurTy)) {
1607 CurTy = CT->getTypeAtIndex(Op1->getOperand(J));
1615 // If not all GEPs are identical we'll have to create a new PHI node.
1616 // Check that the old PHI node has only one use so that it will get
1618 if (DI != -1 && !PN->hasOneUse())
1621 GetElementPtrInst *NewGEP = cast<GetElementPtrInst>(Op1->clone());
1623 // All the GEPs feeding the PHI are identical. Clone one down into our
1624 // BB so that it can be merged with the current GEP.
1625 GEP.getParent()->getInstList().insert(
1626 GEP.getParent()->getFirstInsertionPt(), NewGEP);
1628 // All the GEPs feeding the PHI differ at a single offset. Clone a GEP
1629 // into the current block so it can be merged, and create a new PHI to
1633 IRBuilderBase::InsertPointGuard Guard(Builder);
1634 Builder.SetInsertPoint(PN);
1635 NewPN = Builder.CreatePHI(Op1->getOperand(DI)->getType(),
1636 PN->getNumOperands());
1639 for (auto &I : PN->operands())
1640 NewPN->addIncoming(cast<GEPOperator>(I)->getOperand(DI),
1641 PN->getIncomingBlock(I));
1643 NewGEP->setOperand(DI, NewPN);
1644 GEP.getParent()->getInstList().insert(
1645 GEP.getParent()->getFirstInsertionPt(), NewGEP);
1646 NewGEP->setOperand(DI, NewPN);
1649 GEP.setOperand(0, NewGEP);
1653 // Combine Indices - If the source pointer to this getelementptr instruction
1654 // is a getelementptr instruction, combine the indices of the two
1655 // getelementptr instructions into a single instruction.
1656 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) {
1657 if (!shouldMergeGEPs(*cast<GEPOperator>(&GEP), *Src))
1660 // Note that if our source is a gep chain itself then we wait for that
1661 // chain to be resolved before we perform this transformation. This
1662 // avoids us creating a TON of code in some cases.
1663 if (GEPOperator *SrcGEP =
1664 dyn_cast<GEPOperator>(Src->getOperand(0)))
1665 if (SrcGEP->getNumOperands() == 2 && shouldMergeGEPs(*Src, *SrcGEP))
1666 return nullptr; // Wait until our source is folded to completion.
1668 SmallVector<Value*, 8> Indices;
1670 // Find out whether the last index in the source GEP is a sequential idx.
1671 bool EndsWithSequential = false;
1672 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src);
1674 EndsWithSequential = I.isSequential();
1676 // Can we combine the two pointer arithmetics offsets?
1677 if (EndsWithSequential) {
1678 // Replace: gep (gep %P, long B), long A, ...
1679 // With: T = long A+B; gep %P, T, ...
1680 Value *SO1 = Src->getOperand(Src->getNumOperands()-1);
1681 Value *GO1 = GEP.getOperand(1);
1683 // If they aren't the same type, then the input hasn't been processed
1684 // by the loop above yet (which canonicalizes sequential index types to
1685 // intptr_t). Just avoid transforming this until the input has been
1687 if (SO1->getType() != GO1->getType())
1691 SimplifyAddInst(GO1, SO1, false, false, SQ.getWithInstruction(&GEP));
1692 // Only do the combine when we are sure the cost after the
1693 // merge is never more than that before the merge.
1697 // Update the GEP in place if possible.
1698 if (Src->getNumOperands() == 2) {
1699 GEP.setOperand(0, Src->getOperand(0));
1700 GEP.setOperand(1, Sum);
1703 Indices.append(Src->op_begin()+1, Src->op_end()-1);
1704 Indices.push_back(Sum);
1705 Indices.append(GEP.op_begin()+2, GEP.op_end());
1706 } else if (isa<Constant>(*GEP.idx_begin()) &&
1707 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
1708 Src->getNumOperands() != 1) {
1709 // Otherwise we can do the fold if the first index of the GEP is a zero
1710 Indices.append(Src->op_begin()+1, Src->op_end());
1711 Indices.append(GEP.idx_begin()+1, GEP.idx_end());
1714 if (!Indices.empty())
1715 return GEP.isInBounds() && Src->isInBounds()
1716 ? GetElementPtrInst::CreateInBounds(
1717 Src->getSourceElementType(), Src->getOperand(0), Indices,
1719 : GetElementPtrInst::Create(Src->getSourceElementType(),
1720 Src->getOperand(0), Indices,
1724 if (GEP.getNumIndices() == 1) {
1725 unsigned AS = GEP.getPointerAddressSpace();
1726 if (GEP.getOperand(1)->getType()->getScalarSizeInBits() ==
1727 DL.getPointerSizeInBits(AS)) {
1728 Type *Ty = GEP.getSourceElementType();
1729 uint64_t TyAllocSize = DL.getTypeAllocSize(Ty);
1731 bool Matched = false;
1734 if (TyAllocSize == 1) {
1735 V = GEP.getOperand(1);
1737 } else if (match(GEP.getOperand(1),
1738 m_AShr(m_Value(V), m_ConstantInt(C)))) {
1739 if (TyAllocSize == 1ULL << C)
1741 } else if (match(GEP.getOperand(1),
1742 m_SDiv(m_Value(V), m_ConstantInt(C)))) {
1743 if (TyAllocSize == C)
1748 // Canonicalize (gep i8* X, -(ptrtoint Y))
1749 // to (inttoptr (sub (ptrtoint X), (ptrtoint Y)))
1750 // The GEP pattern is emitted by the SCEV expander for certain kinds of
1751 // pointer arithmetic.
1752 if (match(V, m_Neg(m_PtrToInt(m_Value())))) {
1753 Operator *Index = cast<Operator>(V);
1754 Value *PtrToInt = Builder.CreatePtrToInt(PtrOp, Index->getType());
1755 Value *NewSub = Builder.CreateSub(PtrToInt, Index->getOperand(1));
1756 return CastInst::Create(Instruction::IntToPtr, NewSub, GEP.getType());
1758 // Canonicalize (gep i8* X, (ptrtoint Y)-(ptrtoint X))
1761 if (match(V, m_Sub(m_PtrToInt(m_Value(Y)),
1762 m_PtrToInt(m_Specific(GEP.getOperand(0)))))) {
1763 return CastInst::CreatePointerBitCastOrAddrSpaceCast(Y,
1770 // We do not handle pointer-vector geps here.
1771 if (GEP.getType()->isVectorTy())
1774 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0).
1775 Value *StrippedPtr = PtrOp->stripPointerCasts();
1776 PointerType *StrippedPtrTy = cast<PointerType>(StrippedPtr->getType());
1778 if (StrippedPtr != PtrOp) {
1779 bool HasZeroPointerIndex = false;
1780 if (ConstantInt *C = dyn_cast<ConstantInt>(GEP.getOperand(1)))
1781 HasZeroPointerIndex = C->isZero();
1783 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
1784 // into : GEP [10 x i8]* X, i32 0, ...
1786 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
1787 // into : GEP i8* X, ...
1789 // This occurs when the program declares an array extern like "int X[];"
1790 if (HasZeroPointerIndex) {
1791 if (ArrayType *CATy =
1792 dyn_cast<ArrayType>(GEP.getSourceElementType())) {
1793 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
1794 if (CATy->getElementType() == StrippedPtrTy->getElementType()) {
1795 // -> GEP i8* X, ...
1796 SmallVector<Value*, 8> Idx(GEP.idx_begin()+1, GEP.idx_end());
1797 GetElementPtrInst *Res = GetElementPtrInst::Create(
1798 StrippedPtrTy->getElementType(), StrippedPtr, Idx, GEP.getName());
1799 Res->setIsInBounds(GEP.isInBounds());
1800 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace())
1802 // Insert Res, and create an addrspacecast.
1804 // GEP (addrspacecast i8 addrspace(1)* X to [0 x i8]*), i32 0, ...
1806 // %0 = GEP i8 addrspace(1)* X, ...
1807 // addrspacecast i8 addrspace(1)* %0 to i8*
1808 return new AddrSpaceCastInst(Builder.Insert(Res), GEP.getType());
1811 if (ArrayType *XATy =
1812 dyn_cast<ArrayType>(StrippedPtrTy->getElementType())){
1813 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
1814 if (CATy->getElementType() == XATy->getElementType()) {
1815 // -> GEP [10 x i8]* X, i32 0, ...
1816 // At this point, we know that the cast source type is a pointer
1817 // to an array of the same type as the destination pointer
1818 // array. Because the array type is never stepped over (there
1819 // is a leading zero) we can fold the cast into this GEP.
1820 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace()) {
1821 GEP.setOperand(0, StrippedPtr);
1822 GEP.setSourceElementType(XATy);
1825 // Cannot replace the base pointer directly because StrippedPtr's
1826 // address space is different. Instead, create a new GEP followed by
1827 // an addrspacecast.
1829 // GEP (addrspacecast [10 x i8] addrspace(1)* X to [0 x i8]*),
1832 // %0 = GEP [10 x i8] addrspace(1)* X, ...
1833 // addrspacecast i8 addrspace(1)* %0 to i8*
1834 SmallVector<Value*, 8> Idx(GEP.idx_begin(), GEP.idx_end());
1835 Value *NewGEP = GEP.isInBounds()
1836 ? Builder.CreateInBoundsGEP(
1837 nullptr, StrippedPtr, Idx, GEP.getName())
1838 : Builder.CreateGEP(nullptr, StrippedPtr, Idx,
1840 return new AddrSpaceCastInst(NewGEP, GEP.getType());
1844 } else if (GEP.getNumOperands() == 2) {
1845 // Transform things like:
1846 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
1847 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
1848 Type *SrcElTy = StrippedPtrTy->getElementType();
1849 Type *ResElTy = GEP.getSourceElementType();
1850 if (SrcElTy->isArrayTy() &&
1851 DL.getTypeAllocSize(SrcElTy->getArrayElementType()) ==
1852 DL.getTypeAllocSize(ResElTy)) {
1853 Type *IdxType = DL.getIntPtrType(GEP.getType());
1854 Value *Idx[2] = { Constant::getNullValue(IdxType), GEP.getOperand(1) };
1857 ? Builder.CreateInBoundsGEP(nullptr, StrippedPtr, Idx,
1859 : Builder.CreateGEP(nullptr, StrippedPtr, Idx, GEP.getName());
1861 // V and GEP are both pointer types --> BitCast
1862 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
1866 // Transform things like:
1867 // %V = mul i64 %N, 4
1868 // %t = getelementptr i8* bitcast (i32* %arr to i8*), i32 %V
1869 // into: %t1 = getelementptr i32* %arr, i32 %N; bitcast
1870 if (ResElTy->isSized() && SrcElTy->isSized()) {
1871 // Check that changing the type amounts to dividing the index by a scale
1873 uint64_t ResSize = DL.getTypeAllocSize(ResElTy);
1874 uint64_t SrcSize = DL.getTypeAllocSize(SrcElTy);
1875 if (ResSize && SrcSize % ResSize == 0) {
1876 Value *Idx = GEP.getOperand(1);
1877 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
1878 uint64_t Scale = SrcSize / ResSize;
1880 // Earlier transforms ensure that the index has type IntPtrType, which
1881 // considerably simplifies the logic by eliminating implicit casts.
1882 assert(Idx->getType() == DL.getIntPtrType(GEP.getType()) &&
1883 "Index not cast to pointer width?");
1886 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
1887 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
1888 // If the multiplication NewIdx * Scale may overflow then the new
1889 // GEP may not be "inbounds".
1891 GEP.isInBounds() && NSW
1892 ? Builder.CreateInBoundsGEP(nullptr, StrippedPtr, NewIdx,
1894 : Builder.CreateGEP(nullptr, StrippedPtr, NewIdx,
1897 // The NewGEP must be pointer typed, so must the old one -> BitCast
1898 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
1904 // Similarly, transform things like:
1905 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
1906 // (where tmp = 8*tmp2) into:
1907 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
1908 if (ResElTy->isSized() && SrcElTy->isSized() && SrcElTy->isArrayTy()) {
1909 // Check that changing to the array element type amounts to dividing the
1910 // index by a scale factor.
1911 uint64_t ResSize = DL.getTypeAllocSize(ResElTy);
1912 uint64_t ArrayEltSize =
1913 DL.getTypeAllocSize(SrcElTy->getArrayElementType());
1914 if (ResSize && ArrayEltSize % ResSize == 0) {
1915 Value *Idx = GEP.getOperand(1);
1916 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
1917 uint64_t Scale = ArrayEltSize / ResSize;
1919 // Earlier transforms ensure that the index has type IntPtrType, which
1920 // considerably simplifies the logic by eliminating implicit casts.
1921 assert(Idx->getType() == DL.getIntPtrType(GEP.getType()) &&
1922 "Index not cast to pointer width?");
1925 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
1926 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
1927 // If the multiplication NewIdx * Scale may overflow then the new
1928 // GEP may not be "inbounds".
1930 Constant::getNullValue(DL.getIntPtrType(GEP.getType())),
1933 Value *NewGEP = GEP.isInBounds() && NSW
1934 ? Builder.CreateInBoundsGEP(
1935 SrcElTy, StrippedPtr, Off, GEP.getName())
1936 : Builder.CreateGEP(SrcElTy, StrippedPtr, Off,
1938 // The NewGEP must be pointer typed, so must the old one -> BitCast
1939 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
1947 // addrspacecast between types is canonicalized as a bitcast, then an
1948 // addrspacecast. To take advantage of the below bitcast + struct GEP, look
1949 // through the addrspacecast.
1950 Value *ASCStrippedPtrOp = PtrOp;
1951 if (AddrSpaceCastInst *ASC = dyn_cast<AddrSpaceCastInst>(PtrOp)) {
1952 // X = bitcast A addrspace(1)* to B addrspace(1)*
1953 // Y = addrspacecast A addrspace(1)* to B addrspace(2)*
1954 // Z = gep Y, <...constant indices...>
1955 // Into an addrspacecasted GEP of the struct.
1956 if (BitCastInst *BC = dyn_cast<BitCastInst>(ASC->getOperand(0)))
1957 ASCStrippedPtrOp = BC;
1960 /// See if we can simplify:
1961 /// X = bitcast A* to B*
1962 /// Y = gep X, <...constant indices...>
1963 /// into a gep of the original struct. This is important for SROA and alias
1964 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
1965 if (BitCastInst *BCI = dyn_cast<BitCastInst>(ASCStrippedPtrOp)) {
1966 Value *Operand = BCI->getOperand(0);
1967 PointerType *OpType = cast<PointerType>(Operand->getType());
1968 unsigned OffsetBits = DL.getPointerTypeSizeInBits(GEP.getType());
1969 APInt Offset(OffsetBits, 0);
1970 if (!isa<BitCastInst>(Operand) &&
1971 GEP.accumulateConstantOffset(DL, Offset)) {
1973 // If this GEP instruction doesn't move the pointer, just replace the GEP
1974 // with a bitcast of the real input to the dest type.
1976 // If the bitcast is of an allocation, and the allocation will be
1977 // converted to match the type of the cast, don't touch this.
1978 if (isa<AllocaInst>(Operand) || isAllocationFn(Operand, &TLI)) {
1979 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
1980 if (Instruction *I = visitBitCast(*BCI)) {
1983 BCI->getParent()->getInstList().insert(BCI->getIterator(), I);
1984 replaceInstUsesWith(*BCI, I);
1990 if (Operand->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
1991 return new AddrSpaceCastInst(Operand, GEP.getType());
1992 return new BitCastInst(Operand, GEP.getType());
1995 // Otherwise, if the offset is non-zero, we need to find out if there is a
1996 // field at Offset in 'A's type. If so, we can pull the cast through the
1998 SmallVector<Value*, 8> NewIndices;
1999 if (FindElementAtOffset(OpType, Offset.getSExtValue(), NewIndices)) {
2002 ? Builder.CreateInBoundsGEP(nullptr, Operand, NewIndices)
2003 : Builder.CreateGEP(nullptr, Operand, NewIndices);
2005 if (NGEP->getType() == GEP.getType())
2006 return replaceInstUsesWith(GEP, NGEP);
2007 NGEP->takeName(&GEP);
2009 if (NGEP->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
2010 return new AddrSpaceCastInst(NGEP, GEP.getType());
2011 return new BitCastInst(NGEP, GEP.getType());
2016 if (!GEP.isInBounds()) {
2018 DL.getPointerSizeInBits(PtrOp->getType()->getPointerAddressSpace());
2019 APInt BasePtrOffset(PtrWidth, 0);
2020 Value *UnderlyingPtrOp =
2021 PtrOp->stripAndAccumulateInBoundsConstantOffsets(DL,
2023 if (auto *AI = dyn_cast<AllocaInst>(UnderlyingPtrOp)) {
2024 if (GEP.accumulateConstantOffset(DL, BasePtrOffset) &&
2025 BasePtrOffset.isNonNegative()) {
2026 APInt AllocSize(PtrWidth, DL.getTypeAllocSize(AI->getAllocatedType()));
2027 if (BasePtrOffset.ule(AllocSize)) {
2028 return GetElementPtrInst::CreateInBounds(
2029 PtrOp, makeArrayRef(Ops).slice(1), GEP.getName());
2038 static bool isNeverEqualToUnescapedAlloc(Value *V, const TargetLibraryInfo *TLI,
2040 if (isa<ConstantPointerNull>(V))
2042 if (auto *LI = dyn_cast<LoadInst>(V))
2043 return isa<GlobalVariable>(LI->getPointerOperand());
2044 // Two distinct allocations will never be equal.
2045 // We rely on LookThroughBitCast in isAllocLikeFn being false, since looking
2046 // through bitcasts of V can cause
2047 // the result statement below to be true, even when AI and V (ex:
2048 // i8* ->i32* ->i8* of AI) are the same allocations.
2049 return isAllocLikeFn(V, TLI) && V != AI;
2052 static bool isAllocSiteRemovable(Instruction *AI,
2053 SmallVectorImpl<WeakTrackingVH> &Users,
2054 const TargetLibraryInfo *TLI) {
2055 SmallVector<Instruction*, 4> Worklist;
2056 Worklist.push_back(AI);
2059 Instruction *PI = Worklist.pop_back_val();
2060 for (User *U : PI->users()) {
2061 Instruction *I = cast<Instruction>(U);
2062 switch (I->getOpcode()) {
2064 // Give up the moment we see something we can't handle.
2067 case Instruction::AddrSpaceCast:
2068 case Instruction::BitCast:
2069 case Instruction::GetElementPtr:
2070 Users.emplace_back(I);
2071 Worklist.push_back(I);
2074 case Instruction::ICmp: {
2075 ICmpInst *ICI = cast<ICmpInst>(I);
2076 // We can fold eq/ne comparisons with null to false/true, respectively.
2077 // We also fold comparisons in some conditions provided the alloc has
2078 // not escaped (see isNeverEqualToUnescapedAlloc).
2079 if (!ICI->isEquality())
2081 unsigned OtherIndex = (ICI->getOperand(0) == PI) ? 1 : 0;
2082 if (!isNeverEqualToUnescapedAlloc(ICI->getOperand(OtherIndex), TLI, AI))
2084 Users.emplace_back(I);
2088 case Instruction::Call:
2089 // Ignore no-op and store intrinsics.
2090 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
2091 switch (II->getIntrinsicID()) {
2095 case Intrinsic::memmove:
2096 case Intrinsic::memcpy:
2097 case Intrinsic::memset: {
2098 MemIntrinsic *MI = cast<MemIntrinsic>(II);
2099 if (MI->isVolatile() || MI->getRawDest() != PI)
2103 case Intrinsic::invariant_start:
2104 case Intrinsic::invariant_end:
2105 case Intrinsic::lifetime_start:
2106 case Intrinsic::lifetime_end:
2107 case Intrinsic::objectsize:
2108 Users.emplace_back(I);
2113 if (isFreeCall(I, TLI)) {
2114 Users.emplace_back(I);
2119 case Instruction::Store: {
2120 StoreInst *SI = cast<StoreInst>(I);
2121 if (SI->isVolatile() || SI->getPointerOperand() != PI)
2123 Users.emplace_back(I);
2127 llvm_unreachable("missing a return?");
2129 } while (!Worklist.empty());
2133 Instruction *InstCombiner::visitAllocSite(Instruction &MI) {
2134 // If we have a malloc call which is only used in any amount of comparisons
2135 // to null and free calls, delete the calls and replace the comparisons with
2136 // true or false as appropriate.
2137 SmallVector<WeakTrackingVH, 64> Users;
2139 // If we are removing an alloca with a dbg.declare, insert dbg.value calls
2140 // before each store.
2141 TinyPtrVector<DbgInfoIntrinsic *> DIIs;
2142 std::unique_ptr<DIBuilder> DIB;
2143 if (isa<AllocaInst>(MI)) {
2144 DIIs = FindDbgAddrUses(&MI);
2145 DIB.reset(new DIBuilder(*MI.getModule(), /*AllowUnresolved=*/false));
2148 if (isAllocSiteRemovable(&MI, Users, &TLI)) {
2149 for (unsigned i = 0, e = Users.size(); i != e; ++i) {
2150 // Lowering all @llvm.objectsize calls first because they may
2151 // use a bitcast/GEP of the alloca we are removing.
2155 Instruction *I = cast<Instruction>(&*Users[i]);
2157 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
2158 if (II->getIntrinsicID() == Intrinsic::objectsize) {
2159 ConstantInt *Result = lowerObjectSizeCall(II, DL, &TLI,
2160 /*MustSucceed=*/true);
2161 replaceInstUsesWith(*I, Result);
2162 eraseInstFromFunction(*I);
2163 Users[i] = nullptr; // Skip examining in the next loop.
2167 for (unsigned i = 0, e = Users.size(); i != e; ++i) {
2171 Instruction *I = cast<Instruction>(&*Users[i]);
2173 if (ICmpInst *C = dyn_cast<ICmpInst>(I)) {
2174 replaceInstUsesWith(*C,
2175 ConstantInt::get(Type::getInt1Ty(C->getContext()),
2176 C->isFalseWhenEqual()));
2177 } else if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I) ||
2178 isa<AddrSpaceCastInst>(I)) {
2179 replaceInstUsesWith(*I, UndefValue::get(I->getType()));
2180 } else if (auto *SI = dyn_cast<StoreInst>(I)) {
2181 for (auto *DII : DIIs)
2182 ConvertDebugDeclareToDebugValue(DII, SI, *DIB);
2184 eraseInstFromFunction(*I);
2187 if (InvokeInst *II = dyn_cast<InvokeInst>(&MI)) {
2188 // Replace invoke with a NOP intrinsic to maintain the original CFG
2189 Module *M = II->getModule();
2190 Function *F = Intrinsic::getDeclaration(M, Intrinsic::donothing);
2191 InvokeInst::Create(F, II->getNormalDest(), II->getUnwindDest(),
2192 None, "", II->getParent());
2195 for (auto *DII : DIIs)
2196 eraseInstFromFunction(*DII);
2198 return eraseInstFromFunction(MI);
2203 /// \brief Move the call to free before a NULL test.
2205 /// Check if this free is accessed after its argument has been test
2206 /// against NULL (property 0).
2207 /// If yes, it is legal to move this call in its predecessor block.
2209 /// The move is performed only if the block containing the call to free
2210 /// will be removed, i.e.:
2211 /// 1. it has only one predecessor P, and P has two successors
2212 /// 2. it contains the call and an unconditional branch
2213 /// 3. its successor is the same as its predecessor's successor
2215 /// The profitability is out-of concern here and this function should
2216 /// be called only if the caller knows this transformation would be
2217 /// profitable (e.g., for code size).
2218 static Instruction *
2219 tryToMoveFreeBeforeNullTest(CallInst &FI) {
2220 Value *Op = FI.getArgOperand(0);
2221 BasicBlock *FreeInstrBB = FI.getParent();
2222 BasicBlock *PredBB = FreeInstrBB->getSinglePredecessor();
2224 // Validate part of constraint #1: Only one predecessor
2225 // FIXME: We can extend the number of predecessor, but in that case, we
2226 // would duplicate the call to free in each predecessor and it may
2227 // not be profitable even for code size.
2231 // Validate constraint #2: Does this block contains only the call to
2232 // free and an unconditional branch?
2233 // FIXME: We could check if we can speculate everything in the
2234 // predecessor block
2235 if (FreeInstrBB->size() != 2)
2238 if (!match(FreeInstrBB->getTerminator(), m_UnconditionalBr(SuccBB)))
2241 // Validate the rest of constraint #1 by matching on the pred branch.
2242 TerminatorInst *TI = PredBB->getTerminator();
2243 BasicBlock *TrueBB, *FalseBB;
2244 ICmpInst::Predicate Pred;
2245 if (!match(TI, m_Br(m_ICmp(Pred, m_Specific(Op), m_Zero()), TrueBB, FalseBB)))
2247 if (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE)
2250 // Validate constraint #3: Ensure the null case just falls through.
2251 if (SuccBB != (Pred == ICmpInst::ICMP_EQ ? TrueBB : FalseBB))
2253 assert(FreeInstrBB == (Pred == ICmpInst::ICMP_EQ ? FalseBB : TrueBB) &&
2254 "Broken CFG: missing edge from predecessor to successor");
2260 Instruction *InstCombiner::visitFree(CallInst &FI) {
2261 Value *Op = FI.getArgOperand(0);
2263 // free undef -> unreachable.
2264 if (isa<UndefValue>(Op)) {
2265 // Insert a new store to null because we cannot modify the CFG here.
2266 Builder.CreateStore(ConstantInt::getTrue(FI.getContext()),
2267 UndefValue::get(Type::getInt1PtrTy(FI.getContext())));
2268 return eraseInstFromFunction(FI);
2271 // If we have 'free null' delete the instruction. This can happen in stl code
2272 // when lots of inlining happens.
2273 if (isa<ConstantPointerNull>(Op))
2274 return eraseInstFromFunction(FI);
2276 // If we optimize for code size, try to move the call to free before the null
2277 // test so that simplify cfg can remove the empty block and dead code
2278 // elimination the branch. I.e., helps to turn something like:
2279 // if (foo) free(foo);
2283 if (Instruction *I = tryToMoveFreeBeforeNullTest(FI))
2289 Instruction *InstCombiner::visitReturnInst(ReturnInst &RI) {
2290 if (RI.getNumOperands() == 0) // ret void
2293 Value *ResultOp = RI.getOperand(0);
2294 Type *VTy = ResultOp->getType();
2295 if (!VTy->isIntegerTy())
2298 // There might be assume intrinsics dominating this return that completely
2299 // determine the value. If so, constant fold it.
2300 KnownBits Known = computeKnownBits(ResultOp, 0, &RI);
2301 if (Known.isConstant())
2302 RI.setOperand(0, Constant::getIntegerValue(VTy, Known.getConstant()));
2307 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
2308 // Change br (not X), label True, label False to: br X, label False, True
2310 BasicBlock *TrueDest;
2311 BasicBlock *FalseDest;
2312 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
2313 !isa<Constant>(X)) {
2314 // Swap Destinations and condition...
2316 BI.swapSuccessors();
2320 // If the condition is irrelevant, remove the use so that other
2321 // transforms on the condition become more effective.
2322 if (BI.isConditional() && !isa<ConstantInt>(BI.getCondition()) &&
2323 BI.getSuccessor(0) == BI.getSuccessor(1)) {
2324 BI.setCondition(ConstantInt::getFalse(BI.getCondition()->getType()));
2328 // Canonicalize, for example, icmp_ne -> icmp_eq or fcmp_one -> fcmp_oeq.
2329 CmpInst::Predicate Pred;
2330 if (match(&BI, m_Br(m_OneUse(m_Cmp(Pred, m_Value(), m_Value())), TrueDest,
2332 !isCanonicalPredicate(Pred)) {
2333 // Swap destinations and condition.
2334 CmpInst *Cond = cast<CmpInst>(BI.getCondition());
2335 Cond->setPredicate(CmpInst::getInversePredicate(Pred));
2336 BI.swapSuccessors();
2344 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
2345 Value *Cond = SI.getCondition();
2347 ConstantInt *AddRHS;
2348 if (match(Cond, m_Add(m_Value(Op0), m_ConstantInt(AddRHS)))) {
2349 // Change 'switch (X+4) case 1:' into 'switch (X) case -3'.
2350 for (auto Case : SI.cases()) {
2351 Constant *NewCase = ConstantExpr::getSub(Case.getCaseValue(), AddRHS);
2352 assert(isa<ConstantInt>(NewCase) &&
2353 "Result of expression should be constant");
2354 Case.setValue(cast<ConstantInt>(NewCase));
2356 SI.setCondition(Op0);
2360 KnownBits Known = computeKnownBits(Cond, 0, &SI);
2361 unsigned LeadingKnownZeros = Known.countMinLeadingZeros();
2362 unsigned LeadingKnownOnes = Known.countMinLeadingOnes();
2364 // Compute the number of leading bits we can ignore.
2365 // TODO: A better way to determine this would use ComputeNumSignBits().
2366 for (auto &C : SI.cases()) {
2367 LeadingKnownZeros = std::min(
2368 LeadingKnownZeros, C.getCaseValue()->getValue().countLeadingZeros());
2369 LeadingKnownOnes = std::min(
2370 LeadingKnownOnes, C.getCaseValue()->getValue().countLeadingOnes());
2373 unsigned NewWidth = Known.getBitWidth() - std::max(LeadingKnownZeros, LeadingKnownOnes);
2375 // Shrink the condition operand if the new type is smaller than the old type.
2376 // This may produce a non-standard type for the switch, but that's ok because
2377 // the backend should extend back to a legal type for the target.
2378 if (NewWidth > 0 && NewWidth < Known.getBitWidth()) {
2379 IntegerType *Ty = IntegerType::get(SI.getContext(), NewWidth);
2380 Builder.SetInsertPoint(&SI);
2381 Value *NewCond = Builder.CreateTrunc(Cond, Ty, "trunc");
2382 SI.setCondition(NewCond);
2384 for (auto Case : SI.cases()) {
2385 APInt TruncatedCase = Case.getCaseValue()->getValue().trunc(NewWidth);
2386 Case.setValue(ConstantInt::get(SI.getContext(), TruncatedCase));
2394 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
2395 Value *Agg = EV.getAggregateOperand();
2397 if (!EV.hasIndices())
2398 return replaceInstUsesWith(EV, Agg);
2400 if (Value *V = SimplifyExtractValueInst(Agg, EV.getIndices(),
2401 SQ.getWithInstruction(&EV)))
2402 return replaceInstUsesWith(EV, V);
2404 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
2405 // We're extracting from an insertvalue instruction, compare the indices
2406 const unsigned *exti, *exte, *insi, *inse;
2407 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
2408 exte = EV.idx_end(), inse = IV->idx_end();
2409 exti != exte && insi != inse;
2412 // The insert and extract both reference distinctly different elements.
2413 // This means the extract is not influenced by the insert, and we can
2414 // replace the aggregate operand of the extract with the aggregate
2415 // operand of the insert. i.e., replace
2416 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
2417 // %E = extractvalue { i32, { i32 } } %I, 0
2419 // %E = extractvalue { i32, { i32 } } %A, 0
2420 return ExtractValueInst::Create(IV->getAggregateOperand(),
2423 if (exti == exte && insi == inse)
2424 // Both iterators are at the end: Index lists are identical. Replace
2425 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
2426 // %C = extractvalue { i32, { i32 } } %B, 1, 0
2428 return replaceInstUsesWith(EV, IV->getInsertedValueOperand());
2430 // The extract list is a prefix of the insert list. i.e. replace
2431 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
2432 // %E = extractvalue { i32, { i32 } } %I, 1
2434 // %X = extractvalue { i32, { i32 } } %A, 1
2435 // %E = insertvalue { i32 } %X, i32 42, 0
2436 // by switching the order of the insert and extract (though the
2437 // insertvalue should be left in, since it may have other uses).
2438 Value *NewEV = Builder.CreateExtractValue(IV->getAggregateOperand(),
2440 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
2441 makeArrayRef(insi, inse));
2444 // The insert list is a prefix of the extract list
2445 // We can simply remove the common indices from the extract and make it
2446 // operate on the inserted value instead of the insertvalue result.
2448 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
2449 // %E = extractvalue { i32, { i32 } } %I, 1, 0
2451 // %E extractvalue { i32 } { i32 42 }, 0
2452 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
2453 makeArrayRef(exti, exte));
2455 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) {
2456 // We're extracting from an intrinsic, see if we're the only user, which
2457 // allows us to simplify multiple result intrinsics to simpler things that
2458 // just get one value.
2459 if (II->hasOneUse()) {
2460 // Check if we're grabbing the overflow bit or the result of a 'with
2461 // overflow' intrinsic. If it's the latter we can remove the intrinsic
2462 // and replace it with a traditional binary instruction.
2463 switch (II->getIntrinsicID()) {
2464 case Intrinsic::uadd_with_overflow:
2465 case Intrinsic::sadd_with_overflow:
2466 if (*EV.idx_begin() == 0) { // Normal result.
2467 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2468 replaceInstUsesWith(*II, UndefValue::get(II->getType()));
2469 eraseInstFromFunction(*II);
2470 return BinaryOperator::CreateAdd(LHS, RHS);
2473 // If the normal result of the add is dead, and the RHS is a constant,
2474 // we can transform this into a range comparison.
2475 // overflow = uadd a, -4 --> overflow = icmp ugt a, 3
2476 if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow)
2477 if (ConstantInt *CI = dyn_cast<ConstantInt>(II->getArgOperand(1)))
2478 return new ICmpInst(ICmpInst::ICMP_UGT, II->getArgOperand(0),
2479 ConstantExpr::getNot(CI));
2481 case Intrinsic::usub_with_overflow:
2482 case Intrinsic::ssub_with_overflow:
2483 if (*EV.idx_begin() == 0) { // Normal result.
2484 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2485 replaceInstUsesWith(*II, UndefValue::get(II->getType()));
2486 eraseInstFromFunction(*II);
2487 return BinaryOperator::CreateSub(LHS, RHS);
2490 case Intrinsic::umul_with_overflow:
2491 case Intrinsic::smul_with_overflow:
2492 if (*EV.idx_begin() == 0) { // Normal result.
2493 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2494 replaceInstUsesWith(*II, UndefValue::get(II->getType()));
2495 eraseInstFromFunction(*II);
2496 return BinaryOperator::CreateMul(LHS, RHS);
2504 if (LoadInst *L = dyn_cast<LoadInst>(Agg))
2505 // If the (non-volatile) load only has one use, we can rewrite this to a
2506 // load from a GEP. This reduces the size of the load. If a load is used
2507 // only by extractvalue instructions then this either must have been
2508 // optimized before, or it is a struct with padding, in which case we
2509 // don't want to do the transformation as it loses padding knowledge.
2510 if (L->isSimple() && L->hasOneUse()) {
2511 // extractvalue has integer indices, getelementptr has Value*s. Convert.
2512 SmallVector<Value*, 4> Indices;
2513 // Prefix an i32 0 since we need the first element.
2514 Indices.push_back(Builder.getInt32(0));
2515 for (ExtractValueInst::idx_iterator I = EV.idx_begin(), E = EV.idx_end();
2517 Indices.push_back(Builder.getInt32(*I));
2519 // We need to insert these at the location of the old load, not at that of
2520 // the extractvalue.
2521 Builder.SetInsertPoint(L);
2522 Value *GEP = Builder.CreateInBoundsGEP(L->getType(),
2523 L->getPointerOperand(), Indices);
2524 Instruction *NL = Builder.CreateLoad(GEP);
2525 // Whatever aliasing information we had for the orignal load must also
2526 // hold for the smaller load, so propagate the annotations.
2528 L->getAAMetadata(Nodes);
2529 NL->setAAMetadata(Nodes);
2530 // Returning the load directly will cause the main loop to insert it in
2531 // the wrong spot, so use replaceInstUsesWith().
2532 return replaceInstUsesWith(EV, NL);
2534 // We could simplify extracts from other values. Note that nested extracts may
2535 // already be simplified implicitly by the above: extract (extract (insert) )
2536 // will be translated into extract ( insert ( extract ) ) first and then just
2537 // the value inserted, if appropriate. Similarly for extracts from single-use
2538 // loads: extract (extract (load)) will be translated to extract (load (gep))
2539 // and if again single-use then via load (gep (gep)) to load (gep).
2540 // However, double extracts from e.g. function arguments or return values
2541 // aren't handled yet.
2545 /// Return 'true' if the given typeinfo will match anything.
2546 static bool isCatchAll(EHPersonality Personality, Constant *TypeInfo) {
2547 switch (Personality) {
2548 case EHPersonality::GNU_C:
2549 case EHPersonality::GNU_C_SjLj:
2550 case EHPersonality::Rust:
2551 // The GCC C EH and Rust personality only exists to support cleanups, so
2552 // it's not clear what the semantics of catch clauses are.
2554 case EHPersonality::Unknown:
2556 case EHPersonality::GNU_Ada:
2557 // While __gnat_all_others_value will match any Ada exception, it doesn't
2558 // match foreign exceptions (or didn't, before gcc-4.7).
2560 case EHPersonality::GNU_CXX:
2561 case EHPersonality::GNU_CXX_SjLj:
2562 case EHPersonality::GNU_ObjC:
2563 case EHPersonality::MSVC_X86SEH:
2564 case EHPersonality::MSVC_Win64SEH:
2565 case EHPersonality::MSVC_CXX:
2566 case EHPersonality::CoreCLR:
2567 return TypeInfo->isNullValue();
2569 llvm_unreachable("invalid enum");
2572 static bool shorter_filter(const Value *LHS, const Value *RHS) {
2574 cast<ArrayType>(LHS->getType())->getNumElements()
2576 cast<ArrayType>(RHS->getType())->getNumElements();
2579 Instruction *InstCombiner::visitLandingPadInst(LandingPadInst &LI) {
2580 // The logic here should be correct for any real-world personality function.
2581 // However if that turns out not to be true, the offending logic can always
2582 // be conditioned on the personality function, like the catch-all logic is.
2583 EHPersonality Personality =
2584 classifyEHPersonality(LI.getParent()->getParent()->getPersonalityFn());
2586 // Simplify the list of clauses, eg by removing repeated catch clauses
2587 // (these are often created by inlining).
2588 bool MakeNewInstruction = false; // If true, recreate using the following:
2589 SmallVector<Constant *, 16> NewClauses; // - Clauses for the new instruction;
2590 bool CleanupFlag = LI.isCleanup(); // - The new instruction is a cleanup.
2592 SmallPtrSet<Value *, 16> AlreadyCaught; // Typeinfos known caught already.
2593 for (unsigned i = 0, e = LI.getNumClauses(); i != e; ++i) {
2594 bool isLastClause = i + 1 == e;
2595 if (LI.isCatch(i)) {
2597 Constant *CatchClause = LI.getClause(i);
2598 Constant *TypeInfo = CatchClause->stripPointerCasts();
2600 // If we already saw this clause, there is no point in having a second
2602 if (AlreadyCaught.insert(TypeInfo).second) {
2603 // This catch clause was not already seen.
2604 NewClauses.push_back(CatchClause);
2606 // Repeated catch clause - drop the redundant copy.
2607 MakeNewInstruction = true;
2610 // If this is a catch-all then there is no point in keeping any following
2611 // clauses or marking the landingpad as having a cleanup.
2612 if (isCatchAll(Personality, TypeInfo)) {
2614 MakeNewInstruction = true;
2615 CleanupFlag = false;
2619 // A filter clause. If any of the filter elements were already caught
2620 // then they can be dropped from the filter. It is tempting to try to
2621 // exploit the filter further by saying that any typeinfo that does not
2622 // occur in the filter can't be caught later (and thus can be dropped).
2623 // However this would be wrong, since typeinfos can match without being
2624 // equal (for example if one represents a C++ class, and the other some
2625 // class derived from it).
2626 assert(LI.isFilter(i) && "Unsupported landingpad clause!");
2627 Constant *FilterClause = LI.getClause(i);
2628 ArrayType *FilterType = cast<ArrayType>(FilterClause->getType());
2629 unsigned NumTypeInfos = FilterType->getNumElements();
2631 // An empty filter catches everything, so there is no point in keeping any
2632 // following clauses or marking the landingpad as having a cleanup. By
2633 // dealing with this case here the following code is made a bit simpler.
2634 if (!NumTypeInfos) {
2635 NewClauses.push_back(FilterClause);
2637 MakeNewInstruction = true;
2638 CleanupFlag = false;
2642 bool MakeNewFilter = false; // If true, make a new filter.
2643 SmallVector<Constant *, 16> NewFilterElts; // New elements.
2644 if (isa<ConstantAggregateZero>(FilterClause)) {
2645 // Not an empty filter - it contains at least one null typeinfo.
2646 assert(NumTypeInfos > 0 && "Should have handled empty filter already!");
2647 Constant *TypeInfo =
2648 Constant::getNullValue(FilterType->getElementType());
2649 // If this typeinfo is a catch-all then the filter can never match.
2650 if (isCatchAll(Personality, TypeInfo)) {
2651 // Throw the filter away.
2652 MakeNewInstruction = true;
2656 // There is no point in having multiple copies of this typeinfo, so
2657 // discard all but the first copy if there is more than one.
2658 NewFilterElts.push_back(TypeInfo);
2659 if (NumTypeInfos > 1)
2660 MakeNewFilter = true;
2662 ConstantArray *Filter = cast<ConstantArray>(FilterClause);
2663 SmallPtrSet<Value *, 16> SeenInFilter; // For uniquing the elements.
2664 NewFilterElts.reserve(NumTypeInfos);
2666 // Remove any filter elements that were already caught or that already
2667 // occurred in the filter. While there, see if any of the elements are
2668 // catch-alls. If so, the filter can be discarded.
2669 bool SawCatchAll = false;
2670 for (unsigned j = 0; j != NumTypeInfos; ++j) {
2671 Constant *Elt = Filter->getOperand(j);
2672 Constant *TypeInfo = Elt->stripPointerCasts();
2673 if (isCatchAll(Personality, TypeInfo)) {
2674 // This element is a catch-all. Bail out, noting this fact.
2679 // Even if we've seen a type in a catch clause, we don't want to
2680 // remove it from the filter. An unexpected type handler may be
2681 // set up for a call site which throws an exception of the same
2682 // type caught. In order for the exception thrown by the unexpected
2683 // handler to propagate correctly, the filter must be correctly
2684 // described for the call site.
2688 // void unexpected() { throw 1;}
2689 // void foo() throw (int) {
2690 // std::set_unexpected(unexpected);
2693 // } catch (int i) {}
2696 // There is no point in having multiple copies of the same typeinfo in
2697 // a filter, so only add it if we didn't already.
2698 if (SeenInFilter.insert(TypeInfo).second)
2699 NewFilterElts.push_back(cast<Constant>(Elt));
2701 // A filter containing a catch-all cannot match anything by definition.
2703 // Throw the filter away.
2704 MakeNewInstruction = true;
2708 // If we dropped something from the filter, make a new one.
2709 if (NewFilterElts.size() < NumTypeInfos)
2710 MakeNewFilter = true;
2712 if (MakeNewFilter) {
2713 FilterType = ArrayType::get(FilterType->getElementType(),
2714 NewFilterElts.size());
2715 FilterClause = ConstantArray::get(FilterType, NewFilterElts);
2716 MakeNewInstruction = true;
2719 NewClauses.push_back(FilterClause);
2721 // If the new filter is empty then it will catch everything so there is
2722 // no point in keeping any following clauses or marking the landingpad
2723 // as having a cleanup. The case of the original filter being empty was
2724 // already handled above.
2725 if (MakeNewFilter && !NewFilterElts.size()) {
2726 assert(MakeNewInstruction && "New filter but not a new instruction!");
2727 CleanupFlag = false;
2733 // If several filters occur in a row then reorder them so that the shortest
2734 // filters come first (those with the smallest number of elements). This is
2735 // advantageous because shorter filters are more likely to match, speeding up
2736 // unwinding, but mostly because it increases the effectiveness of the other
2737 // filter optimizations below.
2738 for (unsigned i = 0, e = NewClauses.size(); i + 1 < e; ) {
2740 // Find the maximal 'j' s.t. the range [i, j) consists entirely of filters.
2741 for (j = i; j != e; ++j)
2742 if (!isa<ArrayType>(NewClauses[j]->getType()))
2745 // Check whether the filters are already sorted by length. We need to know
2746 // if sorting them is actually going to do anything so that we only make a
2747 // new landingpad instruction if it does.
2748 for (unsigned k = i; k + 1 < j; ++k)
2749 if (shorter_filter(NewClauses[k+1], NewClauses[k])) {
2750 // Not sorted, so sort the filters now. Doing an unstable sort would be
2751 // correct too but reordering filters pointlessly might confuse users.
2752 std::stable_sort(NewClauses.begin() + i, NewClauses.begin() + j,
2754 MakeNewInstruction = true;
2758 // Look for the next batch of filters.
2762 // If typeinfos matched if and only if equal, then the elements of a filter L
2763 // that occurs later than a filter F could be replaced by the intersection of
2764 // the elements of F and L. In reality two typeinfos can match without being
2765 // equal (for example if one represents a C++ class, and the other some class
2766 // derived from it) so it would be wrong to perform this transform in general.
2767 // However the transform is correct and useful if F is a subset of L. In that
2768 // case L can be replaced by F, and thus removed altogether since repeating a
2769 // filter is pointless. So here we look at all pairs of filters F and L where
2770 // L follows F in the list of clauses, and remove L if every element of F is
2771 // an element of L. This can occur when inlining C++ functions with exception
2773 for (unsigned i = 0; i + 1 < NewClauses.size(); ++i) {
2774 // Examine each filter in turn.
2775 Value *Filter = NewClauses[i];
2776 ArrayType *FTy = dyn_cast<ArrayType>(Filter->getType());
2778 // Not a filter - skip it.
2780 unsigned FElts = FTy->getNumElements();
2781 // Examine each filter following this one. Doing this backwards means that
2782 // we don't have to worry about filters disappearing under us when removed.
2783 for (unsigned j = NewClauses.size() - 1; j != i; --j) {
2784 Value *LFilter = NewClauses[j];
2785 ArrayType *LTy = dyn_cast<ArrayType>(LFilter->getType());
2787 // Not a filter - skip it.
2789 // If Filter is a subset of LFilter, i.e. every element of Filter is also
2790 // an element of LFilter, then discard LFilter.
2791 SmallVectorImpl<Constant *>::iterator J = NewClauses.begin() + j;
2792 // If Filter is empty then it is a subset of LFilter.
2795 NewClauses.erase(J);
2796 MakeNewInstruction = true;
2797 // Move on to the next filter.
2800 unsigned LElts = LTy->getNumElements();
2801 // If Filter is longer than LFilter then it cannot be a subset of it.
2803 // Move on to the next filter.
2805 // At this point we know that LFilter has at least one element.
2806 if (isa<ConstantAggregateZero>(LFilter)) { // LFilter only contains zeros.
2807 // Filter is a subset of LFilter iff Filter contains only zeros (as we
2808 // already know that Filter is not longer than LFilter).
2809 if (isa<ConstantAggregateZero>(Filter)) {
2810 assert(FElts <= LElts && "Should have handled this case earlier!");
2812 NewClauses.erase(J);
2813 MakeNewInstruction = true;
2815 // Move on to the next filter.
2818 ConstantArray *LArray = cast<ConstantArray>(LFilter);
2819 if (isa<ConstantAggregateZero>(Filter)) { // Filter only contains zeros.
2820 // Since Filter is non-empty and contains only zeros, it is a subset of
2821 // LFilter iff LFilter contains a zero.
2822 assert(FElts > 0 && "Should have eliminated the empty filter earlier!");
2823 for (unsigned l = 0; l != LElts; ++l)
2824 if (LArray->getOperand(l)->isNullValue()) {
2825 // LFilter contains a zero - discard it.
2826 NewClauses.erase(J);
2827 MakeNewInstruction = true;
2830 // Move on to the next filter.
2833 // At this point we know that both filters are ConstantArrays. Loop over
2834 // operands to see whether every element of Filter is also an element of
2835 // LFilter. Since filters tend to be short this is probably faster than
2836 // using a method that scales nicely.
2837 ConstantArray *FArray = cast<ConstantArray>(Filter);
2838 bool AllFound = true;
2839 for (unsigned f = 0; f != FElts; ++f) {
2840 Value *FTypeInfo = FArray->getOperand(f)->stripPointerCasts();
2842 for (unsigned l = 0; l != LElts; ++l) {
2843 Value *LTypeInfo = LArray->getOperand(l)->stripPointerCasts();
2844 if (LTypeInfo == FTypeInfo) {
2854 NewClauses.erase(J);
2855 MakeNewInstruction = true;
2857 // Move on to the next filter.
2861 // If we changed any of the clauses, replace the old landingpad instruction
2863 if (MakeNewInstruction) {
2864 LandingPadInst *NLI = LandingPadInst::Create(LI.getType(),
2866 for (unsigned i = 0, e = NewClauses.size(); i != e; ++i)
2867 NLI->addClause(NewClauses[i]);
2868 // A landing pad with no clauses must have the cleanup flag set. It is
2869 // theoretically possible, though highly unlikely, that we eliminated all
2870 // clauses. If so, force the cleanup flag to true.
2871 if (NewClauses.empty())
2873 NLI->setCleanup(CleanupFlag);
2877 // Even if none of the clauses changed, we may nonetheless have understood
2878 // that the cleanup flag is pointless. Clear it if so.
2879 if (LI.isCleanup() != CleanupFlag) {
2880 assert(!CleanupFlag && "Adding a cleanup, not removing one?!");
2881 LI.setCleanup(CleanupFlag);
2888 /// Try to move the specified instruction from its current block into the
2889 /// beginning of DestBlock, which can only happen if it's safe to move the
2890 /// instruction past all of the instructions between it and the end of its
2892 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
2893 assert(I->hasOneUse() && "Invariants didn't hold!");
2895 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
2896 if (isa<PHINode>(I) || I->isEHPad() || I->mayHaveSideEffects() ||
2897 isa<TerminatorInst>(I))
2900 // Do not sink alloca instructions out of the entry block.
2901 if (isa<AllocaInst>(I) && I->getParent() ==
2902 &DestBlock->getParent()->getEntryBlock())
2905 // Do not sink into catchswitch blocks.
2906 if (isa<CatchSwitchInst>(DestBlock->getTerminator()))
2909 // Do not sink convergent call instructions.
2910 if (auto *CI = dyn_cast<CallInst>(I)) {
2911 if (CI->isConvergent())
2914 // We can only sink load instructions if there is nothing between the load and
2915 // the end of block that could change the value.
2916 if (I->mayReadFromMemory()) {
2917 for (BasicBlock::iterator Scan = I->getIterator(),
2918 E = I->getParent()->end();
2920 if (Scan->mayWriteToMemory())
2924 BasicBlock::iterator InsertPos = DestBlock->getFirstInsertionPt();
2925 I->moveBefore(&*InsertPos);
2930 bool InstCombiner::run() {
2931 while (!Worklist.isEmpty()) {
2932 Instruction *I = Worklist.RemoveOne();
2933 if (I == nullptr) continue; // skip null values.
2935 // Check to see if we can DCE the instruction.
2936 if (isInstructionTriviallyDead(I, &TLI)) {
2937 DEBUG(dbgs() << "IC: DCE: " << *I << '\n');
2938 eraseInstFromFunction(*I);
2940 MadeIRChange = true;
2944 if (!DebugCounter::shouldExecute(VisitCounter))
2947 // Instruction isn't dead, see if we can constant propagate it.
2948 if (!I->use_empty() &&
2949 (I->getNumOperands() == 0 || isa<Constant>(I->getOperand(0)))) {
2950 if (Constant *C = ConstantFoldInstruction(I, DL, &TLI)) {
2951 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n');
2953 // Add operands to the worklist.
2954 replaceInstUsesWith(*I, C);
2956 if (isInstructionTriviallyDead(I, &TLI))
2957 eraseInstFromFunction(*I);
2958 MadeIRChange = true;
2963 // In general, it is possible for computeKnownBits to determine all bits in
2964 // a value even when the operands are not all constants.
2965 Type *Ty = I->getType();
2966 if (ExpensiveCombines && !I->use_empty() && Ty->isIntOrIntVectorTy()) {
2967 KnownBits Known = computeKnownBits(I, /*Depth*/0, I);
2968 if (Known.isConstant()) {
2969 Constant *C = ConstantInt::get(Ty, Known.getConstant());
2970 DEBUG(dbgs() << "IC: ConstFold (all bits known) to: " << *C <<
2971 " from: " << *I << '\n');
2973 // Add operands to the worklist.
2974 replaceInstUsesWith(*I, C);
2976 if (isInstructionTriviallyDead(I, &TLI))
2977 eraseInstFromFunction(*I);
2978 MadeIRChange = true;
2983 // See if we can trivially sink this instruction to a successor basic block.
2984 if (I->hasOneUse()) {
2985 BasicBlock *BB = I->getParent();
2986 Instruction *UserInst = cast<Instruction>(*I->user_begin());
2987 BasicBlock *UserParent;
2989 // Get the block the use occurs in.
2990 if (PHINode *PN = dyn_cast<PHINode>(UserInst))
2991 UserParent = PN->getIncomingBlock(*I->use_begin());
2993 UserParent = UserInst->getParent();
2995 if (UserParent != BB) {
2996 bool UserIsSuccessor = false;
2997 // See if the user is one of our successors.
2998 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
2999 if (*SI == UserParent) {
3000 UserIsSuccessor = true;
3004 // If the user is one of our immediate successors, and if that successor
3005 // only has us as a predecessors (we'd have to split the critical edge
3006 // otherwise), we can keep going.
3007 if (UserIsSuccessor && UserParent->getUniquePredecessor()) {
3008 // Okay, the CFG is simple enough, try to sink this instruction.
3009 if (TryToSinkInstruction(I, UserParent)) {
3010 DEBUG(dbgs() << "IC: Sink: " << *I << '\n');
3011 MadeIRChange = true;
3012 // We'll add uses of the sunk instruction below, but since sinking
3013 // can expose opportunities for it's *operands* add them to the
3015 for (Use &U : I->operands())
3016 if (Instruction *OpI = dyn_cast<Instruction>(U.get()))
3023 // Now that we have an instruction, try combining it to simplify it.
3024 Builder.SetInsertPoint(I);
3025 Builder.SetCurrentDebugLocation(I->getDebugLoc());
3030 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str(););
3031 DEBUG(dbgs() << "IC: Visiting: " << OrigI << '\n');
3033 if (Instruction *Result = visit(*I)) {
3035 // Should we replace the old instruction with a new one?
3037 DEBUG(dbgs() << "IC: Old = " << *I << '\n'
3038 << " New = " << *Result << '\n');
3040 if (I->getDebugLoc())
3041 Result->setDebugLoc(I->getDebugLoc());
3042 // Everything uses the new instruction now.
3043 I->replaceAllUsesWith(Result);
3045 // Move the name to the new instruction first.
3046 Result->takeName(I);
3048 // Push the new instruction and any users onto the worklist.
3049 Worklist.AddUsersToWorkList(*Result);
3050 Worklist.Add(Result);
3052 // Insert the new instruction into the basic block...
3053 BasicBlock *InstParent = I->getParent();
3054 BasicBlock::iterator InsertPos = I->getIterator();
3056 // If we replace a PHI with something that isn't a PHI, fix up the
3058 if (!isa<PHINode>(Result) && isa<PHINode>(InsertPos))
3059 InsertPos = InstParent->getFirstInsertionPt();
3061 InstParent->getInstList().insert(InsertPos, Result);
3063 eraseInstFromFunction(*I);
3065 DEBUG(dbgs() << "IC: Mod = " << OrigI << '\n'
3066 << " New = " << *I << '\n');
3068 // If the instruction was modified, it's possible that it is now dead.
3069 // if so, remove it.
3070 if (isInstructionTriviallyDead(I, &TLI)) {
3071 eraseInstFromFunction(*I);
3073 Worklist.AddUsersToWorkList(*I);
3077 MadeIRChange = true;
3082 return MadeIRChange;
3085 /// Walk the function in depth-first order, adding all reachable code to the
3088 /// This has a couple of tricks to make the code faster and more powerful. In
3089 /// particular, we constant fold and DCE instructions as we go, to avoid adding
3090 /// them to the worklist (this significantly speeds up instcombine on code where
3091 /// many instructions are dead or constant). Additionally, if we find a branch
3092 /// whose condition is a known constant, we only visit the reachable successors.
3093 static bool AddReachableCodeToWorklist(BasicBlock *BB, const DataLayout &DL,
3094 SmallPtrSetImpl<BasicBlock *> &Visited,
3095 InstCombineWorklist &ICWorklist,
3096 const TargetLibraryInfo *TLI) {
3097 bool MadeIRChange = false;
3098 SmallVector<BasicBlock*, 256> Worklist;
3099 Worklist.push_back(BB);
3101 SmallVector<Instruction*, 128> InstrsForInstCombineWorklist;
3102 DenseMap<Constant *, Constant *> FoldedConstants;
3105 BB = Worklist.pop_back_val();
3107 // We have now visited this block! If we've already been here, ignore it.
3108 if (!Visited.insert(BB).second)
3111 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
3112 Instruction *Inst = &*BBI++;
3114 // DCE instruction if trivially dead.
3115 if (isInstructionTriviallyDead(Inst, TLI)) {
3117 DEBUG(dbgs() << "IC: DCE: " << *Inst << '\n');
3118 salvageDebugInfo(*Inst);
3119 Inst->eraseFromParent();
3120 MadeIRChange = true;
3124 // ConstantProp instruction if trivially constant.
3125 if (!Inst->use_empty() &&
3126 (Inst->getNumOperands() == 0 || isa<Constant>(Inst->getOperand(0))))
3127 if (Constant *C = ConstantFoldInstruction(Inst, DL, TLI)) {
3128 DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: "
3130 Inst->replaceAllUsesWith(C);
3132 if (isInstructionTriviallyDead(Inst, TLI))
3133 Inst->eraseFromParent();
3134 MadeIRChange = true;
3138 // See if we can constant fold its operands.
3139 for (Use &U : Inst->operands()) {
3140 if (!isa<ConstantVector>(U) && !isa<ConstantExpr>(U))
3143 auto *C = cast<Constant>(U);
3144 Constant *&FoldRes = FoldedConstants[C];
3146 FoldRes = ConstantFoldConstant(C, DL, TLI);
3151 DEBUG(dbgs() << "IC: ConstFold operand of: " << *Inst
3152 << "\n Old = " << *C
3153 << "\n New = " << *FoldRes << '\n');
3155 MadeIRChange = true;
3159 // Skip processing debug intrinsics in InstCombine. Processing these call instructions
3160 // consumes non-trivial amount of time and provides no value for the optimization.
3161 if (!isa<DbgInfoIntrinsic>(Inst))
3162 InstrsForInstCombineWorklist.push_back(Inst);
3165 // Recursively visit successors. If this is a branch or switch on a
3166 // constant, only visit the reachable successor.
3167 TerminatorInst *TI = BB->getTerminator();
3168 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
3169 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
3170 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
3171 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
3172 Worklist.push_back(ReachableBB);
3175 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
3176 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
3177 Worklist.push_back(SI->findCaseValue(Cond)->getCaseSuccessor());
3182 for (BasicBlock *SuccBB : TI->successors())
3183 Worklist.push_back(SuccBB);
3184 } while (!Worklist.empty());
3186 // Once we've found all of the instructions to add to instcombine's worklist,
3187 // add them in reverse order. This way instcombine will visit from the top
3188 // of the function down. This jives well with the way that it adds all uses
3189 // of instructions to the worklist after doing a transformation, thus avoiding
3190 // some N^2 behavior in pathological cases.
3191 ICWorklist.AddInitialGroup(InstrsForInstCombineWorklist);
3193 return MadeIRChange;
3196 /// \brief Populate the IC worklist from a function, and prune any dead basic
3197 /// blocks discovered in the process.
3199 /// This also does basic constant propagation and other forward fixing to make
3200 /// the combiner itself run much faster.
3201 static bool prepareICWorklistFromFunction(Function &F, const DataLayout &DL,
3202 TargetLibraryInfo *TLI,
3203 InstCombineWorklist &ICWorklist) {
3204 bool MadeIRChange = false;
3206 // Do a depth-first traversal of the function, populate the worklist with
3207 // the reachable instructions. Ignore blocks that are not reachable. Keep
3208 // track of which blocks we visit.
3209 SmallPtrSet<BasicBlock *, 32> Visited;
3211 AddReachableCodeToWorklist(&F.front(), DL, Visited, ICWorklist, TLI);
3213 // Do a quick scan over the function. If we find any blocks that are
3214 // unreachable, remove any instructions inside of them. This prevents
3215 // the instcombine code from having to deal with some bad special cases.
3216 for (BasicBlock &BB : F) {
3217 if (Visited.count(&BB))
3220 unsigned NumDeadInstInBB = removeAllNonTerminatorAndEHPadInstructions(&BB);
3221 MadeIRChange |= NumDeadInstInBB > 0;
3222 NumDeadInst += NumDeadInstInBB;
3225 return MadeIRChange;
3228 static bool combineInstructionsOverFunction(
3229 Function &F, InstCombineWorklist &Worklist, AliasAnalysis *AA,
3230 AssumptionCache &AC, TargetLibraryInfo &TLI, DominatorTree &DT,
3231 OptimizationRemarkEmitter &ORE, bool ExpensiveCombines = true,
3232 LoopInfo *LI = nullptr) {
3233 auto &DL = F.getParent()->getDataLayout();
3234 ExpensiveCombines |= EnableExpensiveCombines;
3236 /// Builder - This is an IRBuilder that automatically inserts new
3237 /// instructions into the worklist when they are created.
3238 IRBuilder<TargetFolder, IRBuilderCallbackInserter> Builder(
3239 F.getContext(), TargetFolder(DL),
3240 IRBuilderCallbackInserter([&Worklist, &AC](Instruction *I) {
3242 if (match(I, m_Intrinsic<Intrinsic::assume>()))
3243 AC.registerAssumption(cast<CallInst>(I));
3246 // Lower dbg.declare intrinsics otherwise their value may be clobbered
3248 bool MadeIRChange = false;
3249 if (ShouldLowerDbgDeclare)
3250 MadeIRChange = LowerDbgDeclare(F);
3252 // Iterate while there is work to do.
3256 DEBUG(dbgs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
3257 << F.getName() << "\n");
3259 MadeIRChange |= prepareICWorklistFromFunction(F, DL, &TLI, Worklist);
3261 InstCombiner IC(Worklist, Builder, F.optForMinSize(), ExpensiveCombines, AA,
3262 AC, TLI, DT, ORE, DL, LI);
3263 IC.MaxArraySizeForCombine = MaxArraySize;
3269 return MadeIRChange || Iteration > 1;
3272 PreservedAnalyses InstCombinePass::run(Function &F,
3273 FunctionAnalysisManager &AM) {
3274 auto &AC = AM.getResult<AssumptionAnalysis>(F);
3275 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
3276 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
3277 auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F);
3279 auto *LI = AM.getCachedResult<LoopAnalysis>(F);
3281 auto *AA = &AM.getResult<AAManager>(F);
3282 if (!combineInstructionsOverFunction(F, Worklist, AA, AC, TLI, DT, ORE,
3283 ExpensiveCombines, LI))
3284 // No changes, all analyses are preserved.
3285 return PreservedAnalyses::all();
3287 // Mark all the analyses that instcombine updates as preserved.
3288 PreservedAnalyses PA;
3289 PA.preserveSet<CFGAnalyses>();
3290 PA.preserve<AAManager>();
3291 PA.preserve<BasicAA>();
3292 PA.preserve<GlobalsAA>();
3296 void InstructionCombiningPass::getAnalysisUsage(AnalysisUsage &AU) const {
3297 AU.setPreservesCFG();
3298 AU.addRequired<AAResultsWrapperPass>();
3299 AU.addRequired<AssumptionCacheTracker>();
3300 AU.addRequired<TargetLibraryInfoWrapperPass>();
3301 AU.addRequired<DominatorTreeWrapperPass>();
3302 AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
3303 AU.addPreserved<DominatorTreeWrapperPass>();
3304 AU.addPreserved<AAResultsWrapperPass>();
3305 AU.addPreserved<BasicAAWrapperPass>();
3306 AU.addPreserved<GlobalsAAWrapperPass>();
3309 bool InstructionCombiningPass::runOnFunction(Function &F) {
3310 if (skipFunction(F))
3313 // Required analyses.
3314 auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
3315 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
3316 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
3317 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
3318 auto &ORE = getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
3320 // Optional analyses.
3321 auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>();
3322 auto *LI = LIWP ? &LIWP->getLoopInfo() : nullptr;
3324 return combineInstructionsOverFunction(F, Worklist, AA, AC, TLI, DT, ORE,
3325 ExpensiveCombines, LI);
3328 char InstructionCombiningPass::ID = 0;
3330 INITIALIZE_PASS_BEGIN(InstructionCombiningPass, "instcombine",
3331 "Combine redundant instructions", false, false)
3332 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
3333 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
3334 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
3335 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
3336 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
3337 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)
3338 INITIALIZE_PASS_END(InstructionCombiningPass, "instcombine",
3339 "Combine redundant instructions", false, false)
3341 // Initialization Routines
3342 void llvm::initializeInstCombine(PassRegistry &Registry) {
3343 initializeInstructionCombiningPassPass(Registry);
3346 void LLVMInitializeInstCombine(LLVMPassRegistryRef R) {
3347 initializeInstructionCombiningPassPass(*unwrap(R));
3350 FunctionPass *llvm::createInstructionCombiningPass(bool ExpensiveCombines) {
3351 return new InstructionCombiningPass(ExpensiveCombines);