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-c/Transforms/InstCombine.h"
39 #include "llvm/ADT/APInt.h"
40 #include "llvm/ADT/ArrayRef.h"
41 #include "llvm/ADT/DenseMap.h"
42 #include "llvm/ADT/None.h"
43 #include "llvm/ADT/SmallPtrSet.h"
44 #include "llvm/ADT/SmallVector.h"
45 #include "llvm/ADT/Statistic.h"
46 #include "llvm/ADT/TinyPtrVector.h"
47 #include "llvm/Analysis/AliasAnalysis.h"
48 #include "llvm/Analysis/AssumptionCache.h"
49 #include "llvm/Analysis/BasicAliasAnalysis.h"
50 #include "llvm/Analysis/CFG.h"
51 #include "llvm/Analysis/ConstantFolding.h"
52 #include "llvm/Analysis/EHPersonalities.h"
53 #include "llvm/Analysis/GlobalsModRef.h"
54 #include "llvm/Analysis/InstructionSimplify.h"
55 #include "llvm/Analysis/LoopInfo.h"
56 #include "llvm/Analysis/MemoryBuiltins.h"
57 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
58 #include "llvm/Analysis/TargetFolder.h"
59 #include "llvm/Analysis/TargetLibraryInfo.h"
60 #include "llvm/Analysis/ValueTracking.h"
61 #include "llvm/IR/BasicBlock.h"
62 #include "llvm/IR/CFG.h"
63 #include "llvm/IR/Constant.h"
64 #include "llvm/IR/Constants.h"
65 #include "llvm/IR/DIBuilder.h"
66 #include "llvm/IR/DataLayout.h"
67 #include "llvm/IR/DerivedTypes.h"
68 #include "llvm/IR/Dominators.h"
69 #include "llvm/IR/Function.h"
70 #include "llvm/IR/GetElementPtrTypeIterator.h"
71 #include "llvm/IR/IRBuilder.h"
72 #include "llvm/IR/InstrTypes.h"
73 #include "llvm/IR/Instruction.h"
74 #include "llvm/IR/Instructions.h"
75 #include "llvm/IR/IntrinsicInst.h"
76 #include "llvm/IR/Intrinsics.h"
77 #include "llvm/IR/LegacyPassManager.h"
78 #include "llvm/IR/Metadata.h"
79 #include "llvm/IR/Operator.h"
80 #include "llvm/IR/PassManager.h"
81 #include "llvm/IR/PatternMatch.h"
82 #include "llvm/IR/Type.h"
83 #include "llvm/IR/Use.h"
84 #include "llvm/IR/User.h"
85 #include "llvm/IR/Value.h"
86 #include "llvm/IR/ValueHandle.h"
87 #include "llvm/Pass.h"
88 #include "llvm/Support/CBindingWrapping.h"
89 #include "llvm/Support/Casting.h"
90 #include "llvm/Support/CommandLine.h"
91 #include "llvm/Support/Compiler.h"
92 #include "llvm/Support/Debug.h"
93 #include "llvm/Support/DebugCounter.h"
94 #include "llvm/Support/ErrorHandling.h"
95 #include "llvm/Support/KnownBits.h"
96 #include "llvm/Support/raw_ostream.h"
97 #include "llvm/Transforms/InstCombine/InstCombine.h"
98 #include "llvm/Transforms/InstCombine/InstCombineWorklist.h"
99 #include "llvm/Transforms/Utils/Local.h"
107 using namespace llvm;
108 using namespace llvm::PatternMatch;
110 #define DEBUG_TYPE "instcombine"
112 STATISTIC(NumCombined , "Number of insts combined");
113 STATISTIC(NumConstProp, "Number of constant folds");
114 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
115 STATISTIC(NumSunkInst , "Number of instructions sunk");
116 STATISTIC(NumExpand, "Number of expansions");
117 STATISTIC(NumFactor , "Number of factorizations");
118 STATISTIC(NumReassoc , "Number of reassociations");
119 DEBUG_COUNTER(VisitCounter, "instcombine-visit",
120 "Controls which instructions are visited");
123 EnableCodeSinking("instcombine-code-sinking", cl::desc("Enable code sinking"),
127 EnableExpensiveCombines("expensive-combines",
128 cl::desc("Enable expensive instruction combines"));
130 static cl::opt<unsigned>
131 MaxArraySize("instcombine-maxarray-size", cl::init(1024),
132 cl::desc("Maximum array size considered when doing a combine"));
134 // FIXME: Remove this flag when it is no longer necessary to convert
135 // llvm.dbg.declare to avoid inaccurate debug info. Setting this to false
136 // increases variable availability at the cost of accuracy. Variables that
137 // cannot be promoted by mem2reg or SROA will be described as living in memory
138 // for their entire lifetime. However, passes like DSE and instcombine can
139 // delete stores to the alloca, leading to misleading and inaccurate debug
140 // information. This flag can be removed when those passes are fixed.
141 static cl::opt<unsigned> ShouldLowerDbgDeclare("instcombine-lower-dbg-declare",
142 cl::Hidden, cl::init(true));
144 Value *InstCombiner::EmitGEPOffset(User *GEP) {
145 return llvm::EmitGEPOffset(&Builder, DL, GEP);
148 /// Return true if it is desirable to convert an integer computation from a
149 /// given bit width to a new bit width.
150 /// We don't want to convert from a legal to an illegal type or from a smaller
151 /// to a larger illegal type. A width of '1' is always treated as a legal type
152 /// because i1 is a fundamental type in IR, and there are many specialized
153 /// optimizations for i1 types. Widths of 8, 16 or 32 are equally treated as
154 /// legal to convert to, in order to open up more combining opportunities.
155 /// NOTE: this treats i8, i16 and i32 specially, due to them being so common
156 /// from frontend languages.
157 bool InstCombiner::shouldChangeType(unsigned FromWidth,
158 unsigned ToWidth) const {
159 bool FromLegal = FromWidth == 1 || DL.isLegalInteger(FromWidth);
160 bool ToLegal = ToWidth == 1 || DL.isLegalInteger(ToWidth);
162 // Convert to widths of 8, 16 or 32 even if they are not legal types. Only
163 // shrink types, to prevent infinite loops.
164 if (ToWidth < FromWidth && (ToWidth == 8 || ToWidth == 16 || ToWidth == 32))
167 // If this is a legal integer from type, and the result would be an illegal
168 // type, don't do the transformation.
169 if (FromLegal && !ToLegal)
172 // Otherwise, if both are illegal, do not increase the size of the result. We
173 // do allow things like i160 -> i64, but not i64 -> i160.
174 if (!FromLegal && !ToLegal && ToWidth > FromWidth)
180 /// Return true if it is desirable to convert a computation from 'From' to 'To'.
181 /// We don't want to convert from a legal to an illegal type or from a smaller
182 /// to a larger illegal type. i1 is always treated as a legal type because it is
183 /// a fundamental type in IR, and there are many specialized optimizations for
185 bool InstCombiner::shouldChangeType(Type *From, Type *To) const {
186 // TODO: This could be extended to allow vectors. Datalayout changes might be
187 // needed to properly support that.
188 if (!From->isIntegerTy() || !To->isIntegerTy())
191 unsigned FromWidth = From->getPrimitiveSizeInBits();
192 unsigned ToWidth = To->getPrimitiveSizeInBits();
193 return shouldChangeType(FromWidth, ToWidth);
196 // Return true, if No Signed Wrap should be maintained for I.
197 // The No Signed Wrap flag can be kept if the operation "B (I.getOpcode) C",
198 // where both B and C should be ConstantInts, results in a constant that does
199 // not overflow. This function only handles the Add and Sub opcodes. For
200 // all other opcodes, the function conservatively returns false.
201 static bool MaintainNoSignedWrap(BinaryOperator &I, Value *B, Value *C) {
202 OverflowingBinaryOperator *OBO = dyn_cast<OverflowingBinaryOperator>(&I);
203 if (!OBO || !OBO->hasNoSignedWrap())
206 // We reason about Add and Sub Only.
207 Instruction::BinaryOps Opcode = I.getOpcode();
208 if (Opcode != Instruction::Add && Opcode != Instruction::Sub)
211 const APInt *BVal, *CVal;
212 if (!match(B, m_APInt(BVal)) || !match(C, m_APInt(CVal)))
215 bool Overflow = false;
216 if (Opcode == Instruction::Add)
217 (void)BVal->sadd_ov(*CVal, Overflow);
219 (void)BVal->ssub_ov(*CVal, Overflow);
224 /// Conservatively clears subclassOptionalData after a reassociation or
225 /// commutation. We preserve fast-math flags when applicable as they can be
227 static void ClearSubclassDataAfterReassociation(BinaryOperator &I) {
228 FPMathOperator *FPMO = dyn_cast<FPMathOperator>(&I);
230 I.clearSubclassOptionalData();
234 FastMathFlags FMF = I.getFastMathFlags();
235 I.clearSubclassOptionalData();
236 I.setFastMathFlags(FMF);
239 /// Combine constant operands of associative operations either before or after a
240 /// cast to eliminate one of the associative operations:
241 /// (op (cast (op X, C2)), C1) --> (cast (op X, op (C1, C2)))
242 /// (op (cast (op X, C2)), C1) --> (op (cast X), op (C1, C2))
243 static bool simplifyAssocCastAssoc(BinaryOperator *BinOp1) {
244 auto *Cast = dyn_cast<CastInst>(BinOp1->getOperand(0));
245 if (!Cast || !Cast->hasOneUse())
248 // TODO: Enhance logic for other casts and remove this check.
249 auto CastOpcode = Cast->getOpcode();
250 if (CastOpcode != Instruction::ZExt)
253 // TODO: Enhance logic for other BinOps and remove this check.
254 if (!BinOp1->isBitwiseLogicOp())
257 auto AssocOpcode = BinOp1->getOpcode();
258 auto *BinOp2 = dyn_cast<BinaryOperator>(Cast->getOperand(0));
259 if (!BinOp2 || !BinOp2->hasOneUse() || BinOp2->getOpcode() != AssocOpcode)
263 if (!match(BinOp1->getOperand(1), m_Constant(C1)) ||
264 !match(BinOp2->getOperand(1), m_Constant(C2)))
267 // TODO: This assumes a zext cast.
268 // Eg, if it was a trunc, we'd cast C1 to the source type because casting C2
269 // to the destination type might lose bits.
271 // Fold the constants together in the destination type:
272 // (op (cast (op X, C2)), C1) --> (op (cast X), FoldedC)
273 Type *DestTy = C1->getType();
274 Constant *CastC2 = ConstantExpr::getCast(CastOpcode, C2, DestTy);
275 Constant *FoldedC = ConstantExpr::get(AssocOpcode, C1, CastC2);
276 Cast->setOperand(0, BinOp2->getOperand(0));
277 BinOp1->setOperand(1, FoldedC);
281 /// This performs a few simplifications for operators that are associative or
284 /// Commutative operators:
286 /// 1. Order operands such that they are listed from right (least complex) to
287 /// left (most complex). This puts constants before unary operators before
288 /// binary operators.
290 /// Associative operators:
292 /// 2. Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
293 /// 3. Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
295 /// Associative and commutative operators:
297 /// 4. Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
298 /// 5. Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
299 /// 6. Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
300 /// if C1 and C2 are constants.
301 bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) {
302 Instruction::BinaryOps Opcode = I.getOpcode();
303 bool Changed = false;
306 // Order operands such that they are listed from right (least complex) to
307 // left (most complex). This puts constants before unary operators before
309 if (I.isCommutative() && getComplexity(I.getOperand(0)) <
310 getComplexity(I.getOperand(1)))
311 Changed = !I.swapOperands();
313 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
314 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
316 if (I.isAssociative()) {
317 // Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
318 if (Op0 && Op0->getOpcode() == Opcode) {
319 Value *A = Op0->getOperand(0);
320 Value *B = Op0->getOperand(1);
321 Value *C = I.getOperand(1);
323 // Does "B op C" simplify?
324 if (Value *V = SimplifyBinOp(Opcode, B, C, SQ.getWithInstruction(&I))) {
325 // It simplifies to V. Form "A op V".
328 // Conservatively clear the optional flags, since they may not be
329 // preserved by the reassociation.
330 if (MaintainNoSignedWrap(I, B, C) &&
331 (!Op0 || (isa<BinaryOperator>(Op0) && Op0->hasNoSignedWrap()))) {
332 // Note: this is only valid because SimplifyBinOp doesn't look at
333 // the operands to Op0.
334 I.clearSubclassOptionalData();
335 I.setHasNoSignedWrap(true);
337 ClearSubclassDataAfterReassociation(I);
346 // Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
347 if (Op1 && Op1->getOpcode() == Opcode) {
348 Value *A = I.getOperand(0);
349 Value *B = Op1->getOperand(0);
350 Value *C = Op1->getOperand(1);
352 // Does "A op B" simplify?
353 if (Value *V = SimplifyBinOp(Opcode, A, B, SQ.getWithInstruction(&I))) {
354 // It simplifies to V. Form "V op C".
357 // Conservatively clear the optional flags, since they may not be
358 // preserved by the reassociation.
359 ClearSubclassDataAfterReassociation(I);
367 if (I.isAssociative() && I.isCommutative()) {
368 if (simplifyAssocCastAssoc(&I)) {
374 // Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
375 if (Op0 && Op0->getOpcode() == Opcode) {
376 Value *A = Op0->getOperand(0);
377 Value *B = Op0->getOperand(1);
378 Value *C = I.getOperand(1);
380 // Does "C op A" simplify?
381 if (Value *V = SimplifyBinOp(Opcode, C, A, SQ.getWithInstruction(&I))) {
382 // It simplifies to V. Form "V op B".
385 // Conservatively clear the optional flags, since they may not be
386 // preserved by the reassociation.
387 ClearSubclassDataAfterReassociation(I);
394 // Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
395 if (Op1 && Op1->getOpcode() == Opcode) {
396 Value *A = I.getOperand(0);
397 Value *B = Op1->getOperand(0);
398 Value *C = Op1->getOperand(1);
400 // Does "C op A" simplify?
401 if (Value *V = SimplifyBinOp(Opcode, C, A, SQ.getWithInstruction(&I))) {
402 // It simplifies to V. Form "B op V".
405 // Conservatively clear the optional flags, since they may not be
406 // preserved by the reassociation.
407 ClearSubclassDataAfterReassociation(I);
414 // Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
415 // if C1 and C2 are constants.
419 Op0->getOpcode() == Opcode && Op1->getOpcode() == Opcode &&
420 match(Op0, m_OneUse(m_BinOp(m_Value(A), m_Constant(C1)))) &&
421 match(Op1, m_OneUse(m_BinOp(m_Value(B), m_Constant(C2))))) {
422 BinaryOperator *NewBO = BinaryOperator::Create(Opcode, A, B);
423 if (isa<FPMathOperator>(NewBO)) {
424 FastMathFlags Flags = I.getFastMathFlags();
425 Flags &= Op0->getFastMathFlags();
426 Flags &= Op1->getFastMathFlags();
427 NewBO->setFastMathFlags(Flags);
429 InsertNewInstWith(NewBO, I);
430 NewBO->takeName(Op1);
431 I.setOperand(0, NewBO);
432 I.setOperand(1, ConstantExpr::get(Opcode, C1, C2));
433 // Conservatively clear the optional flags, since they may not be
434 // preserved by the reassociation.
435 ClearSubclassDataAfterReassociation(I);
442 // No further simplifications.
447 /// Return whether "X LOp (Y ROp Z)" is always equal to
448 /// "(X LOp Y) ROp (X LOp Z)".
449 static bool leftDistributesOverRight(Instruction::BinaryOps LOp,
450 Instruction::BinaryOps ROp) {
451 // X & (Y | Z) <--> (X & Y) | (X & Z)
452 // X & (Y ^ Z) <--> (X & Y) ^ (X & Z)
453 if (LOp == Instruction::And)
454 return ROp == Instruction::Or || ROp == Instruction::Xor;
456 // X | (Y & Z) <--> (X | Y) & (X | Z)
457 if (LOp == Instruction::Or)
458 return ROp == Instruction::And;
460 // X * (Y + Z) <--> (X * Y) + (X * Z)
461 // X * (Y - Z) <--> (X * Y) - (X * Z)
462 if (LOp == Instruction::Mul)
463 return ROp == Instruction::Add || ROp == Instruction::Sub;
468 /// Return whether "(X LOp Y) ROp Z" is always equal to
469 /// "(X ROp Z) LOp (Y ROp Z)".
470 static bool rightDistributesOverLeft(Instruction::BinaryOps LOp,
471 Instruction::BinaryOps ROp) {
472 if (Instruction::isCommutative(ROp))
473 return leftDistributesOverRight(ROp, LOp);
475 // (X {&|^} Y) >> Z <--> (X >> Z) {&|^} (Y >> Z) for all shifts.
476 return Instruction::isBitwiseLogicOp(LOp) && Instruction::isShift(ROp);
478 // TODO: It would be nice to handle division, aka "(X + Y)/Z = X/Z + Y/Z",
479 // but this requires knowing that the addition does not overflow and other
483 /// This function returns identity value for given opcode, which can be used to
484 /// factor patterns like (X * 2) + X ==> (X * 2) + (X * 1) ==> X * (2 + 1).
485 static Value *getIdentityValue(Instruction::BinaryOps Opcode, Value *V) {
486 if (isa<Constant>(V))
489 return ConstantExpr::getBinOpIdentity(Opcode, V->getType());
492 /// This function predicates factorization using distributive laws. By default,
493 /// it just returns the 'Op' inputs. But for special-cases like
494 /// 'add(shl(X, 5), ...)', this function will have TopOpcode == Instruction::Add
495 /// and Op = shl(X, 5). The 'shl' is treated as the more general 'mul X, 32' to
496 /// allow more factorization opportunities.
497 static Instruction::BinaryOps
498 getBinOpsForFactorization(Instruction::BinaryOps TopOpcode, BinaryOperator *Op,
499 Value *&LHS, Value *&RHS) {
500 assert(Op && "Expected a binary operator");
501 LHS = Op->getOperand(0);
502 RHS = Op->getOperand(1);
503 if (TopOpcode == Instruction::Add || TopOpcode == Instruction::Sub) {
505 if (match(Op, m_Shl(m_Value(), m_Constant(C)))) {
506 // X << C --> X * (1 << C)
507 RHS = ConstantExpr::getShl(ConstantInt::get(Op->getType(), 1), C);
508 return Instruction::Mul;
510 // TODO: We can add other conversions e.g. shr => div etc.
512 return Op->getOpcode();
515 /// This tries to simplify binary operations by factorizing out common terms
516 /// (e. g. "(A*B)+(A*C)" -> "A*(B+C)").
517 Value *InstCombiner::tryFactorization(BinaryOperator &I,
518 Instruction::BinaryOps InnerOpcode,
519 Value *A, Value *B, Value *C, Value *D) {
520 assert(A && B && C && D && "All values must be provided");
523 Value *SimplifiedInst = nullptr;
524 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
525 Instruction::BinaryOps TopLevelOpcode = I.getOpcode();
527 // Does "X op' Y" always equal "Y op' X"?
528 bool InnerCommutative = Instruction::isCommutative(InnerOpcode);
530 // Does "X op' (Y op Z)" always equal "(X op' Y) op (X op' Z)"?
531 if (leftDistributesOverRight(InnerOpcode, TopLevelOpcode))
532 // Does the instruction have the form "(A op' B) op (A op' D)" or, in the
533 // commutative case, "(A op' B) op (C op' A)"?
534 if (A == C || (InnerCommutative && A == D)) {
537 // Consider forming "A op' (B op D)".
538 // If "B op D" simplifies then it can be formed with no cost.
539 V = SimplifyBinOp(TopLevelOpcode, B, D, SQ.getWithInstruction(&I));
540 // If "B op D" doesn't simplify then only go on if both of the existing
541 // operations "A op' B" and "C op' D" will be zapped as no longer used.
542 if (!V && LHS->hasOneUse() && RHS->hasOneUse())
543 V = Builder.CreateBinOp(TopLevelOpcode, B, D, RHS->getName());
545 SimplifiedInst = Builder.CreateBinOp(InnerOpcode, A, V);
549 // Does "(X op Y) op' Z" always equal "(X op' Z) op (Y op' Z)"?
550 if (!SimplifiedInst && rightDistributesOverLeft(TopLevelOpcode, InnerOpcode))
551 // Does the instruction have the form "(A op' B) op (C op' B)" or, in the
552 // commutative case, "(A op' B) op (B op' D)"?
553 if (B == D || (InnerCommutative && B == C)) {
556 // Consider forming "(A op C) op' B".
557 // If "A op C" simplifies then it can be formed with no cost.
558 V = SimplifyBinOp(TopLevelOpcode, A, C, SQ.getWithInstruction(&I));
560 // If "A op C" doesn't simplify then only go on if both of the existing
561 // operations "A op' B" and "C op' D" will be zapped as no longer used.
562 if (!V && LHS->hasOneUse() && RHS->hasOneUse())
563 V = Builder.CreateBinOp(TopLevelOpcode, A, C, LHS->getName());
565 SimplifiedInst = Builder.CreateBinOp(InnerOpcode, V, B);
569 if (SimplifiedInst) {
571 SimplifiedInst->takeName(&I);
573 // Check if we can add NSW flag to SimplifiedInst. If so, set NSW flag.
574 // TODO: Check for NUW.
575 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(SimplifiedInst)) {
576 if (isa<OverflowingBinaryOperator>(SimplifiedInst)) {
578 if (isa<OverflowingBinaryOperator>(&I))
579 HasNSW = I.hasNoSignedWrap();
581 if (auto *LOBO = dyn_cast<OverflowingBinaryOperator>(LHS))
582 HasNSW &= LOBO->hasNoSignedWrap();
584 if (auto *ROBO = dyn_cast<OverflowingBinaryOperator>(RHS))
585 HasNSW &= ROBO->hasNoSignedWrap();
587 // We can propagate 'nsw' if we know that
588 // %Y = mul nsw i16 %X, C
589 // %Z = add nsw i16 %Y, %X
591 // %Z = mul nsw i16 %X, C+1
593 // iff C+1 isn't INT_MIN
595 if (TopLevelOpcode == Instruction::Add &&
596 InnerOpcode == Instruction::Mul)
597 if (match(V, m_APInt(CInt)) && !CInt->isMinSignedValue())
598 BO->setHasNoSignedWrap(HasNSW);
602 return SimplifiedInst;
605 /// This tries to simplify binary operations which some other binary operation
606 /// distributes over either by factorizing out common terms
607 /// (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this results in
608 /// simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is a win).
609 /// Returns the simplified value, or null if it didn't simplify.
610 Value *InstCombiner::SimplifyUsingDistributiveLaws(BinaryOperator &I) {
611 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
612 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
613 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
614 Instruction::BinaryOps TopLevelOpcode = I.getOpcode();
618 Value *A, *B, *C, *D;
619 Instruction::BinaryOps LHSOpcode, RHSOpcode;
621 LHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op0, A, B);
623 RHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op1, C, D);
625 // The instruction has the form "(A op' B) op (C op' D)". Try to factorize
627 if (Op0 && Op1 && LHSOpcode == RHSOpcode)
628 if (Value *V = tryFactorization(I, LHSOpcode, A, B, C, D))
631 // The instruction has the form "(A op' B) op (C)". Try to factorize common
634 if (Value *Ident = getIdentityValue(LHSOpcode, RHS))
635 if (Value *V = tryFactorization(I, LHSOpcode, A, B, RHS, Ident))
638 // The instruction has the form "(B) op (C op' D)". Try to factorize common
641 if (Value *Ident = getIdentityValue(RHSOpcode, LHS))
642 if (Value *V = tryFactorization(I, RHSOpcode, LHS, Ident, C, D))
647 if (Op0 && rightDistributesOverLeft(Op0->getOpcode(), TopLevelOpcode)) {
648 // The instruction has the form "(A op' B) op C". See if expanding it out
649 // to "(A op C) op' (B op C)" results in simplifications.
650 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
651 Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op'
653 Value *L = SimplifyBinOp(TopLevelOpcode, A, C, SQ.getWithInstruction(&I));
654 Value *R = SimplifyBinOp(TopLevelOpcode, B, C, SQ.getWithInstruction(&I));
656 // Do "A op C" and "B op C" both simplify?
658 // They do! Return "L op' R".
660 C = Builder.CreateBinOp(InnerOpcode, L, R);
665 // Does "A op C" simplify to the identity value for the inner opcode?
666 if (L && L == ConstantExpr::getBinOpIdentity(InnerOpcode, L->getType())) {
667 // They do! Return "B op C".
669 C = Builder.CreateBinOp(TopLevelOpcode, B, C);
674 // Does "B op C" simplify to the identity value for the inner opcode?
675 if (R && R == ConstantExpr::getBinOpIdentity(InnerOpcode, R->getType())) {
676 // They do! Return "A op C".
678 C = Builder.CreateBinOp(TopLevelOpcode, A, C);
684 if (Op1 && leftDistributesOverRight(TopLevelOpcode, Op1->getOpcode())) {
685 // The instruction has the form "A op (B op' C)". See if expanding it out
686 // to "(A op B) op' (A op C)" results in simplifications.
687 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
688 Instruction::BinaryOps InnerOpcode = Op1->getOpcode(); // op'
690 Value *L = SimplifyBinOp(TopLevelOpcode, A, B, SQ.getWithInstruction(&I));
691 Value *R = SimplifyBinOp(TopLevelOpcode, A, C, SQ.getWithInstruction(&I));
693 // Do "A op B" and "A op C" both simplify?
695 // They do! Return "L op' R".
697 A = Builder.CreateBinOp(InnerOpcode, L, R);
702 // Does "A op B" simplify to the identity value for the inner opcode?
703 if (L && L == ConstantExpr::getBinOpIdentity(InnerOpcode, L->getType())) {
704 // They do! Return "A op C".
706 A = Builder.CreateBinOp(TopLevelOpcode, A, C);
711 // Does "A op C" simplify to the identity value for the inner opcode?
712 if (R && R == ConstantExpr::getBinOpIdentity(InnerOpcode, R->getType())) {
713 // They do! Return "A op B".
715 A = Builder.CreateBinOp(TopLevelOpcode, A, B);
721 return SimplifySelectsFeedingBinaryOp(I, LHS, RHS);
724 Value *InstCombiner::SimplifySelectsFeedingBinaryOp(BinaryOperator &I,
725 Value *LHS, Value *RHS) {
726 Instruction::BinaryOps Opcode = I.getOpcode();
727 // (op (select (a, b, c)), (select (a, d, e))) -> (select (a, (op b, d), (op
729 Value *A, *B, *C, *D, *E;
731 if (match(LHS, m_Select(m_Value(A), m_Value(B), m_Value(C))) &&
732 match(RHS, m_Select(m_Specific(A), m_Value(D), m_Value(E)))) {
733 bool SelectsHaveOneUse = LHS->hasOneUse() && RHS->hasOneUse();
734 BuilderTy::FastMathFlagGuard Guard(Builder);
735 if (isa<FPMathOperator>(&I))
736 Builder.setFastMathFlags(I.getFastMathFlags());
738 Value *V1 = SimplifyBinOp(Opcode, C, E, SQ.getWithInstruction(&I));
739 Value *V2 = SimplifyBinOp(Opcode, B, D, SQ.getWithInstruction(&I));
741 SI = Builder.CreateSelect(A, V2, V1);
742 else if (V2 && SelectsHaveOneUse)
743 SI = Builder.CreateSelect(A, V2, Builder.CreateBinOp(Opcode, C, E));
744 else if (V1 && SelectsHaveOneUse)
745 SI = Builder.CreateSelect(A, Builder.CreateBinOp(Opcode, B, D), V1);
754 /// Given a 'sub' instruction, return the RHS of the instruction if the LHS is a
755 /// constant zero (which is the 'negate' form).
756 Value *InstCombiner::dyn_castNegVal(Value *V) const {
758 if (match(V, m_Neg(m_Value(NegV))))
761 // Constants can be considered to be negated values if they can be folded.
762 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
763 return ConstantExpr::getNeg(C);
765 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
766 if (C->getType()->getElementType()->isIntegerTy())
767 return ConstantExpr::getNeg(C);
769 if (ConstantVector *CV = dyn_cast<ConstantVector>(V)) {
770 for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) {
771 Constant *Elt = CV->getAggregateElement(i);
775 if (isa<UndefValue>(Elt))
778 if (!isa<ConstantInt>(Elt))
781 return ConstantExpr::getNeg(CV);
787 static Value *foldOperationIntoSelectOperand(Instruction &I, Value *SO,
788 InstCombiner::BuilderTy &Builder) {
789 if (auto *Cast = dyn_cast<CastInst>(&I))
790 return Builder.CreateCast(Cast->getOpcode(), SO, I.getType());
792 assert(I.isBinaryOp() && "Unexpected opcode for select folding");
794 // Figure out if the constant is the left or the right argument.
795 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
796 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
798 if (auto *SOC = dyn_cast<Constant>(SO)) {
800 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
801 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
804 Value *Op0 = SO, *Op1 = ConstOperand;
808 auto *BO = cast<BinaryOperator>(&I);
809 Value *RI = Builder.CreateBinOp(BO->getOpcode(), Op0, Op1,
810 SO->getName() + ".op");
811 auto *FPInst = dyn_cast<Instruction>(RI);
812 if (FPInst && isa<FPMathOperator>(FPInst))
813 FPInst->copyFastMathFlags(BO);
817 Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) {
818 // Don't modify shared select instructions.
819 if (!SI->hasOneUse())
822 Value *TV = SI->getTrueValue();
823 Value *FV = SI->getFalseValue();
824 if (!(isa<Constant>(TV) || isa<Constant>(FV)))
827 // Bool selects with constant operands can be folded to logical ops.
828 if (SI->getType()->isIntOrIntVectorTy(1))
831 // If it's a bitcast involving vectors, make sure it has the same number of
832 // elements on both sides.
833 if (auto *BC = dyn_cast<BitCastInst>(&Op)) {
834 VectorType *DestTy = dyn_cast<VectorType>(BC->getDestTy());
835 VectorType *SrcTy = dyn_cast<VectorType>(BC->getSrcTy());
837 // Verify that either both or neither are vectors.
838 if ((SrcTy == nullptr) != (DestTy == nullptr))
841 // If vectors, verify that they have the same number of elements.
842 if (SrcTy && SrcTy->getNumElements() != DestTy->getNumElements())
846 // Test if a CmpInst instruction is used exclusively by a select as
847 // part of a minimum or maximum operation. If so, refrain from doing
848 // any other folding. This helps out other analyses which understand
849 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
850 // and CodeGen. And in this case, at least one of the comparison
851 // operands has at least one user besides the compare (the select),
852 // which would often largely negate the benefit of folding anyway.
853 if (auto *CI = dyn_cast<CmpInst>(SI->getCondition())) {
854 if (CI->hasOneUse()) {
855 Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1);
856 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
857 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
862 Value *NewTV = foldOperationIntoSelectOperand(Op, TV, Builder);
863 Value *NewFV = foldOperationIntoSelectOperand(Op, FV, Builder);
864 return SelectInst::Create(SI->getCondition(), NewTV, NewFV, "", nullptr, SI);
867 static Value *foldOperationIntoPhiValue(BinaryOperator *I, Value *InV,
868 InstCombiner::BuilderTy &Builder) {
869 bool ConstIsRHS = isa<Constant>(I->getOperand(1));
870 Constant *C = cast<Constant>(I->getOperand(ConstIsRHS));
872 if (auto *InC = dyn_cast<Constant>(InV)) {
874 return ConstantExpr::get(I->getOpcode(), InC, C);
875 return ConstantExpr::get(I->getOpcode(), C, InC);
878 Value *Op0 = InV, *Op1 = C;
882 Value *RI = Builder.CreateBinOp(I->getOpcode(), Op0, Op1, "phitmp");
883 auto *FPInst = dyn_cast<Instruction>(RI);
884 if (FPInst && isa<FPMathOperator>(FPInst))
885 FPInst->copyFastMathFlags(I);
889 Instruction *InstCombiner::foldOpIntoPhi(Instruction &I, PHINode *PN) {
890 unsigned NumPHIValues = PN->getNumIncomingValues();
891 if (NumPHIValues == 0)
894 // We normally only transform phis with a single use. However, if a PHI has
895 // multiple uses and they are all the same operation, we can fold *all* of the
896 // uses into the PHI.
897 if (!PN->hasOneUse()) {
898 // Walk the use list for the instruction, comparing them to I.
899 for (User *U : PN->users()) {
900 Instruction *UI = cast<Instruction>(U);
901 if (UI != &I && !I.isIdenticalTo(UI))
904 // Otherwise, we can replace *all* users with the new PHI we form.
907 // Check to see if all of the operands of the PHI are simple constants
908 // (constantint/constantfp/undef). If there is one non-constant value,
909 // remember the BB it is in. If there is more than one or if *it* is a PHI,
910 // bail out. We don't do arbitrary constant expressions here because moving
911 // their computation can be expensive without a cost model.
912 BasicBlock *NonConstBB = nullptr;
913 for (unsigned i = 0; i != NumPHIValues; ++i) {
914 Value *InVal = PN->getIncomingValue(i);
915 if (isa<Constant>(InVal) && !isa<ConstantExpr>(InVal))
918 if (isa<PHINode>(InVal)) return nullptr; // Itself a phi.
919 if (NonConstBB) return nullptr; // More than one non-const value.
921 NonConstBB = PN->getIncomingBlock(i);
923 // If the InVal is an invoke at the end of the pred block, then we can't
924 // insert a computation after it without breaking the edge.
925 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal))
926 if (II->getParent() == NonConstBB)
929 // If the incoming non-constant value is in I's block, we will remove one
930 // instruction, but insert another equivalent one, leading to infinite
932 if (isPotentiallyReachable(I.getParent(), NonConstBB, &DT, LI))
936 // If there is exactly one non-constant value, we can insert a copy of the
937 // operation in that block. However, if this is a critical edge, we would be
938 // inserting the computation on some other paths (e.g. inside a loop). Only
939 // do this if the pred block is unconditionally branching into the phi block.
940 if (NonConstBB != nullptr) {
941 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
942 if (!BI || !BI->isUnconditional()) return nullptr;
945 // Okay, we can do the transformation: create the new PHI node.
946 PHINode *NewPN = PHINode::Create(I.getType(), PN->getNumIncomingValues());
947 InsertNewInstBefore(NewPN, *PN);
950 // If we are going to have to insert a new computation, do so right before the
951 // predecessor's terminator.
953 Builder.SetInsertPoint(NonConstBB->getTerminator());
955 // Next, add all of the operands to the PHI.
956 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) {
957 // We only currently try to fold the condition of a select when it is a phi,
958 // not the true/false values.
959 Value *TrueV = SI->getTrueValue();
960 Value *FalseV = SI->getFalseValue();
961 BasicBlock *PhiTransBB = PN->getParent();
962 for (unsigned i = 0; i != NumPHIValues; ++i) {
963 BasicBlock *ThisBB = PN->getIncomingBlock(i);
964 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB);
965 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB);
966 Value *InV = nullptr;
967 // Beware of ConstantExpr: it may eventually evaluate to getNullValue,
968 // even if currently isNullValue gives false.
969 Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i));
970 // For vector constants, we cannot use isNullValue to fold into
971 // FalseVInPred versus TrueVInPred. When we have individual nonzero
972 // elements in the vector, we will incorrectly fold InC to
974 if (InC && !isa<ConstantExpr>(InC) && isa<ConstantInt>(InC))
975 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred;
977 // Generate the select in the same block as PN's current incoming block.
978 // Note: ThisBB need not be the NonConstBB because vector constants
979 // which are constants by definition are handled here.
980 // FIXME: This can lead to an increase in IR generation because we might
981 // generate selects for vector constant phi operand, that could not be
982 // folded to TrueVInPred or FalseVInPred as done for ConstantInt. For
983 // non-vector phis, this transformation was always profitable because
984 // the select would be generated exactly once in the NonConstBB.
985 Builder.SetInsertPoint(ThisBB->getTerminator());
986 InV = Builder.CreateSelect(PN->getIncomingValue(i), TrueVInPred,
987 FalseVInPred, "phitmp");
989 NewPN->addIncoming(InV, ThisBB);
991 } else if (CmpInst *CI = dyn_cast<CmpInst>(&I)) {
992 Constant *C = cast<Constant>(I.getOperand(1));
993 for (unsigned i = 0; i != NumPHIValues; ++i) {
994 Value *InV = nullptr;
995 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
996 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
997 else if (isa<ICmpInst>(CI))
998 InV = Builder.CreateICmp(CI->getPredicate(), PN->getIncomingValue(i),
1001 InV = Builder.CreateFCmp(CI->getPredicate(), PN->getIncomingValue(i),
1003 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1005 } else if (auto *BO = dyn_cast<BinaryOperator>(&I)) {
1006 for (unsigned i = 0; i != NumPHIValues; ++i) {
1007 Value *InV = foldOperationIntoPhiValue(BO, PN->getIncomingValue(i),
1009 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1012 CastInst *CI = cast<CastInst>(&I);
1013 Type *RetTy = CI->getType();
1014 for (unsigned i = 0; i != NumPHIValues; ++i) {
1016 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
1017 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1019 InV = Builder.CreateCast(CI->getOpcode(), PN->getIncomingValue(i),
1020 I.getType(), "phitmp");
1021 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1025 for (auto UI = PN->user_begin(), E = PN->user_end(); UI != E;) {
1026 Instruction *User = cast<Instruction>(*UI++);
1027 if (User == &I) continue;
1028 replaceInstUsesWith(*User, NewPN);
1029 eraseInstFromFunction(*User);
1031 return replaceInstUsesWith(I, NewPN);
1034 Instruction *InstCombiner::foldBinOpIntoSelectOrPhi(BinaryOperator &I) {
1035 if (!isa<Constant>(I.getOperand(1)))
1038 if (auto *Sel = dyn_cast<SelectInst>(I.getOperand(0))) {
1039 if (Instruction *NewSel = FoldOpIntoSelect(I, Sel))
1041 } else if (auto *PN = dyn_cast<PHINode>(I.getOperand(0))) {
1042 if (Instruction *NewPhi = foldOpIntoPhi(I, PN))
1048 /// Given a pointer type and a constant offset, determine whether or not there
1049 /// is a sequence of GEP indices into the pointed type that will land us at the
1050 /// specified offset. If so, fill them into NewIndices and return the resultant
1051 /// element type, otherwise return null.
1052 Type *InstCombiner::FindElementAtOffset(PointerType *PtrTy, int64_t Offset,
1053 SmallVectorImpl<Value *> &NewIndices) {
1054 Type *Ty = PtrTy->getElementType();
1058 // Start with the index over the outer type. Note that the type size
1059 // might be zero (even if the offset isn't zero) if the indexed type
1060 // is something like [0 x {int, int}]
1061 Type *IndexTy = DL.getIndexType(PtrTy);
1062 int64_t FirstIdx = 0;
1063 if (int64_t TySize = DL.getTypeAllocSize(Ty)) {
1064 FirstIdx = Offset/TySize;
1065 Offset -= FirstIdx*TySize;
1067 // Handle hosts where % returns negative instead of values [0..TySize).
1071 assert(Offset >= 0);
1073 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
1076 NewIndices.push_back(ConstantInt::get(IndexTy, FirstIdx));
1078 // Index into the types. If we fail, set OrigBase to null.
1080 // Indexing into tail padding between struct/array elements.
1081 if (uint64_t(Offset * 8) >= DL.getTypeSizeInBits(Ty))
1084 if (StructType *STy = dyn_cast<StructType>(Ty)) {
1085 const StructLayout *SL = DL.getStructLayout(STy);
1086 assert(Offset < (int64_t)SL->getSizeInBytes() &&
1087 "Offset must stay within the indexed type");
1089 unsigned Elt = SL->getElementContainingOffset(Offset);
1090 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
1093 Offset -= SL->getElementOffset(Elt);
1094 Ty = STy->getElementType(Elt);
1095 } else if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
1096 uint64_t EltSize = DL.getTypeAllocSize(AT->getElementType());
1097 assert(EltSize && "Cannot index into a zero-sized array");
1098 NewIndices.push_back(ConstantInt::get(IndexTy,Offset/EltSize));
1100 Ty = AT->getElementType();
1102 // Otherwise, we can't index into the middle of this atomic type, bail.
1110 static bool shouldMergeGEPs(GEPOperator &GEP, GEPOperator &Src) {
1111 // If this GEP has only 0 indices, it is the same pointer as
1112 // Src. If Src is not a trivial GEP too, don't combine
1114 if (GEP.hasAllZeroIndices() && !Src.hasAllZeroIndices() &&
1120 /// Return a value X such that Val = X * Scale, or null if none.
1121 /// If the multiplication is known not to overflow, then NoSignedWrap is set.
1122 Value *InstCombiner::Descale(Value *Val, APInt Scale, bool &NoSignedWrap) {
1123 assert(isa<IntegerType>(Val->getType()) && "Can only descale integers!");
1124 assert(cast<IntegerType>(Val->getType())->getBitWidth() ==
1125 Scale.getBitWidth() && "Scale not compatible with value!");
1127 // If Val is zero or Scale is one then Val = Val * Scale.
1128 if (match(Val, m_Zero()) || Scale == 1) {
1129 NoSignedWrap = true;
1133 // If Scale is zero then it does not divide Val.
1134 if (Scale.isMinValue())
1137 // Look through chains of multiplications, searching for a constant that is
1138 // divisible by Scale. For example, descaling X*(Y*(Z*4)) by a factor of 4
1139 // will find the constant factor 4 and produce X*(Y*Z). Descaling X*(Y*8) by
1140 // a factor of 4 will produce X*(Y*2). The principle of operation is to bore
1143 // Val = M1 * X || Analysis starts here and works down
1144 // M1 = M2 * Y || Doesn't descend into terms with more
1145 // M2 = Z * 4 \/ than one use
1147 // Then to modify a term at the bottom:
1150 // M1 = Z * Y || Replaced M2 with Z
1152 // Then to work back up correcting nsw flags.
1154 // Op - the term we are currently analyzing. Starts at Val then drills down.
1155 // Replaced with its descaled value before exiting from the drill down loop.
1158 // Parent - initially null, but after drilling down notes where Op came from.
1159 // In the example above, Parent is (Val, 0) when Op is M1, because M1 is the
1160 // 0'th operand of Val.
1161 std::pair<Instruction *, unsigned> Parent;
1163 // Set if the transform requires a descaling at deeper levels that doesn't
1165 bool RequireNoSignedWrap = false;
1167 // Log base 2 of the scale. Negative if not a power of 2.
1168 int32_t logScale = Scale.exactLogBase2();
1170 for (;; Op = Parent.first->getOperand(Parent.second)) { // Drill down
1171 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
1172 // If Op is a constant divisible by Scale then descale to the quotient.
1173 APInt Quotient(Scale), Remainder(Scale); // Init ensures right bitwidth.
1174 APInt::sdivrem(CI->getValue(), Scale, Quotient, Remainder);
1175 if (!Remainder.isMinValue())
1176 // Not divisible by Scale.
1178 // Replace with the quotient in the parent.
1179 Op = ConstantInt::get(CI->getType(), Quotient);
1180 NoSignedWrap = true;
1184 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op)) {
1185 if (BO->getOpcode() == Instruction::Mul) {
1187 NoSignedWrap = BO->hasNoSignedWrap();
1188 if (RequireNoSignedWrap && !NoSignedWrap)
1191 // There are three cases for multiplication: multiplication by exactly
1192 // the scale, multiplication by a constant different to the scale, and
1193 // multiplication by something else.
1194 Value *LHS = BO->getOperand(0);
1195 Value *RHS = BO->getOperand(1);
1197 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
1198 // Multiplication by a constant.
1199 if (CI->getValue() == Scale) {
1200 // Multiplication by exactly the scale, replace the multiplication
1201 // by its left-hand side in the parent.
1206 // Otherwise drill down into the constant.
1207 if (!Op->hasOneUse())
1210 Parent = std::make_pair(BO, 1);
1214 // Multiplication by something else. Drill down into the left-hand side
1215 // since that's where the reassociate pass puts the good stuff.
1216 if (!Op->hasOneUse())
1219 Parent = std::make_pair(BO, 0);
1223 if (logScale > 0 && BO->getOpcode() == Instruction::Shl &&
1224 isa<ConstantInt>(BO->getOperand(1))) {
1225 // Multiplication by a power of 2.
1226 NoSignedWrap = BO->hasNoSignedWrap();
1227 if (RequireNoSignedWrap && !NoSignedWrap)
1230 Value *LHS = BO->getOperand(0);
1231 int32_t Amt = cast<ConstantInt>(BO->getOperand(1))->
1232 getLimitedValue(Scale.getBitWidth());
1235 if (Amt == logScale) {
1236 // Multiplication by exactly the scale, replace the multiplication
1237 // by its left-hand side in the parent.
1241 if (Amt < logScale || !Op->hasOneUse())
1244 // Multiplication by more than the scale. Reduce the multiplying amount
1245 // by the scale in the parent.
1246 Parent = std::make_pair(BO, 1);
1247 Op = ConstantInt::get(BO->getType(), Amt - logScale);
1252 if (!Op->hasOneUse())
1255 if (CastInst *Cast = dyn_cast<CastInst>(Op)) {
1256 if (Cast->getOpcode() == Instruction::SExt) {
1257 // Op is sign-extended from a smaller type, descale in the smaller type.
1258 unsigned SmallSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1259 APInt SmallScale = Scale.trunc(SmallSize);
1260 // Suppose Op = sext X, and we descale X as Y * SmallScale. We want to
1261 // descale Op as (sext Y) * Scale. In order to have
1262 // sext (Y * SmallScale) = (sext Y) * Scale
1263 // some conditions need to hold however: SmallScale must sign-extend to
1264 // Scale and the multiplication Y * SmallScale should not overflow.
1265 if (SmallScale.sext(Scale.getBitWidth()) != Scale)
1266 // SmallScale does not sign-extend to Scale.
1268 assert(SmallScale.exactLogBase2() == logScale);
1269 // Require that Y * SmallScale must not overflow.
1270 RequireNoSignedWrap = true;
1272 // Drill down through the cast.
1273 Parent = std::make_pair(Cast, 0);
1278 if (Cast->getOpcode() == Instruction::Trunc) {
1279 // Op is truncated from a larger type, descale in the larger type.
1280 // Suppose Op = trunc X, and we descale X as Y * sext Scale. Then
1281 // trunc (Y * sext Scale) = (trunc Y) * Scale
1282 // always holds. However (trunc Y) * Scale may overflow even if
1283 // trunc (Y * sext Scale) does not, so nsw flags need to be cleared
1284 // from this point up in the expression (see later).
1285 if (RequireNoSignedWrap)
1288 // Drill down through the cast.
1289 unsigned LargeSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1290 Parent = std::make_pair(Cast, 0);
1291 Scale = Scale.sext(LargeSize);
1292 if (logScale + 1 == (int32_t)Cast->getType()->getPrimitiveSizeInBits())
1294 assert(Scale.exactLogBase2() == logScale);
1299 // Unsupported expression, bail out.
1303 // If Op is zero then Val = Op * Scale.
1304 if (match(Op, m_Zero())) {
1305 NoSignedWrap = true;
1309 // We know that we can successfully descale, so from here on we can safely
1310 // modify the IR. Op holds the descaled version of the deepest term in the
1311 // expression. NoSignedWrap is 'true' if multiplying Op by Scale is known
1315 // The expression only had one term.
1318 // Rewrite the parent using the descaled version of its operand.
1319 assert(Parent.first->hasOneUse() && "Drilled down when more than one use!");
1320 assert(Op != Parent.first->getOperand(Parent.second) &&
1321 "Descaling was a no-op?");
1322 Parent.first->setOperand(Parent.second, Op);
1323 Worklist.Add(Parent.first);
1325 // Now work back up the expression correcting nsw flags. The logic is based
1326 // on the following observation: if X * Y is known not to overflow as a signed
1327 // multiplication, and Y is replaced by a value Z with smaller absolute value,
1328 // then X * Z will not overflow as a signed multiplication either. As we work
1329 // our way up, having NoSignedWrap 'true' means that the descaled value at the
1330 // current level has strictly smaller absolute value than the original.
1331 Instruction *Ancestor = Parent.first;
1333 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Ancestor)) {
1334 // If the multiplication wasn't nsw then we can't say anything about the
1335 // value of the descaled multiplication, and we have to clear nsw flags
1336 // from this point on up.
1337 bool OpNoSignedWrap = BO->hasNoSignedWrap();
1338 NoSignedWrap &= OpNoSignedWrap;
1339 if (NoSignedWrap != OpNoSignedWrap) {
1340 BO->setHasNoSignedWrap(NoSignedWrap);
1341 Worklist.Add(Ancestor);
1343 } else if (Ancestor->getOpcode() == Instruction::Trunc) {
1344 // The fact that the descaled input to the trunc has smaller absolute
1345 // value than the original input doesn't tell us anything useful about
1346 // the absolute values of the truncations.
1347 NoSignedWrap = false;
1349 assert((Ancestor->getOpcode() != Instruction::SExt || NoSignedWrap) &&
1350 "Failed to keep proper track of nsw flags while drilling down?");
1352 if (Ancestor == Val)
1353 // Got to the top, all done!
1356 // Move up one level in the expression.
1357 assert(Ancestor->hasOneUse() && "Drilled down when more than one use!");
1358 Ancestor = Ancestor->user_back();
1362 Instruction *InstCombiner::foldVectorBinop(BinaryOperator &Inst) {
1363 if (!Inst.getType()->isVectorTy()) return nullptr;
1365 BinaryOperator::BinaryOps Opcode = Inst.getOpcode();
1366 unsigned NumElts = cast<VectorType>(Inst.getType())->getNumElements();
1367 Value *LHS = Inst.getOperand(0), *RHS = Inst.getOperand(1);
1368 assert(cast<VectorType>(LHS->getType())->getNumElements() == NumElts);
1369 assert(cast<VectorType>(RHS->getType())->getNumElements() == NumElts);
1371 // If both operands of the binop are vector concatenations, then perform the
1372 // narrow binop on each pair of the source operands followed by concatenation
1374 Value *L0, *L1, *R0, *R1;
1376 if (match(LHS, m_ShuffleVector(m_Value(L0), m_Value(L1), m_Constant(Mask))) &&
1377 match(RHS, m_ShuffleVector(m_Value(R0), m_Value(R1), m_Specific(Mask))) &&
1378 LHS->hasOneUse() && RHS->hasOneUse() &&
1379 cast<ShuffleVectorInst>(LHS)->isConcat()) {
1380 // This transform does not have the speculative execution constraint as
1381 // below because the shuffle is a concatenation. The new binops are
1382 // operating on exactly the same elements as the existing binop.
1383 // TODO: We could ease the mask requirement to allow different undef lanes,
1384 // but that requires an analysis of the binop-with-undef output value.
1385 Value *NewBO0 = Builder.CreateBinOp(Opcode, L0, R0);
1386 if (auto *BO = dyn_cast<BinaryOperator>(NewBO0))
1387 BO->copyIRFlags(&Inst);
1388 Value *NewBO1 = Builder.CreateBinOp(Opcode, L1, R1);
1389 if (auto *BO = dyn_cast<BinaryOperator>(NewBO1))
1390 BO->copyIRFlags(&Inst);
1391 return new ShuffleVectorInst(NewBO0, NewBO1, Mask);
1394 // It may not be safe to reorder shuffles and things like div, urem, etc.
1395 // because we may trap when executing those ops on unknown vector elements.
1397 if (!isSafeToSpeculativelyExecute(&Inst))
1400 auto createBinOpShuffle = [&](Value *X, Value *Y, Constant *M) {
1401 Value *XY = Builder.CreateBinOp(Opcode, X, Y);
1402 if (auto *BO = dyn_cast<BinaryOperator>(XY))
1403 BO->copyIRFlags(&Inst);
1404 return new ShuffleVectorInst(XY, UndefValue::get(XY->getType()), M);
1407 // If both arguments of the binary operation are shuffles that use the same
1408 // mask and shuffle within a single vector, move the shuffle after the binop.
1410 if (match(LHS, m_ShuffleVector(m_Value(V1), m_Undef(), m_Constant(Mask))) &&
1411 match(RHS, m_ShuffleVector(m_Value(V2), m_Undef(), m_Specific(Mask))) &&
1412 V1->getType() == V2->getType() &&
1413 (LHS->hasOneUse() || RHS->hasOneUse() || LHS == RHS)) {
1414 // Op(shuffle(V1, Mask), shuffle(V2, Mask)) -> shuffle(Op(V1, V2), Mask)
1415 return createBinOpShuffle(V1, V2, Mask);
1418 // If one argument is a shuffle within one vector and the other is a constant,
1419 // try moving the shuffle after the binary operation. This canonicalization
1420 // intends to move shuffles closer to other shuffles and binops closer to
1421 // other binops, so they can be folded. It may also enable demanded elements
1424 if (match(&Inst, m_c_BinOp(
1425 m_OneUse(m_ShuffleVector(m_Value(V1), m_Undef(), m_Constant(Mask))),
1427 V1->getType()->getVectorNumElements() <= NumElts) {
1428 assert(Inst.getType()->getScalarType() == V1->getType()->getScalarType() &&
1429 "Shuffle should not change scalar type");
1431 // Find constant NewC that has property:
1432 // shuffle(NewC, ShMask) = C
1433 // If such constant does not exist (example: ShMask=<0,0> and C=<1,2>)
1434 // reorder is not possible. A 1-to-1 mapping is not required. Example:
1435 // ShMask = <1,1,2,2> and C = <5,5,6,6> --> NewC = <undef,5,6,undef>
1436 bool ConstOp1 = isa<Constant>(RHS);
1437 SmallVector<int, 16> ShMask;
1438 ShuffleVectorInst::getShuffleMask(Mask, ShMask);
1439 unsigned SrcVecNumElts = V1->getType()->getVectorNumElements();
1440 UndefValue *UndefScalar = UndefValue::get(C->getType()->getScalarType());
1441 SmallVector<Constant *, 16> NewVecC(SrcVecNumElts, UndefScalar);
1442 bool MayChange = true;
1443 for (unsigned I = 0; I < NumElts; ++I) {
1444 Constant *CElt = C->getAggregateElement(I);
1445 if (ShMask[I] >= 0) {
1446 assert(ShMask[I] < (int)NumElts && "Not expecting narrowing shuffle");
1447 Constant *NewCElt = NewVecC[ShMask[I]];
1449 // 1. The constant vector contains a constant expression.
1450 // 2. The shuffle needs an element of the constant vector that can't
1451 // be mapped to a new constant vector.
1452 // 3. This is a widening shuffle that copies elements of V1 into the
1453 // extended elements (extending with undef is allowed).
1454 if (!CElt || (!isa<UndefValue>(NewCElt) && NewCElt != CElt) ||
1455 I >= SrcVecNumElts) {
1459 NewVecC[ShMask[I]] = CElt;
1461 // If this is a widening shuffle, we must be able to extend with undef
1462 // elements. If the original binop does not produce an undef in the high
1463 // lanes, then this transform is not safe.
1464 // TODO: We could shuffle those non-undef constant values into the
1465 // result by using a constant vector (rather than an undef vector)
1466 // as operand 1 of the new binop, but that might be too aggressive
1467 // for target-independent shuffle creation.
1468 if (I >= SrcVecNumElts) {
1469 Constant *MaybeUndef =
1470 ConstOp1 ? ConstantExpr::get(Opcode, UndefScalar, CElt)
1471 : ConstantExpr::get(Opcode, CElt, UndefScalar);
1472 if (!isa<UndefValue>(MaybeUndef)) {
1479 Constant *NewC = ConstantVector::get(NewVecC);
1480 // It may not be safe to execute a binop on a vector with undef elements
1481 // because the entire instruction can be folded to undef or create poison
1482 // that did not exist in the original code.
1483 if (Inst.isIntDivRem() || (Inst.isShift() && ConstOp1))
1484 NewC = getSafeVectorConstantForBinop(Opcode, NewC, ConstOp1);
1486 // Op(shuffle(V1, Mask), C) -> shuffle(Op(V1, NewC), Mask)
1487 // Op(C, shuffle(V1, Mask)) -> shuffle(Op(NewC, V1), Mask)
1488 Value *NewLHS = ConstOp1 ? V1 : NewC;
1489 Value *NewRHS = ConstOp1 ? NewC : V1;
1490 return createBinOpShuffle(NewLHS, NewRHS, Mask);
1497 /// Try to narrow the width of a binop if at least 1 operand is an extend of
1498 /// of a value. This requires a potentially expensive known bits check to make
1499 /// sure the narrow op does not overflow.
1500 Instruction *InstCombiner::narrowMathIfNoOverflow(BinaryOperator &BO) {
1501 // We need at least one extended operand.
1502 Value *Op0 = BO.getOperand(0), *Op1 = BO.getOperand(1);
1504 // If this is a sub, we swap the operands since we always want an extension
1505 // on the RHS. The LHS can be an extension or a constant.
1506 if (BO.getOpcode() == Instruction::Sub)
1507 std::swap(Op0, Op1);
1510 bool IsSext = match(Op0, m_SExt(m_Value(X)));
1511 if (!IsSext && !match(Op0, m_ZExt(m_Value(X))))
1514 // If both operands are the same extension from the same source type and we
1515 // can eliminate at least one (hasOneUse), this might work.
1516 CastInst::CastOps CastOpc = IsSext ? Instruction::SExt : Instruction::ZExt;
1518 if (!(match(Op1, m_ZExtOrSExt(m_Value(Y))) && X->getType() == Y->getType() &&
1519 cast<Operator>(Op1)->getOpcode() == CastOpc &&
1520 (Op0->hasOneUse() || Op1->hasOneUse()))) {
1521 // If that did not match, see if we have a suitable constant operand.
1522 // Truncating and extending must produce the same constant.
1524 if (!Op0->hasOneUse() || !match(Op1, m_Constant(WideC)))
1526 Constant *NarrowC = ConstantExpr::getTrunc(WideC, X->getType());
1527 if (ConstantExpr::getCast(CastOpc, NarrowC, BO.getType()) != WideC)
1532 // Swap back now that we found our operands.
1533 if (BO.getOpcode() == Instruction::Sub)
1536 // Both operands have narrow versions. Last step: the math must not overflow
1537 // in the narrow width.
1538 if (!willNotOverflow(BO.getOpcode(), X, Y, BO, IsSext))
1541 // bo (ext X), (ext Y) --> ext (bo X, Y)
1542 // bo (ext X), C --> ext (bo X, C')
1543 Value *NarrowBO = Builder.CreateBinOp(BO.getOpcode(), X, Y, "narrow");
1544 if (auto *NewBinOp = dyn_cast<BinaryOperator>(NarrowBO)) {
1546 NewBinOp->setHasNoSignedWrap();
1548 NewBinOp->setHasNoUnsignedWrap();
1550 return CastInst::Create(CastOpc, NarrowBO, BO.getType());
1553 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
1554 SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end());
1555 Type *GEPType = GEP.getType();
1556 Type *GEPEltType = GEP.getSourceElementType();
1557 if (Value *V = SimplifyGEPInst(GEPEltType, Ops, SQ.getWithInstruction(&GEP)))
1558 return replaceInstUsesWith(GEP, V);
1560 Value *PtrOp = GEP.getOperand(0);
1562 // Eliminate unneeded casts for indices, and replace indices which displace
1563 // by multiples of a zero size type with zero.
1564 bool MadeChange = false;
1566 // Index width may not be the same width as pointer width.
1567 // Data layout chooses the right type based on supported integer types.
1568 Type *NewScalarIndexTy =
1569 DL.getIndexType(GEP.getPointerOperandType()->getScalarType());
1571 gep_type_iterator GTI = gep_type_begin(GEP);
1572 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end(); I != E;
1574 // Skip indices into struct types.
1578 Type *IndexTy = (*I)->getType();
1579 Type *NewIndexType =
1580 IndexTy->isVectorTy()
1581 ? VectorType::get(NewScalarIndexTy, IndexTy->getVectorNumElements())
1584 // If the element type has zero size then any index over it is equivalent
1585 // to an index of zero, so replace it with zero if it is not zero already.
1586 Type *EltTy = GTI.getIndexedType();
1587 if (EltTy->isSized() && DL.getTypeAllocSize(EltTy) == 0)
1588 if (!isa<Constant>(*I) || !cast<Constant>(*I)->isNullValue()) {
1589 *I = Constant::getNullValue(NewIndexType);
1593 if (IndexTy != NewIndexType) {
1594 // If we are using a wider index than needed for this platform, shrink
1595 // it to what we need. If narrower, sign-extend it to what we need.
1596 // This explicit cast can make subsequent optimizations more obvious.
1597 *I = Builder.CreateIntCast(*I, NewIndexType, true);
1604 // Check to see if the inputs to the PHI node are getelementptr instructions.
1605 if (auto *PN = dyn_cast<PHINode>(PtrOp)) {
1606 auto *Op1 = dyn_cast<GetElementPtrInst>(PN->getOperand(0));
1610 // Don't fold a GEP into itself through a PHI node. This can only happen
1611 // through the back-edge of a loop. Folding a GEP into itself means that
1612 // the value of the previous iteration needs to be stored in the meantime,
1613 // thus requiring an additional register variable to be live, but not
1614 // actually achieving anything (the GEP still needs to be executed once per
1621 for (auto I = PN->op_begin()+1, E = PN->op_end(); I !=E; ++I) {
1622 auto *Op2 = dyn_cast<GetElementPtrInst>(*I);
1623 if (!Op2 || Op1->getNumOperands() != Op2->getNumOperands())
1626 // As for Op1 above, don't try to fold a GEP into itself.
1630 // Keep track of the type as we walk the GEP.
1631 Type *CurTy = nullptr;
1633 for (unsigned J = 0, F = Op1->getNumOperands(); J != F; ++J) {
1634 if (Op1->getOperand(J)->getType() != Op2->getOperand(J)->getType())
1637 if (Op1->getOperand(J) != Op2->getOperand(J)) {
1639 // We have not seen any differences yet in the GEPs feeding the
1640 // PHI yet, so we record this one if it is allowed to be a
1643 // The first two arguments can vary for any GEP, the rest have to be
1644 // static for struct slots
1645 if (J > 1 && CurTy->isStructTy())
1650 // The GEP is different by more than one input. While this could be
1651 // extended to support GEPs that vary by more than one variable it
1652 // doesn't make sense since it greatly increases the complexity and
1653 // would result in an R+R+R addressing mode which no backend
1654 // directly supports and would need to be broken into several
1655 // simpler instructions anyway.
1660 // Sink down a layer of the type for the next iteration.
1663 CurTy = Op1->getSourceElementType();
1664 } else if (auto *CT = dyn_cast<CompositeType>(CurTy)) {
1665 CurTy = CT->getTypeAtIndex(Op1->getOperand(J));
1673 // If not all GEPs are identical we'll have to create a new PHI node.
1674 // Check that the old PHI node has only one use so that it will get
1676 if (DI != -1 && !PN->hasOneUse())
1679 auto *NewGEP = cast<GetElementPtrInst>(Op1->clone());
1681 // All the GEPs feeding the PHI are identical. Clone one down into our
1682 // BB so that it can be merged with the current GEP.
1683 GEP.getParent()->getInstList().insert(
1684 GEP.getParent()->getFirstInsertionPt(), NewGEP);
1686 // All the GEPs feeding the PHI differ at a single offset. Clone a GEP
1687 // into the current block so it can be merged, and create a new PHI to
1691 IRBuilderBase::InsertPointGuard Guard(Builder);
1692 Builder.SetInsertPoint(PN);
1693 NewPN = Builder.CreatePHI(Op1->getOperand(DI)->getType(),
1694 PN->getNumOperands());
1697 for (auto &I : PN->operands())
1698 NewPN->addIncoming(cast<GEPOperator>(I)->getOperand(DI),
1699 PN->getIncomingBlock(I));
1701 NewGEP->setOperand(DI, NewPN);
1702 GEP.getParent()->getInstList().insert(
1703 GEP.getParent()->getFirstInsertionPt(), NewGEP);
1704 NewGEP->setOperand(DI, NewPN);
1707 GEP.setOperand(0, NewGEP);
1711 // Combine Indices - If the source pointer to this getelementptr instruction
1712 // is a getelementptr instruction, combine the indices of the two
1713 // getelementptr instructions into a single instruction.
1714 if (auto *Src = dyn_cast<GEPOperator>(PtrOp)) {
1715 if (!shouldMergeGEPs(*cast<GEPOperator>(&GEP), *Src))
1718 // Try to reassociate loop invariant GEP chains to enable LICM.
1719 if (LI && Src->getNumOperands() == 2 && GEP.getNumOperands() == 2 &&
1721 if (Loop *L = LI->getLoopFor(GEP.getParent())) {
1722 Value *GO1 = GEP.getOperand(1);
1723 Value *SO1 = Src->getOperand(1);
1724 // Reassociate the two GEPs if SO1 is variant in the loop and GO1 is
1725 // invariant: this breaks the dependence between GEPs and allows LICM
1726 // to hoist the invariant part out of the loop.
1727 if (L->isLoopInvariant(GO1) && !L->isLoopInvariant(SO1)) {
1728 // We have to be careful here.
1729 // We have something like:
1730 // %src = getelementptr <ty>, <ty>* %base, <ty> %idx
1731 // %gep = getelementptr <ty>, <ty>* %src, <ty> %idx2
1732 // If we just swap idx & idx2 then we could inadvertantly
1733 // change %src from a vector to a scalar, or vice versa.
1735 // 1) %base a scalar & idx a scalar & idx2 a vector
1736 // => Swapping idx & idx2 turns %src into a vector type.
1737 // 2) %base a scalar & idx a vector & idx2 a scalar
1738 // => Swapping idx & idx2 turns %src in a scalar type
1739 // 3) %base, %idx, and %idx2 are scalars
1740 // => %src & %gep are scalars
1741 // => swapping idx & idx2 is safe
1742 // 4) %base a vector
1743 // => %src is a vector
1744 // => swapping idx & idx2 is safe.
1745 auto *SO0 = Src->getOperand(0);
1746 auto *SO0Ty = SO0->getType();
1747 if (!isa<VectorType>(GEPType) || // case 3
1748 isa<VectorType>(SO0Ty)) { // case 4
1749 Src->setOperand(1, GO1);
1750 GEP.setOperand(1, SO1);
1754 // -- have to recreate %src & %gep
1755 // put NewSrc at same location as %src
1756 Builder.SetInsertPoint(cast<Instruction>(PtrOp));
1757 auto *NewSrc = cast<GetElementPtrInst>(
1758 Builder.CreateGEP(SO0, GO1, Src->getName()));
1759 NewSrc->setIsInBounds(Src->isInBounds());
1760 auto *NewGEP = GetElementPtrInst::Create(nullptr, NewSrc, {SO1});
1761 NewGEP->setIsInBounds(GEP.isInBounds());
1768 // Note that if our source is a gep chain itself then we wait for that
1769 // chain to be resolved before we perform this transformation. This
1770 // avoids us creating a TON of code in some cases.
1771 if (auto *SrcGEP = dyn_cast<GEPOperator>(Src->getOperand(0)))
1772 if (SrcGEP->getNumOperands() == 2 && shouldMergeGEPs(*Src, *SrcGEP))
1773 return nullptr; // Wait until our source is folded to completion.
1775 SmallVector<Value*, 8> Indices;
1777 // Find out whether the last index in the source GEP is a sequential idx.
1778 bool EndsWithSequential = false;
1779 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src);
1781 EndsWithSequential = I.isSequential();
1783 // Can we combine the two pointer arithmetics offsets?
1784 if (EndsWithSequential) {
1785 // Replace: gep (gep %P, long B), long A, ...
1786 // With: T = long A+B; gep %P, T, ...
1787 Value *SO1 = Src->getOperand(Src->getNumOperands()-1);
1788 Value *GO1 = GEP.getOperand(1);
1790 // If they aren't the same type, then the input hasn't been processed
1791 // by the loop above yet (which canonicalizes sequential index types to
1792 // intptr_t). Just avoid transforming this until the input has been
1794 if (SO1->getType() != GO1->getType())
1798 SimplifyAddInst(GO1, SO1, false, false, SQ.getWithInstruction(&GEP));
1799 // Only do the combine when we are sure the cost after the
1800 // merge is never more than that before the merge.
1804 // Update the GEP in place if possible.
1805 if (Src->getNumOperands() == 2) {
1806 GEP.setOperand(0, Src->getOperand(0));
1807 GEP.setOperand(1, Sum);
1810 Indices.append(Src->op_begin()+1, Src->op_end()-1);
1811 Indices.push_back(Sum);
1812 Indices.append(GEP.op_begin()+2, GEP.op_end());
1813 } else if (isa<Constant>(*GEP.idx_begin()) &&
1814 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
1815 Src->getNumOperands() != 1) {
1816 // Otherwise we can do the fold if the first index of the GEP is a zero
1817 Indices.append(Src->op_begin()+1, Src->op_end());
1818 Indices.append(GEP.idx_begin()+1, GEP.idx_end());
1821 if (!Indices.empty())
1822 return GEP.isInBounds() && Src->isInBounds()
1823 ? GetElementPtrInst::CreateInBounds(
1824 Src->getSourceElementType(), Src->getOperand(0), Indices,
1826 : GetElementPtrInst::Create(Src->getSourceElementType(),
1827 Src->getOperand(0), Indices,
1831 if (GEP.getNumIndices() == 1) {
1832 unsigned AS = GEP.getPointerAddressSpace();
1833 if (GEP.getOperand(1)->getType()->getScalarSizeInBits() ==
1834 DL.getIndexSizeInBits(AS)) {
1835 uint64_t TyAllocSize = DL.getTypeAllocSize(GEPEltType);
1837 bool Matched = false;
1840 if (TyAllocSize == 1) {
1841 V = GEP.getOperand(1);
1843 } else if (match(GEP.getOperand(1),
1844 m_AShr(m_Value(V), m_ConstantInt(C)))) {
1845 if (TyAllocSize == 1ULL << C)
1847 } else if (match(GEP.getOperand(1),
1848 m_SDiv(m_Value(V), m_ConstantInt(C)))) {
1849 if (TyAllocSize == C)
1854 // Canonicalize (gep i8* X, -(ptrtoint Y))
1855 // to (inttoptr (sub (ptrtoint X), (ptrtoint Y)))
1856 // The GEP pattern is emitted by the SCEV expander for certain kinds of
1857 // pointer arithmetic.
1858 if (match(V, m_Neg(m_PtrToInt(m_Value())))) {
1859 Operator *Index = cast<Operator>(V);
1860 Value *PtrToInt = Builder.CreatePtrToInt(PtrOp, Index->getType());
1861 Value *NewSub = Builder.CreateSub(PtrToInt, Index->getOperand(1));
1862 return CastInst::Create(Instruction::IntToPtr, NewSub, GEPType);
1864 // Canonicalize (gep i8* X, (ptrtoint Y)-(ptrtoint X))
1867 if (match(V, m_Sub(m_PtrToInt(m_Value(Y)),
1868 m_PtrToInt(m_Specific(GEP.getOperand(0))))))
1869 return CastInst::CreatePointerBitCastOrAddrSpaceCast(Y, GEPType);
1874 // We do not handle pointer-vector geps here.
1875 if (GEPType->isVectorTy())
1878 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0).
1879 Value *StrippedPtr = PtrOp->stripPointerCasts();
1880 PointerType *StrippedPtrTy = cast<PointerType>(StrippedPtr->getType());
1882 if (StrippedPtr != PtrOp) {
1883 bool HasZeroPointerIndex = false;
1884 if (auto *C = dyn_cast<ConstantInt>(GEP.getOperand(1)))
1885 HasZeroPointerIndex = C->isZero();
1887 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
1888 // into : GEP [10 x i8]* X, i32 0, ...
1890 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
1891 // into : GEP i8* X, ...
1893 // This occurs when the program declares an array extern like "int X[];"
1894 if (HasZeroPointerIndex) {
1895 if (auto *CATy = dyn_cast<ArrayType>(GEPEltType)) {
1896 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
1897 if (CATy->getElementType() == StrippedPtrTy->getElementType()) {
1898 // -> GEP i8* X, ...
1899 SmallVector<Value*, 8> Idx(GEP.idx_begin()+1, GEP.idx_end());
1900 GetElementPtrInst *Res = GetElementPtrInst::Create(
1901 StrippedPtrTy->getElementType(), StrippedPtr, Idx, GEP.getName());
1902 Res->setIsInBounds(GEP.isInBounds());
1903 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace())
1905 // Insert Res, and create an addrspacecast.
1907 // GEP (addrspacecast i8 addrspace(1)* X to [0 x i8]*), i32 0, ...
1909 // %0 = GEP i8 addrspace(1)* X, ...
1910 // addrspacecast i8 addrspace(1)* %0 to i8*
1911 return new AddrSpaceCastInst(Builder.Insert(Res), GEPType);
1914 if (auto *XATy = dyn_cast<ArrayType>(StrippedPtrTy->getElementType())) {
1915 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
1916 if (CATy->getElementType() == XATy->getElementType()) {
1917 // -> GEP [10 x i8]* X, i32 0, ...
1918 // At this point, we know that the cast source type is a pointer
1919 // to an array of the same type as the destination pointer
1920 // array. Because the array type is never stepped over (there
1921 // is a leading zero) we can fold the cast into this GEP.
1922 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace()) {
1923 GEP.setOperand(0, StrippedPtr);
1924 GEP.setSourceElementType(XATy);
1927 // Cannot replace the base pointer directly because StrippedPtr's
1928 // address space is different. Instead, create a new GEP followed by
1929 // an addrspacecast.
1931 // GEP (addrspacecast [10 x i8] addrspace(1)* X to [0 x i8]*),
1934 // %0 = GEP [10 x i8] addrspace(1)* X, ...
1935 // addrspacecast i8 addrspace(1)* %0 to i8*
1936 SmallVector<Value*, 8> Idx(GEP.idx_begin(), GEP.idx_end());
1937 Value *NewGEP = GEP.isInBounds()
1938 ? Builder.CreateInBoundsGEP(
1939 nullptr, StrippedPtr, Idx, GEP.getName())
1940 : Builder.CreateGEP(nullptr, StrippedPtr, Idx,
1942 return new AddrSpaceCastInst(NewGEP, GEPType);
1946 } else if (GEP.getNumOperands() == 2) {
1947 // Transform things like:
1948 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
1949 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
1950 Type *SrcEltTy = StrippedPtrTy->getElementType();
1951 if (SrcEltTy->isArrayTy() &&
1952 DL.getTypeAllocSize(SrcEltTy->getArrayElementType()) ==
1953 DL.getTypeAllocSize(GEPEltType)) {
1954 Type *IdxType = DL.getIndexType(GEPType);
1955 Value *Idx[2] = { Constant::getNullValue(IdxType), GEP.getOperand(1) };
1958 ? Builder.CreateInBoundsGEP(nullptr, StrippedPtr, Idx,
1960 : Builder.CreateGEP(nullptr, StrippedPtr, Idx, GEP.getName());
1962 // V and GEP are both pointer types --> BitCast
1963 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP, GEPType);
1966 // Transform things like:
1967 // %V = mul i64 %N, 4
1968 // %t = getelementptr i8* bitcast (i32* %arr to i8*), i32 %V
1969 // into: %t1 = getelementptr i32* %arr, i32 %N; bitcast
1970 if (GEPEltType->isSized() && SrcEltTy->isSized()) {
1971 // Check that changing the type amounts to dividing the index by a scale
1973 uint64_t ResSize = DL.getTypeAllocSize(GEPEltType);
1974 uint64_t SrcSize = DL.getTypeAllocSize(SrcEltTy);
1975 if (ResSize && SrcSize % ResSize == 0) {
1976 Value *Idx = GEP.getOperand(1);
1977 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
1978 uint64_t Scale = SrcSize / ResSize;
1980 // Earlier transforms ensure that the index has the right type
1981 // according to Data Layout, which considerably simplifies the
1982 // logic by eliminating implicit casts.
1983 assert(Idx->getType() == DL.getIndexType(GEPType) &&
1984 "Index type does not match the Data Layout preferences");
1987 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
1988 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
1989 // If the multiplication NewIdx * Scale may overflow then the new
1990 // GEP may not be "inbounds".
1992 GEP.isInBounds() && NSW
1993 ? Builder.CreateInBoundsGEP(nullptr, StrippedPtr, NewIdx,
1995 : Builder.CreateGEP(nullptr, StrippedPtr, NewIdx,
1998 // The NewGEP must be pointer typed, so must the old one -> BitCast
1999 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
2005 // Similarly, transform things like:
2006 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
2007 // (where tmp = 8*tmp2) into:
2008 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
2009 if (GEPEltType->isSized() && SrcEltTy->isSized() &&
2010 SrcEltTy->isArrayTy()) {
2011 // Check that changing to the array element type amounts to dividing the
2012 // index by a scale factor.
2013 uint64_t ResSize = DL.getTypeAllocSize(GEPEltType);
2014 uint64_t ArrayEltSize =
2015 DL.getTypeAllocSize(SrcEltTy->getArrayElementType());
2016 if (ResSize && ArrayEltSize % ResSize == 0) {
2017 Value *Idx = GEP.getOperand(1);
2018 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
2019 uint64_t Scale = ArrayEltSize / ResSize;
2021 // Earlier transforms ensure that the index has the right type
2022 // according to the Data Layout, which considerably simplifies
2023 // the logic by eliminating implicit casts.
2024 assert(Idx->getType() == DL.getIndexType(GEPType) &&
2025 "Index type does not match the Data Layout preferences");
2028 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
2029 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
2030 // If the multiplication NewIdx * Scale may overflow then the new
2031 // GEP may not be "inbounds".
2032 Type *IndTy = DL.getIndexType(GEPType);
2033 Value *Off[2] = {Constant::getNullValue(IndTy), NewIdx};
2035 Value *NewGEP = GEP.isInBounds() && NSW
2036 ? Builder.CreateInBoundsGEP(
2037 SrcEltTy, StrippedPtr, Off, GEP.getName())
2038 : Builder.CreateGEP(SrcEltTy, StrippedPtr, Off,
2040 // The NewGEP must be pointer typed, so must the old one -> BitCast
2041 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
2049 // addrspacecast between types is canonicalized as a bitcast, then an
2050 // addrspacecast. To take advantage of the below bitcast + struct GEP, look
2051 // through the addrspacecast.
2052 Value *ASCStrippedPtrOp = PtrOp;
2053 if (auto *ASC = dyn_cast<AddrSpaceCastInst>(PtrOp)) {
2054 // X = bitcast A addrspace(1)* to B addrspace(1)*
2055 // Y = addrspacecast A addrspace(1)* to B addrspace(2)*
2056 // Z = gep Y, <...constant indices...>
2057 // Into an addrspacecasted GEP of the struct.
2058 if (auto *BC = dyn_cast<BitCastInst>(ASC->getOperand(0)))
2059 ASCStrippedPtrOp = BC;
2062 if (auto *BCI = dyn_cast<BitCastInst>(ASCStrippedPtrOp)) {
2063 Value *SrcOp = BCI->getOperand(0);
2064 PointerType *SrcType = cast<PointerType>(BCI->getSrcTy());
2065 Type *SrcEltType = SrcType->getElementType();
2067 // GEP directly using the source operand if this GEP is accessing an element
2068 // of a bitcasted pointer to vector or array of the same dimensions:
2069 // gep (bitcast <c x ty>* X to [c x ty]*), Y, Z --> gep X, Y, Z
2070 // gep (bitcast [c x ty]* X to <c x ty>*), Y, Z --> gep X, Y, Z
2071 auto areMatchingArrayAndVecTypes = [](Type *ArrTy, Type *VecTy) {
2072 return ArrTy->getArrayElementType() == VecTy->getVectorElementType() &&
2073 ArrTy->getArrayNumElements() == VecTy->getVectorNumElements();
2075 if (GEP.getNumOperands() == 3 &&
2076 ((GEPEltType->isArrayTy() && SrcEltType->isVectorTy() &&
2077 areMatchingArrayAndVecTypes(GEPEltType, SrcEltType)) ||
2078 (GEPEltType->isVectorTy() && SrcEltType->isArrayTy() &&
2079 areMatchingArrayAndVecTypes(SrcEltType, GEPEltType)))) {
2081 // Create a new GEP here, as using `setOperand()` followed by
2082 // `setSourceElementType()` won't actually update the type of the
2083 // existing GEP Value. Causing issues if this Value is accessed when
2084 // constructing an AddrSpaceCastInst
2087 ? Builder.CreateInBoundsGEP(nullptr, SrcOp, {Ops[1], Ops[2]})
2088 : Builder.CreateGEP(nullptr, SrcOp, {Ops[1], Ops[2]});
2089 NGEP->takeName(&GEP);
2091 // Preserve GEP address space to satisfy users
2092 if (NGEP->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
2093 return new AddrSpaceCastInst(NGEP, GEPType);
2095 return replaceInstUsesWith(GEP, NGEP);
2098 // See if we can simplify:
2099 // X = bitcast A* to B*
2100 // Y = gep X, <...constant indices...>
2101 // into a gep of the original struct. This is important for SROA and alias
2102 // analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
2103 unsigned OffsetBits = DL.getIndexTypeSizeInBits(GEPType);
2104 APInt Offset(OffsetBits, 0);
2105 if (!isa<BitCastInst>(SrcOp) && GEP.accumulateConstantOffset(DL, Offset)) {
2106 // If this GEP instruction doesn't move the pointer, just replace the GEP
2107 // with a bitcast of the real input to the dest type.
2109 // If the bitcast is of an allocation, and the allocation will be
2110 // converted to match the type of the cast, don't touch this.
2111 if (isa<AllocaInst>(SrcOp) || isAllocationFn(SrcOp, &TLI)) {
2112 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
2113 if (Instruction *I = visitBitCast(*BCI)) {
2116 BCI->getParent()->getInstList().insert(BCI->getIterator(), I);
2117 replaceInstUsesWith(*BCI, I);
2123 if (SrcType->getPointerAddressSpace() != GEP.getAddressSpace())
2124 return new AddrSpaceCastInst(SrcOp, GEPType);
2125 return new BitCastInst(SrcOp, GEPType);
2128 // Otherwise, if the offset is non-zero, we need to find out if there is a
2129 // field at Offset in 'A's type. If so, we can pull the cast through the
2131 SmallVector<Value*, 8> NewIndices;
2132 if (FindElementAtOffset(SrcType, Offset.getSExtValue(), NewIndices)) {
2135 ? Builder.CreateInBoundsGEP(nullptr, SrcOp, NewIndices)
2136 : Builder.CreateGEP(nullptr, SrcOp, NewIndices);
2138 if (NGEP->getType() == GEPType)
2139 return replaceInstUsesWith(GEP, NGEP);
2140 NGEP->takeName(&GEP);
2142 if (NGEP->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
2143 return new AddrSpaceCastInst(NGEP, GEPType);
2144 return new BitCastInst(NGEP, GEPType);
2149 if (!GEP.isInBounds()) {
2151 DL.getIndexSizeInBits(PtrOp->getType()->getPointerAddressSpace());
2152 APInt BasePtrOffset(IdxWidth, 0);
2153 Value *UnderlyingPtrOp =
2154 PtrOp->stripAndAccumulateInBoundsConstantOffsets(DL,
2156 if (auto *AI = dyn_cast<AllocaInst>(UnderlyingPtrOp)) {
2157 if (GEP.accumulateConstantOffset(DL, BasePtrOffset) &&
2158 BasePtrOffset.isNonNegative()) {
2159 APInt AllocSize(IdxWidth, DL.getTypeAllocSize(AI->getAllocatedType()));
2160 if (BasePtrOffset.ule(AllocSize)) {
2161 return GetElementPtrInst::CreateInBounds(
2162 PtrOp, makeArrayRef(Ops).slice(1), GEP.getName());
2171 static bool isNeverEqualToUnescapedAlloc(Value *V, const TargetLibraryInfo *TLI,
2173 if (isa<ConstantPointerNull>(V))
2175 if (auto *LI = dyn_cast<LoadInst>(V))
2176 return isa<GlobalVariable>(LI->getPointerOperand());
2177 // Two distinct allocations will never be equal.
2178 // We rely on LookThroughBitCast in isAllocLikeFn being false, since looking
2179 // through bitcasts of V can cause
2180 // the result statement below to be true, even when AI and V (ex:
2181 // i8* ->i32* ->i8* of AI) are the same allocations.
2182 return isAllocLikeFn(V, TLI) && V != AI;
2185 static bool isAllocSiteRemovable(Instruction *AI,
2186 SmallVectorImpl<WeakTrackingVH> &Users,
2187 const TargetLibraryInfo *TLI) {
2188 SmallVector<Instruction*, 4> Worklist;
2189 Worklist.push_back(AI);
2192 Instruction *PI = Worklist.pop_back_val();
2193 for (User *U : PI->users()) {
2194 Instruction *I = cast<Instruction>(U);
2195 switch (I->getOpcode()) {
2197 // Give up the moment we see something we can't handle.
2200 case Instruction::AddrSpaceCast:
2201 case Instruction::BitCast:
2202 case Instruction::GetElementPtr:
2203 Users.emplace_back(I);
2204 Worklist.push_back(I);
2207 case Instruction::ICmp: {
2208 ICmpInst *ICI = cast<ICmpInst>(I);
2209 // We can fold eq/ne comparisons with null to false/true, respectively.
2210 // We also fold comparisons in some conditions provided the alloc has
2211 // not escaped (see isNeverEqualToUnescapedAlloc).
2212 if (!ICI->isEquality())
2214 unsigned OtherIndex = (ICI->getOperand(0) == PI) ? 1 : 0;
2215 if (!isNeverEqualToUnescapedAlloc(ICI->getOperand(OtherIndex), TLI, AI))
2217 Users.emplace_back(I);
2221 case Instruction::Call:
2222 // Ignore no-op and store intrinsics.
2223 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
2224 switch (II->getIntrinsicID()) {
2228 case Intrinsic::memmove:
2229 case Intrinsic::memcpy:
2230 case Intrinsic::memset: {
2231 MemIntrinsic *MI = cast<MemIntrinsic>(II);
2232 if (MI->isVolatile() || MI->getRawDest() != PI)
2236 case Intrinsic::invariant_start:
2237 case Intrinsic::invariant_end:
2238 case Intrinsic::lifetime_start:
2239 case Intrinsic::lifetime_end:
2240 case Intrinsic::objectsize:
2241 Users.emplace_back(I);
2246 if (isFreeCall(I, TLI)) {
2247 Users.emplace_back(I);
2252 case Instruction::Store: {
2253 StoreInst *SI = cast<StoreInst>(I);
2254 if (SI->isVolatile() || SI->getPointerOperand() != PI)
2256 Users.emplace_back(I);
2260 llvm_unreachable("missing a return?");
2262 } while (!Worklist.empty());
2266 Instruction *InstCombiner::visitAllocSite(Instruction &MI) {
2267 // If we have a malloc call which is only used in any amount of comparisons to
2268 // null and free calls, delete the calls and replace the comparisons with true
2269 // or false as appropriate.
2271 // This is based on the principle that we can substitute our own allocation
2272 // function (which will never return null) rather than knowledge of the
2273 // specific function being called. In some sense this can change the permitted
2274 // outputs of a program (when we convert a malloc to an alloca, the fact that
2275 // the allocation is now on the stack is potentially visible, for example),
2276 // but we believe in a permissible manner.
2277 SmallVector<WeakTrackingVH, 64> Users;
2279 // If we are removing an alloca with a dbg.declare, insert dbg.value calls
2280 // before each store.
2281 TinyPtrVector<DbgVariableIntrinsic *> DIIs;
2282 std::unique_ptr<DIBuilder> DIB;
2283 if (isa<AllocaInst>(MI)) {
2284 DIIs = FindDbgAddrUses(&MI);
2285 DIB.reset(new DIBuilder(*MI.getModule(), /*AllowUnresolved=*/false));
2288 if (isAllocSiteRemovable(&MI, Users, &TLI)) {
2289 for (unsigned i = 0, e = Users.size(); i != e; ++i) {
2290 // Lowering all @llvm.objectsize calls first because they may
2291 // use a bitcast/GEP of the alloca we are removing.
2295 Instruction *I = cast<Instruction>(&*Users[i]);
2297 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
2298 if (II->getIntrinsicID() == Intrinsic::objectsize) {
2299 ConstantInt *Result = lowerObjectSizeCall(II, DL, &TLI,
2300 /*MustSucceed=*/true);
2301 replaceInstUsesWith(*I, Result);
2302 eraseInstFromFunction(*I);
2303 Users[i] = nullptr; // Skip examining in the next loop.
2307 for (unsigned i = 0, e = Users.size(); i != e; ++i) {
2311 Instruction *I = cast<Instruction>(&*Users[i]);
2313 if (ICmpInst *C = dyn_cast<ICmpInst>(I)) {
2314 replaceInstUsesWith(*C,
2315 ConstantInt::get(Type::getInt1Ty(C->getContext()),
2316 C->isFalseWhenEqual()));
2317 } else if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I) ||
2318 isa<AddrSpaceCastInst>(I)) {
2319 replaceInstUsesWith(*I, UndefValue::get(I->getType()));
2320 } else if (auto *SI = dyn_cast<StoreInst>(I)) {
2321 for (auto *DII : DIIs)
2322 ConvertDebugDeclareToDebugValue(DII, SI, *DIB);
2324 eraseInstFromFunction(*I);
2327 if (InvokeInst *II = dyn_cast<InvokeInst>(&MI)) {
2328 // Replace invoke with a NOP intrinsic to maintain the original CFG
2329 Module *M = II->getModule();
2330 Function *F = Intrinsic::getDeclaration(M, Intrinsic::donothing);
2331 InvokeInst::Create(F, II->getNormalDest(), II->getUnwindDest(),
2332 None, "", II->getParent());
2335 for (auto *DII : DIIs)
2336 eraseInstFromFunction(*DII);
2338 return eraseInstFromFunction(MI);
2343 /// Move the call to free before a NULL test.
2345 /// Check if this free is accessed after its argument has been test
2346 /// against NULL (property 0).
2347 /// If yes, it is legal to move this call in its predecessor block.
2349 /// The move is performed only if the block containing the call to free
2350 /// will be removed, i.e.:
2351 /// 1. it has only one predecessor P, and P has two successors
2352 /// 2. it contains the call, noops, and an unconditional branch
2353 /// 3. its successor is the same as its predecessor's successor
2355 /// The profitability is out-of concern here and this function should
2356 /// be called only if the caller knows this transformation would be
2357 /// profitable (e.g., for code size).
2358 static Instruction *tryToMoveFreeBeforeNullTest(CallInst &FI,
2359 const DataLayout &DL) {
2360 Value *Op = FI.getArgOperand(0);
2361 BasicBlock *FreeInstrBB = FI.getParent();
2362 BasicBlock *PredBB = FreeInstrBB->getSinglePredecessor();
2364 // Validate part of constraint #1: Only one predecessor
2365 // FIXME: We can extend the number of predecessor, but in that case, we
2366 // would duplicate the call to free in each predecessor and it may
2367 // not be profitable even for code size.
2371 // Validate constraint #2: Does this block contains only the call to
2372 // free, noops, and an unconditional branch?
2374 Instruction *FreeInstrBBTerminator = FreeInstrBB->getTerminator();
2375 if (!match(FreeInstrBBTerminator, m_UnconditionalBr(SuccBB)))
2378 // If there are only 2 instructions in the block, at this point,
2379 // this is the call to free and unconditional.
2380 // If there are more than 2 instructions, check that they are noops
2381 // i.e., they won't hurt the performance of the generated code.
2382 if (FreeInstrBB->size() != 2) {
2383 for (const Instruction &Inst : *FreeInstrBB) {
2384 if (&Inst == &FI || &Inst == FreeInstrBBTerminator)
2386 auto *Cast = dyn_cast<CastInst>(&Inst);
2387 if (!Cast || !Cast->isNoopCast(DL))
2391 // Validate the rest of constraint #1 by matching on the pred branch.
2392 Instruction *TI = PredBB->getTerminator();
2393 BasicBlock *TrueBB, *FalseBB;
2394 ICmpInst::Predicate Pred;
2395 if (!match(TI, m_Br(m_ICmp(Pred,
2396 m_CombineOr(m_Specific(Op),
2397 m_Specific(Op->stripPointerCasts())),
2401 if (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE)
2404 // Validate constraint #3: Ensure the null case just falls through.
2405 if (SuccBB != (Pred == ICmpInst::ICMP_EQ ? TrueBB : FalseBB))
2407 assert(FreeInstrBB == (Pred == ICmpInst::ICMP_EQ ? FalseBB : TrueBB) &&
2408 "Broken CFG: missing edge from predecessor to successor");
2410 // At this point, we know that everything in FreeInstrBB can be moved
2412 for (BasicBlock::iterator It = FreeInstrBB->begin(), End = FreeInstrBB->end();
2414 Instruction &Instr = *It++;
2415 if (&Instr == FreeInstrBBTerminator)
2417 Instr.moveBefore(TI);
2419 assert(FreeInstrBB->size() == 1 &&
2420 "Only the branch instruction should remain");
2424 Instruction *InstCombiner::visitFree(CallInst &FI) {
2425 Value *Op = FI.getArgOperand(0);
2427 // free undef -> unreachable.
2428 if (isa<UndefValue>(Op)) {
2429 // Insert a new store to null because we cannot modify the CFG here.
2430 Builder.CreateStore(ConstantInt::getTrue(FI.getContext()),
2431 UndefValue::get(Type::getInt1PtrTy(FI.getContext())));
2432 return eraseInstFromFunction(FI);
2435 // If we have 'free null' delete the instruction. This can happen in stl code
2436 // when lots of inlining happens.
2437 if (isa<ConstantPointerNull>(Op))
2438 return eraseInstFromFunction(FI);
2440 // If we optimize for code size, try to move the call to free before the null
2441 // test so that simplify cfg can remove the empty block and dead code
2442 // elimination the branch. I.e., helps to turn something like:
2443 // if (foo) free(foo);
2447 if (Instruction *I = tryToMoveFreeBeforeNullTest(FI, DL))
2453 Instruction *InstCombiner::visitReturnInst(ReturnInst &RI) {
2454 if (RI.getNumOperands() == 0) // ret void
2457 Value *ResultOp = RI.getOperand(0);
2458 Type *VTy = ResultOp->getType();
2459 if (!VTy->isIntegerTy())
2462 // There might be assume intrinsics dominating this return that completely
2463 // determine the value. If so, constant fold it.
2464 KnownBits Known = computeKnownBits(ResultOp, 0, &RI);
2465 if (Known.isConstant())
2466 RI.setOperand(0, Constant::getIntegerValue(VTy, Known.getConstant()));
2471 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
2472 // Change br (not X), label True, label False to: br X, label False, True
2474 BasicBlock *TrueDest;
2475 BasicBlock *FalseDest;
2476 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
2477 !isa<Constant>(X)) {
2478 // Swap Destinations and condition...
2480 BI.swapSuccessors();
2484 // If the condition is irrelevant, remove the use so that other
2485 // transforms on the condition become more effective.
2486 if (BI.isConditional() && !isa<ConstantInt>(BI.getCondition()) &&
2487 BI.getSuccessor(0) == BI.getSuccessor(1)) {
2488 BI.setCondition(ConstantInt::getFalse(BI.getCondition()->getType()));
2492 // Canonicalize, for example, icmp_ne -> icmp_eq or fcmp_one -> fcmp_oeq.
2493 CmpInst::Predicate Pred;
2494 if (match(&BI, m_Br(m_OneUse(m_Cmp(Pred, m_Value(), m_Value())), TrueDest,
2496 !isCanonicalPredicate(Pred)) {
2497 // Swap destinations and condition.
2498 CmpInst *Cond = cast<CmpInst>(BI.getCondition());
2499 Cond->setPredicate(CmpInst::getInversePredicate(Pred));
2500 BI.swapSuccessors();
2508 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
2509 Value *Cond = SI.getCondition();
2511 ConstantInt *AddRHS;
2512 if (match(Cond, m_Add(m_Value(Op0), m_ConstantInt(AddRHS)))) {
2513 // Change 'switch (X+4) case 1:' into 'switch (X) case -3'.
2514 for (auto Case : SI.cases()) {
2515 Constant *NewCase = ConstantExpr::getSub(Case.getCaseValue(), AddRHS);
2516 assert(isa<ConstantInt>(NewCase) &&
2517 "Result of expression should be constant");
2518 Case.setValue(cast<ConstantInt>(NewCase));
2520 SI.setCondition(Op0);
2524 KnownBits Known = computeKnownBits(Cond, 0, &SI);
2525 unsigned LeadingKnownZeros = Known.countMinLeadingZeros();
2526 unsigned LeadingKnownOnes = Known.countMinLeadingOnes();
2528 // Compute the number of leading bits we can ignore.
2529 // TODO: A better way to determine this would use ComputeNumSignBits().
2530 for (auto &C : SI.cases()) {
2531 LeadingKnownZeros = std::min(
2532 LeadingKnownZeros, C.getCaseValue()->getValue().countLeadingZeros());
2533 LeadingKnownOnes = std::min(
2534 LeadingKnownOnes, C.getCaseValue()->getValue().countLeadingOnes());
2537 unsigned NewWidth = Known.getBitWidth() - std::max(LeadingKnownZeros, LeadingKnownOnes);
2539 // Shrink the condition operand if the new type is smaller than the old type.
2540 // But do not shrink to a non-standard type, because backend can't generate
2541 // good code for that yet.
2542 // TODO: We can make it aggressive again after fixing PR39569.
2543 if (NewWidth > 0 && NewWidth < Known.getBitWidth() &&
2544 shouldChangeType(Known.getBitWidth(), NewWidth)) {
2545 IntegerType *Ty = IntegerType::get(SI.getContext(), NewWidth);
2546 Builder.SetInsertPoint(&SI);
2547 Value *NewCond = Builder.CreateTrunc(Cond, Ty, "trunc");
2548 SI.setCondition(NewCond);
2550 for (auto Case : SI.cases()) {
2551 APInt TruncatedCase = Case.getCaseValue()->getValue().trunc(NewWidth);
2552 Case.setValue(ConstantInt::get(SI.getContext(), TruncatedCase));
2560 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
2561 Value *Agg = EV.getAggregateOperand();
2563 if (!EV.hasIndices())
2564 return replaceInstUsesWith(EV, Agg);
2566 if (Value *V = SimplifyExtractValueInst(Agg, EV.getIndices(),
2567 SQ.getWithInstruction(&EV)))
2568 return replaceInstUsesWith(EV, V);
2570 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
2571 // We're extracting from an insertvalue instruction, compare the indices
2572 const unsigned *exti, *exte, *insi, *inse;
2573 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
2574 exte = EV.idx_end(), inse = IV->idx_end();
2575 exti != exte && insi != inse;
2578 // The insert and extract both reference distinctly different elements.
2579 // This means the extract is not influenced by the insert, and we can
2580 // replace the aggregate operand of the extract with the aggregate
2581 // operand of the insert. i.e., replace
2582 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
2583 // %E = extractvalue { i32, { i32 } } %I, 0
2585 // %E = extractvalue { i32, { i32 } } %A, 0
2586 return ExtractValueInst::Create(IV->getAggregateOperand(),
2589 if (exti == exte && insi == inse)
2590 // Both iterators are at the end: Index lists are identical. Replace
2591 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
2592 // %C = extractvalue { i32, { i32 } } %B, 1, 0
2594 return replaceInstUsesWith(EV, IV->getInsertedValueOperand());
2596 // The extract list is a prefix of the insert list. i.e. replace
2597 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
2598 // %E = extractvalue { i32, { i32 } } %I, 1
2600 // %X = extractvalue { i32, { i32 } } %A, 1
2601 // %E = insertvalue { i32 } %X, i32 42, 0
2602 // by switching the order of the insert and extract (though the
2603 // insertvalue should be left in, since it may have other uses).
2604 Value *NewEV = Builder.CreateExtractValue(IV->getAggregateOperand(),
2606 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
2607 makeArrayRef(insi, inse));
2610 // The insert list is a prefix of the extract list
2611 // We can simply remove the common indices from the extract and make it
2612 // operate on the inserted value instead of the insertvalue result.
2614 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
2615 // %E = extractvalue { i32, { i32 } } %I, 1, 0
2617 // %E extractvalue { i32 } { i32 42 }, 0
2618 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
2619 makeArrayRef(exti, exte));
2621 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) {
2622 // We're extracting from an intrinsic, see if we're the only user, which
2623 // allows us to simplify multiple result intrinsics to simpler things that
2624 // just get one value.
2625 if (II->hasOneUse()) {
2626 // Check if we're grabbing the overflow bit or the result of a 'with
2627 // overflow' intrinsic. If it's the latter we can remove the intrinsic
2628 // and replace it with a traditional binary instruction.
2629 switch (II->getIntrinsicID()) {
2630 case Intrinsic::uadd_with_overflow:
2631 case Intrinsic::sadd_with_overflow:
2632 if (*EV.idx_begin() == 0) { // Normal result.
2633 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2634 replaceInstUsesWith(*II, UndefValue::get(II->getType()));
2635 eraseInstFromFunction(*II);
2636 return BinaryOperator::CreateAdd(LHS, RHS);
2639 // If the normal result of the add is dead, and the RHS is a constant,
2640 // we can transform this into a range comparison.
2641 // overflow = uadd a, -4 --> overflow = icmp ugt a, 3
2642 if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow)
2643 if (ConstantInt *CI = dyn_cast<ConstantInt>(II->getArgOperand(1)))
2644 return new ICmpInst(ICmpInst::ICMP_UGT, II->getArgOperand(0),
2645 ConstantExpr::getNot(CI));
2647 case Intrinsic::usub_with_overflow:
2648 case Intrinsic::ssub_with_overflow:
2649 if (*EV.idx_begin() == 0) { // Normal result.
2650 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2651 replaceInstUsesWith(*II, UndefValue::get(II->getType()));
2652 eraseInstFromFunction(*II);
2653 return BinaryOperator::CreateSub(LHS, RHS);
2656 case Intrinsic::umul_with_overflow:
2657 case Intrinsic::smul_with_overflow:
2658 if (*EV.idx_begin() == 0) { // Normal result.
2659 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2660 replaceInstUsesWith(*II, UndefValue::get(II->getType()));
2661 eraseInstFromFunction(*II);
2662 return BinaryOperator::CreateMul(LHS, RHS);
2670 if (LoadInst *L = dyn_cast<LoadInst>(Agg))
2671 // If the (non-volatile) load only has one use, we can rewrite this to a
2672 // load from a GEP. This reduces the size of the load. If a load is used
2673 // only by extractvalue instructions then this either must have been
2674 // optimized before, or it is a struct with padding, in which case we
2675 // don't want to do the transformation as it loses padding knowledge.
2676 if (L->isSimple() && L->hasOneUse()) {
2677 // extractvalue has integer indices, getelementptr has Value*s. Convert.
2678 SmallVector<Value*, 4> Indices;
2679 // Prefix an i32 0 since we need the first element.
2680 Indices.push_back(Builder.getInt32(0));
2681 for (ExtractValueInst::idx_iterator I = EV.idx_begin(), E = EV.idx_end();
2683 Indices.push_back(Builder.getInt32(*I));
2685 // We need to insert these at the location of the old load, not at that of
2686 // the extractvalue.
2687 Builder.SetInsertPoint(L);
2688 Value *GEP = Builder.CreateInBoundsGEP(L->getType(),
2689 L->getPointerOperand(), Indices);
2690 Instruction *NL = Builder.CreateLoad(GEP);
2691 // Whatever aliasing information we had for the orignal load must also
2692 // hold for the smaller load, so propagate the annotations.
2694 L->getAAMetadata(Nodes);
2695 NL->setAAMetadata(Nodes);
2696 // Returning the load directly will cause the main loop to insert it in
2697 // the wrong spot, so use replaceInstUsesWith().
2698 return replaceInstUsesWith(EV, NL);
2700 // We could simplify extracts from other values. Note that nested extracts may
2701 // already be simplified implicitly by the above: extract (extract (insert) )
2702 // will be translated into extract ( insert ( extract ) ) first and then just
2703 // the value inserted, if appropriate. Similarly for extracts from single-use
2704 // loads: extract (extract (load)) will be translated to extract (load (gep))
2705 // and if again single-use then via load (gep (gep)) to load (gep).
2706 // However, double extracts from e.g. function arguments or return values
2707 // aren't handled yet.
2711 /// Return 'true' if the given typeinfo will match anything.
2712 static bool isCatchAll(EHPersonality Personality, Constant *TypeInfo) {
2713 switch (Personality) {
2714 case EHPersonality::GNU_C:
2715 case EHPersonality::GNU_C_SjLj:
2716 case EHPersonality::Rust:
2717 // The GCC C EH and Rust personality only exists to support cleanups, so
2718 // it's not clear what the semantics of catch clauses are.
2720 case EHPersonality::Unknown:
2722 case EHPersonality::GNU_Ada:
2723 // While __gnat_all_others_value will match any Ada exception, it doesn't
2724 // match foreign exceptions (or didn't, before gcc-4.7).
2726 case EHPersonality::GNU_CXX:
2727 case EHPersonality::GNU_CXX_SjLj:
2728 case EHPersonality::GNU_ObjC:
2729 case EHPersonality::MSVC_X86SEH:
2730 case EHPersonality::MSVC_Win64SEH:
2731 case EHPersonality::MSVC_CXX:
2732 case EHPersonality::CoreCLR:
2733 case EHPersonality::Wasm_CXX:
2734 return TypeInfo->isNullValue();
2736 llvm_unreachable("invalid enum");
2739 static bool shorter_filter(const Value *LHS, const Value *RHS) {
2741 cast<ArrayType>(LHS->getType())->getNumElements()
2743 cast<ArrayType>(RHS->getType())->getNumElements();
2746 Instruction *InstCombiner::visitLandingPadInst(LandingPadInst &LI) {
2747 // The logic here should be correct for any real-world personality function.
2748 // However if that turns out not to be true, the offending logic can always
2749 // be conditioned on the personality function, like the catch-all logic is.
2750 EHPersonality Personality =
2751 classifyEHPersonality(LI.getParent()->getParent()->getPersonalityFn());
2753 // Simplify the list of clauses, eg by removing repeated catch clauses
2754 // (these are often created by inlining).
2755 bool MakeNewInstruction = false; // If true, recreate using the following:
2756 SmallVector<Constant *, 16> NewClauses; // - Clauses for the new instruction;
2757 bool CleanupFlag = LI.isCleanup(); // - The new instruction is a cleanup.
2759 SmallPtrSet<Value *, 16> AlreadyCaught; // Typeinfos known caught already.
2760 for (unsigned i = 0, e = LI.getNumClauses(); i != e; ++i) {
2761 bool isLastClause = i + 1 == e;
2762 if (LI.isCatch(i)) {
2764 Constant *CatchClause = LI.getClause(i);
2765 Constant *TypeInfo = CatchClause->stripPointerCasts();
2767 // If we already saw this clause, there is no point in having a second
2769 if (AlreadyCaught.insert(TypeInfo).second) {
2770 // This catch clause was not already seen.
2771 NewClauses.push_back(CatchClause);
2773 // Repeated catch clause - drop the redundant copy.
2774 MakeNewInstruction = true;
2777 // If this is a catch-all then there is no point in keeping any following
2778 // clauses or marking the landingpad as having a cleanup.
2779 if (isCatchAll(Personality, TypeInfo)) {
2781 MakeNewInstruction = true;
2782 CleanupFlag = false;
2786 // A filter clause. If any of the filter elements were already caught
2787 // then they can be dropped from the filter. It is tempting to try to
2788 // exploit the filter further by saying that any typeinfo that does not
2789 // occur in the filter can't be caught later (and thus can be dropped).
2790 // However this would be wrong, since typeinfos can match without being
2791 // equal (for example if one represents a C++ class, and the other some
2792 // class derived from it).
2793 assert(LI.isFilter(i) && "Unsupported landingpad clause!");
2794 Constant *FilterClause = LI.getClause(i);
2795 ArrayType *FilterType = cast<ArrayType>(FilterClause->getType());
2796 unsigned NumTypeInfos = FilterType->getNumElements();
2798 // An empty filter catches everything, so there is no point in keeping any
2799 // following clauses or marking the landingpad as having a cleanup. By
2800 // dealing with this case here the following code is made a bit simpler.
2801 if (!NumTypeInfos) {
2802 NewClauses.push_back(FilterClause);
2804 MakeNewInstruction = true;
2805 CleanupFlag = false;
2809 bool MakeNewFilter = false; // If true, make a new filter.
2810 SmallVector<Constant *, 16> NewFilterElts; // New elements.
2811 if (isa<ConstantAggregateZero>(FilterClause)) {
2812 // Not an empty filter - it contains at least one null typeinfo.
2813 assert(NumTypeInfos > 0 && "Should have handled empty filter already!");
2814 Constant *TypeInfo =
2815 Constant::getNullValue(FilterType->getElementType());
2816 // If this typeinfo is a catch-all then the filter can never match.
2817 if (isCatchAll(Personality, TypeInfo)) {
2818 // Throw the filter away.
2819 MakeNewInstruction = true;
2823 // There is no point in having multiple copies of this typeinfo, so
2824 // discard all but the first copy if there is more than one.
2825 NewFilterElts.push_back(TypeInfo);
2826 if (NumTypeInfos > 1)
2827 MakeNewFilter = true;
2829 ConstantArray *Filter = cast<ConstantArray>(FilterClause);
2830 SmallPtrSet<Value *, 16> SeenInFilter; // For uniquing the elements.
2831 NewFilterElts.reserve(NumTypeInfos);
2833 // Remove any filter elements that were already caught or that already
2834 // occurred in the filter. While there, see if any of the elements are
2835 // catch-alls. If so, the filter can be discarded.
2836 bool SawCatchAll = false;
2837 for (unsigned j = 0; j != NumTypeInfos; ++j) {
2838 Constant *Elt = Filter->getOperand(j);
2839 Constant *TypeInfo = Elt->stripPointerCasts();
2840 if (isCatchAll(Personality, TypeInfo)) {
2841 // This element is a catch-all. Bail out, noting this fact.
2846 // Even if we've seen a type in a catch clause, we don't want to
2847 // remove it from the filter. An unexpected type handler may be
2848 // set up for a call site which throws an exception of the same
2849 // type caught. In order for the exception thrown by the unexpected
2850 // handler to propagate correctly, the filter must be correctly
2851 // described for the call site.
2855 // void unexpected() { throw 1;}
2856 // void foo() throw (int) {
2857 // std::set_unexpected(unexpected);
2860 // } catch (int i) {}
2863 // There is no point in having multiple copies of the same typeinfo in
2864 // a filter, so only add it if we didn't already.
2865 if (SeenInFilter.insert(TypeInfo).second)
2866 NewFilterElts.push_back(cast<Constant>(Elt));
2868 // A filter containing a catch-all cannot match anything by definition.
2870 // Throw the filter away.
2871 MakeNewInstruction = true;
2875 // If we dropped something from the filter, make a new one.
2876 if (NewFilterElts.size() < NumTypeInfos)
2877 MakeNewFilter = true;
2879 if (MakeNewFilter) {
2880 FilterType = ArrayType::get(FilterType->getElementType(),
2881 NewFilterElts.size());
2882 FilterClause = ConstantArray::get(FilterType, NewFilterElts);
2883 MakeNewInstruction = true;
2886 NewClauses.push_back(FilterClause);
2888 // If the new filter is empty then it will catch everything so there is
2889 // no point in keeping any following clauses or marking the landingpad
2890 // as having a cleanup. The case of the original filter being empty was
2891 // already handled above.
2892 if (MakeNewFilter && !NewFilterElts.size()) {
2893 assert(MakeNewInstruction && "New filter but not a new instruction!");
2894 CleanupFlag = false;
2900 // If several filters occur in a row then reorder them so that the shortest
2901 // filters come first (those with the smallest number of elements). This is
2902 // advantageous because shorter filters are more likely to match, speeding up
2903 // unwinding, but mostly because it increases the effectiveness of the other
2904 // filter optimizations below.
2905 for (unsigned i = 0, e = NewClauses.size(); i + 1 < e; ) {
2907 // Find the maximal 'j' s.t. the range [i, j) consists entirely of filters.
2908 for (j = i; j != e; ++j)
2909 if (!isa<ArrayType>(NewClauses[j]->getType()))
2912 // Check whether the filters are already sorted by length. We need to know
2913 // if sorting them is actually going to do anything so that we only make a
2914 // new landingpad instruction if it does.
2915 for (unsigned k = i; k + 1 < j; ++k)
2916 if (shorter_filter(NewClauses[k+1], NewClauses[k])) {
2917 // Not sorted, so sort the filters now. Doing an unstable sort would be
2918 // correct too but reordering filters pointlessly might confuse users.
2919 std::stable_sort(NewClauses.begin() + i, NewClauses.begin() + j,
2921 MakeNewInstruction = true;
2925 // Look for the next batch of filters.
2929 // If typeinfos matched if and only if equal, then the elements of a filter L
2930 // that occurs later than a filter F could be replaced by the intersection of
2931 // the elements of F and L. In reality two typeinfos can match without being
2932 // equal (for example if one represents a C++ class, and the other some class
2933 // derived from it) so it would be wrong to perform this transform in general.
2934 // However the transform is correct and useful if F is a subset of L. In that
2935 // case L can be replaced by F, and thus removed altogether since repeating a
2936 // filter is pointless. So here we look at all pairs of filters F and L where
2937 // L follows F in the list of clauses, and remove L if every element of F is
2938 // an element of L. This can occur when inlining C++ functions with exception
2940 for (unsigned i = 0; i + 1 < NewClauses.size(); ++i) {
2941 // Examine each filter in turn.
2942 Value *Filter = NewClauses[i];
2943 ArrayType *FTy = dyn_cast<ArrayType>(Filter->getType());
2945 // Not a filter - skip it.
2947 unsigned FElts = FTy->getNumElements();
2948 // Examine each filter following this one. Doing this backwards means that
2949 // we don't have to worry about filters disappearing under us when removed.
2950 for (unsigned j = NewClauses.size() - 1; j != i; --j) {
2951 Value *LFilter = NewClauses[j];
2952 ArrayType *LTy = dyn_cast<ArrayType>(LFilter->getType());
2954 // Not a filter - skip it.
2956 // If Filter is a subset of LFilter, i.e. every element of Filter is also
2957 // an element of LFilter, then discard LFilter.
2958 SmallVectorImpl<Constant *>::iterator J = NewClauses.begin() + j;
2959 // If Filter is empty then it is a subset of LFilter.
2962 NewClauses.erase(J);
2963 MakeNewInstruction = true;
2964 // Move on to the next filter.
2967 unsigned LElts = LTy->getNumElements();
2968 // If Filter is longer than LFilter then it cannot be a subset of it.
2970 // Move on to the next filter.
2972 // At this point we know that LFilter has at least one element.
2973 if (isa<ConstantAggregateZero>(LFilter)) { // LFilter only contains zeros.
2974 // Filter is a subset of LFilter iff Filter contains only zeros (as we
2975 // already know that Filter is not longer than LFilter).
2976 if (isa<ConstantAggregateZero>(Filter)) {
2977 assert(FElts <= LElts && "Should have handled this case earlier!");
2979 NewClauses.erase(J);
2980 MakeNewInstruction = true;
2982 // Move on to the next filter.
2985 ConstantArray *LArray = cast<ConstantArray>(LFilter);
2986 if (isa<ConstantAggregateZero>(Filter)) { // Filter only contains zeros.
2987 // Since Filter is non-empty and contains only zeros, it is a subset of
2988 // LFilter iff LFilter contains a zero.
2989 assert(FElts > 0 && "Should have eliminated the empty filter earlier!");
2990 for (unsigned l = 0; l != LElts; ++l)
2991 if (LArray->getOperand(l)->isNullValue()) {
2992 // LFilter contains a zero - discard it.
2993 NewClauses.erase(J);
2994 MakeNewInstruction = true;
2997 // Move on to the next filter.
3000 // At this point we know that both filters are ConstantArrays. Loop over
3001 // operands to see whether every element of Filter is also an element of
3002 // LFilter. Since filters tend to be short this is probably faster than
3003 // using a method that scales nicely.
3004 ConstantArray *FArray = cast<ConstantArray>(Filter);
3005 bool AllFound = true;
3006 for (unsigned f = 0; f != FElts; ++f) {
3007 Value *FTypeInfo = FArray->getOperand(f)->stripPointerCasts();
3009 for (unsigned l = 0; l != LElts; ++l) {
3010 Value *LTypeInfo = LArray->getOperand(l)->stripPointerCasts();
3011 if (LTypeInfo == FTypeInfo) {
3021 NewClauses.erase(J);
3022 MakeNewInstruction = true;
3024 // Move on to the next filter.
3028 // If we changed any of the clauses, replace the old landingpad instruction
3030 if (MakeNewInstruction) {
3031 LandingPadInst *NLI = LandingPadInst::Create(LI.getType(),
3033 for (unsigned i = 0, e = NewClauses.size(); i != e; ++i)
3034 NLI->addClause(NewClauses[i]);
3035 // A landing pad with no clauses must have the cleanup flag set. It is
3036 // theoretically possible, though highly unlikely, that we eliminated all
3037 // clauses. If so, force the cleanup flag to true.
3038 if (NewClauses.empty())
3040 NLI->setCleanup(CleanupFlag);
3044 // Even if none of the clauses changed, we may nonetheless have understood
3045 // that the cleanup flag is pointless. Clear it if so.
3046 if (LI.isCleanup() != CleanupFlag) {
3047 assert(!CleanupFlag && "Adding a cleanup, not removing one?!");
3048 LI.setCleanup(CleanupFlag);
3055 /// Try to move the specified instruction from its current block into the
3056 /// beginning of DestBlock, which can only happen if it's safe to move the
3057 /// instruction past all of the instructions between it and the end of its
3059 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
3060 assert(I->hasOneUse() && "Invariants didn't hold!");
3061 BasicBlock *SrcBlock = I->getParent();
3063 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
3064 if (isa<PHINode>(I) || I->isEHPad() || I->mayHaveSideEffects() ||
3068 // Do not sink alloca instructions out of the entry block.
3069 if (isa<AllocaInst>(I) && I->getParent() ==
3070 &DestBlock->getParent()->getEntryBlock())
3073 // Do not sink into catchswitch blocks.
3074 if (isa<CatchSwitchInst>(DestBlock->getTerminator()))
3077 // Do not sink convergent call instructions.
3078 if (auto *CI = dyn_cast<CallInst>(I)) {
3079 if (CI->isConvergent())
3082 // We can only sink load instructions if there is nothing between the load and
3083 // the end of block that could change the value.
3084 if (I->mayReadFromMemory()) {
3085 for (BasicBlock::iterator Scan = I->getIterator(),
3086 E = I->getParent()->end();
3088 if (Scan->mayWriteToMemory())
3091 BasicBlock::iterator InsertPos = DestBlock->getFirstInsertionPt();
3092 I->moveBefore(&*InsertPos);
3095 // Also sink all related debug uses from the source basic block. Otherwise we
3096 // get debug use before the def.
3097 SmallVector<DbgVariableIntrinsic *, 1> DbgUsers;
3098 findDbgUsers(DbgUsers, I);
3099 for (auto *DII : DbgUsers) {
3100 if (DII->getParent() == SrcBlock) {
3101 DII->moveBefore(&*InsertPos);
3102 LLVM_DEBUG(dbgs() << "SINK: " << *DII << '\n');
3108 bool InstCombiner::run() {
3109 while (!Worklist.isEmpty()) {
3110 Instruction *I = Worklist.RemoveOne();
3111 if (I == nullptr) continue; // skip null values.
3113 // Check to see if we can DCE the instruction.
3114 if (isInstructionTriviallyDead(I, &TLI)) {
3115 LLVM_DEBUG(dbgs() << "IC: DCE: " << *I << '\n');
3116 eraseInstFromFunction(*I);
3118 MadeIRChange = true;
3122 if (!DebugCounter::shouldExecute(VisitCounter))
3125 // Instruction isn't dead, see if we can constant propagate it.
3126 if (!I->use_empty() &&
3127 (I->getNumOperands() == 0 || isa<Constant>(I->getOperand(0)))) {
3128 if (Constant *C = ConstantFoldInstruction(I, DL, &TLI)) {
3129 LLVM_DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " << *I
3132 // Add operands to the worklist.
3133 replaceInstUsesWith(*I, C);
3135 if (isInstructionTriviallyDead(I, &TLI))
3136 eraseInstFromFunction(*I);
3137 MadeIRChange = true;
3142 // In general, it is possible for computeKnownBits to determine all bits in
3143 // a value even when the operands are not all constants.
3144 Type *Ty = I->getType();
3145 if (ExpensiveCombines && !I->use_empty() && Ty->isIntOrIntVectorTy()) {
3146 KnownBits Known = computeKnownBits(I, /*Depth*/0, I);
3147 if (Known.isConstant()) {
3148 Constant *C = ConstantInt::get(Ty, Known.getConstant());
3149 LLVM_DEBUG(dbgs() << "IC: ConstFold (all bits known) to: " << *C
3150 << " from: " << *I << '\n');
3152 // Add operands to the worklist.
3153 replaceInstUsesWith(*I, C);
3155 if (isInstructionTriviallyDead(I, &TLI))
3156 eraseInstFromFunction(*I);
3157 MadeIRChange = true;
3162 // See if we can trivially sink this instruction to a successor basic block.
3163 if (EnableCodeSinking && I->hasOneUse()) {
3164 BasicBlock *BB = I->getParent();
3165 Instruction *UserInst = cast<Instruction>(*I->user_begin());
3166 BasicBlock *UserParent;
3168 // Get the block the use occurs in.
3169 if (PHINode *PN = dyn_cast<PHINode>(UserInst))
3170 UserParent = PN->getIncomingBlock(*I->use_begin());
3172 UserParent = UserInst->getParent();
3174 if (UserParent != BB) {
3175 bool UserIsSuccessor = false;
3176 // See if the user is one of our successors.
3177 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
3178 if (*SI == UserParent) {
3179 UserIsSuccessor = true;
3183 // If the user is one of our immediate successors, and if that successor
3184 // only has us as a predecessors (we'd have to split the critical edge
3185 // otherwise), we can keep going.
3186 if (UserIsSuccessor && UserParent->getUniquePredecessor()) {
3187 // Okay, the CFG is simple enough, try to sink this instruction.
3188 if (TryToSinkInstruction(I, UserParent)) {
3189 LLVM_DEBUG(dbgs() << "IC: Sink: " << *I << '\n');
3190 MadeIRChange = true;
3191 // We'll add uses of the sunk instruction below, but since sinking
3192 // can expose opportunities for it's *operands* add them to the
3194 for (Use &U : I->operands())
3195 if (Instruction *OpI = dyn_cast<Instruction>(U.get()))
3202 // Now that we have an instruction, try combining it to simplify it.
3203 Builder.SetInsertPoint(I);
3204 Builder.SetCurrentDebugLocation(I->getDebugLoc());
3209 LLVM_DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str(););
3210 LLVM_DEBUG(dbgs() << "IC: Visiting: " << OrigI << '\n');
3212 if (Instruction *Result = visit(*I)) {
3214 // Should we replace the old instruction with a new one?
3216 LLVM_DEBUG(dbgs() << "IC: Old = " << *I << '\n'
3217 << " New = " << *Result << '\n');
3219 if (I->getDebugLoc())
3220 Result->setDebugLoc(I->getDebugLoc());
3221 // Everything uses the new instruction now.
3222 I->replaceAllUsesWith(Result);
3224 // Move the name to the new instruction first.
3225 Result->takeName(I);
3227 // Push the new instruction and any users onto the worklist.
3228 Worklist.AddUsersToWorkList(*Result);
3229 Worklist.Add(Result);
3231 // Insert the new instruction into the basic block...
3232 BasicBlock *InstParent = I->getParent();
3233 BasicBlock::iterator InsertPos = I->getIterator();
3235 // If we replace a PHI with something that isn't a PHI, fix up the
3237 if (!isa<PHINode>(Result) && isa<PHINode>(InsertPos))
3238 InsertPos = InstParent->getFirstInsertionPt();
3240 InstParent->getInstList().insert(InsertPos, Result);
3242 eraseInstFromFunction(*I);
3244 LLVM_DEBUG(dbgs() << "IC: Mod = " << OrigI << '\n'
3245 << " New = " << *I << '\n');
3247 // If the instruction was modified, it's possible that it is now dead.
3248 // if so, remove it.
3249 if (isInstructionTriviallyDead(I, &TLI)) {
3250 eraseInstFromFunction(*I);
3252 Worklist.AddUsersToWorkList(*I);
3256 MadeIRChange = true;
3261 return MadeIRChange;
3264 /// Walk the function in depth-first order, adding all reachable code to the
3267 /// This has a couple of tricks to make the code faster and more powerful. In
3268 /// particular, we constant fold and DCE instructions as we go, to avoid adding
3269 /// them to the worklist (this significantly speeds up instcombine on code where
3270 /// many instructions are dead or constant). Additionally, if we find a branch
3271 /// whose condition is a known constant, we only visit the reachable successors.
3272 static bool AddReachableCodeToWorklist(BasicBlock *BB, const DataLayout &DL,
3273 SmallPtrSetImpl<BasicBlock *> &Visited,
3274 InstCombineWorklist &ICWorklist,
3275 const TargetLibraryInfo *TLI) {
3276 bool MadeIRChange = false;
3277 SmallVector<BasicBlock*, 256> Worklist;
3278 Worklist.push_back(BB);
3280 SmallVector<Instruction*, 128> InstrsForInstCombineWorklist;
3281 DenseMap<Constant *, Constant *> FoldedConstants;
3284 BB = Worklist.pop_back_val();
3286 // We have now visited this block! If we've already been here, ignore it.
3287 if (!Visited.insert(BB).second)
3290 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
3291 Instruction *Inst = &*BBI++;
3293 // DCE instruction if trivially dead.
3294 if (isInstructionTriviallyDead(Inst, TLI)) {
3296 LLVM_DEBUG(dbgs() << "IC: DCE: " << *Inst << '\n');
3297 salvageDebugInfo(*Inst);
3298 Inst->eraseFromParent();
3299 MadeIRChange = true;
3303 // ConstantProp instruction if trivially constant.
3304 if (!Inst->use_empty() &&
3305 (Inst->getNumOperands() == 0 || isa<Constant>(Inst->getOperand(0))))
3306 if (Constant *C = ConstantFoldInstruction(Inst, DL, TLI)) {
3307 LLVM_DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " << *Inst
3309 Inst->replaceAllUsesWith(C);
3311 if (isInstructionTriviallyDead(Inst, TLI))
3312 Inst->eraseFromParent();
3313 MadeIRChange = true;
3317 // See if we can constant fold its operands.
3318 for (Use &U : Inst->operands()) {
3319 if (!isa<ConstantVector>(U) && !isa<ConstantExpr>(U))
3322 auto *C = cast<Constant>(U);
3323 Constant *&FoldRes = FoldedConstants[C];
3325 FoldRes = ConstantFoldConstant(C, DL, TLI);
3330 LLVM_DEBUG(dbgs() << "IC: ConstFold operand of: " << *Inst
3331 << "\n Old = " << *C
3332 << "\n New = " << *FoldRes << '\n');
3334 MadeIRChange = true;
3338 // Skip processing debug intrinsics in InstCombine. Processing these call instructions
3339 // consumes non-trivial amount of time and provides no value for the optimization.
3340 if (!isa<DbgInfoIntrinsic>(Inst))
3341 InstrsForInstCombineWorklist.push_back(Inst);
3344 // Recursively visit successors. If this is a branch or switch on a
3345 // constant, only visit the reachable successor.
3346 Instruction *TI = BB->getTerminator();
3347 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
3348 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
3349 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
3350 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
3351 Worklist.push_back(ReachableBB);
3354 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
3355 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
3356 Worklist.push_back(SI->findCaseValue(Cond)->getCaseSuccessor());
3361 for (BasicBlock *SuccBB : successors(TI))
3362 Worklist.push_back(SuccBB);
3363 } while (!Worklist.empty());
3365 // Once we've found all of the instructions to add to instcombine's worklist,
3366 // add them in reverse order. This way instcombine will visit from the top
3367 // of the function down. This jives well with the way that it adds all uses
3368 // of instructions to the worklist after doing a transformation, thus avoiding
3369 // some N^2 behavior in pathological cases.
3370 ICWorklist.AddInitialGroup(InstrsForInstCombineWorklist);
3372 return MadeIRChange;
3375 /// Populate the IC worklist from a function, and prune any dead basic
3376 /// blocks discovered in the process.
3378 /// This also does basic constant propagation and other forward fixing to make
3379 /// the combiner itself run much faster.
3380 static bool prepareICWorklistFromFunction(Function &F, const DataLayout &DL,
3381 TargetLibraryInfo *TLI,
3382 InstCombineWorklist &ICWorklist) {
3383 bool MadeIRChange = false;
3385 // Do a depth-first traversal of the function, populate the worklist with
3386 // the reachable instructions. Ignore blocks that are not reachable. Keep
3387 // track of which blocks we visit.
3388 SmallPtrSet<BasicBlock *, 32> Visited;
3390 AddReachableCodeToWorklist(&F.front(), DL, Visited, ICWorklist, TLI);
3392 // Do a quick scan over the function. If we find any blocks that are
3393 // unreachable, remove any instructions inside of them. This prevents
3394 // the instcombine code from having to deal with some bad special cases.
3395 for (BasicBlock &BB : F) {
3396 if (Visited.count(&BB))
3399 unsigned NumDeadInstInBB = removeAllNonTerminatorAndEHPadInstructions(&BB);
3400 MadeIRChange |= NumDeadInstInBB > 0;
3401 NumDeadInst += NumDeadInstInBB;
3404 return MadeIRChange;
3407 static bool combineInstructionsOverFunction(
3408 Function &F, InstCombineWorklist &Worklist, AliasAnalysis *AA,
3409 AssumptionCache &AC, TargetLibraryInfo &TLI, DominatorTree &DT,
3410 OptimizationRemarkEmitter &ORE, bool ExpensiveCombines = true,
3411 LoopInfo *LI = nullptr) {
3412 auto &DL = F.getParent()->getDataLayout();
3413 ExpensiveCombines |= EnableExpensiveCombines;
3415 /// Builder - This is an IRBuilder that automatically inserts new
3416 /// instructions into the worklist when they are created.
3417 IRBuilder<TargetFolder, IRBuilderCallbackInserter> Builder(
3418 F.getContext(), TargetFolder(DL),
3419 IRBuilderCallbackInserter([&Worklist, &AC](Instruction *I) {
3421 if (match(I, m_Intrinsic<Intrinsic::assume>()))
3422 AC.registerAssumption(cast<CallInst>(I));
3425 // Lower dbg.declare intrinsics otherwise their value may be clobbered
3427 bool MadeIRChange = false;
3428 if (ShouldLowerDbgDeclare)
3429 MadeIRChange = LowerDbgDeclare(F);
3431 // Iterate while there is work to do.
3435 LLVM_DEBUG(dbgs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
3436 << F.getName() << "\n");
3438 MadeIRChange |= prepareICWorklistFromFunction(F, DL, &TLI, Worklist);
3440 InstCombiner IC(Worklist, Builder, F.optForMinSize(), ExpensiveCombines, AA,
3441 AC, TLI, DT, ORE, DL, LI);
3442 IC.MaxArraySizeForCombine = MaxArraySize;
3448 return MadeIRChange || Iteration > 1;
3451 PreservedAnalyses InstCombinePass::run(Function &F,
3452 FunctionAnalysisManager &AM) {
3453 auto &AC = AM.getResult<AssumptionAnalysis>(F);
3454 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
3455 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
3456 auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F);
3458 auto *LI = AM.getCachedResult<LoopAnalysis>(F);
3460 auto *AA = &AM.getResult<AAManager>(F);
3461 if (!combineInstructionsOverFunction(F, Worklist, AA, AC, TLI, DT, ORE,
3462 ExpensiveCombines, LI))
3463 // No changes, all analyses are preserved.
3464 return PreservedAnalyses::all();
3466 // Mark all the analyses that instcombine updates as preserved.
3467 PreservedAnalyses PA;
3468 PA.preserveSet<CFGAnalyses>();
3469 PA.preserve<AAManager>();
3470 PA.preserve<BasicAA>();
3471 PA.preserve<GlobalsAA>();
3475 void InstructionCombiningPass::getAnalysisUsage(AnalysisUsage &AU) const {
3476 AU.setPreservesCFG();
3477 AU.addRequired<AAResultsWrapperPass>();
3478 AU.addRequired<AssumptionCacheTracker>();
3479 AU.addRequired<TargetLibraryInfoWrapperPass>();
3480 AU.addRequired<DominatorTreeWrapperPass>();
3481 AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
3482 AU.addPreserved<DominatorTreeWrapperPass>();
3483 AU.addPreserved<AAResultsWrapperPass>();
3484 AU.addPreserved<BasicAAWrapperPass>();
3485 AU.addPreserved<GlobalsAAWrapperPass>();
3488 bool InstructionCombiningPass::runOnFunction(Function &F) {
3489 if (skipFunction(F))
3492 // Required analyses.
3493 auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
3494 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
3495 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
3496 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
3497 auto &ORE = getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
3499 // Optional analyses.
3500 auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>();
3501 auto *LI = LIWP ? &LIWP->getLoopInfo() : nullptr;
3503 return combineInstructionsOverFunction(F, Worklist, AA, AC, TLI, DT, ORE,
3504 ExpensiveCombines, LI);
3507 char InstructionCombiningPass::ID = 0;
3509 INITIALIZE_PASS_BEGIN(InstructionCombiningPass, "instcombine",
3510 "Combine redundant instructions", false, false)
3511 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
3512 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
3513 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
3514 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
3515 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
3516 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)
3517 INITIALIZE_PASS_END(InstructionCombiningPass, "instcombine",
3518 "Combine redundant instructions", false, false)
3520 // Initialization Routines
3521 void llvm::initializeInstCombine(PassRegistry &Registry) {
3522 initializeInstructionCombiningPassPass(Registry);
3525 void LLVMInitializeInstCombine(LLVMPassRegistryRef R) {
3526 initializeInstructionCombiningPassPass(*unwrap(R));
3529 FunctionPass *llvm::createInstructionCombiningPass(bool ExpensiveCombines) {
3530 return new InstructionCombiningPass(ExpensiveCombines);
3533 void LLVMAddInstructionCombiningPass(LLVMPassManagerRef PM) {
3534 unwrap(PM)->add(createInstructionCombiningPass());