1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 // InstructionCombining - Combine instructions to form fewer, simple
10 // instructions. This pass does not modify the CFG. This pass is where
11 // algebraic simplification happens.
13 // This pass combines things like:
19 // This is a simple worklist driven algorithm.
21 // This pass guarantees that the following canonicalizations are performed on
23 // 1. If a binary operator has a constant operand, it is moved to the RHS
24 // 2. Bitwise operators with constant operands are always grouped so that
25 // shifts are performed first, then or's, then and's, then xor's.
26 // 3. Compare instructions are converted from <,>,<=,>= to ==,!= if possible
27 // 4. All cmp instructions on boolean values are replaced with logical ops
28 // 5. add X, X is represented as (X*2) => (X << 1)
29 // 6. Multiplies with a power-of-two constant argument are transformed into
33 //===----------------------------------------------------------------------===//
35 #include "InstCombineInternal.h"
36 #include "llvm-c/Initialization.h"
37 #include "llvm-c/Transforms/InstCombine.h"
38 #include "llvm/ADT/APInt.h"
39 #include "llvm/ADT/ArrayRef.h"
40 #include "llvm/ADT/DenseMap.h"
41 #include "llvm/ADT/None.h"
42 #include "llvm/ADT/SmallPtrSet.h"
43 #include "llvm/ADT/SmallVector.h"
44 #include "llvm/ADT/Statistic.h"
45 #include "llvm/ADT/TinyPtrVector.h"
46 #include "llvm/Analysis/AliasAnalysis.h"
47 #include "llvm/Analysis/AssumptionCache.h"
48 #include "llvm/Analysis/BasicAliasAnalysis.h"
49 #include "llvm/Analysis/BlockFrequencyInfo.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/LazyBlockFrequencyInfo.h"
56 #include "llvm/Analysis/LoopInfo.h"
57 #include "llvm/Analysis/MemoryBuiltins.h"
58 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
59 #include "llvm/Analysis/ProfileSummaryInfo.h"
60 #include "llvm/Analysis/TargetFolder.h"
61 #include "llvm/Analysis/TargetLibraryInfo.h"
62 #include "llvm/Analysis/ValueTracking.h"
63 #include "llvm/Analysis/VectorUtils.h"
64 #include "llvm/IR/BasicBlock.h"
65 #include "llvm/IR/CFG.h"
66 #include "llvm/IR/Constant.h"
67 #include "llvm/IR/Constants.h"
68 #include "llvm/IR/DIBuilder.h"
69 #include "llvm/IR/DataLayout.h"
70 #include "llvm/IR/DerivedTypes.h"
71 #include "llvm/IR/Dominators.h"
72 #include "llvm/IR/Function.h"
73 #include "llvm/IR/GetElementPtrTypeIterator.h"
74 #include "llvm/IR/IRBuilder.h"
75 #include "llvm/IR/InstrTypes.h"
76 #include "llvm/IR/Instruction.h"
77 #include "llvm/IR/Instructions.h"
78 #include "llvm/IR/IntrinsicInst.h"
79 #include "llvm/IR/Intrinsics.h"
80 #include "llvm/IR/LegacyPassManager.h"
81 #include "llvm/IR/Metadata.h"
82 #include "llvm/IR/Operator.h"
83 #include "llvm/IR/PassManager.h"
84 #include "llvm/IR/PatternMatch.h"
85 #include "llvm/IR/Type.h"
86 #include "llvm/IR/Use.h"
87 #include "llvm/IR/User.h"
88 #include "llvm/IR/Value.h"
89 #include "llvm/IR/ValueHandle.h"
90 #include "llvm/InitializePasses.h"
91 #include "llvm/Pass.h"
92 #include "llvm/Support/CBindingWrapping.h"
93 #include "llvm/Support/Casting.h"
94 #include "llvm/Support/CommandLine.h"
95 #include "llvm/Support/Compiler.h"
96 #include "llvm/Support/Debug.h"
97 #include "llvm/Support/DebugCounter.h"
98 #include "llvm/Support/ErrorHandling.h"
99 #include "llvm/Support/KnownBits.h"
100 #include "llvm/Support/raw_ostream.h"
101 #include "llvm/Transforms/InstCombine/InstCombine.h"
102 #include "llvm/Transforms/InstCombine/InstCombineWorklist.h"
103 #include "llvm/Transforms/Utils/Local.h"
111 using namespace llvm;
112 using namespace llvm::PatternMatch;
114 #define DEBUG_TYPE "instcombine"
116 STATISTIC(NumCombined , "Number of insts combined");
117 STATISTIC(NumConstProp, "Number of constant folds");
118 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
119 STATISTIC(NumSunkInst , "Number of instructions sunk");
120 STATISTIC(NumExpand, "Number of expansions");
121 STATISTIC(NumFactor , "Number of factorizations");
122 STATISTIC(NumReassoc , "Number of reassociations");
123 DEBUG_COUNTER(VisitCounter, "instcombine-visit",
124 "Controls which instructions are visited");
126 static constexpr unsigned InstCombineDefaultMaxIterations = 1000;
127 static constexpr unsigned InstCombineDefaultInfiniteLoopThreshold = 1000;
130 EnableCodeSinking("instcombine-code-sinking", cl::desc("Enable code sinking"),
133 static cl::opt<unsigned> LimitMaxIterations(
134 "instcombine-max-iterations",
135 cl::desc("Limit the maximum number of instruction combining iterations"),
136 cl::init(InstCombineDefaultMaxIterations));
138 static cl::opt<unsigned> InfiniteLoopDetectionThreshold(
139 "instcombine-infinite-loop-threshold",
140 cl::desc("Number of instruction combining iterations considered an "
142 cl::init(InstCombineDefaultInfiniteLoopThreshold), cl::Hidden);
144 static cl::opt<unsigned>
145 MaxArraySize("instcombine-maxarray-size", cl::init(1024),
146 cl::desc("Maximum array size considered when doing a combine"));
148 // FIXME: Remove this flag when it is no longer necessary to convert
149 // llvm.dbg.declare to avoid inaccurate debug info. Setting this to false
150 // increases variable availability at the cost of accuracy. Variables that
151 // cannot be promoted by mem2reg or SROA will be described as living in memory
152 // for their entire lifetime. However, passes like DSE and instcombine can
153 // delete stores to the alloca, leading to misleading and inaccurate debug
154 // information. This flag can be removed when those passes are fixed.
155 static cl::opt<unsigned> ShouldLowerDbgDeclare("instcombine-lower-dbg-declare",
156 cl::Hidden, cl::init(true));
158 Value *InstCombiner::EmitGEPOffset(User *GEP) {
159 return llvm::EmitGEPOffset(&Builder, DL, GEP);
162 /// Return true if it is desirable to convert an integer computation from a
163 /// given bit width to a new bit width.
164 /// We don't want to convert from a legal to an illegal type or from a smaller
165 /// to a larger illegal type. A width of '1' is always treated as a legal type
166 /// because i1 is a fundamental type in IR, and there are many specialized
167 /// optimizations for i1 types. Widths of 8, 16 or 32 are equally treated as
168 /// legal to convert to, in order to open up more combining opportunities.
169 /// NOTE: this treats i8, i16 and i32 specially, due to them being so common
170 /// from frontend languages.
171 bool InstCombiner::shouldChangeType(unsigned FromWidth,
172 unsigned ToWidth) const {
173 bool FromLegal = FromWidth == 1 || DL.isLegalInteger(FromWidth);
174 bool ToLegal = ToWidth == 1 || DL.isLegalInteger(ToWidth);
176 // Convert to widths of 8, 16 or 32 even if they are not legal types. Only
177 // shrink types, to prevent infinite loops.
178 if (ToWidth < FromWidth && (ToWidth == 8 || ToWidth == 16 || ToWidth == 32))
181 // If this is a legal integer from type, and the result would be an illegal
182 // type, don't do the transformation.
183 if (FromLegal && !ToLegal)
186 // Otherwise, if both are illegal, do not increase the size of the result. We
187 // do allow things like i160 -> i64, but not i64 -> i160.
188 if (!FromLegal && !ToLegal && ToWidth > FromWidth)
194 /// Return true if it is desirable to convert a computation from 'From' to 'To'.
195 /// We don't want to convert from a legal to an illegal type or from a smaller
196 /// to a larger illegal type. i1 is always treated as a legal type because it is
197 /// a fundamental type in IR, and there are many specialized optimizations for
199 bool InstCombiner::shouldChangeType(Type *From, Type *To) const {
200 // TODO: This could be extended to allow vectors. Datalayout changes might be
201 // needed to properly support that.
202 if (!From->isIntegerTy() || !To->isIntegerTy())
205 unsigned FromWidth = From->getPrimitiveSizeInBits();
206 unsigned ToWidth = To->getPrimitiveSizeInBits();
207 return shouldChangeType(FromWidth, ToWidth);
210 // Return true, if No Signed Wrap should be maintained for I.
211 // The No Signed Wrap flag can be kept if the operation "B (I.getOpcode) C",
212 // where both B and C should be ConstantInts, results in a constant that does
213 // not overflow. This function only handles the Add and Sub opcodes. For
214 // all other opcodes, the function conservatively returns false.
215 static bool maintainNoSignedWrap(BinaryOperator &I, Value *B, Value *C) {
216 auto *OBO = dyn_cast<OverflowingBinaryOperator>(&I);
217 if (!OBO || !OBO->hasNoSignedWrap())
220 // We reason about Add and Sub Only.
221 Instruction::BinaryOps Opcode = I.getOpcode();
222 if (Opcode != Instruction::Add && Opcode != Instruction::Sub)
225 const APInt *BVal, *CVal;
226 if (!match(B, m_APInt(BVal)) || !match(C, m_APInt(CVal)))
229 bool Overflow = false;
230 if (Opcode == Instruction::Add)
231 (void)BVal->sadd_ov(*CVal, Overflow);
233 (void)BVal->ssub_ov(*CVal, Overflow);
238 static bool hasNoUnsignedWrap(BinaryOperator &I) {
239 auto *OBO = dyn_cast<OverflowingBinaryOperator>(&I);
240 return OBO && OBO->hasNoUnsignedWrap();
243 static bool hasNoSignedWrap(BinaryOperator &I) {
244 auto *OBO = dyn_cast<OverflowingBinaryOperator>(&I);
245 return OBO && OBO->hasNoSignedWrap();
248 /// Conservatively clears subclassOptionalData after a reassociation or
249 /// commutation. We preserve fast-math flags when applicable as they can be
251 static void ClearSubclassDataAfterReassociation(BinaryOperator &I) {
252 FPMathOperator *FPMO = dyn_cast<FPMathOperator>(&I);
254 I.clearSubclassOptionalData();
258 FastMathFlags FMF = I.getFastMathFlags();
259 I.clearSubclassOptionalData();
260 I.setFastMathFlags(FMF);
263 /// Combine constant operands of associative operations either before or after a
264 /// cast to eliminate one of the associative operations:
265 /// (op (cast (op X, C2)), C1) --> (cast (op X, op (C1, C2)))
266 /// (op (cast (op X, C2)), C1) --> (op (cast X), op (C1, C2))
267 static bool simplifyAssocCastAssoc(BinaryOperator *BinOp1, InstCombiner &IC) {
268 auto *Cast = dyn_cast<CastInst>(BinOp1->getOperand(0));
269 if (!Cast || !Cast->hasOneUse())
272 // TODO: Enhance logic for other casts and remove this check.
273 auto CastOpcode = Cast->getOpcode();
274 if (CastOpcode != Instruction::ZExt)
277 // TODO: Enhance logic for other BinOps and remove this check.
278 if (!BinOp1->isBitwiseLogicOp())
281 auto AssocOpcode = BinOp1->getOpcode();
282 auto *BinOp2 = dyn_cast<BinaryOperator>(Cast->getOperand(0));
283 if (!BinOp2 || !BinOp2->hasOneUse() || BinOp2->getOpcode() != AssocOpcode)
287 if (!match(BinOp1->getOperand(1), m_Constant(C1)) ||
288 !match(BinOp2->getOperand(1), m_Constant(C2)))
291 // TODO: This assumes a zext cast.
292 // Eg, if it was a trunc, we'd cast C1 to the source type because casting C2
293 // to the destination type might lose bits.
295 // Fold the constants together in the destination type:
296 // (op (cast (op X, C2)), C1) --> (op (cast X), FoldedC)
297 Type *DestTy = C1->getType();
298 Constant *CastC2 = ConstantExpr::getCast(CastOpcode, C2, DestTy);
299 Constant *FoldedC = ConstantExpr::get(AssocOpcode, C1, CastC2);
300 IC.replaceOperand(*Cast, 0, BinOp2->getOperand(0));
301 IC.replaceOperand(*BinOp1, 1, FoldedC);
305 /// This performs a few simplifications for operators that are associative or
308 /// Commutative operators:
310 /// 1. Order operands such that they are listed from right (least complex) to
311 /// left (most complex). This puts constants before unary operators before
312 /// binary operators.
314 /// Associative operators:
316 /// 2. Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
317 /// 3. Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
319 /// Associative and commutative operators:
321 /// 4. Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
322 /// 5. Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
323 /// 6. Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
324 /// if C1 and C2 are constants.
325 bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) {
326 Instruction::BinaryOps Opcode = I.getOpcode();
327 bool Changed = false;
330 // Order operands such that they are listed from right (least complex) to
331 // left (most complex). This puts constants before unary operators before
333 if (I.isCommutative() && getComplexity(I.getOperand(0)) <
334 getComplexity(I.getOperand(1)))
335 Changed = !I.swapOperands();
337 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
338 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
340 if (I.isAssociative()) {
341 // Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
342 if (Op0 && Op0->getOpcode() == Opcode) {
343 Value *A = Op0->getOperand(0);
344 Value *B = Op0->getOperand(1);
345 Value *C = I.getOperand(1);
347 // Does "B op C" simplify?
348 if (Value *V = SimplifyBinOp(Opcode, B, C, SQ.getWithInstruction(&I))) {
349 // It simplifies to V. Form "A op V".
350 replaceOperand(I, 0, A);
351 replaceOperand(I, 1, V);
352 bool IsNUW = hasNoUnsignedWrap(I) && hasNoUnsignedWrap(*Op0);
353 bool IsNSW = maintainNoSignedWrap(I, B, C) && hasNoSignedWrap(*Op0);
355 // Conservatively clear all optional flags since they may not be
356 // preserved by the reassociation. Reset nsw/nuw based on the above
358 ClearSubclassDataAfterReassociation(I);
360 // Note: this is only valid because SimplifyBinOp doesn't look at
361 // the operands to Op0.
363 I.setHasNoUnsignedWrap(true);
366 I.setHasNoSignedWrap(true);
374 // Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
375 if (Op1 && Op1->getOpcode() == Opcode) {
376 Value *A = I.getOperand(0);
377 Value *B = Op1->getOperand(0);
378 Value *C = Op1->getOperand(1);
380 // Does "A op B" simplify?
381 if (Value *V = SimplifyBinOp(Opcode, A, B, SQ.getWithInstruction(&I))) {
382 // It simplifies to V. Form "V op C".
383 replaceOperand(I, 0, V);
384 replaceOperand(I, 1, C);
385 // Conservatively clear the optional flags, since they may not be
386 // preserved by the reassociation.
387 ClearSubclassDataAfterReassociation(I);
395 if (I.isAssociative() && I.isCommutative()) {
396 if (simplifyAssocCastAssoc(&I, *this)) {
402 // Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
403 if (Op0 && Op0->getOpcode() == Opcode) {
404 Value *A = Op0->getOperand(0);
405 Value *B = Op0->getOperand(1);
406 Value *C = I.getOperand(1);
408 // Does "C op A" simplify?
409 if (Value *V = SimplifyBinOp(Opcode, C, A, SQ.getWithInstruction(&I))) {
410 // It simplifies to V. Form "V op B".
411 replaceOperand(I, 0, V);
412 replaceOperand(I, 1, B);
413 // Conservatively clear the optional flags, since they may not be
414 // preserved by the reassociation.
415 ClearSubclassDataAfterReassociation(I);
422 // Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
423 if (Op1 && Op1->getOpcode() == Opcode) {
424 Value *A = I.getOperand(0);
425 Value *B = Op1->getOperand(0);
426 Value *C = Op1->getOperand(1);
428 // Does "C op A" simplify?
429 if (Value *V = SimplifyBinOp(Opcode, C, A, SQ.getWithInstruction(&I))) {
430 // It simplifies to V. Form "B op V".
431 replaceOperand(I, 0, B);
432 replaceOperand(I, 1, V);
433 // Conservatively clear the optional flags, since they may not be
434 // preserved by the reassociation.
435 ClearSubclassDataAfterReassociation(I);
442 // Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
443 // if C1 and C2 are constants.
447 Op0->getOpcode() == Opcode && Op1->getOpcode() == Opcode &&
448 match(Op0, m_OneUse(m_BinOp(m_Value(A), m_Constant(C1)))) &&
449 match(Op1, m_OneUse(m_BinOp(m_Value(B), m_Constant(C2))))) {
450 bool IsNUW = hasNoUnsignedWrap(I) &&
451 hasNoUnsignedWrap(*Op0) &&
452 hasNoUnsignedWrap(*Op1);
453 BinaryOperator *NewBO = (IsNUW && Opcode == Instruction::Add) ?
454 BinaryOperator::CreateNUW(Opcode, A, B) :
455 BinaryOperator::Create(Opcode, A, B);
457 if (isa<FPMathOperator>(NewBO)) {
458 FastMathFlags Flags = I.getFastMathFlags();
459 Flags &= Op0->getFastMathFlags();
460 Flags &= Op1->getFastMathFlags();
461 NewBO->setFastMathFlags(Flags);
463 InsertNewInstWith(NewBO, I);
464 NewBO->takeName(Op1);
465 replaceOperand(I, 0, NewBO);
466 replaceOperand(I, 1, ConstantExpr::get(Opcode, C1, C2));
467 // Conservatively clear the optional flags, since they may not be
468 // preserved by the reassociation.
469 ClearSubclassDataAfterReassociation(I);
471 I.setHasNoUnsignedWrap(true);
478 // No further simplifications.
483 /// Return whether "X LOp (Y ROp Z)" is always equal to
484 /// "(X LOp Y) ROp (X LOp Z)".
485 static bool leftDistributesOverRight(Instruction::BinaryOps LOp,
486 Instruction::BinaryOps ROp) {
487 // X & (Y | Z) <--> (X & Y) | (X & Z)
488 // X & (Y ^ Z) <--> (X & Y) ^ (X & Z)
489 if (LOp == Instruction::And)
490 return ROp == Instruction::Or || ROp == Instruction::Xor;
492 // X | (Y & Z) <--> (X | Y) & (X | Z)
493 if (LOp == Instruction::Or)
494 return ROp == Instruction::And;
496 // X * (Y + Z) <--> (X * Y) + (X * Z)
497 // X * (Y - Z) <--> (X * Y) - (X * Z)
498 if (LOp == Instruction::Mul)
499 return ROp == Instruction::Add || ROp == Instruction::Sub;
504 /// Return whether "(X LOp Y) ROp Z" is always equal to
505 /// "(X ROp Z) LOp (Y ROp Z)".
506 static bool rightDistributesOverLeft(Instruction::BinaryOps LOp,
507 Instruction::BinaryOps ROp) {
508 if (Instruction::isCommutative(ROp))
509 return leftDistributesOverRight(ROp, LOp);
511 // (X {&|^} Y) >> Z <--> (X >> Z) {&|^} (Y >> Z) for all shifts.
512 return Instruction::isBitwiseLogicOp(LOp) && Instruction::isShift(ROp);
514 // TODO: It would be nice to handle division, aka "(X + Y)/Z = X/Z + Y/Z",
515 // but this requires knowing that the addition does not overflow and other
519 /// This function returns identity value for given opcode, which can be used to
520 /// factor patterns like (X * 2) + X ==> (X * 2) + (X * 1) ==> X * (2 + 1).
521 static Value *getIdentityValue(Instruction::BinaryOps Opcode, Value *V) {
522 if (isa<Constant>(V))
525 return ConstantExpr::getBinOpIdentity(Opcode, V->getType());
528 /// This function predicates factorization using distributive laws. By default,
529 /// it just returns the 'Op' inputs. But for special-cases like
530 /// 'add(shl(X, 5), ...)', this function will have TopOpcode == Instruction::Add
531 /// and Op = shl(X, 5). The 'shl' is treated as the more general 'mul X, 32' to
532 /// allow more factorization opportunities.
533 static Instruction::BinaryOps
534 getBinOpsForFactorization(Instruction::BinaryOps TopOpcode, BinaryOperator *Op,
535 Value *&LHS, Value *&RHS) {
536 assert(Op && "Expected a binary operator");
537 LHS = Op->getOperand(0);
538 RHS = Op->getOperand(1);
539 if (TopOpcode == Instruction::Add || TopOpcode == Instruction::Sub) {
541 if (match(Op, m_Shl(m_Value(), m_Constant(C)))) {
542 // X << C --> X * (1 << C)
543 RHS = ConstantExpr::getShl(ConstantInt::get(Op->getType(), 1), C);
544 return Instruction::Mul;
546 // TODO: We can add other conversions e.g. shr => div etc.
548 return Op->getOpcode();
551 /// This tries to simplify binary operations by factorizing out common terms
552 /// (e. g. "(A*B)+(A*C)" -> "A*(B+C)").
553 Value *InstCombiner::tryFactorization(BinaryOperator &I,
554 Instruction::BinaryOps InnerOpcode,
555 Value *A, Value *B, Value *C, Value *D) {
556 assert(A && B && C && D && "All values must be provided");
559 Value *SimplifiedInst = nullptr;
560 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
561 Instruction::BinaryOps TopLevelOpcode = I.getOpcode();
563 // Does "X op' Y" always equal "Y op' X"?
564 bool InnerCommutative = Instruction::isCommutative(InnerOpcode);
566 // Does "X op' (Y op Z)" always equal "(X op' Y) op (X op' Z)"?
567 if (leftDistributesOverRight(InnerOpcode, TopLevelOpcode))
568 // Does the instruction have the form "(A op' B) op (A op' D)" or, in the
569 // commutative case, "(A op' B) op (C op' A)"?
570 if (A == C || (InnerCommutative && A == D)) {
573 // Consider forming "A op' (B op D)".
574 // If "B op D" simplifies then it can be formed with no cost.
575 V = SimplifyBinOp(TopLevelOpcode, B, D, SQ.getWithInstruction(&I));
576 // If "B op D" doesn't simplify then only go on if both of the existing
577 // operations "A op' B" and "C op' D" will be zapped as no longer used.
578 if (!V && LHS->hasOneUse() && RHS->hasOneUse())
579 V = Builder.CreateBinOp(TopLevelOpcode, B, D, RHS->getName());
581 SimplifiedInst = Builder.CreateBinOp(InnerOpcode, A, V);
585 // Does "(X op Y) op' Z" always equal "(X op' Z) op (Y op' Z)"?
586 if (!SimplifiedInst && rightDistributesOverLeft(TopLevelOpcode, InnerOpcode))
587 // Does the instruction have the form "(A op' B) op (C op' B)" or, in the
588 // commutative case, "(A op' B) op (B op' D)"?
589 if (B == D || (InnerCommutative && B == C)) {
592 // Consider forming "(A op C) op' B".
593 // If "A op C" simplifies then it can be formed with no cost.
594 V = SimplifyBinOp(TopLevelOpcode, A, C, SQ.getWithInstruction(&I));
596 // If "A op C" doesn't simplify then only go on if both of the existing
597 // operations "A op' B" and "C op' D" will be zapped as no longer used.
598 if (!V && LHS->hasOneUse() && RHS->hasOneUse())
599 V = Builder.CreateBinOp(TopLevelOpcode, A, C, LHS->getName());
601 SimplifiedInst = Builder.CreateBinOp(InnerOpcode, V, B);
605 if (SimplifiedInst) {
607 SimplifiedInst->takeName(&I);
609 // Check if we can add NSW/NUW flags to SimplifiedInst. If so, set them.
610 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(SimplifiedInst)) {
611 if (isa<OverflowingBinaryOperator>(SimplifiedInst)) {
614 if (isa<OverflowingBinaryOperator>(&I)) {
615 HasNSW = I.hasNoSignedWrap();
616 HasNUW = I.hasNoUnsignedWrap();
619 if (auto *LOBO = dyn_cast<OverflowingBinaryOperator>(LHS)) {
620 HasNSW &= LOBO->hasNoSignedWrap();
621 HasNUW &= LOBO->hasNoUnsignedWrap();
624 if (auto *ROBO = dyn_cast<OverflowingBinaryOperator>(RHS)) {
625 HasNSW &= ROBO->hasNoSignedWrap();
626 HasNUW &= ROBO->hasNoUnsignedWrap();
629 if (TopLevelOpcode == Instruction::Add &&
630 InnerOpcode == Instruction::Mul) {
631 // We can propagate 'nsw' if we know that
632 // %Y = mul nsw i16 %X, C
633 // %Z = add nsw i16 %Y, %X
635 // %Z = mul nsw i16 %X, C+1
637 // iff C+1 isn't INT_MIN
639 if (match(V, m_APInt(CInt))) {
640 if (!CInt->isMinSignedValue())
641 BO->setHasNoSignedWrap(HasNSW);
644 // nuw can be propagated with any constant or nuw value.
645 BO->setHasNoUnsignedWrap(HasNUW);
650 return SimplifiedInst;
653 /// This tries to simplify binary operations which some other binary operation
654 /// distributes over either by factorizing out common terms
655 /// (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this results in
656 /// simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is a win).
657 /// Returns the simplified value, or null if it didn't simplify.
658 Value *InstCombiner::SimplifyUsingDistributiveLaws(BinaryOperator &I) {
659 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
660 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
661 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
662 Instruction::BinaryOps TopLevelOpcode = I.getOpcode();
666 Value *A, *B, *C, *D;
667 Instruction::BinaryOps LHSOpcode, RHSOpcode;
669 LHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op0, A, B);
671 RHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op1, C, D);
673 // The instruction has the form "(A op' B) op (C op' D)". Try to factorize
675 if (Op0 && Op1 && LHSOpcode == RHSOpcode)
676 if (Value *V = tryFactorization(I, LHSOpcode, A, B, C, D))
679 // The instruction has the form "(A op' B) op (C)". Try to factorize common
682 if (Value *Ident = getIdentityValue(LHSOpcode, RHS))
683 if (Value *V = tryFactorization(I, LHSOpcode, A, B, RHS, Ident))
686 // The instruction has the form "(B) op (C op' D)". Try to factorize common
689 if (Value *Ident = getIdentityValue(RHSOpcode, LHS))
690 if (Value *V = tryFactorization(I, RHSOpcode, LHS, Ident, C, D))
695 if (Op0 && rightDistributesOverLeft(Op0->getOpcode(), TopLevelOpcode)) {
696 // The instruction has the form "(A op' B) op C". See if expanding it out
697 // to "(A op C) op' (B op C)" results in simplifications.
698 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
699 Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op'
701 Value *L = SimplifyBinOp(TopLevelOpcode, A, C, SQ.getWithInstruction(&I));
702 Value *R = SimplifyBinOp(TopLevelOpcode, B, C, SQ.getWithInstruction(&I));
704 // Do "A op C" and "B op C" both simplify?
706 // They do! Return "L op' R".
708 C = Builder.CreateBinOp(InnerOpcode, L, R);
713 // Does "A op C" simplify to the identity value for the inner opcode?
714 if (L && L == ConstantExpr::getBinOpIdentity(InnerOpcode, L->getType())) {
715 // They do! Return "B op C".
717 C = Builder.CreateBinOp(TopLevelOpcode, B, C);
722 // Does "B op C" simplify to the identity value for the inner opcode?
723 if (R && R == ConstantExpr::getBinOpIdentity(InnerOpcode, R->getType())) {
724 // They do! Return "A op C".
726 C = Builder.CreateBinOp(TopLevelOpcode, A, C);
732 if (Op1 && leftDistributesOverRight(TopLevelOpcode, Op1->getOpcode())) {
733 // The instruction has the form "A op (B op' C)". See if expanding it out
734 // to "(A op B) op' (A op C)" results in simplifications.
735 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
736 Instruction::BinaryOps InnerOpcode = Op1->getOpcode(); // op'
738 Value *L = SimplifyBinOp(TopLevelOpcode, A, B, SQ.getWithInstruction(&I));
739 Value *R = SimplifyBinOp(TopLevelOpcode, A, C, SQ.getWithInstruction(&I));
741 // Do "A op B" and "A op C" both simplify?
743 // They do! Return "L op' R".
745 A = Builder.CreateBinOp(InnerOpcode, L, R);
750 // Does "A op B" simplify to the identity value for the inner opcode?
751 if (L && L == ConstantExpr::getBinOpIdentity(InnerOpcode, L->getType())) {
752 // They do! Return "A op C".
754 A = Builder.CreateBinOp(TopLevelOpcode, A, C);
759 // Does "A op C" simplify to the identity value for the inner opcode?
760 if (R && R == ConstantExpr::getBinOpIdentity(InnerOpcode, R->getType())) {
761 // They do! Return "A op B".
763 A = Builder.CreateBinOp(TopLevelOpcode, A, B);
769 return SimplifySelectsFeedingBinaryOp(I, LHS, RHS);
772 Value *InstCombiner::SimplifySelectsFeedingBinaryOp(BinaryOperator &I,
773 Value *LHS, Value *RHS) {
774 Value *A, *B, *C, *D, *E, *F;
775 bool LHSIsSelect = match(LHS, m_Select(m_Value(A), m_Value(B), m_Value(C)));
776 bool RHSIsSelect = match(RHS, m_Select(m_Value(D), m_Value(E), m_Value(F)));
777 if (!LHSIsSelect && !RHSIsSelect)
781 BuilderTy::FastMathFlagGuard Guard(Builder);
782 if (isa<FPMathOperator>(&I)) {
783 FMF = I.getFastMathFlags();
784 Builder.setFastMathFlags(FMF);
787 Instruction::BinaryOps Opcode = I.getOpcode();
788 SimplifyQuery Q = SQ.getWithInstruction(&I);
790 Value *Cond, *True = nullptr, *False = nullptr;
791 if (LHSIsSelect && RHSIsSelect && A == D) {
792 // (A ? B : C) op (A ? E : F) -> A ? (B op E) : (C op F)
794 True = SimplifyBinOp(Opcode, B, E, FMF, Q);
795 False = SimplifyBinOp(Opcode, C, F, FMF, Q);
797 if (LHS->hasOneUse() && RHS->hasOneUse()) {
799 True = Builder.CreateBinOp(Opcode, B, E);
800 else if (True && !False)
801 False = Builder.CreateBinOp(Opcode, C, F);
803 } else if (LHSIsSelect && LHS->hasOneUse()) {
804 // (A ? B : C) op Y -> A ? (B op Y) : (C op Y)
806 True = SimplifyBinOp(Opcode, B, RHS, FMF, Q);
807 False = SimplifyBinOp(Opcode, C, RHS, FMF, Q);
808 } else if (RHSIsSelect && RHS->hasOneUse()) {
809 // X op (D ? E : F) -> D ? (X op E) : (X op F)
811 True = SimplifyBinOp(Opcode, LHS, E, FMF, Q);
812 False = SimplifyBinOp(Opcode, LHS, F, FMF, Q);
818 Value *SI = Builder.CreateSelect(Cond, True, False);
823 /// Given a 'sub' instruction, return the RHS of the instruction if the LHS is a
824 /// constant zero (which is the 'negate' form).
825 Value *InstCombiner::dyn_castNegVal(Value *V) const {
827 if (match(V, m_Neg(m_Value(NegV))))
830 // Constants can be considered to be negated values if they can be folded.
831 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
832 return ConstantExpr::getNeg(C);
834 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
835 if (C->getType()->getElementType()->isIntegerTy())
836 return ConstantExpr::getNeg(C);
838 if (ConstantVector *CV = dyn_cast<ConstantVector>(V)) {
839 for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) {
840 Constant *Elt = CV->getAggregateElement(i);
844 if (isa<UndefValue>(Elt))
847 if (!isa<ConstantInt>(Elt))
850 return ConstantExpr::getNeg(CV);
856 static Value *foldOperationIntoSelectOperand(Instruction &I, Value *SO,
857 InstCombiner::BuilderTy &Builder) {
858 if (auto *Cast = dyn_cast<CastInst>(&I))
859 return Builder.CreateCast(Cast->getOpcode(), SO, I.getType());
861 assert(I.isBinaryOp() && "Unexpected opcode for select folding");
863 // Figure out if the constant is the left or the right argument.
864 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
865 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
867 if (auto *SOC = dyn_cast<Constant>(SO)) {
869 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
870 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
873 Value *Op0 = SO, *Op1 = ConstOperand;
877 auto *BO = cast<BinaryOperator>(&I);
878 Value *RI = Builder.CreateBinOp(BO->getOpcode(), Op0, Op1,
879 SO->getName() + ".op");
880 auto *FPInst = dyn_cast<Instruction>(RI);
881 if (FPInst && isa<FPMathOperator>(FPInst))
882 FPInst->copyFastMathFlags(BO);
886 Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) {
887 // Don't modify shared select instructions.
888 if (!SI->hasOneUse())
891 Value *TV = SI->getTrueValue();
892 Value *FV = SI->getFalseValue();
893 if (!(isa<Constant>(TV) || isa<Constant>(FV)))
896 // Bool selects with constant operands can be folded to logical ops.
897 if (SI->getType()->isIntOrIntVectorTy(1))
900 // If it's a bitcast involving vectors, make sure it has the same number of
901 // elements on both sides.
902 if (auto *BC = dyn_cast<BitCastInst>(&Op)) {
903 VectorType *DestTy = dyn_cast<VectorType>(BC->getDestTy());
904 VectorType *SrcTy = dyn_cast<VectorType>(BC->getSrcTy());
906 // Verify that either both or neither are vectors.
907 if ((SrcTy == nullptr) != (DestTy == nullptr))
910 // If vectors, verify that they have the same number of elements.
911 if (SrcTy && SrcTy->getNumElements() != DestTy->getNumElements())
915 // Test if a CmpInst instruction is used exclusively by a select as
916 // part of a minimum or maximum operation. If so, refrain from doing
917 // any other folding. This helps out other analyses which understand
918 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
919 // and CodeGen. And in this case, at least one of the comparison
920 // operands has at least one user besides the compare (the select),
921 // which would often largely negate the benefit of folding anyway.
922 if (auto *CI = dyn_cast<CmpInst>(SI->getCondition())) {
923 if (CI->hasOneUse()) {
924 Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1);
926 // FIXME: This is a hack to avoid infinite looping with min/max patterns.
927 // We have to ensure that vector constants that only differ with
928 // undef elements are treated as equivalent.
929 auto areLooselyEqual = [](Value *A, Value *B) {
933 // Test for vector constants.
934 Constant *ConstA, *ConstB;
935 if (!match(A, m_Constant(ConstA)) || !match(B, m_Constant(ConstB)))
938 // TODO: Deal with FP constants?
939 if (!A->getType()->isIntOrIntVectorTy() || A->getType() != B->getType())
942 // Compare for equality including undefs as equal.
943 auto *Cmp = ConstantExpr::getCompare(ICmpInst::ICMP_EQ, ConstA, ConstB);
945 return match(Cmp, m_APIntAllowUndef(C)) && C->isOneValue();
948 if ((areLooselyEqual(TV, Op0) && areLooselyEqual(FV, Op1)) ||
949 (areLooselyEqual(FV, Op0) && areLooselyEqual(TV, Op1)))
954 Value *NewTV = foldOperationIntoSelectOperand(Op, TV, Builder);
955 Value *NewFV = foldOperationIntoSelectOperand(Op, FV, Builder);
956 return SelectInst::Create(SI->getCondition(), NewTV, NewFV, "", nullptr, SI);
959 static Value *foldOperationIntoPhiValue(BinaryOperator *I, Value *InV,
960 InstCombiner::BuilderTy &Builder) {
961 bool ConstIsRHS = isa<Constant>(I->getOperand(1));
962 Constant *C = cast<Constant>(I->getOperand(ConstIsRHS));
964 if (auto *InC = dyn_cast<Constant>(InV)) {
966 return ConstantExpr::get(I->getOpcode(), InC, C);
967 return ConstantExpr::get(I->getOpcode(), C, InC);
970 Value *Op0 = InV, *Op1 = C;
974 Value *RI = Builder.CreateBinOp(I->getOpcode(), Op0, Op1, "phi.bo");
975 auto *FPInst = dyn_cast<Instruction>(RI);
976 if (FPInst && isa<FPMathOperator>(FPInst))
977 FPInst->copyFastMathFlags(I);
981 Instruction *InstCombiner::foldOpIntoPhi(Instruction &I, PHINode *PN) {
982 unsigned NumPHIValues = PN->getNumIncomingValues();
983 if (NumPHIValues == 0)
986 // We normally only transform phis with a single use. However, if a PHI has
987 // multiple uses and they are all the same operation, we can fold *all* of the
988 // uses into the PHI.
989 if (!PN->hasOneUse()) {
990 // Walk the use list for the instruction, comparing them to I.
991 for (User *U : PN->users()) {
992 Instruction *UI = cast<Instruction>(U);
993 if (UI != &I && !I.isIdenticalTo(UI))
996 // Otherwise, we can replace *all* users with the new PHI we form.
999 // Check to see if all of the operands of the PHI are simple constants
1000 // (constantint/constantfp/undef). If there is one non-constant value,
1001 // remember the BB it is in. If there is more than one or if *it* is a PHI,
1002 // bail out. We don't do arbitrary constant expressions here because moving
1003 // their computation can be expensive without a cost model.
1004 BasicBlock *NonConstBB = nullptr;
1005 for (unsigned i = 0; i != NumPHIValues; ++i) {
1006 Value *InVal = PN->getIncomingValue(i);
1007 if (isa<Constant>(InVal) && !isa<ConstantExpr>(InVal))
1010 if (isa<PHINode>(InVal)) return nullptr; // Itself a phi.
1011 if (NonConstBB) return nullptr; // More than one non-const value.
1013 NonConstBB = PN->getIncomingBlock(i);
1015 // If the InVal is an invoke at the end of the pred block, then we can't
1016 // insert a computation after it without breaking the edge.
1017 if (isa<InvokeInst>(InVal))
1018 if (cast<Instruction>(InVal)->getParent() == NonConstBB)
1021 // If the incoming non-constant value is in I's block, we will remove one
1022 // instruction, but insert another equivalent one, leading to infinite
1024 if (isPotentiallyReachable(I.getParent(), NonConstBB, &DT, LI))
1028 // If there is exactly one non-constant value, we can insert a copy of the
1029 // operation in that block. However, if this is a critical edge, we would be
1030 // inserting the computation on some other paths (e.g. inside a loop). Only
1031 // do this if the pred block is unconditionally branching into the phi block.
1032 if (NonConstBB != nullptr) {
1033 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1034 if (!BI || !BI->isUnconditional()) return nullptr;
1037 // Okay, we can do the transformation: create the new PHI node.
1038 PHINode *NewPN = PHINode::Create(I.getType(), PN->getNumIncomingValues());
1039 InsertNewInstBefore(NewPN, *PN);
1040 NewPN->takeName(PN);
1042 // If we are going to have to insert a new computation, do so right before the
1043 // predecessor's terminator.
1045 Builder.SetInsertPoint(NonConstBB->getTerminator());
1047 // Next, add all of the operands to the PHI.
1048 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) {
1049 // We only currently try to fold the condition of a select when it is a phi,
1050 // not the true/false values.
1051 Value *TrueV = SI->getTrueValue();
1052 Value *FalseV = SI->getFalseValue();
1053 BasicBlock *PhiTransBB = PN->getParent();
1054 for (unsigned i = 0; i != NumPHIValues; ++i) {
1055 BasicBlock *ThisBB = PN->getIncomingBlock(i);
1056 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB);
1057 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB);
1058 Value *InV = nullptr;
1059 // Beware of ConstantExpr: it may eventually evaluate to getNullValue,
1060 // even if currently isNullValue gives false.
1061 Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i));
1062 // For vector constants, we cannot use isNullValue to fold into
1063 // FalseVInPred versus TrueVInPred. When we have individual nonzero
1064 // elements in the vector, we will incorrectly fold InC to
1066 if (InC && !isa<ConstantExpr>(InC) && isa<ConstantInt>(InC))
1067 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred;
1069 // Generate the select in the same block as PN's current incoming block.
1070 // Note: ThisBB need not be the NonConstBB because vector constants
1071 // which are constants by definition are handled here.
1072 // FIXME: This can lead to an increase in IR generation because we might
1073 // generate selects for vector constant phi operand, that could not be
1074 // folded to TrueVInPred or FalseVInPred as done for ConstantInt. For
1075 // non-vector phis, this transformation was always profitable because
1076 // the select would be generated exactly once in the NonConstBB.
1077 Builder.SetInsertPoint(ThisBB->getTerminator());
1078 InV = Builder.CreateSelect(PN->getIncomingValue(i), TrueVInPred,
1079 FalseVInPred, "phi.sel");
1081 NewPN->addIncoming(InV, ThisBB);
1083 } else if (CmpInst *CI = dyn_cast<CmpInst>(&I)) {
1084 Constant *C = cast<Constant>(I.getOperand(1));
1085 for (unsigned i = 0; i != NumPHIValues; ++i) {
1086 Value *InV = nullptr;
1087 if (auto *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
1088 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1090 InV = Builder.CreateCmp(CI->getPredicate(), PN->getIncomingValue(i),
1092 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1094 } else if (auto *BO = dyn_cast<BinaryOperator>(&I)) {
1095 for (unsigned i = 0; i != NumPHIValues; ++i) {
1096 Value *InV = foldOperationIntoPhiValue(BO, PN->getIncomingValue(i),
1098 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1101 CastInst *CI = cast<CastInst>(&I);
1102 Type *RetTy = CI->getType();
1103 for (unsigned i = 0; i != NumPHIValues; ++i) {
1105 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
1106 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1108 InV = Builder.CreateCast(CI->getOpcode(), PN->getIncomingValue(i),
1109 I.getType(), "phi.cast");
1110 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1114 for (auto UI = PN->user_begin(), E = PN->user_end(); UI != E;) {
1115 Instruction *User = cast<Instruction>(*UI++);
1116 if (User == &I) continue;
1117 replaceInstUsesWith(*User, NewPN);
1118 eraseInstFromFunction(*User);
1120 return replaceInstUsesWith(I, NewPN);
1123 Instruction *InstCombiner::foldBinOpIntoSelectOrPhi(BinaryOperator &I) {
1124 if (!isa<Constant>(I.getOperand(1)))
1127 if (auto *Sel = dyn_cast<SelectInst>(I.getOperand(0))) {
1128 if (Instruction *NewSel = FoldOpIntoSelect(I, Sel))
1130 } else if (auto *PN = dyn_cast<PHINode>(I.getOperand(0))) {
1131 if (Instruction *NewPhi = foldOpIntoPhi(I, PN))
1137 /// Given a pointer type and a constant offset, determine whether or not there
1138 /// is a sequence of GEP indices into the pointed type that will land us at the
1139 /// specified offset. If so, fill them into NewIndices and return the resultant
1140 /// element type, otherwise return null.
1141 Type *InstCombiner::FindElementAtOffset(PointerType *PtrTy, int64_t Offset,
1142 SmallVectorImpl<Value *> &NewIndices) {
1143 Type *Ty = PtrTy->getElementType();
1147 // Start with the index over the outer type. Note that the type size
1148 // might be zero (even if the offset isn't zero) if the indexed type
1149 // is something like [0 x {int, int}]
1150 Type *IndexTy = DL.getIndexType(PtrTy);
1151 int64_t FirstIdx = 0;
1152 if (int64_t TySize = DL.getTypeAllocSize(Ty)) {
1153 FirstIdx = Offset/TySize;
1154 Offset -= FirstIdx*TySize;
1156 // Handle hosts where % returns negative instead of values [0..TySize).
1160 assert(Offset >= 0);
1162 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
1165 NewIndices.push_back(ConstantInt::get(IndexTy, FirstIdx));
1167 // Index into the types. If we fail, set OrigBase to null.
1169 // Indexing into tail padding between struct/array elements.
1170 if (uint64_t(Offset * 8) >= DL.getTypeSizeInBits(Ty))
1173 if (StructType *STy = dyn_cast<StructType>(Ty)) {
1174 const StructLayout *SL = DL.getStructLayout(STy);
1175 assert(Offset < (int64_t)SL->getSizeInBytes() &&
1176 "Offset must stay within the indexed type");
1178 unsigned Elt = SL->getElementContainingOffset(Offset);
1179 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
1182 Offset -= SL->getElementOffset(Elt);
1183 Ty = STy->getElementType(Elt);
1184 } else if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
1185 uint64_t EltSize = DL.getTypeAllocSize(AT->getElementType());
1186 assert(EltSize && "Cannot index into a zero-sized array");
1187 NewIndices.push_back(ConstantInt::get(IndexTy,Offset/EltSize));
1189 Ty = AT->getElementType();
1191 // Otherwise, we can't index into the middle of this atomic type, bail.
1199 static bool shouldMergeGEPs(GEPOperator &GEP, GEPOperator &Src) {
1200 // If this GEP has only 0 indices, it is the same pointer as
1201 // Src. If Src is not a trivial GEP too, don't combine
1203 if (GEP.hasAllZeroIndices() && !Src.hasAllZeroIndices() &&
1209 /// Return a value X such that Val = X * Scale, or null if none.
1210 /// If the multiplication is known not to overflow, then NoSignedWrap is set.
1211 Value *InstCombiner::Descale(Value *Val, APInt Scale, bool &NoSignedWrap) {
1212 assert(isa<IntegerType>(Val->getType()) && "Can only descale integers!");
1213 assert(cast<IntegerType>(Val->getType())->getBitWidth() ==
1214 Scale.getBitWidth() && "Scale not compatible with value!");
1216 // If Val is zero or Scale is one then Val = Val * Scale.
1217 if (match(Val, m_Zero()) || Scale == 1) {
1218 NoSignedWrap = true;
1222 // If Scale is zero then it does not divide Val.
1223 if (Scale.isMinValue())
1226 // Look through chains of multiplications, searching for a constant that is
1227 // divisible by Scale. For example, descaling X*(Y*(Z*4)) by a factor of 4
1228 // will find the constant factor 4 and produce X*(Y*Z). Descaling X*(Y*8) by
1229 // a factor of 4 will produce X*(Y*2). The principle of operation is to bore
1232 // Val = M1 * X || Analysis starts here and works down
1233 // M1 = M2 * Y || Doesn't descend into terms with more
1234 // M2 = Z * 4 \/ than one use
1236 // Then to modify a term at the bottom:
1239 // M1 = Z * Y || Replaced M2 with Z
1241 // Then to work back up correcting nsw flags.
1243 // Op - the term we are currently analyzing. Starts at Val then drills down.
1244 // Replaced with its descaled value before exiting from the drill down loop.
1247 // Parent - initially null, but after drilling down notes where Op came from.
1248 // In the example above, Parent is (Val, 0) when Op is M1, because M1 is the
1249 // 0'th operand of Val.
1250 std::pair<Instruction *, unsigned> Parent;
1252 // Set if the transform requires a descaling at deeper levels that doesn't
1254 bool RequireNoSignedWrap = false;
1256 // Log base 2 of the scale. Negative if not a power of 2.
1257 int32_t logScale = Scale.exactLogBase2();
1259 for (;; Op = Parent.first->getOperand(Parent.second)) { // Drill down
1260 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
1261 // If Op is a constant divisible by Scale then descale to the quotient.
1262 APInt Quotient(Scale), Remainder(Scale); // Init ensures right bitwidth.
1263 APInt::sdivrem(CI->getValue(), Scale, Quotient, Remainder);
1264 if (!Remainder.isMinValue())
1265 // Not divisible by Scale.
1267 // Replace with the quotient in the parent.
1268 Op = ConstantInt::get(CI->getType(), Quotient);
1269 NoSignedWrap = true;
1273 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op)) {
1274 if (BO->getOpcode() == Instruction::Mul) {
1276 NoSignedWrap = BO->hasNoSignedWrap();
1277 if (RequireNoSignedWrap && !NoSignedWrap)
1280 // There are three cases for multiplication: multiplication by exactly
1281 // the scale, multiplication by a constant different to the scale, and
1282 // multiplication by something else.
1283 Value *LHS = BO->getOperand(0);
1284 Value *RHS = BO->getOperand(1);
1286 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
1287 // Multiplication by a constant.
1288 if (CI->getValue() == Scale) {
1289 // Multiplication by exactly the scale, replace the multiplication
1290 // by its left-hand side in the parent.
1295 // Otherwise drill down into the constant.
1296 if (!Op->hasOneUse())
1299 Parent = std::make_pair(BO, 1);
1303 // Multiplication by something else. Drill down into the left-hand side
1304 // since that's where the reassociate pass puts the good stuff.
1305 if (!Op->hasOneUse())
1308 Parent = std::make_pair(BO, 0);
1312 if (logScale > 0 && BO->getOpcode() == Instruction::Shl &&
1313 isa<ConstantInt>(BO->getOperand(1))) {
1314 // Multiplication by a power of 2.
1315 NoSignedWrap = BO->hasNoSignedWrap();
1316 if (RequireNoSignedWrap && !NoSignedWrap)
1319 Value *LHS = BO->getOperand(0);
1320 int32_t Amt = cast<ConstantInt>(BO->getOperand(1))->
1321 getLimitedValue(Scale.getBitWidth());
1324 if (Amt == logScale) {
1325 // Multiplication by exactly the scale, replace the multiplication
1326 // by its left-hand side in the parent.
1330 if (Amt < logScale || !Op->hasOneUse())
1333 // Multiplication by more than the scale. Reduce the multiplying amount
1334 // by the scale in the parent.
1335 Parent = std::make_pair(BO, 1);
1336 Op = ConstantInt::get(BO->getType(), Amt - logScale);
1341 if (!Op->hasOneUse())
1344 if (CastInst *Cast = dyn_cast<CastInst>(Op)) {
1345 if (Cast->getOpcode() == Instruction::SExt) {
1346 // Op is sign-extended from a smaller type, descale in the smaller type.
1347 unsigned SmallSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1348 APInt SmallScale = Scale.trunc(SmallSize);
1349 // Suppose Op = sext X, and we descale X as Y * SmallScale. We want to
1350 // descale Op as (sext Y) * Scale. In order to have
1351 // sext (Y * SmallScale) = (sext Y) * Scale
1352 // some conditions need to hold however: SmallScale must sign-extend to
1353 // Scale and the multiplication Y * SmallScale should not overflow.
1354 if (SmallScale.sext(Scale.getBitWidth()) != Scale)
1355 // SmallScale does not sign-extend to Scale.
1357 assert(SmallScale.exactLogBase2() == logScale);
1358 // Require that Y * SmallScale must not overflow.
1359 RequireNoSignedWrap = true;
1361 // Drill down through the cast.
1362 Parent = std::make_pair(Cast, 0);
1367 if (Cast->getOpcode() == Instruction::Trunc) {
1368 // Op is truncated from a larger type, descale in the larger type.
1369 // Suppose Op = trunc X, and we descale X as Y * sext Scale. Then
1370 // trunc (Y * sext Scale) = (trunc Y) * Scale
1371 // always holds. However (trunc Y) * Scale may overflow even if
1372 // trunc (Y * sext Scale) does not, so nsw flags need to be cleared
1373 // from this point up in the expression (see later).
1374 if (RequireNoSignedWrap)
1377 // Drill down through the cast.
1378 unsigned LargeSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1379 Parent = std::make_pair(Cast, 0);
1380 Scale = Scale.sext(LargeSize);
1381 if (logScale + 1 == (int32_t)Cast->getType()->getPrimitiveSizeInBits())
1383 assert(Scale.exactLogBase2() == logScale);
1388 // Unsupported expression, bail out.
1392 // If Op is zero then Val = Op * Scale.
1393 if (match(Op, m_Zero())) {
1394 NoSignedWrap = true;
1398 // We know that we can successfully descale, so from here on we can safely
1399 // modify the IR. Op holds the descaled version of the deepest term in the
1400 // expression. NoSignedWrap is 'true' if multiplying Op by Scale is known
1404 // The expression only had one term.
1407 // Rewrite the parent using the descaled version of its operand.
1408 assert(Parent.first->hasOneUse() && "Drilled down when more than one use!");
1409 assert(Op != Parent.first->getOperand(Parent.second) &&
1410 "Descaling was a no-op?");
1411 replaceOperand(*Parent.first, Parent.second, Op);
1412 Worklist.push(Parent.first);
1414 // Now work back up the expression correcting nsw flags. The logic is based
1415 // on the following observation: if X * Y is known not to overflow as a signed
1416 // multiplication, and Y is replaced by a value Z with smaller absolute value,
1417 // then X * Z will not overflow as a signed multiplication either. As we work
1418 // our way up, having NoSignedWrap 'true' means that the descaled value at the
1419 // current level has strictly smaller absolute value than the original.
1420 Instruction *Ancestor = Parent.first;
1422 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Ancestor)) {
1423 // If the multiplication wasn't nsw then we can't say anything about the
1424 // value of the descaled multiplication, and we have to clear nsw flags
1425 // from this point on up.
1426 bool OpNoSignedWrap = BO->hasNoSignedWrap();
1427 NoSignedWrap &= OpNoSignedWrap;
1428 if (NoSignedWrap != OpNoSignedWrap) {
1429 BO->setHasNoSignedWrap(NoSignedWrap);
1430 Worklist.push(Ancestor);
1432 } else if (Ancestor->getOpcode() == Instruction::Trunc) {
1433 // The fact that the descaled input to the trunc has smaller absolute
1434 // value than the original input doesn't tell us anything useful about
1435 // the absolute values of the truncations.
1436 NoSignedWrap = false;
1438 assert((Ancestor->getOpcode() != Instruction::SExt || NoSignedWrap) &&
1439 "Failed to keep proper track of nsw flags while drilling down?");
1441 if (Ancestor == Val)
1442 // Got to the top, all done!
1445 // Move up one level in the expression.
1446 assert(Ancestor->hasOneUse() && "Drilled down when more than one use!");
1447 Ancestor = Ancestor->user_back();
1451 Instruction *InstCombiner::foldVectorBinop(BinaryOperator &Inst) {
1452 // FIXME: some of this is likely fine for scalable vectors
1453 if (!isa<FixedVectorType>(Inst.getType()))
1456 BinaryOperator::BinaryOps Opcode = Inst.getOpcode();
1457 Value *LHS = Inst.getOperand(0), *RHS = Inst.getOperand(1);
1458 assert(cast<VectorType>(LHS->getType())->getElementCount() ==
1459 cast<VectorType>(Inst.getType())->getElementCount());
1460 assert(cast<VectorType>(RHS->getType())->getElementCount() ==
1461 cast<VectorType>(Inst.getType())->getElementCount());
1463 // If both operands of the binop are vector concatenations, then perform the
1464 // narrow binop on each pair of the source operands followed by concatenation
1466 Value *L0, *L1, *R0, *R1;
1468 if (match(LHS, m_Shuffle(m_Value(L0), m_Value(L1), m_Mask(Mask))) &&
1469 match(RHS, m_Shuffle(m_Value(R0), m_Value(R1), m_SpecificMask(Mask))) &&
1470 LHS->hasOneUse() && RHS->hasOneUse() &&
1471 cast<ShuffleVectorInst>(LHS)->isConcat() &&
1472 cast<ShuffleVectorInst>(RHS)->isConcat()) {
1473 // This transform does not have the speculative execution constraint as
1474 // below because the shuffle is a concatenation. The new binops are
1475 // operating on exactly the same elements as the existing binop.
1476 // TODO: We could ease the mask requirement to allow different undef lanes,
1477 // but that requires an analysis of the binop-with-undef output value.
1478 Value *NewBO0 = Builder.CreateBinOp(Opcode, L0, R0);
1479 if (auto *BO = dyn_cast<BinaryOperator>(NewBO0))
1480 BO->copyIRFlags(&Inst);
1481 Value *NewBO1 = Builder.CreateBinOp(Opcode, L1, R1);
1482 if (auto *BO = dyn_cast<BinaryOperator>(NewBO1))
1483 BO->copyIRFlags(&Inst);
1484 return new ShuffleVectorInst(NewBO0, NewBO1, Mask);
1487 // It may not be safe to reorder shuffles and things like div, urem, etc.
1488 // because we may trap when executing those ops on unknown vector elements.
1490 if (!isSafeToSpeculativelyExecute(&Inst))
1493 auto createBinOpShuffle = [&](Value *X, Value *Y, ArrayRef<int> M) {
1494 Value *XY = Builder.CreateBinOp(Opcode, X, Y);
1495 if (auto *BO = dyn_cast<BinaryOperator>(XY))
1496 BO->copyIRFlags(&Inst);
1497 return new ShuffleVectorInst(XY, UndefValue::get(XY->getType()), M);
1500 // If both arguments of the binary operation are shuffles that use the same
1501 // mask and shuffle within a single vector, move the shuffle after the binop.
1503 if (match(LHS, m_Shuffle(m_Value(V1), m_Undef(), m_Mask(Mask))) &&
1504 match(RHS, m_Shuffle(m_Value(V2), m_Undef(), m_SpecificMask(Mask))) &&
1505 V1->getType() == V2->getType() &&
1506 (LHS->hasOneUse() || RHS->hasOneUse() || LHS == RHS)) {
1507 // Op(shuffle(V1, Mask), shuffle(V2, Mask)) -> shuffle(Op(V1, V2), Mask)
1508 return createBinOpShuffle(V1, V2, Mask);
1511 // If both arguments of a commutative binop are select-shuffles that use the
1512 // same mask with commuted operands, the shuffles are unnecessary.
1513 if (Inst.isCommutative() &&
1514 match(LHS, m_Shuffle(m_Value(V1), m_Value(V2), m_Mask(Mask))) &&
1516 m_Shuffle(m_Specific(V2), m_Specific(V1), m_SpecificMask(Mask)))) {
1517 auto *LShuf = cast<ShuffleVectorInst>(LHS);
1518 auto *RShuf = cast<ShuffleVectorInst>(RHS);
1519 // TODO: Allow shuffles that contain undefs in the mask?
1520 // That is legal, but it reduces undef knowledge.
1521 // TODO: Allow arbitrary shuffles by shuffling after binop?
1522 // That might be legal, but we have to deal with poison.
1523 if (LShuf->isSelect() &&
1524 !is_contained(LShuf->getShuffleMask(), UndefMaskElem) &&
1525 RShuf->isSelect() &&
1526 !is_contained(RShuf->getShuffleMask(), UndefMaskElem)) {
1528 // LHS = shuffle V1, V2, <0, 5, 6, 3>
1529 // RHS = shuffle V2, V1, <0, 5, 6, 3>
1530 // LHS + RHS --> (V10+V20, V21+V11, V22+V12, V13+V23) --> V1 + V2
1531 Instruction *NewBO = BinaryOperator::Create(Opcode, V1, V2);
1532 NewBO->copyIRFlags(&Inst);
1537 // If one argument is a shuffle within one vector and the other is a constant,
1538 // try moving the shuffle after the binary operation. This canonicalization
1539 // intends to move shuffles closer to other shuffles and binops closer to
1540 // other binops, so they can be folded. It may also enable demanded elements
1542 unsigned NumElts = cast<FixedVectorType>(Inst.getType())->getNumElements();
1545 m_c_BinOp(m_OneUse(m_Shuffle(m_Value(V1), m_Undef(), m_Mask(Mask))),
1546 m_Constant(C))) && !isa<ConstantExpr>(C) &&
1547 cast<FixedVectorType>(V1->getType())->getNumElements() <= NumElts) {
1548 assert(Inst.getType()->getScalarType() == V1->getType()->getScalarType() &&
1549 "Shuffle should not change scalar type");
1551 // Find constant NewC that has property:
1552 // shuffle(NewC, ShMask) = C
1553 // If such constant does not exist (example: ShMask=<0,0> and C=<1,2>)
1554 // reorder is not possible. A 1-to-1 mapping is not required. Example:
1555 // ShMask = <1,1,2,2> and C = <5,5,6,6> --> NewC = <undef,5,6,undef>
1556 bool ConstOp1 = isa<Constant>(RHS);
1557 ArrayRef<int> ShMask = Mask;
1558 unsigned SrcVecNumElts =
1559 cast<FixedVectorType>(V1->getType())->getNumElements();
1560 UndefValue *UndefScalar = UndefValue::get(C->getType()->getScalarType());
1561 SmallVector<Constant *, 16> NewVecC(SrcVecNumElts, UndefScalar);
1562 bool MayChange = true;
1563 for (unsigned I = 0; I < NumElts; ++I) {
1564 Constant *CElt = C->getAggregateElement(I);
1565 if (ShMask[I] >= 0) {
1566 assert(ShMask[I] < (int)NumElts && "Not expecting narrowing shuffle");
1567 Constant *NewCElt = NewVecC[ShMask[I]];
1569 // 1. The constant vector contains a constant expression.
1570 // 2. The shuffle needs an element of the constant vector that can't
1571 // be mapped to a new constant vector.
1572 // 3. This is a widening shuffle that copies elements of V1 into the
1573 // extended elements (extending with undef is allowed).
1574 if (!CElt || (!isa<UndefValue>(NewCElt) && NewCElt != CElt) ||
1575 I >= SrcVecNumElts) {
1579 NewVecC[ShMask[I]] = CElt;
1581 // If this is a widening shuffle, we must be able to extend with undef
1582 // elements. If the original binop does not produce an undef in the high
1583 // lanes, then this transform is not safe.
1584 // Similarly for undef lanes due to the shuffle mask, we can only
1585 // transform binops that preserve undef.
1586 // TODO: We could shuffle those non-undef constant values into the
1587 // result by using a constant vector (rather than an undef vector)
1588 // as operand 1 of the new binop, but that might be too aggressive
1589 // for target-independent shuffle creation.
1590 if (I >= SrcVecNumElts || ShMask[I] < 0) {
1591 Constant *MaybeUndef =
1592 ConstOp1 ? ConstantExpr::get(Opcode, UndefScalar, CElt)
1593 : ConstantExpr::get(Opcode, CElt, UndefScalar);
1594 if (!isa<UndefValue>(MaybeUndef)) {
1601 Constant *NewC = ConstantVector::get(NewVecC);
1602 // It may not be safe to execute a binop on a vector with undef elements
1603 // because the entire instruction can be folded to undef or create poison
1604 // that did not exist in the original code.
1605 if (Inst.isIntDivRem() || (Inst.isShift() && ConstOp1))
1606 NewC = getSafeVectorConstantForBinop(Opcode, NewC, ConstOp1);
1608 // Op(shuffle(V1, Mask), C) -> shuffle(Op(V1, NewC), Mask)
1609 // Op(C, shuffle(V1, Mask)) -> shuffle(Op(NewC, V1), Mask)
1610 Value *NewLHS = ConstOp1 ? V1 : NewC;
1611 Value *NewRHS = ConstOp1 ? NewC : V1;
1612 return createBinOpShuffle(NewLHS, NewRHS, Mask);
1616 // Try to reassociate to sink a splat shuffle after a binary operation.
1617 if (Inst.isAssociative() && Inst.isCommutative()) {
1618 // Canonicalize shuffle operand as LHS.
1619 if (isa<ShuffleVectorInst>(RHS))
1620 std::swap(LHS, RHS);
1623 ArrayRef<int> MaskC;
1627 m_OneUse(m_Shuffle(m_Value(X), m_Undef(), m_Mask(MaskC)))) ||
1628 !match(MaskC, m_SplatOrUndefMask(SplatIndex)) ||
1629 X->getType() != Inst.getType() || !match(RHS, m_OneUse(m_BinOp(BO))) ||
1630 BO->getOpcode() != Opcode)
1633 // FIXME: This may not be safe if the analysis allows undef elements. By
1634 // moving 'Y' before the splat shuffle, we are implicitly assuming
1635 // that it is not undef/poison at the splat index.
1637 if (isSplatValue(BO->getOperand(0), SplatIndex)) {
1638 Y = BO->getOperand(0);
1639 OtherOp = BO->getOperand(1);
1640 } else if (isSplatValue(BO->getOperand(1), SplatIndex)) {
1641 Y = BO->getOperand(1);
1642 OtherOp = BO->getOperand(0);
1647 // X and Y are splatted values, so perform the binary operation on those
1648 // values followed by a splat followed by the 2nd binary operation:
1649 // bo (splat X), (bo Y, OtherOp) --> bo (splat (bo X, Y)), OtherOp
1650 Value *NewBO = Builder.CreateBinOp(Opcode, X, Y);
1651 UndefValue *Undef = UndefValue::get(Inst.getType());
1652 SmallVector<int, 8> NewMask(MaskC.size(), SplatIndex);
1653 Value *NewSplat = Builder.CreateShuffleVector(NewBO, Undef, NewMask);
1654 Instruction *R = BinaryOperator::Create(Opcode, NewSplat, OtherOp);
1656 // Intersect FMF on both new binops. Other (poison-generating) flags are
1657 // dropped to be safe.
1658 if (isa<FPMathOperator>(R)) {
1659 R->copyFastMathFlags(&Inst);
1662 if (auto *NewInstBO = dyn_cast<BinaryOperator>(NewBO))
1663 NewInstBO->copyIRFlags(R);
1670 /// Try to narrow the width of a binop if at least 1 operand is an extend of
1671 /// of a value. This requires a potentially expensive known bits check to make
1672 /// sure the narrow op does not overflow.
1673 Instruction *InstCombiner::narrowMathIfNoOverflow(BinaryOperator &BO) {
1674 // We need at least one extended operand.
1675 Value *Op0 = BO.getOperand(0), *Op1 = BO.getOperand(1);
1677 // If this is a sub, we swap the operands since we always want an extension
1678 // on the RHS. The LHS can be an extension or a constant.
1679 if (BO.getOpcode() == Instruction::Sub)
1680 std::swap(Op0, Op1);
1683 bool IsSext = match(Op0, m_SExt(m_Value(X)));
1684 if (!IsSext && !match(Op0, m_ZExt(m_Value(X))))
1687 // If both operands are the same extension from the same source type and we
1688 // can eliminate at least one (hasOneUse), this might work.
1689 CastInst::CastOps CastOpc = IsSext ? Instruction::SExt : Instruction::ZExt;
1691 if (!(match(Op1, m_ZExtOrSExt(m_Value(Y))) && X->getType() == Y->getType() &&
1692 cast<Operator>(Op1)->getOpcode() == CastOpc &&
1693 (Op0->hasOneUse() || Op1->hasOneUse()))) {
1694 // If that did not match, see if we have a suitable constant operand.
1695 // Truncating and extending must produce the same constant.
1697 if (!Op0->hasOneUse() || !match(Op1, m_Constant(WideC)))
1699 Constant *NarrowC = ConstantExpr::getTrunc(WideC, X->getType());
1700 if (ConstantExpr::getCast(CastOpc, NarrowC, BO.getType()) != WideC)
1705 // Swap back now that we found our operands.
1706 if (BO.getOpcode() == Instruction::Sub)
1709 // Both operands have narrow versions. Last step: the math must not overflow
1710 // in the narrow width.
1711 if (!willNotOverflow(BO.getOpcode(), X, Y, BO, IsSext))
1714 // bo (ext X), (ext Y) --> ext (bo X, Y)
1715 // bo (ext X), C --> ext (bo X, C')
1716 Value *NarrowBO = Builder.CreateBinOp(BO.getOpcode(), X, Y, "narrow");
1717 if (auto *NewBinOp = dyn_cast<BinaryOperator>(NarrowBO)) {
1719 NewBinOp->setHasNoSignedWrap();
1721 NewBinOp->setHasNoUnsignedWrap();
1723 return CastInst::Create(CastOpc, NarrowBO, BO.getType());
1726 static bool isMergedGEPInBounds(GEPOperator &GEP1, GEPOperator &GEP2) {
1727 // At least one GEP must be inbounds.
1728 if (!GEP1.isInBounds() && !GEP2.isInBounds())
1731 return (GEP1.isInBounds() || GEP1.hasAllZeroIndices()) &&
1732 (GEP2.isInBounds() || GEP2.hasAllZeroIndices());
1735 /// Thread a GEP operation with constant indices through the constant true/false
1736 /// arms of a select.
1737 static Instruction *foldSelectGEP(GetElementPtrInst &GEP,
1738 InstCombiner::BuilderTy &Builder) {
1739 if (!GEP.hasAllConstantIndices())
1744 Constant *TrueC, *FalseC;
1745 if (!match(GEP.getPointerOperand(), m_Instruction(Sel)) ||
1747 m_Select(m_Value(Cond), m_Constant(TrueC), m_Constant(FalseC))))
1750 // gep (select Cond, TrueC, FalseC), IndexC --> select Cond, TrueC', FalseC'
1751 // Propagate 'inbounds' and metadata from existing instructions.
1752 // Note: using IRBuilder to create the constants for efficiency.
1753 SmallVector<Value *, 4> IndexC(GEP.idx_begin(), GEP.idx_end());
1754 bool IsInBounds = GEP.isInBounds();
1755 Value *NewTrueC = IsInBounds ? Builder.CreateInBoundsGEP(TrueC, IndexC)
1756 : Builder.CreateGEP(TrueC, IndexC);
1757 Value *NewFalseC = IsInBounds ? Builder.CreateInBoundsGEP(FalseC, IndexC)
1758 : Builder.CreateGEP(FalseC, IndexC);
1759 return SelectInst::Create(Cond, NewTrueC, NewFalseC, "", nullptr, Sel);
1762 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
1763 SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end());
1764 Type *GEPType = GEP.getType();
1765 Type *GEPEltType = GEP.getSourceElementType();
1766 bool IsGEPSrcEleScalable = isa<ScalableVectorType>(GEPEltType);
1767 if (Value *V = SimplifyGEPInst(GEPEltType, Ops, SQ.getWithInstruction(&GEP)))
1768 return replaceInstUsesWith(GEP, V);
1770 // For vector geps, use the generic demanded vector support.
1771 // Skip if GEP return type is scalable. The number of elements is unknown at
1773 if (auto *GEPFVTy = dyn_cast<FixedVectorType>(GEPType)) {
1774 auto VWidth = GEPFVTy->getNumElements();
1775 APInt UndefElts(VWidth, 0);
1776 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
1777 if (Value *V = SimplifyDemandedVectorElts(&GEP, AllOnesEltMask,
1780 return replaceInstUsesWith(GEP, V);
1784 // TODO: 1) Scalarize splat operands, 2) scalarize entire instruction if
1785 // possible (decide on canonical form for pointer broadcast), 3) exploit
1786 // undef elements to decrease demanded bits
1789 Value *PtrOp = GEP.getOperand(0);
1791 // Eliminate unneeded casts for indices, and replace indices which displace
1792 // by multiples of a zero size type with zero.
1793 bool MadeChange = false;
1795 // Index width may not be the same width as pointer width.
1796 // Data layout chooses the right type based on supported integer types.
1797 Type *NewScalarIndexTy =
1798 DL.getIndexType(GEP.getPointerOperandType()->getScalarType());
1800 gep_type_iterator GTI = gep_type_begin(GEP);
1801 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end(); I != E;
1803 // Skip indices into struct types.
1807 Type *IndexTy = (*I)->getType();
1808 Type *NewIndexType =
1809 IndexTy->isVectorTy()
1810 ? VectorType::get(NewScalarIndexTy,
1811 cast<VectorType>(IndexTy)->getElementCount())
1814 // If the element type has zero size then any index over it is equivalent
1815 // to an index of zero, so replace it with zero if it is not zero already.
1816 Type *EltTy = GTI.getIndexedType();
1817 if (EltTy->isSized() && DL.getTypeAllocSize(EltTy).isZero())
1818 if (!isa<Constant>(*I) || !match(I->get(), m_Zero())) {
1819 *I = Constant::getNullValue(NewIndexType);
1823 if (IndexTy != NewIndexType) {
1824 // If we are using a wider index than needed for this platform, shrink
1825 // it to what we need. If narrower, sign-extend it to what we need.
1826 // This explicit cast can make subsequent optimizations more obvious.
1827 *I = Builder.CreateIntCast(*I, NewIndexType, true);
1834 // Check to see if the inputs to the PHI node are getelementptr instructions.
1835 if (auto *PN = dyn_cast<PHINode>(PtrOp)) {
1836 auto *Op1 = dyn_cast<GetElementPtrInst>(PN->getOperand(0));
1840 // Don't fold a GEP into itself through a PHI node. This can only happen
1841 // through the back-edge of a loop. Folding a GEP into itself means that
1842 // the value of the previous iteration needs to be stored in the meantime,
1843 // thus requiring an additional register variable to be live, but not
1844 // actually achieving anything (the GEP still needs to be executed once per
1851 for (auto I = PN->op_begin()+1, E = PN->op_end(); I !=E; ++I) {
1852 auto *Op2 = dyn_cast<GetElementPtrInst>(*I);
1853 if (!Op2 || Op1->getNumOperands() != Op2->getNumOperands())
1856 // As for Op1 above, don't try to fold a GEP into itself.
1860 // Keep track of the type as we walk the GEP.
1861 Type *CurTy = nullptr;
1863 for (unsigned J = 0, F = Op1->getNumOperands(); J != F; ++J) {
1864 if (Op1->getOperand(J)->getType() != Op2->getOperand(J)->getType())
1867 if (Op1->getOperand(J) != Op2->getOperand(J)) {
1869 // We have not seen any differences yet in the GEPs feeding the
1870 // PHI yet, so we record this one if it is allowed to be a
1873 // The first two arguments can vary for any GEP, the rest have to be
1874 // static for struct slots
1876 assert(CurTy && "No current type?");
1877 if (CurTy->isStructTy())
1883 // The GEP is different by more than one input. While this could be
1884 // extended to support GEPs that vary by more than one variable it
1885 // doesn't make sense since it greatly increases the complexity and
1886 // would result in an R+R+R addressing mode which no backend
1887 // directly supports and would need to be broken into several
1888 // simpler instructions anyway.
1893 // Sink down a layer of the type for the next iteration.
1896 CurTy = Op1->getSourceElementType();
1899 GetElementPtrInst::getTypeAtIndex(CurTy, Op1->getOperand(J));
1905 // If not all GEPs are identical we'll have to create a new PHI node.
1906 // Check that the old PHI node has only one use so that it will get
1908 if (DI != -1 && !PN->hasOneUse())
1911 auto *NewGEP = cast<GetElementPtrInst>(Op1->clone());
1913 // All the GEPs feeding the PHI are identical. Clone one down into our
1914 // BB so that it can be merged with the current GEP.
1916 // All the GEPs feeding the PHI differ at a single offset. Clone a GEP
1917 // into the current block so it can be merged, and create a new PHI to
1921 IRBuilderBase::InsertPointGuard Guard(Builder);
1922 Builder.SetInsertPoint(PN);
1923 NewPN = Builder.CreatePHI(Op1->getOperand(DI)->getType(),
1924 PN->getNumOperands());
1927 for (auto &I : PN->operands())
1928 NewPN->addIncoming(cast<GEPOperator>(I)->getOperand(DI),
1929 PN->getIncomingBlock(I));
1931 NewGEP->setOperand(DI, NewPN);
1934 GEP.getParent()->getInstList().insert(
1935 GEP.getParent()->getFirstInsertionPt(), NewGEP);
1936 replaceOperand(GEP, 0, NewGEP);
1940 // Combine Indices - If the source pointer to this getelementptr instruction
1941 // is a getelementptr instruction, combine the indices of the two
1942 // getelementptr instructions into a single instruction.
1943 if (auto *Src = dyn_cast<GEPOperator>(PtrOp)) {
1944 if (!shouldMergeGEPs(*cast<GEPOperator>(&GEP), *Src))
1947 // Try to reassociate loop invariant GEP chains to enable LICM.
1948 if (LI && Src->getNumOperands() == 2 && GEP.getNumOperands() == 2 &&
1950 if (Loop *L = LI->getLoopFor(GEP.getParent())) {
1951 Value *GO1 = GEP.getOperand(1);
1952 Value *SO1 = Src->getOperand(1);
1953 // Reassociate the two GEPs if SO1 is variant in the loop and GO1 is
1954 // invariant: this breaks the dependence between GEPs and allows LICM
1955 // to hoist the invariant part out of the loop.
1956 if (L->isLoopInvariant(GO1) && !L->isLoopInvariant(SO1)) {
1957 // We have to be careful here.
1958 // We have something like:
1959 // %src = getelementptr <ty>, <ty>* %base, <ty> %idx
1960 // %gep = getelementptr <ty>, <ty>* %src, <ty> %idx2
1961 // If we just swap idx & idx2 then we could inadvertantly
1962 // change %src from a vector to a scalar, or vice versa.
1964 // 1) %base a scalar & idx a scalar & idx2 a vector
1965 // => Swapping idx & idx2 turns %src into a vector type.
1966 // 2) %base a scalar & idx a vector & idx2 a scalar
1967 // => Swapping idx & idx2 turns %src in a scalar type
1968 // 3) %base, %idx, and %idx2 are scalars
1969 // => %src & %gep are scalars
1970 // => swapping idx & idx2 is safe
1971 // 4) %base a vector
1972 // => %src is a vector
1973 // => swapping idx & idx2 is safe.
1974 auto *SO0 = Src->getOperand(0);
1975 auto *SO0Ty = SO0->getType();
1976 if (!isa<VectorType>(GEPType) || // case 3
1977 isa<VectorType>(SO0Ty)) { // case 4
1978 Src->setOperand(1, GO1);
1979 GEP.setOperand(1, SO1);
1983 // -- have to recreate %src & %gep
1984 // put NewSrc at same location as %src
1985 Builder.SetInsertPoint(cast<Instruction>(PtrOp));
1986 auto *NewSrc = cast<GetElementPtrInst>(
1987 Builder.CreateGEP(GEPEltType, SO0, GO1, Src->getName()));
1988 NewSrc->setIsInBounds(Src->isInBounds());
1989 auto *NewGEP = GetElementPtrInst::Create(GEPEltType, NewSrc, {SO1});
1990 NewGEP->setIsInBounds(GEP.isInBounds());
1997 // Note that if our source is a gep chain itself then we wait for that
1998 // chain to be resolved before we perform this transformation. This
1999 // avoids us creating a TON of code in some cases.
2000 if (auto *SrcGEP = dyn_cast<GEPOperator>(Src->getOperand(0)))
2001 if (SrcGEP->getNumOperands() == 2 && shouldMergeGEPs(*Src, *SrcGEP))
2002 return nullptr; // Wait until our source is folded to completion.
2004 SmallVector<Value*, 8> Indices;
2006 // Find out whether the last index in the source GEP is a sequential idx.
2007 bool EndsWithSequential = false;
2008 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src);
2010 EndsWithSequential = I.isSequential();
2012 // Can we combine the two pointer arithmetics offsets?
2013 if (EndsWithSequential) {
2014 // Replace: gep (gep %P, long B), long A, ...
2015 // With: T = long A+B; gep %P, T, ...
2016 Value *SO1 = Src->getOperand(Src->getNumOperands()-1);
2017 Value *GO1 = GEP.getOperand(1);
2019 // If they aren't the same type, then the input hasn't been processed
2020 // by the loop above yet (which canonicalizes sequential index types to
2021 // intptr_t). Just avoid transforming this until the input has been
2023 if (SO1->getType() != GO1->getType())
2027 SimplifyAddInst(GO1, SO1, false, false, SQ.getWithInstruction(&GEP));
2028 // Only do the combine when we are sure the cost after the
2029 // merge is never more than that before the merge.
2033 // Update the GEP in place if possible.
2034 if (Src->getNumOperands() == 2) {
2035 GEP.setIsInBounds(isMergedGEPInBounds(*Src, *cast<GEPOperator>(&GEP)));
2036 replaceOperand(GEP, 0, Src->getOperand(0));
2037 replaceOperand(GEP, 1, Sum);
2040 Indices.append(Src->op_begin()+1, Src->op_end()-1);
2041 Indices.push_back(Sum);
2042 Indices.append(GEP.op_begin()+2, GEP.op_end());
2043 } else if (isa<Constant>(*GEP.idx_begin()) &&
2044 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
2045 Src->getNumOperands() != 1) {
2046 // Otherwise we can do the fold if the first index of the GEP is a zero
2047 Indices.append(Src->op_begin()+1, Src->op_end());
2048 Indices.append(GEP.idx_begin()+1, GEP.idx_end());
2051 if (!Indices.empty())
2052 return isMergedGEPInBounds(*Src, *cast<GEPOperator>(&GEP))
2053 ? GetElementPtrInst::CreateInBounds(
2054 Src->getSourceElementType(), Src->getOperand(0), Indices,
2056 : GetElementPtrInst::Create(Src->getSourceElementType(),
2057 Src->getOperand(0), Indices,
2061 // Skip if GEP source element type is scalable. The type alloc size is unknown
2063 if (GEP.getNumIndices() == 1 && !IsGEPSrcEleScalable) {
2064 unsigned AS = GEP.getPointerAddressSpace();
2065 if (GEP.getOperand(1)->getType()->getScalarSizeInBits() ==
2066 DL.getIndexSizeInBits(AS)) {
2067 uint64_t TyAllocSize = DL.getTypeAllocSize(GEPEltType).getFixedSize();
2069 bool Matched = false;
2072 if (TyAllocSize == 1) {
2073 V = GEP.getOperand(1);
2075 } else if (match(GEP.getOperand(1),
2076 m_AShr(m_Value(V), m_ConstantInt(C)))) {
2077 if (TyAllocSize == 1ULL << C)
2079 } else if (match(GEP.getOperand(1),
2080 m_SDiv(m_Value(V), m_ConstantInt(C)))) {
2081 if (TyAllocSize == C)
2086 // Canonicalize (gep i8* X, -(ptrtoint Y))
2087 // to (inttoptr (sub (ptrtoint X), (ptrtoint Y)))
2088 // The GEP pattern is emitted by the SCEV expander for certain kinds of
2089 // pointer arithmetic.
2090 if (match(V, m_Neg(m_PtrToInt(m_Value())))) {
2091 Operator *Index = cast<Operator>(V);
2092 Value *PtrToInt = Builder.CreatePtrToInt(PtrOp, Index->getType());
2093 Value *NewSub = Builder.CreateSub(PtrToInt, Index->getOperand(1));
2094 return CastInst::Create(Instruction::IntToPtr, NewSub, GEPType);
2096 // Canonicalize (gep i8* X, (ptrtoint Y)-(ptrtoint X))
2099 if (match(V, m_Sub(m_PtrToInt(m_Value(Y)),
2100 m_PtrToInt(m_Specific(GEP.getOperand(0))))))
2101 return CastInst::CreatePointerBitCastOrAddrSpaceCast(Y, GEPType);
2106 // We do not handle pointer-vector geps here.
2107 if (GEPType->isVectorTy())
2110 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0).
2111 Value *StrippedPtr = PtrOp->stripPointerCasts();
2112 PointerType *StrippedPtrTy = cast<PointerType>(StrippedPtr->getType());
2114 if (StrippedPtr != PtrOp) {
2115 bool HasZeroPointerIndex = false;
2116 Type *StrippedPtrEltTy = StrippedPtrTy->getElementType();
2118 if (auto *C = dyn_cast<ConstantInt>(GEP.getOperand(1)))
2119 HasZeroPointerIndex = C->isZero();
2121 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
2122 // into : GEP [10 x i8]* X, i32 0, ...
2124 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
2125 // into : GEP i8* X, ...
2127 // This occurs when the program declares an array extern like "int X[];"
2128 if (HasZeroPointerIndex) {
2129 if (auto *CATy = dyn_cast<ArrayType>(GEPEltType)) {
2130 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
2131 if (CATy->getElementType() == StrippedPtrEltTy) {
2132 // -> GEP i8* X, ...
2133 SmallVector<Value*, 8> Idx(GEP.idx_begin()+1, GEP.idx_end());
2134 GetElementPtrInst *Res = GetElementPtrInst::Create(
2135 StrippedPtrEltTy, StrippedPtr, Idx, GEP.getName());
2136 Res->setIsInBounds(GEP.isInBounds());
2137 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace())
2139 // Insert Res, and create an addrspacecast.
2141 // GEP (addrspacecast i8 addrspace(1)* X to [0 x i8]*), i32 0, ...
2143 // %0 = GEP i8 addrspace(1)* X, ...
2144 // addrspacecast i8 addrspace(1)* %0 to i8*
2145 return new AddrSpaceCastInst(Builder.Insert(Res), GEPType);
2148 if (auto *XATy = dyn_cast<ArrayType>(StrippedPtrEltTy)) {
2149 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
2150 if (CATy->getElementType() == XATy->getElementType()) {
2151 // -> GEP [10 x i8]* X, i32 0, ...
2152 // At this point, we know that the cast source type is a pointer
2153 // to an array of the same type as the destination pointer
2154 // array. Because the array type is never stepped over (there
2155 // is a leading zero) we can fold the cast into this GEP.
2156 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace()) {
2157 GEP.setSourceElementType(XATy);
2158 return replaceOperand(GEP, 0, StrippedPtr);
2160 // Cannot replace the base pointer directly because StrippedPtr's
2161 // address space is different. Instead, create a new GEP followed by
2162 // an addrspacecast.
2164 // GEP (addrspacecast [10 x i8] addrspace(1)* X to [0 x i8]*),
2167 // %0 = GEP [10 x i8] addrspace(1)* X, ...
2168 // addrspacecast i8 addrspace(1)* %0 to i8*
2169 SmallVector<Value*, 8> Idx(GEP.idx_begin(), GEP.idx_end());
2172 ? Builder.CreateInBoundsGEP(StrippedPtrEltTy, StrippedPtr,
2174 : Builder.CreateGEP(StrippedPtrEltTy, StrippedPtr, Idx,
2176 return new AddrSpaceCastInst(NewGEP, GEPType);
2180 } else if (GEP.getNumOperands() == 2 && !IsGEPSrcEleScalable) {
2181 // Skip if GEP source element type is scalable. The type alloc size is
2182 // unknown at compile-time.
2183 // Transform things like: %t = getelementptr i32*
2184 // bitcast ([2 x i32]* %str to i32*), i32 %V into: %t1 = getelementptr [2
2185 // x i32]* %str, i32 0, i32 %V; bitcast
2186 if (StrippedPtrEltTy->isArrayTy() &&
2187 DL.getTypeAllocSize(StrippedPtrEltTy->getArrayElementType()) ==
2188 DL.getTypeAllocSize(GEPEltType)) {
2189 Type *IdxType = DL.getIndexType(GEPType);
2190 Value *Idx[2] = { Constant::getNullValue(IdxType), GEP.getOperand(1) };
2193 ? Builder.CreateInBoundsGEP(StrippedPtrEltTy, StrippedPtr, Idx,
2195 : Builder.CreateGEP(StrippedPtrEltTy, StrippedPtr, Idx,
2198 // V and GEP are both pointer types --> BitCast
2199 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP, GEPType);
2202 // Transform things like:
2203 // %V = mul i64 %N, 4
2204 // %t = getelementptr i8* bitcast (i32* %arr to i8*), i32 %V
2205 // into: %t1 = getelementptr i32* %arr, i32 %N; bitcast
2206 if (GEPEltType->isSized() && StrippedPtrEltTy->isSized()) {
2207 // Check that changing the type amounts to dividing the index by a scale
2209 uint64_t ResSize = DL.getTypeAllocSize(GEPEltType).getFixedSize();
2210 uint64_t SrcSize = DL.getTypeAllocSize(StrippedPtrEltTy).getFixedSize();
2211 if (ResSize && SrcSize % ResSize == 0) {
2212 Value *Idx = GEP.getOperand(1);
2213 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
2214 uint64_t Scale = SrcSize / ResSize;
2216 // Earlier transforms ensure that the index has the right type
2217 // according to Data Layout, which considerably simplifies the
2218 // logic by eliminating implicit casts.
2219 assert(Idx->getType() == DL.getIndexType(GEPType) &&
2220 "Index type does not match the Data Layout preferences");
2223 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
2224 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
2225 // If the multiplication NewIdx * Scale may overflow then the new
2226 // GEP may not be "inbounds".
2228 GEP.isInBounds() && NSW
2229 ? Builder.CreateInBoundsGEP(StrippedPtrEltTy, StrippedPtr,
2230 NewIdx, GEP.getName())
2231 : Builder.CreateGEP(StrippedPtrEltTy, StrippedPtr, NewIdx,
2234 // The NewGEP must be pointer typed, so must the old one -> BitCast
2235 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
2241 // Similarly, transform things like:
2242 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
2243 // (where tmp = 8*tmp2) into:
2244 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
2245 if (GEPEltType->isSized() && StrippedPtrEltTy->isSized() &&
2246 StrippedPtrEltTy->isArrayTy()) {
2247 // Check that changing to the array element type amounts to dividing the
2248 // index by a scale factor.
2249 uint64_t ResSize = DL.getTypeAllocSize(GEPEltType).getFixedSize();
2250 uint64_t ArrayEltSize =
2251 DL.getTypeAllocSize(StrippedPtrEltTy->getArrayElementType())
2253 if (ResSize && ArrayEltSize % ResSize == 0) {
2254 Value *Idx = GEP.getOperand(1);
2255 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
2256 uint64_t Scale = ArrayEltSize / ResSize;
2258 // Earlier transforms ensure that the index has the right type
2259 // according to the Data Layout, which considerably simplifies
2260 // the logic by eliminating implicit casts.
2261 assert(Idx->getType() == DL.getIndexType(GEPType) &&
2262 "Index type does not match the Data Layout preferences");
2265 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
2266 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
2267 // If the multiplication NewIdx * Scale may overflow then the new
2268 // GEP may not be "inbounds".
2269 Type *IndTy = DL.getIndexType(GEPType);
2270 Value *Off[2] = {Constant::getNullValue(IndTy), NewIdx};
2273 GEP.isInBounds() && NSW
2274 ? Builder.CreateInBoundsGEP(StrippedPtrEltTy, StrippedPtr,
2276 : Builder.CreateGEP(StrippedPtrEltTy, StrippedPtr, Off,
2278 // The NewGEP must be pointer typed, so must the old one -> BitCast
2279 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
2287 // addrspacecast between types is canonicalized as a bitcast, then an
2288 // addrspacecast. To take advantage of the below bitcast + struct GEP, look
2289 // through the addrspacecast.
2290 Value *ASCStrippedPtrOp = PtrOp;
2291 if (auto *ASC = dyn_cast<AddrSpaceCastInst>(PtrOp)) {
2292 // X = bitcast A addrspace(1)* to B addrspace(1)*
2293 // Y = addrspacecast A addrspace(1)* to B addrspace(2)*
2294 // Z = gep Y, <...constant indices...>
2295 // Into an addrspacecasted GEP of the struct.
2296 if (auto *BC = dyn_cast<BitCastInst>(ASC->getOperand(0)))
2297 ASCStrippedPtrOp = BC;
2300 if (auto *BCI = dyn_cast<BitCastInst>(ASCStrippedPtrOp)) {
2301 Value *SrcOp = BCI->getOperand(0);
2302 PointerType *SrcType = cast<PointerType>(BCI->getSrcTy());
2303 Type *SrcEltType = SrcType->getElementType();
2305 // GEP directly using the source operand if this GEP is accessing an element
2306 // of a bitcasted pointer to vector or array of the same dimensions:
2307 // gep (bitcast <c x ty>* X to [c x ty]*), Y, Z --> gep X, Y, Z
2308 // gep (bitcast [c x ty]* X to <c x ty>*), Y, Z --> gep X, Y, Z
2309 auto areMatchingArrayAndVecTypes = [](Type *ArrTy, Type *VecTy,
2310 const DataLayout &DL) {
2311 auto *VecVTy = cast<VectorType>(VecTy);
2312 return ArrTy->getArrayElementType() == VecVTy->getElementType() &&
2313 ArrTy->getArrayNumElements() == VecVTy->getNumElements() &&
2314 DL.getTypeAllocSize(ArrTy) == DL.getTypeAllocSize(VecTy);
2316 if (GEP.getNumOperands() == 3 &&
2317 ((GEPEltType->isArrayTy() && SrcEltType->isVectorTy() &&
2318 areMatchingArrayAndVecTypes(GEPEltType, SrcEltType, DL)) ||
2319 (GEPEltType->isVectorTy() && SrcEltType->isArrayTy() &&
2320 areMatchingArrayAndVecTypes(SrcEltType, GEPEltType, DL)))) {
2322 // Create a new GEP here, as using `setOperand()` followed by
2323 // `setSourceElementType()` won't actually update the type of the
2324 // existing GEP Value. Causing issues if this Value is accessed when
2325 // constructing an AddrSpaceCastInst
2328 ? Builder.CreateInBoundsGEP(SrcEltType, SrcOp, {Ops[1], Ops[2]})
2329 : Builder.CreateGEP(SrcEltType, SrcOp, {Ops[1], Ops[2]});
2330 NGEP->takeName(&GEP);
2332 // Preserve GEP address space to satisfy users
2333 if (NGEP->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
2334 return new AddrSpaceCastInst(NGEP, GEPType);
2336 return replaceInstUsesWith(GEP, NGEP);
2339 // See if we can simplify:
2340 // X = bitcast A* to B*
2341 // Y = gep X, <...constant indices...>
2342 // into a gep of the original struct. This is important for SROA and alias
2343 // analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
2344 unsigned OffsetBits = DL.getIndexTypeSizeInBits(GEPType);
2345 APInt Offset(OffsetBits, 0);
2346 if (!isa<BitCastInst>(SrcOp) && GEP.accumulateConstantOffset(DL, Offset)) {
2347 // If this GEP instruction doesn't move the pointer, just replace the GEP
2348 // with a bitcast of the real input to the dest type.
2350 // If the bitcast is of an allocation, and the allocation will be
2351 // converted to match the type of the cast, don't touch this.
2352 if (isa<AllocaInst>(SrcOp) || isAllocationFn(SrcOp, &TLI)) {
2353 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
2354 if (Instruction *I = visitBitCast(*BCI)) {
2357 BCI->getParent()->getInstList().insert(BCI->getIterator(), I);
2358 replaceInstUsesWith(*BCI, I);
2364 if (SrcType->getPointerAddressSpace() != GEP.getAddressSpace())
2365 return new AddrSpaceCastInst(SrcOp, GEPType);
2366 return new BitCastInst(SrcOp, GEPType);
2369 // Otherwise, if the offset is non-zero, we need to find out if there is a
2370 // field at Offset in 'A's type. If so, we can pull the cast through the
2372 SmallVector<Value*, 8> NewIndices;
2373 if (FindElementAtOffset(SrcType, Offset.getSExtValue(), NewIndices)) {
2376 ? Builder.CreateInBoundsGEP(SrcEltType, SrcOp, NewIndices)
2377 : Builder.CreateGEP(SrcEltType, SrcOp, NewIndices);
2379 if (NGEP->getType() == GEPType)
2380 return replaceInstUsesWith(GEP, NGEP);
2381 NGEP->takeName(&GEP);
2383 if (NGEP->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
2384 return new AddrSpaceCastInst(NGEP, GEPType);
2385 return new BitCastInst(NGEP, GEPType);
2390 if (!GEP.isInBounds()) {
2392 DL.getIndexSizeInBits(PtrOp->getType()->getPointerAddressSpace());
2393 APInt BasePtrOffset(IdxWidth, 0);
2394 Value *UnderlyingPtrOp =
2395 PtrOp->stripAndAccumulateInBoundsConstantOffsets(DL,
2397 if (auto *AI = dyn_cast<AllocaInst>(UnderlyingPtrOp)) {
2398 if (GEP.accumulateConstantOffset(DL, BasePtrOffset) &&
2399 BasePtrOffset.isNonNegative()) {
2402 DL.getTypeAllocSize(AI->getAllocatedType()).getKnownMinSize());
2403 if (BasePtrOffset.ule(AllocSize)) {
2404 return GetElementPtrInst::CreateInBounds(
2405 GEP.getSourceElementType(), PtrOp, makeArrayRef(Ops).slice(1),
2412 if (Instruction *R = foldSelectGEP(GEP, Builder))
2418 static bool isNeverEqualToUnescapedAlloc(Value *V, const TargetLibraryInfo *TLI,
2420 if (isa<ConstantPointerNull>(V))
2422 if (auto *LI = dyn_cast<LoadInst>(V))
2423 return isa<GlobalVariable>(LI->getPointerOperand());
2424 // Two distinct allocations will never be equal.
2425 // We rely on LookThroughBitCast in isAllocLikeFn being false, since looking
2426 // through bitcasts of V can cause
2427 // the result statement below to be true, even when AI and V (ex:
2428 // i8* ->i32* ->i8* of AI) are the same allocations.
2429 return isAllocLikeFn(V, TLI) && V != AI;
2432 static bool isAllocSiteRemovable(Instruction *AI,
2433 SmallVectorImpl<WeakTrackingVH> &Users,
2434 const TargetLibraryInfo *TLI) {
2435 SmallVector<Instruction*, 4> Worklist;
2436 Worklist.push_back(AI);
2439 Instruction *PI = Worklist.pop_back_val();
2440 for (User *U : PI->users()) {
2441 Instruction *I = cast<Instruction>(U);
2442 switch (I->getOpcode()) {
2444 // Give up the moment we see something we can't handle.
2447 case Instruction::AddrSpaceCast:
2448 case Instruction::BitCast:
2449 case Instruction::GetElementPtr:
2450 Users.emplace_back(I);
2451 Worklist.push_back(I);
2454 case Instruction::ICmp: {
2455 ICmpInst *ICI = cast<ICmpInst>(I);
2456 // We can fold eq/ne comparisons with null to false/true, respectively.
2457 // We also fold comparisons in some conditions provided the alloc has
2458 // not escaped (see isNeverEqualToUnescapedAlloc).
2459 if (!ICI->isEquality())
2461 unsigned OtherIndex = (ICI->getOperand(0) == PI) ? 1 : 0;
2462 if (!isNeverEqualToUnescapedAlloc(ICI->getOperand(OtherIndex), TLI, AI))
2464 Users.emplace_back(I);
2468 case Instruction::Call:
2469 // Ignore no-op and store intrinsics.
2470 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
2471 switch (II->getIntrinsicID()) {
2475 case Intrinsic::memmove:
2476 case Intrinsic::memcpy:
2477 case Intrinsic::memset: {
2478 MemIntrinsic *MI = cast<MemIntrinsic>(II);
2479 if (MI->isVolatile() || MI->getRawDest() != PI)
2483 case Intrinsic::assume:
2484 case Intrinsic::invariant_start:
2485 case Intrinsic::invariant_end:
2486 case Intrinsic::lifetime_start:
2487 case Intrinsic::lifetime_end:
2488 case Intrinsic::objectsize:
2489 Users.emplace_back(I);
2494 if (isFreeCall(I, TLI)) {
2495 Users.emplace_back(I);
2500 case Instruction::Store: {
2501 StoreInst *SI = cast<StoreInst>(I);
2502 if (SI->isVolatile() || SI->getPointerOperand() != PI)
2504 Users.emplace_back(I);
2508 llvm_unreachable("missing a return?");
2510 } while (!Worklist.empty());
2514 Instruction *InstCombiner::visitAllocSite(Instruction &MI) {
2515 // If we have a malloc call which is only used in any amount of comparisons to
2516 // null and free calls, delete the calls and replace the comparisons with true
2517 // or false as appropriate.
2519 // This is based on the principle that we can substitute our own allocation
2520 // function (which will never return null) rather than knowledge of the
2521 // specific function being called. In some sense this can change the permitted
2522 // outputs of a program (when we convert a malloc to an alloca, the fact that
2523 // the allocation is now on the stack is potentially visible, for example),
2524 // but we believe in a permissible manner.
2525 SmallVector<WeakTrackingVH, 64> Users;
2527 // If we are removing an alloca with a dbg.declare, insert dbg.value calls
2528 // before each store.
2529 TinyPtrVector<DbgVariableIntrinsic *> DIIs;
2530 std::unique_ptr<DIBuilder> DIB;
2531 if (isa<AllocaInst>(MI)) {
2532 DIIs = FindDbgAddrUses(&MI);
2533 DIB.reset(new DIBuilder(*MI.getModule(), /*AllowUnresolved=*/false));
2536 if (isAllocSiteRemovable(&MI, Users, &TLI)) {
2537 for (unsigned i = 0, e = Users.size(); i != e; ++i) {
2538 // Lowering all @llvm.objectsize calls first because they may
2539 // use a bitcast/GEP of the alloca we are removing.
2543 Instruction *I = cast<Instruction>(&*Users[i]);
2545 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
2546 if (II->getIntrinsicID() == Intrinsic::objectsize) {
2548 lowerObjectSizeCall(II, DL, &TLI, /*MustSucceed=*/true);
2549 replaceInstUsesWith(*I, Result);
2550 eraseInstFromFunction(*I);
2551 Users[i] = nullptr; // Skip examining in the next loop.
2555 for (unsigned i = 0, e = Users.size(); i != e; ++i) {
2559 Instruction *I = cast<Instruction>(&*Users[i]);
2561 if (ICmpInst *C = dyn_cast<ICmpInst>(I)) {
2562 replaceInstUsesWith(*C,
2563 ConstantInt::get(Type::getInt1Ty(C->getContext()),
2564 C->isFalseWhenEqual()));
2565 } else if (auto *SI = dyn_cast<StoreInst>(I)) {
2566 for (auto *DII : DIIs)
2567 ConvertDebugDeclareToDebugValue(DII, SI, *DIB);
2569 // Casts, GEP, or anything else: we're about to delete this instruction,
2570 // so it can not have any valid uses.
2571 replaceInstUsesWith(*I, UndefValue::get(I->getType()));
2573 eraseInstFromFunction(*I);
2576 if (InvokeInst *II = dyn_cast<InvokeInst>(&MI)) {
2577 // Replace invoke with a NOP intrinsic to maintain the original CFG
2578 Module *M = II->getModule();
2579 Function *F = Intrinsic::getDeclaration(M, Intrinsic::donothing);
2580 InvokeInst::Create(F, II->getNormalDest(), II->getUnwindDest(),
2581 None, "", II->getParent());
2584 for (auto *DII : DIIs)
2585 eraseInstFromFunction(*DII);
2587 return eraseInstFromFunction(MI);
2592 /// Move the call to free before a NULL test.
2594 /// Check if this free is accessed after its argument has been test
2595 /// against NULL (property 0).
2596 /// If yes, it is legal to move this call in its predecessor block.
2598 /// The move is performed only if the block containing the call to free
2599 /// will be removed, i.e.:
2600 /// 1. it has only one predecessor P, and P has two successors
2601 /// 2. it contains the call, noops, and an unconditional branch
2602 /// 3. its successor is the same as its predecessor's successor
2604 /// The profitability is out-of concern here and this function should
2605 /// be called only if the caller knows this transformation would be
2606 /// profitable (e.g., for code size).
2607 static Instruction *tryToMoveFreeBeforeNullTest(CallInst &FI,
2608 const DataLayout &DL) {
2609 Value *Op = FI.getArgOperand(0);
2610 BasicBlock *FreeInstrBB = FI.getParent();
2611 BasicBlock *PredBB = FreeInstrBB->getSinglePredecessor();
2613 // Validate part of constraint #1: Only one predecessor
2614 // FIXME: We can extend the number of predecessor, but in that case, we
2615 // would duplicate the call to free in each predecessor and it may
2616 // not be profitable even for code size.
2620 // Validate constraint #2: Does this block contains only the call to
2621 // free, noops, and an unconditional branch?
2623 Instruction *FreeInstrBBTerminator = FreeInstrBB->getTerminator();
2624 if (!match(FreeInstrBBTerminator, m_UnconditionalBr(SuccBB)))
2627 // If there are only 2 instructions in the block, at this point,
2628 // this is the call to free and unconditional.
2629 // If there are more than 2 instructions, check that they are noops
2630 // i.e., they won't hurt the performance of the generated code.
2631 if (FreeInstrBB->size() != 2) {
2632 for (const Instruction &Inst : FreeInstrBB->instructionsWithoutDebug()) {
2633 if (&Inst == &FI || &Inst == FreeInstrBBTerminator)
2635 auto *Cast = dyn_cast<CastInst>(&Inst);
2636 if (!Cast || !Cast->isNoopCast(DL))
2640 // Validate the rest of constraint #1 by matching on the pred branch.
2641 Instruction *TI = PredBB->getTerminator();
2642 BasicBlock *TrueBB, *FalseBB;
2643 ICmpInst::Predicate Pred;
2644 if (!match(TI, m_Br(m_ICmp(Pred,
2645 m_CombineOr(m_Specific(Op),
2646 m_Specific(Op->stripPointerCasts())),
2650 if (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE)
2653 // Validate constraint #3: Ensure the null case just falls through.
2654 if (SuccBB != (Pred == ICmpInst::ICMP_EQ ? TrueBB : FalseBB))
2656 assert(FreeInstrBB == (Pred == ICmpInst::ICMP_EQ ? FalseBB : TrueBB) &&
2657 "Broken CFG: missing edge from predecessor to successor");
2659 // At this point, we know that everything in FreeInstrBB can be moved
2661 for (BasicBlock::iterator It = FreeInstrBB->begin(), End = FreeInstrBB->end();
2663 Instruction &Instr = *It++;
2664 if (&Instr == FreeInstrBBTerminator)
2666 Instr.moveBefore(TI);
2668 assert(FreeInstrBB->size() == 1 &&
2669 "Only the branch instruction should remain");
2673 Instruction *InstCombiner::visitFree(CallInst &FI) {
2674 Value *Op = FI.getArgOperand(0);
2676 // free undef -> unreachable.
2677 if (isa<UndefValue>(Op)) {
2678 // Leave a marker since we can't modify the CFG here.
2679 CreateNonTerminatorUnreachable(&FI);
2680 return eraseInstFromFunction(FI);
2683 // If we have 'free null' delete the instruction. This can happen in stl code
2684 // when lots of inlining happens.
2685 if (isa<ConstantPointerNull>(Op))
2686 return eraseInstFromFunction(FI);
2688 // If we optimize for code size, try to move the call to free before the null
2689 // test so that simplify cfg can remove the empty block and dead code
2690 // elimination the branch. I.e., helps to turn something like:
2691 // if (foo) free(foo);
2695 // Note that we can only do this for 'free' and not for any flavor of
2696 // 'operator delete'; there is no 'operator delete' symbol for which we are
2697 // permitted to invent a call, even if we're passing in a null pointer.
2700 if (TLI.getLibFunc(FI, Func) && TLI.has(Func) && Func == LibFunc_free)
2701 if (Instruction *I = tryToMoveFreeBeforeNullTest(FI, DL))
2708 static bool isMustTailCall(Value *V) {
2709 if (auto *CI = dyn_cast<CallInst>(V))
2710 return CI->isMustTailCall();
2714 Instruction *InstCombiner::visitReturnInst(ReturnInst &RI) {
2715 if (RI.getNumOperands() == 0) // ret void
2718 Value *ResultOp = RI.getOperand(0);
2719 Type *VTy = ResultOp->getType();
2720 if (!VTy->isIntegerTy() || isa<Constant>(ResultOp))
2723 // Don't replace result of musttail calls.
2724 if (isMustTailCall(ResultOp))
2727 // There might be assume intrinsics dominating this return that completely
2728 // determine the value. If so, constant fold it.
2729 KnownBits Known = computeKnownBits(ResultOp, 0, &RI);
2730 if (Known.isConstant())
2731 return replaceOperand(RI, 0,
2732 Constant::getIntegerValue(VTy, Known.getConstant()));
2737 Instruction *InstCombiner::visitUnconditionalBranchInst(BranchInst &BI) {
2738 assert(BI.isUnconditional() && "Only for unconditional branches.");
2740 // If this store is the second-to-last instruction in the basic block
2741 // (excluding debug info and bitcasts of pointers) and if the block ends with
2742 // an unconditional branch, try to move the store to the successor block.
2744 auto GetLastSinkableStore = [](BasicBlock::iterator BBI) {
2745 auto IsNoopInstrForStoreMerging = [](BasicBlock::iterator BBI) {
2746 return isa<DbgInfoIntrinsic>(BBI) ||
2747 (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy());
2750 BasicBlock::iterator FirstInstr = BBI->getParent()->begin();
2752 if (BBI != FirstInstr)
2754 } while (BBI != FirstInstr && IsNoopInstrForStoreMerging(BBI));
2756 return dyn_cast<StoreInst>(BBI);
2759 if (StoreInst *SI = GetLastSinkableStore(BasicBlock::iterator(BI)))
2760 if (mergeStoreIntoSuccessor(*SI))
2766 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
2767 if (BI.isUnconditional())
2768 return visitUnconditionalBranchInst(BI);
2770 // Change br (not X), label True, label False to: br X, label False, True
2772 if (match(&BI, m_Br(m_Not(m_Value(X)), m_BasicBlock(), m_BasicBlock())) &&
2773 !isa<Constant>(X)) {
2774 // Swap Destinations and condition...
2775 BI.swapSuccessors();
2776 return replaceOperand(BI, 0, X);
2779 // If the condition is irrelevant, remove the use so that other
2780 // transforms on the condition become more effective.
2781 if (!isa<ConstantInt>(BI.getCondition()) &&
2782 BI.getSuccessor(0) == BI.getSuccessor(1))
2783 return replaceOperand(
2784 BI, 0, ConstantInt::getFalse(BI.getCondition()->getType()));
2786 // Canonicalize, for example, fcmp_one -> fcmp_oeq.
2787 CmpInst::Predicate Pred;
2788 if (match(&BI, m_Br(m_OneUse(m_FCmp(Pred, m_Value(), m_Value())),
2789 m_BasicBlock(), m_BasicBlock())) &&
2790 !isCanonicalPredicate(Pred)) {
2791 // Swap destinations and condition.
2792 CmpInst *Cond = cast<CmpInst>(BI.getCondition());
2793 Cond->setPredicate(CmpInst::getInversePredicate(Pred));
2794 BI.swapSuccessors();
2795 Worklist.push(Cond);
2802 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
2803 Value *Cond = SI.getCondition();
2805 ConstantInt *AddRHS;
2806 if (match(Cond, m_Add(m_Value(Op0), m_ConstantInt(AddRHS)))) {
2807 // Change 'switch (X+4) case 1:' into 'switch (X) case -3'.
2808 for (auto Case : SI.cases()) {
2809 Constant *NewCase = ConstantExpr::getSub(Case.getCaseValue(), AddRHS);
2810 assert(isa<ConstantInt>(NewCase) &&
2811 "Result of expression should be constant");
2812 Case.setValue(cast<ConstantInt>(NewCase));
2814 return replaceOperand(SI, 0, Op0);
2817 KnownBits Known = computeKnownBits(Cond, 0, &SI);
2818 unsigned LeadingKnownZeros = Known.countMinLeadingZeros();
2819 unsigned LeadingKnownOnes = Known.countMinLeadingOnes();
2821 // Compute the number of leading bits we can ignore.
2822 // TODO: A better way to determine this would use ComputeNumSignBits().
2823 for (auto &C : SI.cases()) {
2824 LeadingKnownZeros = std::min(
2825 LeadingKnownZeros, C.getCaseValue()->getValue().countLeadingZeros());
2826 LeadingKnownOnes = std::min(
2827 LeadingKnownOnes, C.getCaseValue()->getValue().countLeadingOnes());
2830 unsigned NewWidth = Known.getBitWidth() - std::max(LeadingKnownZeros, LeadingKnownOnes);
2832 // Shrink the condition operand if the new type is smaller than the old type.
2833 // But do not shrink to a non-standard type, because backend can't generate
2834 // good code for that yet.
2835 // TODO: We can make it aggressive again after fixing PR39569.
2836 if (NewWidth > 0 && NewWidth < Known.getBitWidth() &&
2837 shouldChangeType(Known.getBitWidth(), NewWidth)) {
2838 IntegerType *Ty = IntegerType::get(SI.getContext(), NewWidth);
2839 Builder.SetInsertPoint(&SI);
2840 Value *NewCond = Builder.CreateTrunc(Cond, Ty, "trunc");
2842 for (auto Case : SI.cases()) {
2843 APInt TruncatedCase = Case.getCaseValue()->getValue().trunc(NewWidth);
2844 Case.setValue(ConstantInt::get(SI.getContext(), TruncatedCase));
2846 return replaceOperand(SI, 0, NewCond);
2852 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
2853 Value *Agg = EV.getAggregateOperand();
2855 if (!EV.hasIndices())
2856 return replaceInstUsesWith(EV, Agg);
2858 if (Value *V = SimplifyExtractValueInst(Agg, EV.getIndices(),
2859 SQ.getWithInstruction(&EV)))
2860 return replaceInstUsesWith(EV, V);
2862 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
2863 // We're extracting from an insertvalue instruction, compare the indices
2864 const unsigned *exti, *exte, *insi, *inse;
2865 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
2866 exte = EV.idx_end(), inse = IV->idx_end();
2867 exti != exte && insi != inse;
2870 // The insert and extract both reference distinctly different elements.
2871 // This means the extract is not influenced by the insert, and we can
2872 // replace the aggregate operand of the extract with the aggregate
2873 // operand of the insert. i.e., replace
2874 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
2875 // %E = extractvalue { i32, { i32 } } %I, 0
2877 // %E = extractvalue { i32, { i32 } } %A, 0
2878 return ExtractValueInst::Create(IV->getAggregateOperand(),
2881 if (exti == exte && insi == inse)
2882 // Both iterators are at the end: Index lists are identical. Replace
2883 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
2884 // %C = extractvalue { i32, { i32 } } %B, 1, 0
2886 return replaceInstUsesWith(EV, IV->getInsertedValueOperand());
2888 // The extract list is a prefix of the insert list. i.e. replace
2889 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
2890 // %E = extractvalue { i32, { i32 } } %I, 1
2892 // %X = extractvalue { i32, { i32 } } %A, 1
2893 // %E = insertvalue { i32 } %X, i32 42, 0
2894 // by switching the order of the insert and extract (though the
2895 // insertvalue should be left in, since it may have other uses).
2896 Value *NewEV = Builder.CreateExtractValue(IV->getAggregateOperand(),
2898 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
2899 makeArrayRef(insi, inse));
2902 // The insert list is a prefix of the extract list
2903 // We can simply remove the common indices from the extract and make it
2904 // operate on the inserted value instead of the insertvalue result.
2906 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
2907 // %E = extractvalue { i32, { i32 } } %I, 1, 0
2909 // %E extractvalue { i32 } { i32 42 }, 0
2910 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
2911 makeArrayRef(exti, exte));
2913 if (WithOverflowInst *WO = dyn_cast<WithOverflowInst>(Agg)) {
2914 // We're extracting from an overflow intrinsic, see if we're the only user,
2915 // which allows us to simplify multiple result intrinsics to simpler
2916 // things that just get one value.
2917 if (WO->hasOneUse()) {
2918 // Check if we're grabbing only the result of a 'with overflow' intrinsic
2919 // and replace it with a traditional binary instruction.
2920 if (*EV.idx_begin() == 0) {
2921 Instruction::BinaryOps BinOp = WO->getBinaryOp();
2922 Value *LHS = WO->getLHS(), *RHS = WO->getRHS();
2923 replaceInstUsesWith(*WO, UndefValue::get(WO->getType()));
2924 eraseInstFromFunction(*WO);
2925 return BinaryOperator::Create(BinOp, LHS, RHS);
2928 // If the normal result of the add is dead, and the RHS is a constant,
2929 // we can transform this into a range comparison.
2930 // overflow = uadd a, -4 --> overflow = icmp ugt a, 3
2931 if (WO->getIntrinsicID() == Intrinsic::uadd_with_overflow)
2932 if (ConstantInt *CI = dyn_cast<ConstantInt>(WO->getRHS()))
2933 return new ICmpInst(ICmpInst::ICMP_UGT, WO->getLHS(),
2934 ConstantExpr::getNot(CI));
2937 if (LoadInst *L = dyn_cast<LoadInst>(Agg))
2938 // If the (non-volatile) load only has one use, we can rewrite this to a
2939 // load from a GEP. This reduces the size of the load. If a load is used
2940 // only by extractvalue instructions then this either must have been
2941 // optimized before, or it is a struct with padding, in which case we
2942 // don't want to do the transformation as it loses padding knowledge.
2943 if (L->isSimple() && L->hasOneUse()) {
2944 // extractvalue has integer indices, getelementptr has Value*s. Convert.
2945 SmallVector<Value*, 4> Indices;
2946 // Prefix an i32 0 since we need the first element.
2947 Indices.push_back(Builder.getInt32(0));
2948 for (ExtractValueInst::idx_iterator I = EV.idx_begin(), E = EV.idx_end();
2950 Indices.push_back(Builder.getInt32(*I));
2952 // We need to insert these at the location of the old load, not at that of
2953 // the extractvalue.
2954 Builder.SetInsertPoint(L);
2955 Value *GEP = Builder.CreateInBoundsGEP(L->getType(),
2956 L->getPointerOperand(), Indices);
2957 Instruction *NL = Builder.CreateLoad(EV.getType(), GEP);
2958 // Whatever aliasing information we had for the orignal load must also
2959 // hold for the smaller load, so propagate the annotations.
2961 L->getAAMetadata(Nodes);
2962 NL->setAAMetadata(Nodes);
2963 // Returning the load directly will cause the main loop to insert it in
2964 // the wrong spot, so use replaceInstUsesWith().
2965 return replaceInstUsesWith(EV, NL);
2967 // We could simplify extracts from other values. Note that nested extracts may
2968 // already be simplified implicitly by the above: extract (extract (insert) )
2969 // will be translated into extract ( insert ( extract ) ) first and then just
2970 // the value inserted, if appropriate. Similarly for extracts from single-use
2971 // loads: extract (extract (load)) will be translated to extract (load (gep))
2972 // and if again single-use then via load (gep (gep)) to load (gep).
2973 // However, double extracts from e.g. function arguments or return values
2974 // aren't handled yet.
2978 /// Return 'true' if the given typeinfo will match anything.
2979 static bool isCatchAll(EHPersonality Personality, Constant *TypeInfo) {
2980 switch (Personality) {
2981 case EHPersonality::GNU_C:
2982 case EHPersonality::GNU_C_SjLj:
2983 case EHPersonality::Rust:
2984 // The GCC C EH and Rust personality only exists to support cleanups, so
2985 // it's not clear what the semantics of catch clauses are.
2987 case EHPersonality::Unknown:
2989 case EHPersonality::GNU_Ada:
2990 // While __gnat_all_others_value will match any Ada exception, it doesn't
2991 // match foreign exceptions (or didn't, before gcc-4.7).
2993 case EHPersonality::GNU_CXX:
2994 case EHPersonality::GNU_CXX_SjLj:
2995 case EHPersonality::GNU_ObjC:
2996 case EHPersonality::MSVC_X86SEH:
2997 case EHPersonality::MSVC_Win64SEH:
2998 case EHPersonality::MSVC_CXX:
2999 case EHPersonality::CoreCLR:
3000 case EHPersonality::Wasm_CXX:
3001 return TypeInfo->isNullValue();
3003 llvm_unreachable("invalid enum");
3006 static bool shorter_filter(const Value *LHS, const Value *RHS) {
3008 cast<ArrayType>(LHS->getType())->getNumElements()
3010 cast<ArrayType>(RHS->getType())->getNumElements();
3013 Instruction *InstCombiner::visitLandingPadInst(LandingPadInst &LI) {
3014 // The logic here should be correct for any real-world personality function.
3015 // However if that turns out not to be true, the offending logic can always
3016 // be conditioned on the personality function, like the catch-all logic is.
3017 EHPersonality Personality =
3018 classifyEHPersonality(LI.getParent()->getParent()->getPersonalityFn());
3020 // Simplify the list of clauses, eg by removing repeated catch clauses
3021 // (these are often created by inlining).
3022 bool MakeNewInstruction = false; // If true, recreate using the following:
3023 SmallVector<Constant *, 16> NewClauses; // - Clauses for the new instruction;
3024 bool CleanupFlag = LI.isCleanup(); // - The new instruction is a cleanup.
3026 SmallPtrSet<Value *, 16> AlreadyCaught; // Typeinfos known caught already.
3027 for (unsigned i = 0, e = LI.getNumClauses(); i != e; ++i) {
3028 bool isLastClause = i + 1 == e;
3029 if (LI.isCatch(i)) {
3031 Constant *CatchClause = LI.getClause(i);
3032 Constant *TypeInfo = CatchClause->stripPointerCasts();
3034 // If we already saw this clause, there is no point in having a second
3036 if (AlreadyCaught.insert(TypeInfo).second) {
3037 // This catch clause was not already seen.
3038 NewClauses.push_back(CatchClause);
3040 // Repeated catch clause - drop the redundant copy.
3041 MakeNewInstruction = true;
3044 // If this is a catch-all then there is no point in keeping any following
3045 // clauses or marking the landingpad as having a cleanup.
3046 if (isCatchAll(Personality, TypeInfo)) {
3048 MakeNewInstruction = true;
3049 CleanupFlag = false;
3053 // A filter clause. If any of the filter elements were already caught
3054 // then they can be dropped from the filter. It is tempting to try to
3055 // exploit the filter further by saying that any typeinfo that does not
3056 // occur in the filter can't be caught later (and thus can be dropped).
3057 // However this would be wrong, since typeinfos can match without being
3058 // equal (for example if one represents a C++ class, and the other some
3059 // class derived from it).
3060 assert(LI.isFilter(i) && "Unsupported landingpad clause!");
3061 Constant *FilterClause = LI.getClause(i);
3062 ArrayType *FilterType = cast<ArrayType>(FilterClause->getType());
3063 unsigned NumTypeInfos = FilterType->getNumElements();
3065 // An empty filter catches everything, so there is no point in keeping any
3066 // following clauses or marking the landingpad as having a cleanup. By
3067 // dealing with this case here the following code is made a bit simpler.
3068 if (!NumTypeInfos) {
3069 NewClauses.push_back(FilterClause);
3071 MakeNewInstruction = true;
3072 CleanupFlag = false;
3076 bool MakeNewFilter = false; // If true, make a new filter.
3077 SmallVector<Constant *, 16> NewFilterElts; // New elements.
3078 if (isa<ConstantAggregateZero>(FilterClause)) {
3079 // Not an empty filter - it contains at least one null typeinfo.
3080 assert(NumTypeInfos > 0 && "Should have handled empty filter already!");
3081 Constant *TypeInfo =
3082 Constant::getNullValue(FilterType->getElementType());
3083 // If this typeinfo is a catch-all then the filter can never match.
3084 if (isCatchAll(Personality, TypeInfo)) {
3085 // Throw the filter away.
3086 MakeNewInstruction = true;
3090 // There is no point in having multiple copies of this typeinfo, so
3091 // discard all but the first copy if there is more than one.
3092 NewFilterElts.push_back(TypeInfo);
3093 if (NumTypeInfos > 1)
3094 MakeNewFilter = true;
3096 ConstantArray *Filter = cast<ConstantArray>(FilterClause);
3097 SmallPtrSet<Value *, 16> SeenInFilter; // For uniquing the elements.
3098 NewFilterElts.reserve(NumTypeInfos);
3100 // Remove any filter elements that were already caught or that already
3101 // occurred in the filter. While there, see if any of the elements are
3102 // catch-alls. If so, the filter can be discarded.
3103 bool SawCatchAll = false;
3104 for (unsigned j = 0; j != NumTypeInfos; ++j) {
3105 Constant *Elt = Filter->getOperand(j);
3106 Constant *TypeInfo = Elt->stripPointerCasts();
3107 if (isCatchAll(Personality, TypeInfo)) {
3108 // This element is a catch-all. Bail out, noting this fact.
3113 // Even if we've seen a type in a catch clause, we don't want to
3114 // remove it from the filter. An unexpected type handler may be
3115 // set up for a call site which throws an exception of the same
3116 // type caught. In order for the exception thrown by the unexpected
3117 // handler to propagate correctly, the filter must be correctly
3118 // described for the call site.
3122 // void unexpected() { throw 1;}
3123 // void foo() throw (int) {
3124 // std::set_unexpected(unexpected);
3127 // } catch (int i) {}
3130 // There is no point in having multiple copies of the same typeinfo in
3131 // a filter, so only add it if we didn't already.
3132 if (SeenInFilter.insert(TypeInfo).second)
3133 NewFilterElts.push_back(cast<Constant>(Elt));
3135 // A filter containing a catch-all cannot match anything by definition.
3137 // Throw the filter away.
3138 MakeNewInstruction = true;
3142 // If we dropped something from the filter, make a new one.
3143 if (NewFilterElts.size() < NumTypeInfos)
3144 MakeNewFilter = true;
3146 if (MakeNewFilter) {
3147 FilterType = ArrayType::get(FilterType->getElementType(),
3148 NewFilterElts.size());
3149 FilterClause = ConstantArray::get(FilterType, NewFilterElts);
3150 MakeNewInstruction = true;
3153 NewClauses.push_back(FilterClause);
3155 // If the new filter is empty then it will catch everything so there is
3156 // no point in keeping any following clauses or marking the landingpad
3157 // as having a cleanup. The case of the original filter being empty was
3158 // already handled above.
3159 if (MakeNewFilter && !NewFilterElts.size()) {
3160 assert(MakeNewInstruction && "New filter but not a new instruction!");
3161 CleanupFlag = false;
3167 // If several filters occur in a row then reorder them so that the shortest
3168 // filters come first (those with the smallest number of elements). This is
3169 // advantageous because shorter filters are more likely to match, speeding up
3170 // unwinding, but mostly because it increases the effectiveness of the other
3171 // filter optimizations below.
3172 for (unsigned i = 0, e = NewClauses.size(); i + 1 < e; ) {
3174 // Find the maximal 'j' s.t. the range [i, j) consists entirely of filters.
3175 for (j = i; j != e; ++j)
3176 if (!isa<ArrayType>(NewClauses[j]->getType()))
3179 // Check whether the filters are already sorted by length. We need to know
3180 // if sorting them is actually going to do anything so that we only make a
3181 // new landingpad instruction if it does.
3182 for (unsigned k = i; k + 1 < j; ++k)
3183 if (shorter_filter(NewClauses[k+1], NewClauses[k])) {
3184 // Not sorted, so sort the filters now. Doing an unstable sort would be
3185 // correct too but reordering filters pointlessly might confuse users.
3186 std::stable_sort(NewClauses.begin() + i, NewClauses.begin() + j,
3188 MakeNewInstruction = true;
3192 // Look for the next batch of filters.
3196 // If typeinfos matched if and only if equal, then the elements of a filter L
3197 // that occurs later than a filter F could be replaced by the intersection of
3198 // the elements of F and L. In reality two typeinfos can match without being
3199 // equal (for example if one represents a C++ class, and the other some class
3200 // derived from it) so it would be wrong to perform this transform in general.
3201 // However the transform is correct and useful if F is a subset of L. In that
3202 // case L can be replaced by F, and thus removed altogether since repeating a
3203 // filter is pointless. So here we look at all pairs of filters F and L where
3204 // L follows F in the list of clauses, and remove L if every element of F is
3205 // an element of L. This can occur when inlining C++ functions with exception
3207 for (unsigned i = 0; i + 1 < NewClauses.size(); ++i) {
3208 // Examine each filter in turn.
3209 Value *Filter = NewClauses[i];
3210 ArrayType *FTy = dyn_cast<ArrayType>(Filter->getType());
3212 // Not a filter - skip it.
3214 unsigned FElts = FTy->getNumElements();
3215 // Examine each filter following this one. Doing this backwards means that
3216 // we don't have to worry about filters disappearing under us when removed.
3217 for (unsigned j = NewClauses.size() - 1; j != i; --j) {
3218 Value *LFilter = NewClauses[j];
3219 ArrayType *LTy = dyn_cast<ArrayType>(LFilter->getType());
3221 // Not a filter - skip it.
3223 // If Filter is a subset of LFilter, i.e. every element of Filter is also
3224 // an element of LFilter, then discard LFilter.
3225 SmallVectorImpl<Constant *>::iterator J = NewClauses.begin() + j;
3226 // If Filter is empty then it is a subset of LFilter.
3229 NewClauses.erase(J);
3230 MakeNewInstruction = true;
3231 // Move on to the next filter.
3234 unsigned LElts = LTy->getNumElements();
3235 // If Filter is longer than LFilter then it cannot be a subset of it.
3237 // Move on to the next filter.
3239 // At this point we know that LFilter has at least one element.
3240 if (isa<ConstantAggregateZero>(LFilter)) { // LFilter only contains zeros.
3241 // Filter is a subset of LFilter iff Filter contains only zeros (as we
3242 // already know that Filter is not longer than LFilter).
3243 if (isa<ConstantAggregateZero>(Filter)) {
3244 assert(FElts <= LElts && "Should have handled this case earlier!");
3246 NewClauses.erase(J);
3247 MakeNewInstruction = true;
3249 // Move on to the next filter.
3252 ConstantArray *LArray = cast<ConstantArray>(LFilter);
3253 if (isa<ConstantAggregateZero>(Filter)) { // Filter only contains zeros.
3254 // Since Filter is non-empty and contains only zeros, it is a subset of
3255 // LFilter iff LFilter contains a zero.
3256 assert(FElts > 0 && "Should have eliminated the empty filter earlier!");
3257 for (unsigned l = 0; l != LElts; ++l)
3258 if (LArray->getOperand(l)->isNullValue()) {
3259 // LFilter contains a zero - discard it.
3260 NewClauses.erase(J);
3261 MakeNewInstruction = true;
3264 // Move on to the next filter.
3267 // At this point we know that both filters are ConstantArrays. Loop over
3268 // operands to see whether every element of Filter is also an element of
3269 // LFilter. Since filters tend to be short this is probably faster than
3270 // using a method that scales nicely.
3271 ConstantArray *FArray = cast<ConstantArray>(Filter);
3272 bool AllFound = true;
3273 for (unsigned f = 0; f != FElts; ++f) {
3274 Value *FTypeInfo = FArray->getOperand(f)->stripPointerCasts();
3276 for (unsigned l = 0; l != LElts; ++l) {
3277 Value *LTypeInfo = LArray->getOperand(l)->stripPointerCasts();
3278 if (LTypeInfo == FTypeInfo) {
3288 NewClauses.erase(J);
3289 MakeNewInstruction = true;
3291 // Move on to the next filter.
3295 // If we changed any of the clauses, replace the old landingpad instruction
3297 if (MakeNewInstruction) {
3298 LandingPadInst *NLI = LandingPadInst::Create(LI.getType(),
3300 for (unsigned i = 0, e = NewClauses.size(); i != e; ++i)
3301 NLI->addClause(NewClauses[i]);
3302 // A landing pad with no clauses must have the cleanup flag set. It is
3303 // theoretically possible, though highly unlikely, that we eliminated all
3304 // clauses. If so, force the cleanup flag to true.
3305 if (NewClauses.empty())
3307 NLI->setCleanup(CleanupFlag);
3311 // Even if none of the clauses changed, we may nonetheless have understood
3312 // that the cleanup flag is pointless. Clear it if so.
3313 if (LI.isCleanup() != CleanupFlag) {
3314 assert(!CleanupFlag && "Adding a cleanup, not removing one?!");
3315 LI.setCleanup(CleanupFlag);
3322 Instruction *InstCombiner::visitFreeze(FreezeInst &I) {
3323 Value *Op0 = I.getOperand(0);
3325 if (Value *V = SimplifyFreezeInst(Op0, SQ.getWithInstruction(&I)))
3326 return replaceInstUsesWith(I, V);
3331 /// Try to move the specified instruction from its current block into the
3332 /// beginning of DestBlock, which can only happen if it's safe to move the
3333 /// instruction past all of the instructions between it and the end of its
3335 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
3336 assert(I->getSingleUndroppableUse() && "Invariants didn't hold!");
3337 BasicBlock *SrcBlock = I->getParent();
3339 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
3340 if (isa<PHINode>(I) || I->isEHPad() || I->mayHaveSideEffects() ||
3344 // Do not sink static or dynamic alloca instructions. Static allocas must
3345 // remain in the entry block, and dynamic allocas must not be sunk in between
3346 // a stacksave / stackrestore pair, which would incorrectly shorten its
3348 if (isa<AllocaInst>(I))
3351 // Do not sink into catchswitch blocks.
3352 if (isa<CatchSwitchInst>(DestBlock->getTerminator()))
3355 // Do not sink convergent call instructions.
3356 if (auto *CI = dyn_cast<CallInst>(I)) {
3357 if (CI->isConvergent())
3360 // We can only sink load instructions if there is nothing between the load and
3361 // the end of block that could change the value.
3362 if (I->mayReadFromMemory()) {
3363 // We don't want to do any sophisticated alias analysis, so we only check
3364 // the instructions after I in I's parent block if we try to sink to its
3366 if (DestBlock->getUniquePredecessor() != I->getParent())
3368 for (BasicBlock::iterator Scan = I->getIterator(),
3369 E = I->getParent()->end();
3371 if (Scan->mayWriteToMemory())
3375 I->dropDroppableUses([DestBlock](const Use *U) {
3376 if (auto *I = dyn_cast<Instruction>(U->getUser()))
3377 return I->getParent() != DestBlock;
3380 /// FIXME: We could remove droppable uses that are not dominated by
3381 /// the new position.
3383 BasicBlock::iterator InsertPos = DestBlock->getFirstInsertionPt();
3384 I->moveBefore(&*InsertPos);
3387 // Also sink all related debug uses from the source basic block. Otherwise we
3388 // get debug use before the def. Attempt to salvage debug uses first, to
3389 // maximise the range variables have location for. If we cannot salvage, then
3390 // mark the location undef: we know it was supposed to receive a new location
3391 // here, but that computation has been sunk.
3392 SmallVector<DbgVariableIntrinsic *, 2> DbgUsers;
3393 findDbgUsers(DbgUsers, I);
3395 // Update the arguments of a dbg.declare instruction, so that it
3396 // does not point into a sunk instruction.
3397 auto updateDbgDeclare = [&I](DbgVariableIntrinsic *DII) {
3398 if (!isa<DbgDeclareInst>(DII))
3401 if (isa<CastInst>(I))
3403 0, MetadataAsValue::get(I->getContext(),
3404 ValueAsMetadata::get(I->getOperand(0))));
3408 SmallVector<DbgVariableIntrinsic *, 2> DIIClones;
3409 for (auto User : DbgUsers) {
3410 // A dbg.declare instruction should not be cloned, since there can only be
3411 // one per variable fragment. It should be left in the original place
3412 // because the sunk instruction is not an alloca (otherwise we could not be
3414 if (User->getParent() != SrcBlock || updateDbgDeclare(User))
3417 DIIClones.emplace_back(cast<DbgVariableIntrinsic>(User->clone()));
3418 LLVM_DEBUG(dbgs() << "CLONE: " << *DIIClones.back() << '\n');
3421 // Perform salvaging without the clones, then sink the clones.
3422 if (!DIIClones.empty()) {
3423 salvageDebugInfoForDbgValues(*I, DbgUsers);
3424 for (auto &DIIClone : DIIClones) {
3425 DIIClone->insertBefore(&*InsertPos);
3426 LLVM_DEBUG(dbgs() << "SINK: " << *DIIClone << '\n');
3433 bool InstCombiner::run() {
3434 while (!Worklist.isEmpty()) {
3435 // Walk deferred instructions in reverse order, and push them to the
3436 // worklist, which means they'll end up popped from the worklist in-order.
3437 while (Instruction *I = Worklist.popDeferred()) {
3438 // Check to see if we can DCE the instruction. We do this already here to
3439 // reduce the number of uses and thus allow other folds to trigger.
3440 // Note that eraseInstFromFunction() may push additional instructions on
3441 // the deferred worklist, so this will DCE whole instruction chains.
3442 if (isInstructionTriviallyDead(I, &TLI)) {
3443 eraseInstFromFunction(*I);
3451 Instruction *I = Worklist.removeOne();
3452 if (I == nullptr) continue; // skip null values.
3454 // Check to see if we can DCE the instruction.
3455 if (isInstructionTriviallyDead(I, &TLI)) {
3456 eraseInstFromFunction(*I);
3461 if (!DebugCounter::shouldExecute(VisitCounter))
3464 // Instruction isn't dead, see if we can constant propagate it.
3465 if (!I->use_empty() &&
3466 (I->getNumOperands() == 0 || isa<Constant>(I->getOperand(0)))) {
3467 if (Constant *C = ConstantFoldInstruction(I, DL, &TLI)) {
3468 LLVM_DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " << *I
3471 // Add operands to the worklist.
3472 replaceInstUsesWith(*I, C);
3474 if (isInstructionTriviallyDead(I, &TLI))
3475 eraseInstFromFunction(*I);
3476 MadeIRChange = true;
3481 // See if we can trivially sink this instruction to its user if we can
3482 // prove that the successor is not executed more frequently than our block.
3483 if (EnableCodeSinking)
3484 if (Use *SingleUse = I->getSingleUndroppableUse()) {
3485 BasicBlock *BB = I->getParent();
3486 Instruction *UserInst = cast<Instruction>(SingleUse->getUser());
3487 BasicBlock *UserParent;
3489 // Get the block the use occurs in.
3490 if (PHINode *PN = dyn_cast<PHINode>(UserInst))
3491 UserParent = PN->getIncomingBlock(*SingleUse);
3493 UserParent = UserInst->getParent();
3495 if (UserParent != BB) {
3496 // See if the user is one of our successors that has only one
3497 // predecessor, so that we don't have to split the critical edge.
3498 bool ShouldSink = UserParent->getUniquePredecessor() == BB;
3499 // Another option where we can sink is a block that ends with a
3500 // terminator that does not pass control to other block (such as
3501 // return or unreachable). In this case:
3502 // - I dominates the User (by SSA form);
3503 // - the User will be executed at most once.
3504 // So sinking I down to User is always profitable or neutral.
3506 auto *Term = UserParent->getTerminator();
3507 ShouldSink = isa<ReturnInst>(Term) || isa<UnreachableInst>(Term);
3510 assert(DT.dominates(BB, UserParent) &&
3511 "Dominance relation broken?");
3512 // Okay, the CFG is simple enough, try to sink this instruction.
3513 if (TryToSinkInstruction(I, UserParent)) {
3514 LLVM_DEBUG(dbgs() << "IC: Sink: " << *I << '\n');
3515 MadeIRChange = true;
3516 // We'll add uses of the sunk instruction below, but since sinking
3517 // can expose opportunities for it's *operands* add them to the
3519 for (Use &U : I->operands())
3520 if (Instruction *OpI = dyn_cast<Instruction>(U.get()))
3527 // Now that we have an instruction, try combining it to simplify it.
3528 Builder.SetInsertPoint(I);
3529 Builder.SetCurrentDebugLocation(I->getDebugLoc());
3534 LLVM_DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str(););
3535 LLVM_DEBUG(dbgs() << "IC: Visiting: " << OrigI << '\n');
3537 if (Instruction *Result = visit(*I)) {
3539 // Should we replace the old instruction with a new one?
3541 LLVM_DEBUG(dbgs() << "IC: Old = " << *I << '\n'
3542 << " New = " << *Result << '\n');
3544 if (I->getDebugLoc())
3545 Result->setDebugLoc(I->getDebugLoc());
3546 // Everything uses the new instruction now.
3547 I->replaceAllUsesWith(Result);
3549 // Move the name to the new instruction first.
3550 Result->takeName(I);
3552 // Insert the new instruction into the basic block...
3553 BasicBlock *InstParent = I->getParent();
3554 BasicBlock::iterator InsertPos = I->getIterator();
3556 // If we replace a PHI with something that isn't a PHI, fix up the
3558 if (!isa<PHINode>(Result) && isa<PHINode>(InsertPos))
3559 InsertPos = InstParent->getFirstInsertionPt();
3561 InstParent->getInstList().insert(InsertPos, Result);
3563 // Push the new instruction and any users onto the worklist.
3564 Worklist.pushUsersToWorkList(*Result);
3565 Worklist.push(Result);
3567 eraseInstFromFunction(*I);
3569 LLVM_DEBUG(dbgs() << "IC: Mod = " << OrigI << '\n'
3570 << " New = " << *I << '\n');
3572 // If the instruction was modified, it's possible that it is now dead.
3573 // if so, remove it.
3574 if (isInstructionTriviallyDead(I, &TLI)) {
3575 eraseInstFromFunction(*I);
3577 Worklist.pushUsersToWorkList(*I);
3581 MadeIRChange = true;
3586 return MadeIRChange;
3589 /// Populate the IC worklist from a function, by walking it in depth-first
3590 /// order and adding all reachable code to the worklist.
3592 /// This has a couple of tricks to make the code faster and more powerful. In
3593 /// particular, we constant fold and DCE instructions as we go, to avoid adding
3594 /// them to the worklist (this significantly speeds up instcombine on code where
3595 /// many instructions are dead or constant). Additionally, if we find a branch
3596 /// whose condition is a known constant, we only visit the reachable successors.
3597 static bool prepareICWorklistFromFunction(Function &F, const DataLayout &DL,
3598 const TargetLibraryInfo *TLI,
3599 InstCombineWorklist &ICWorklist) {
3600 bool MadeIRChange = false;
3601 SmallPtrSet<BasicBlock *, 32> Visited;
3602 SmallVector<BasicBlock*, 256> Worklist;
3603 Worklist.push_back(&F.front());
3605 SmallVector<Instruction*, 128> InstrsForInstCombineWorklist;
3606 DenseMap<Constant *, Constant *> FoldedConstants;
3609 BasicBlock *BB = Worklist.pop_back_val();
3611 // We have now visited this block! If we've already been here, ignore it.
3612 if (!Visited.insert(BB).second)
3615 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
3616 Instruction *Inst = &*BBI++;
3618 // ConstantProp instruction if trivially constant.
3619 if (!Inst->use_empty() &&
3620 (Inst->getNumOperands() == 0 || isa<Constant>(Inst->getOperand(0))))
3621 if (Constant *C = ConstantFoldInstruction(Inst, DL, TLI)) {
3622 LLVM_DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " << *Inst
3624 Inst->replaceAllUsesWith(C);
3626 if (isInstructionTriviallyDead(Inst, TLI))
3627 Inst->eraseFromParent();
3628 MadeIRChange = true;
3632 // See if we can constant fold its operands.
3633 for (Use &U : Inst->operands()) {
3634 if (!isa<ConstantVector>(U) && !isa<ConstantExpr>(U))
3637 auto *C = cast<Constant>(U);
3638 Constant *&FoldRes = FoldedConstants[C];
3640 FoldRes = ConstantFoldConstant(C, DL, TLI);
3643 LLVM_DEBUG(dbgs() << "IC: ConstFold operand of: " << *Inst
3644 << "\n Old = " << *C
3645 << "\n New = " << *FoldRes << '\n');
3647 MadeIRChange = true;
3651 // Skip processing debug intrinsics in InstCombine. Processing these call instructions
3652 // consumes non-trivial amount of time and provides no value for the optimization.
3653 if (!isa<DbgInfoIntrinsic>(Inst))
3654 InstrsForInstCombineWorklist.push_back(Inst);
3657 // Recursively visit successors. If this is a branch or switch on a
3658 // constant, only visit the reachable successor.
3659 Instruction *TI = BB->getTerminator();
3660 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
3661 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
3662 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
3663 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
3664 Worklist.push_back(ReachableBB);
3667 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
3668 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
3669 Worklist.push_back(SI->findCaseValue(Cond)->getCaseSuccessor());
3674 for (BasicBlock *SuccBB : successors(TI))
3675 Worklist.push_back(SuccBB);
3676 } while (!Worklist.empty());
3678 // Remove instructions inside unreachable blocks. This prevents the
3679 // instcombine code from having to deal with some bad special cases, and
3680 // reduces use counts of instructions.
3681 for (BasicBlock &BB : F) {
3682 if (Visited.count(&BB))
3685 unsigned NumDeadInstInBB = removeAllNonTerminatorAndEHPadInstructions(&BB);
3686 MadeIRChange |= NumDeadInstInBB > 0;
3687 NumDeadInst += NumDeadInstInBB;
3690 // Once we've found all of the instructions to add to instcombine's worklist,
3691 // add them in reverse order. This way instcombine will visit from the top
3692 // of the function down. This jives well with the way that it adds all uses
3693 // of instructions to the worklist after doing a transformation, thus avoiding
3694 // some N^2 behavior in pathological cases.
3695 ICWorklist.reserve(InstrsForInstCombineWorklist.size());
3696 for (Instruction *Inst : reverse(InstrsForInstCombineWorklist)) {
3697 // DCE instruction if trivially dead. As we iterate in reverse program
3698 // order here, we will clean up whole chains of dead instructions.
3699 if (isInstructionTriviallyDead(Inst, TLI)) {
3701 LLVM_DEBUG(dbgs() << "IC: DCE: " << *Inst << '\n');
3702 salvageDebugInfo(*Inst);
3703 Inst->eraseFromParent();
3704 MadeIRChange = true;
3708 ICWorklist.push(Inst);
3711 return MadeIRChange;
3714 static bool combineInstructionsOverFunction(
3715 Function &F, InstCombineWorklist &Worklist, AliasAnalysis *AA,
3716 AssumptionCache &AC, TargetLibraryInfo &TLI, DominatorTree &DT,
3717 OptimizationRemarkEmitter &ORE, BlockFrequencyInfo *BFI,
3718 ProfileSummaryInfo *PSI, unsigned MaxIterations, LoopInfo *LI) {
3719 auto &DL = F.getParent()->getDataLayout();
3720 MaxIterations = std::min(MaxIterations, LimitMaxIterations.getValue());
3722 /// Builder - This is an IRBuilder that automatically inserts new
3723 /// instructions into the worklist when they are created.
3724 IRBuilder<TargetFolder, IRBuilderCallbackInserter> Builder(
3725 F.getContext(), TargetFolder(DL),
3726 IRBuilderCallbackInserter([&Worklist, &AC](Instruction *I) {
3728 if (match(I, m_Intrinsic<Intrinsic::assume>()))
3729 AC.registerAssumption(cast<CallInst>(I));
3732 // Lower dbg.declare intrinsics otherwise their value may be clobbered
3734 bool MadeIRChange = false;
3735 if (ShouldLowerDbgDeclare)
3736 MadeIRChange = LowerDbgDeclare(F);
3738 // Iterate while there is work to do.
3739 unsigned Iteration = 0;
3743 if (Iteration > InfiniteLoopDetectionThreshold) {
3745 "Instruction Combining seems stuck in an infinite loop after " +
3746 Twine(InfiniteLoopDetectionThreshold) + " iterations.");
3749 if (Iteration > MaxIterations) {
3750 LLVM_DEBUG(dbgs() << "\n\n[IC] Iteration limit #" << MaxIterations
3751 << " on " << F.getName()
3752 << " reached; stopping before reaching a fixpoint\n");
3756 LLVM_DEBUG(dbgs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
3757 << F.getName() << "\n");
3759 MadeIRChange |= prepareICWorklistFromFunction(F, DL, &TLI, Worklist);
3761 InstCombiner IC(Worklist, Builder, F.hasMinSize(), AA,
3762 AC, TLI, DT, ORE, BFI, PSI, DL, LI);
3763 IC.MaxArraySizeForCombine = MaxArraySize;
3768 MadeIRChange = true;
3771 return MadeIRChange;
3774 InstCombinePass::InstCombinePass() : MaxIterations(LimitMaxIterations) {}
3776 InstCombinePass::InstCombinePass(unsigned MaxIterations)
3777 : MaxIterations(MaxIterations) {}
3779 PreservedAnalyses InstCombinePass::run(Function &F,
3780 FunctionAnalysisManager &AM) {
3781 auto &AC = AM.getResult<AssumptionAnalysis>(F);
3782 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
3783 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
3784 auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F);
3786 auto *LI = AM.getCachedResult<LoopAnalysis>(F);
3788 auto *AA = &AM.getResult<AAManager>(F);
3789 auto &MAMProxy = AM.getResult<ModuleAnalysisManagerFunctionProxy>(F);
3790 ProfileSummaryInfo *PSI =
3791 MAMProxy.getCachedResult<ProfileSummaryAnalysis>(*F.getParent());
3792 auto *BFI = (PSI && PSI->hasProfileSummary()) ?
3793 &AM.getResult<BlockFrequencyAnalysis>(F) : nullptr;
3795 if (!combineInstructionsOverFunction(F, Worklist, AA, AC, TLI, DT, ORE, BFI,
3796 PSI, MaxIterations, LI))
3797 // No changes, all analyses are preserved.
3798 return PreservedAnalyses::all();
3800 // Mark all the analyses that instcombine updates as preserved.
3801 PreservedAnalyses PA;
3802 PA.preserveSet<CFGAnalyses>();
3803 PA.preserve<AAManager>();
3804 PA.preserve<BasicAA>();
3805 PA.preserve<GlobalsAA>();
3809 void InstructionCombiningPass::getAnalysisUsage(AnalysisUsage &AU) const {
3810 AU.setPreservesCFG();
3811 AU.addRequired<AAResultsWrapperPass>();
3812 AU.addRequired<AssumptionCacheTracker>();
3813 AU.addRequired<TargetLibraryInfoWrapperPass>();
3814 AU.addRequired<DominatorTreeWrapperPass>();
3815 AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
3816 AU.addPreserved<DominatorTreeWrapperPass>();
3817 AU.addPreserved<AAResultsWrapperPass>();
3818 AU.addPreserved<BasicAAWrapperPass>();
3819 AU.addPreserved<GlobalsAAWrapperPass>();
3820 AU.addRequired<ProfileSummaryInfoWrapperPass>();
3821 LazyBlockFrequencyInfoPass::getLazyBFIAnalysisUsage(AU);
3824 bool InstructionCombiningPass::runOnFunction(Function &F) {
3825 if (skipFunction(F))
3828 // Required analyses.
3829 auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
3830 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
3831 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
3832 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
3833 auto &ORE = getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
3835 // Optional analyses.
3836 auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>();
3837 auto *LI = LIWP ? &LIWP->getLoopInfo() : nullptr;
3838 ProfileSummaryInfo *PSI =
3839 &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
3840 BlockFrequencyInfo *BFI =
3841 (PSI && PSI->hasProfileSummary()) ?
3842 &getAnalysis<LazyBlockFrequencyInfoPass>().getBFI() :
3845 return combineInstructionsOverFunction(F, Worklist, AA, AC, TLI, DT, ORE, BFI,
3846 PSI, MaxIterations, LI);
3849 char InstructionCombiningPass::ID = 0;
3851 InstructionCombiningPass::InstructionCombiningPass()
3852 : FunctionPass(ID), MaxIterations(InstCombineDefaultMaxIterations) {
3853 initializeInstructionCombiningPassPass(*PassRegistry::getPassRegistry());
3856 InstructionCombiningPass::InstructionCombiningPass(unsigned MaxIterations)
3857 : FunctionPass(ID), MaxIterations(MaxIterations) {
3858 initializeInstructionCombiningPassPass(*PassRegistry::getPassRegistry());
3861 INITIALIZE_PASS_BEGIN(InstructionCombiningPass, "instcombine",
3862 "Combine redundant instructions", false, false)
3863 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
3864 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
3865 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
3866 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
3867 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
3868 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)
3869 INITIALIZE_PASS_DEPENDENCY(LazyBlockFrequencyInfoPass)
3870 INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)
3871 INITIALIZE_PASS_END(InstructionCombiningPass, "instcombine",
3872 "Combine redundant instructions", false, false)
3874 // Initialization Routines
3875 void llvm::initializeInstCombine(PassRegistry &Registry) {
3876 initializeInstructionCombiningPassPass(Registry);
3879 void LLVMInitializeInstCombine(LLVMPassRegistryRef R) {
3880 initializeInstructionCombiningPassPass(*unwrap(R));
3883 FunctionPass *llvm::createInstructionCombiningPass() {
3884 return new InstructionCombiningPass();
3887 FunctionPass *llvm::createInstructionCombiningPass(unsigned MaxIterations) {
3888 return new InstructionCombiningPass(MaxIterations);
3891 void LLVMAddInstructionCombiningPass(LLVMPassManagerRef PM) {
3892 unwrap(PM)->add(createInstructionCombiningPass());