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/TargetTransformInfo.h"
63 #include "llvm/Analysis/ValueTracking.h"
64 #include "llvm/Analysis/VectorUtils.h"
65 #include "llvm/IR/BasicBlock.h"
66 #include "llvm/IR/CFG.h"
67 #include "llvm/IR/Constant.h"
68 #include "llvm/IR/Constants.h"
69 #include "llvm/IR/DIBuilder.h"
70 #include "llvm/IR/DataLayout.h"
71 #include "llvm/IR/DebugInfo.h"
72 #include "llvm/IR/DerivedTypes.h"
73 #include "llvm/IR/Dominators.h"
74 #include "llvm/IR/Function.h"
75 #include "llvm/IR/GetElementPtrTypeIterator.h"
76 #include "llvm/IR/IRBuilder.h"
77 #include "llvm/IR/InstrTypes.h"
78 #include "llvm/IR/Instruction.h"
79 #include "llvm/IR/Instructions.h"
80 #include "llvm/IR/IntrinsicInst.h"
81 #include "llvm/IR/Intrinsics.h"
82 #include "llvm/IR/LegacyPassManager.h"
83 #include "llvm/IR/Metadata.h"
84 #include "llvm/IR/Operator.h"
85 #include "llvm/IR/PassManager.h"
86 #include "llvm/IR/PatternMatch.h"
87 #include "llvm/IR/Type.h"
88 #include "llvm/IR/Use.h"
89 #include "llvm/IR/User.h"
90 #include "llvm/IR/Value.h"
91 #include "llvm/IR/ValueHandle.h"
92 #include "llvm/InitializePasses.h"
93 #include "llvm/Pass.h"
94 #include "llvm/Support/CBindingWrapping.h"
95 #include "llvm/Support/Casting.h"
96 #include "llvm/Support/CommandLine.h"
97 #include "llvm/Support/Compiler.h"
98 #include "llvm/Support/Debug.h"
99 #include "llvm/Support/DebugCounter.h"
100 #include "llvm/Support/ErrorHandling.h"
101 #include "llvm/Support/KnownBits.h"
102 #include "llvm/Support/raw_ostream.h"
103 #include "llvm/Transforms/InstCombine/InstCombine.h"
104 #include "llvm/Transforms/Utils/Local.h"
112 #define DEBUG_TYPE "instcombine"
113 #include "llvm/Transforms/Utils/InstructionWorklist.h"
115 using namespace llvm;
116 using namespace llvm::PatternMatch;
118 STATISTIC(NumWorklistIterations,
119 "Number of instruction combining iterations performed");
121 STATISTIC(NumCombined , "Number of insts combined");
122 STATISTIC(NumConstProp, "Number of constant folds");
123 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
124 STATISTIC(NumSunkInst , "Number of instructions sunk");
125 STATISTIC(NumExpand, "Number of expansions");
126 STATISTIC(NumFactor , "Number of factorizations");
127 STATISTIC(NumReassoc , "Number of reassociations");
128 DEBUG_COUNTER(VisitCounter, "instcombine-visit",
129 "Controls which instructions are visited");
131 // FIXME: these limits eventually should be as low as 2.
132 static constexpr unsigned InstCombineDefaultMaxIterations = 1000;
134 static constexpr unsigned InstCombineDefaultInfiniteLoopThreshold = 100;
136 static constexpr unsigned InstCombineDefaultInfiniteLoopThreshold = 1000;
140 EnableCodeSinking("instcombine-code-sinking", cl::desc("Enable code sinking"),
143 static cl::opt<unsigned> LimitMaxIterations(
144 "instcombine-max-iterations",
145 cl::desc("Limit the maximum number of instruction combining iterations"),
146 cl::init(InstCombineDefaultMaxIterations));
148 static cl::opt<unsigned> InfiniteLoopDetectionThreshold(
149 "instcombine-infinite-loop-threshold",
150 cl::desc("Number of instruction combining iterations considered an "
152 cl::init(InstCombineDefaultInfiniteLoopThreshold), cl::Hidden);
154 static cl::opt<unsigned>
155 MaxArraySize("instcombine-maxarray-size", cl::init(1024),
156 cl::desc("Maximum array size considered when doing a combine"));
158 // FIXME: Remove this flag when it is no longer necessary to convert
159 // llvm.dbg.declare to avoid inaccurate debug info. Setting this to false
160 // increases variable availability at the cost of accuracy. Variables that
161 // cannot be promoted by mem2reg or SROA will be described as living in memory
162 // for their entire lifetime. However, passes like DSE and instcombine can
163 // delete stores to the alloca, leading to misleading and inaccurate debug
164 // information. This flag can be removed when those passes are fixed.
165 static cl::opt<unsigned> ShouldLowerDbgDeclare("instcombine-lower-dbg-declare",
166 cl::Hidden, cl::init(true));
168 Optional<Instruction *>
169 InstCombiner::targetInstCombineIntrinsic(IntrinsicInst &II) {
170 // Handle target specific intrinsics
171 if (II.getCalledFunction()->isTargetIntrinsic()) {
172 return TTI.instCombineIntrinsic(*this, II);
177 Optional<Value *> InstCombiner::targetSimplifyDemandedUseBitsIntrinsic(
178 IntrinsicInst &II, APInt DemandedMask, KnownBits &Known,
179 bool &KnownBitsComputed) {
180 // Handle target specific intrinsics
181 if (II.getCalledFunction()->isTargetIntrinsic()) {
182 return TTI.simplifyDemandedUseBitsIntrinsic(*this, II, DemandedMask, Known,
188 Optional<Value *> InstCombiner::targetSimplifyDemandedVectorEltsIntrinsic(
189 IntrinsicInst &II, APInt DemandedElts, APInt &UndefElts, APInt &UndefElts2,
191 std::function<void(Instruction *, unsigned, APInt, APInt &)>
193 // Handle target specific intrinsics
194 if (II.getCalledFunction()->isTargetIntrinsic()) {
195 return TTI.simplifyDemandedVectorEltsIntrinsic(
196 *this, II, DemandedElts, UndefElts, UndefElts2, UndefElts3,
202 Value *InstCombinerImpl::EmitGEPOffset(User *GEP) {
203 return llvm::EmitGEPOffset(&Builder, DL, GEP);
206 /// Legal integers and common types are considered desirable. This is used to
207 /// avoid creating instructions with types that may not be supported well by the
209 /// NOTE: This treats i8, i16 and i32 specially because they are common
210 /// types in frontend languages.
211 bool InstCombinerImpl::isDesirableIntType(unsigned BitWidth) const {
218 return DL.isLegalInteger(BitWidth);
222 /// Return true if it is desirable to convert an integer computation from a
223 /// given bit width to a new bit width.
224 /// We don't want to convert from a legal to an illegal type or from a smaller
225 /// to a larger illegal type. A width of '1' is always treated as a desirable
226 /// type because i1 is a fundamental type in IR, and there are many specialized
227 /// optimizations for i1 types. Common/desirable widths are equally treated as
228 /// legal to convert to, in order to open up more combining opportunities.
229 bool InstCombinerImpl::shouldChangeType(unsigned FromWidth,
230 unsigned ToWidth) const {
231 bool FromLegal = FromWidth == 1 || DL.isLegalInteger(FromWidth);
232 bool ToLegal = ToWidth == 1 || DL.isLegalInteger(ToWidth);
234 // Convert to desirable widths even if they are not legal types.
235 // Only shrink types, to prevent infinite loops.
236 if (ToWidth < FromWidth && isDesirableIntType(ToWidth))
239 // If this is a legal integer from type, and the result would be an illegal
240 // type, don't do the transformation.
241 if (FromLegal && !ToLegal)
244 // Otherwise, if both are illegal, do not increase the size of the result. We
245 // do allow things like i160 -> i64, but not i64 -> i160.
246 if (!FromLegal && !ToLegal && ToWidth > FromWidth)
252 /// Return true if it is desirable to convert a computation from 'From' to 'To'.
253 /// We don't want to convert from a legal to an illegal type or from a smaller
254 /// to a larger illegal type. i1 is always treated as a legal type because it is
255 /// a fundamental type in IR, and there are many specialized optimizations for
257 bool InstCombinerImpl::shouldChangeType(Type *From, Type *To) const {
258 // TODO: This could be extended to allow vectors. Datalayout changes might be
259 // needed to properly support that.
260 if (!From->isIntegerTy() || !To->isIntegerTy())
263 unsigned FromWidth = From->getPrimitiveSizeInBits();
264 unsigned ToWidth = To->getPrimitiveSizeInBits();
265 return shouldChangeType(FromWidth, ToWidth);
268 // Return true, if No Signed Wrap should be maintained for I.
269 // The No Signed Wrap flag can be kept if the operation "B (I.getOpcode) C",
270 // where both B and C should be ConstantInts, results in a constant that does
271 // not overflow. This function only handles the Add and Sub opcodes. For
272 // all other opcodes, the function conservatively returns false.
273 static bool maintainNoSignedWrap(BinaryOperator &I, Value *B, Value *C) {
274 auto *OBO = dyn_cast<OverflowingBinaryOperator>(&I);
275 if (!OBO || !OBO->hasNoSignedWrap())
278 // We reason about Add and Sub Only.
279 Instruction::BinaryOps Opcode = I.getOpcode();
280 if (Opcode != Instruction::Add && Opcode != Instruction::Sub)
283 const APInt *BVal, *CVal;
284 if (!match(B, m_APInt(BVal)) || !match(C, m_APInt(CVal)))
287 bool Overflow = false;
288 if (Opcode == Instruction::Add)
289 (void)BVal->sadd_ov(*CVal, Overflow);
291 (void)BVal->ssub_ov(*CVal, Overflow);
296 static bool hasNoUnsignedWrap(BinaryOperator &I) {
297 auto *OBO = dyn_cast<OverflowingBinaryOperator>(&I);
298 return OBO && OBO->hasNoUnsignedWrap();
301 static bool hasNoSignedWrap(BinaryOperator &I) {
302 auto *OBO = dyn_cast<OverflowingBinaryOperator>(&I);
303 return OBO && OBO->hasNoSignedWrap();
306 /// Conservatively clears subclassOptionalData after a reassociation or
307 /// commutation. We preserve fast-math flags when applicable as they can be
309 static void ClearSubclassDataAfterReassociation(BinaryOperator &I) {
310 FPMathOperator *FPMO = dyn_cast<FPMathOperator>(&I);
312 I.clearSubclassOptionalData();
316 FastMathFlags FMF = I.getFastMathFlags();
317 I.clearSubclassOptionalData();
318 I.setFastMathFlags(FMF);
321 /// Combine constant operands of associative operations either before or after a
322 /// cast to eliminate one of the associative operations:
323 /// (op (cast (op X, C2)), C1) --> (cast (op X, op (C1, C2)))
324 /// (op (cast (op X, C2)), C1) --> (op (cast X), op (C1, C2))
325 static bool simplifyAssocCastAssoc(BinaryOperator *BinOp1,
326 InstCombinerImpl &IC) {
327 auto *Cast = dyn_cast<CastInst>(BinOp1->getOperand(0));
328 if (!Cast || !Cast->hasOneUse())
331 // TODO: Enhance logic for other casts and remove this check.
332 auto CastOpcode = Cast->getOpcode();
333 if (CastOpcode != Instruction::ZExt)
336 // TODO: Enhance logic for other BinOps and remove this check.
337 if (!BinOp1->isBitwiseLogicOp())
340 auto AssocOpcode = BinOp1->getOpcode();
341 auto *BinOp2 = dyn_cast<BinaryOperator>(Cast->getOperand(0));
342 if (!BinOp2 || !BinOp2->hasOneUse() || BinOp2->getOpcode() != AssocOpcode)
346 if (!match(BinOp1->getOperand(1), m_Constant(C1)) ||
347 !match(BinOp2->getOperand(1), m_Constant(C2)))
350 // TODO: This assumes a zext cast.
351 // Eg, if it was a trunc, we'd cast C1 to the source type because casting C2
352 // to the destination type might lose bits.
354 // Fold the constants together in the destination type:
355 // (op (cast (op X, C2)), C1) --> (op (cast X), FoldedC)
356 Type *DestTy = C1->getType();
357 Constant *CastC2 = ConstantExpr::getCast(CastOpcode, C2, DestTy);
358 Constant *FoldedC = ConstantExpr::get(AssocOpcode, C1, CastC2);
359 IC.replaceOperand(*Cast, 0, BinOp2->getOperand(0));
360 IC.replaceOperand(*BinOp1, 1, FoldedC);
364 // Simplifies IntToPtr/PtrToInt RoundTrip Cast To BitCast.
365 // inttoptr ( ptrtoint (x) ) --> x
366 Value *InstCombinerImpl::simplifyIntToPtrRoundTripCast(Value *Val) {
367 auto *IntToPtr = dyn_cast<IntToPtrInst>(Val);
368 if (IntToPtr && DL.getPointerTypeSizeInBits(IntToPtr->getDestTy()) ==
369 DL.getTypeSizeInBits(IntToPtr->getSrcTy())) {
370 auto *PtrToInt = dyn_cast<PtrToIntInst>(IntToPtr->getOperand(0));
371 Type *CastTy = IntToPtr->getDestTy();
373 CastTy->getPointerAddressSpace() ==
374 PtrToInt->getSrcTy()->getPointerAddressSpace() &&
375 DL.getPointerTypeSizeInBits(PtrToInt->getSrcTy()) ==
376 DL.getTypeSizeInBits(PtrToInt->getDestTy())) {
377 return CastInst::CreateBitOrPointerCast(PtrToInt->getOperand(0), CastTy,
384 /// This performs a few simplifications for operators that are associative or
387 /// Commutative operators:
389 /// 1. Order operands such that they are listed from right (least complex) to
390 /// left (most complex). This puts constants before unary operators before
391 /// binary operators.
393 /// Associative operators:
395 /// 2. Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
396 /// 3. Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
398 /// Associative and commutative operators:
400 /// 4. Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
401 /// 5. Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
402 /// 6. Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
403 /// if C1 and C2 are constants.
404 bool InstCombinerImpl::SimplifyAssociativeOrCommutative(BinaryOperator &I) {
405 Instruction::BinaryOps Opcode = I.getOpcode();
406 bool Changed = false;
409 // Order operands such that they are listed from right (least complex) to
410 // left (most complex). This puts constants before unary operators before
412 if (I.isCommutative() && getComplexity(I.getOperand(0)) <
413 getComplexity(I.getOperand(1)))
414 Changed = !I.swapOperands();
416 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
417 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
419 if (I.isAssociative()) {
420 // Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
421 if (Op0 && Op0->getOpcode() == Opcode) {
422 Value *A = Op0->getOperand(0);
423 Value *B = Op0->getOperand(1);
424 Value *C = I.getOperand(1);
426 // Does "B op C" simplify?
427 if (Value *V = SimplifyBinOp(Opcode, B, C, SQ.getWithInstruction(&I))) {
428 // It simplifies to V. Form "A op V".
429 replaceOperand(I, 0, A);
430 replaceOperand(I, 1, V);
431 bool IsNUW = hasNoUnsignedWrap(I) && hasNoUnsignedWrap(*Op0);
432 bool IsNSW = maintainNoSignedWrap(I, B, C) && hasNoSignedWrap(*Op0);
434 // Conservatively clear all optional flags since they may not be
435 // preserved by the reassociation. Reset nsw/nuw based on the above
437 ClearSubclassDataAfterReassociation(I);
439 // Note: this is only valid because SimplifyBinOp doesn't look at
440 // the operands to Op0.
442 I.setHasNoUnsignedWrap(true);
445 I.setHasNoSignedWrap(true);
453 // Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
454 if (Op1 && Op1->getOpcode() == Opcode) {
455 Value *A = I.getOperand(0);
456 Value *B = Op1->getOperand(0);
457 Value *C = Op1->getOperand(1);
459 // Does "A op B" simplify?
460 if (Value *V = SimplifyBinOp(Opcode, A, B, SQ.getWithInstruction(&I))) {
461 // It simplifies to V. Form "V op C".
462 replaceOperand(I, 0, V);
463 replaceOperand(I, 1, C);
464 // Conservatively clear the optional flags, since they may not be
465 // preserved by the reassociation.
466 ClearSubclassDataAfterReassociation(I);
474 if (I.isAssociative() && I.isCommutative()) {
475 if (simplifyAssocCastAssoc(&I, *this)) {
481 // Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
482 if (Op0 && Op0->getOpcode() == Opcode) {
483 Value *A = Op0->getOperand(0);
484 Value *B = Op0->getOperand(1);
485 Value *C = I.getOperand(1);
487 // Does "C op A" simplify?
488 if (Value *V = SimplifyBinOp(Opcode, C, A, SQ.getWithInstruction(&I))) {
489 // It simplifies to V. Form "V op B".
490 replaceOperand(I, 0, V);
491 replaceOperand(I, 1, B);
492 // Conservatively clear the optional flags, since they may not be
493 // preserved by the reassociation.
494 ClearSubclassDataAfterReassociation(I);
501 // Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
502 if (Op1 && Op1->getOpcode() == Opcode) {
503 Value *A = I.getOperand(0);
504 Value *B = Op1->getOperand(0);
505 Value *C = Op1->getOperand(1);
507 // Does "C op A" simplify?
508 if (Value *V = SimplifyBinOp(Opcode, C, A, SQ.getWithInstruction(&I))) {
509 // It simplifies to V. Form "B op V".
510 replaceOperand(I, 0, B);
511 replaceOperand(I, 1, V);
512 // Conservatively clear the optional flags, since they may not be
513 // preserved by the reassociation.
514 ClearSubclassDataAfterReassociation(I);
521 // Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
522 // if C1 and C2 are constants.
526 Op0->getOpcode() == Opcode && Op1->getOpcode() == Opcode &&
527 match(Op0, m_OneUse(m_BinOp(m_Value(A), m_Constant(C1)))) &&
528 match(Op1, m_OneUse(m_BinOp(m_Value(B), m_Constant(C2))))) {
529 bool IsNUW = hasNoUnsignedWrap(I) &&
530 hasNoUnsignedWrap(*Op0) &&
531 hasNoUnsignedWrap(*Op1);
532 BinaryOperator *NewBO = (IsNUW && Opcode == Instruction::Add) ?
533 BinaryOperator::CreateNUW(Opcode, A, B) :
534 BinaryOperator::Create(Opcode, A, B);
536 if (isa<FPMathOperator>(NewBO)) {
537 FastMathFlags Flags = I.getFastMathFlags();
538 Flags &= Op0->getFastMathFlags();
539 Flags &= Op1->getFastMathFlags();
540 NewBO->setFastMathFlags(Flags);
542 InsertNewInstWith(NewBO, I);
543 NewBO->takeName(Op1);
544 replaceOperand(I, 0, NewBO);
545 replaceOperand(I, 1, ConstantExpr::get(Opcode, C1, C2));
546 // Conservatively clear the optional flags, since they may not be
547 // preserved by the reassociation.
548 ClearSubclassDataAfterReassociation(I);
550 I.setHasNoUnsignedWrap(true);
557 // No further simplifications.
562 /// Return whether "X LOp (Y ROp Z)" is always equal to
563 /// "(X LOp Y) ROp (X LOp Z)".
564 static bool leftDistributesOverRight(Instruction::BinaryOps LOp,
565 Instruction::BinaryOps ROp) {
566 // X & (Y | Z) <--> (X & Y) | (X & Z)
567 // X & (Y ^ Z) <--> (X & Y) ^ (X & Z)
568 if (LOp == Instruction::And)
569 return ROp == Instruction::Or || ROp == Instruction::Xor;
571 // X | (Y & Z) <--> (X | Y) & (X | Z)
572 if (LOp == Instruction::Or)
573 return ROp == Instruction::And;
575 // X * (Y + Z) <--> (X * Y) + (X * Z)
576 // X * (Y - Z) <--> (X * Y) - (X * Z)
577 if (LOp == Instruction::Mul)
578 return ROp == Instruction::Add || ROp == Instruction::Sub;
583 /// Return whether "(X LOp Y) ROp Z" is always equal to
584 /// "(X ROp Z) LOp (Y ROp Z)".
585 static bool rightDistributesOverLeft(Instruction::BinaryOps LOp,
586 Instruction::BinaryOps ROp) {
587 if (Instruction::isCommutative(ROp))
588 return leftDistributesOverRight(ROp, LOp);
590 // (X {&|^} Y) >> Z <--> (X >> Z) {&|^} (Y >> Z) for all shifts.
591 return Instruction::isBitwiseLogicOp(LOp) && Instruction::isShift(ROp);
593 // TODO: It would be nice to handle division, aka "(X + Y)/Z = X/Z + Y/Z",
594 // but this requires knowing that the addition does not overflow and other
598 /// This function returns identity value for given opcode, which can be used to
599 /// factor patterns like (X * 2) + X ==> (X * 2) + (X * 1) ==> X * (2 + 1).
600 static Value *getIdentityValue(Instruction::BinaryOps Opcode, Value *V) {
601 if (isa<Constant>(V))
604 return ConstantExpr::getBinOpIdentity(Opcode, V->getType());
607 /// This function predicates factorization using distributive laws. By default,
608 /// it just returns the 'Op' inputs. But for special-cases like
609 /// 'add(shl(X, 5), ...)', this function will have TopOpcode == Instruction::Add
610 /// and Op = shl(X, 5). The 'shl' is treated as the more general 'mul X, 32' to
611 /// allow more factorization opportunities.
612 static Instruction::BinaryOps
613 getBinOpsForFactorization(Instruction::BinaryOps TopOpcode, BinaryOperator *Op,
614 Value *&LHS, Value *&RHS) {
615 assert(Op && "Expected a binary operator");
616 LHS = Op->getOperand(0);
617 RHS = Op->getOperand(1);
618 if (TopOpcode == Instruction::Add || TopOpcode == Instruction::Sub) {
620 if (match(Op, m_Shl(m_Value(), m_Constant(C)))) {
621 // X << C --> X * (1 << C)
622 RHS = ConstantExpr::getShl(ConstantInt::get(Op->getType(), 1), C);
623 return Instruction::Mul;
625 // TODO: We can add other conversions e.g. shr => div etc.
627 return Op->getOpcode();
630 /// This tries to simplify binary operations by factorizing out common terms
631 /// (e. g. "(A*B)+(A*C)" -> "A*(B+C)").
632 Value *InstCombinerImpl::tryFactorization(BinaryOperator &I,
633 Instruction::BinaryOps InnerOpcode,
634 Value *A, Value *B, Value *C,
636 assert(A && B && C && D && "All values must be provided");
639 Value *SimplifiedInst = nullptr;
640 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
641 Instruction::BinaryOps TopLevelOpcode = I.getOpcode();
643 // Does "X op' Y" always equal "Y op' X"?
644 bool InnerCommutative = Instruction::isCommutative(InnerOpcode);
646 // Does "X op' (Y op Z)" always equal "(X op' Y) op (X op' Z)"?
647 if (leftDistributesOverRight(InnerOpcode, TopLevelOpcode))
648 // Does the instruction have the form "(A op' B) op (A op' D)" or, in the
649 // commutative case, "(A op' B) op (C op' A)"?
650 if (A == C || (InnerCommutative && A == D)) {
653 // Consider forming "A op' (B op D)".
654 // If "B op D" simplifies then it can be formed with no cost.
655 V = SimplifyBinOp(TopLevelOpcode, B, D, SQ.getWithInstruction(&I));
656 // If "B op D" doesn't simplify then only go on if both of the existing
657 // operations "A op' B" and "C op' D" will be zapped as no longer used.
658 if (!V && LHS->hasOneUse() && RHS->hasOneUse())
659 V = Builder.CreateBinOp(TopLevelOpcode, B, D, RHS->getName());
661 SimplifiedInst = Builder.CreateBinOp(InnerOpcode, A, V);
665 // Does "(X op Y) op' Z" always equal "(X op' Z) op (Y op' Z)"?
666 if (!SimplifiedInst && rightDistributesOverLeft(TopLevelOpcode, InnerOpcode))
667 // Does the instruction have the form "(A op' B) op (C op' B)" or, in the
668 // commutative case, "(A op' B) op (B op' D)"?
669 if (B == D || (InnerCommutative && B == C)) {
672 // Consider forming "(A op C) op' B".
673 // If "A op C" simplifies then it can be formed with no cost.
674 V = SimplifyBinOp(TopLevelOpcode, A, C, SQ.getWithInstruction(&I));
676 // If "A op C" doesn't simplify then only go on if both of the existing
677 // operations "A op' B" and "C op' D" will be zapped as no longer used.
678 if (!V && LHS->hasOneUse() && RHS->hasOneUse())
679 V = Builder.CreateBinOp(TopLevelOpcode, A, C, LHS->getName());
681 SimplifiedInst = Builder.CreateBinOp(InnerOpcode, V, B);
685 if (SimplifiedInst) {
687 SimplifiedInst->takeName(&I);
689 // Check if we can add NSW/NUW flags to SimplifiedInst. If so, set them.
690 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(SimplifiedInst)) {
691 if (isa<OverflowingBinaryOperator>(SimplifiedInst)) {
694 if (isa<OverflowingBinaryOperator>(&I)) {
695 HasNSW = I.hasNoSignedWrap();
696 HasNUW = I.hasNoUnsignedWrap();
699 if (auto *LOBO = dyn_cast<OverflowingBinaryOperator>(LHS)) {
700 HasNSW &= LOBO->hasNoSignedWrap();
701 HasNUW &= LOBO->hasNoUnsignedWrap();
704 if (auto *ROBO = dyn_cast<OverflowingBinaryOperator>(RHS)) {
705 HasNSW &= ROBO->hasNoSignedWrap();
706 HasNUW &= ROBO->hasNoUnsignedWrap();
709 if (TopLevelOpcode == Instruction::Add &&
710 InnerOpcode == Instruction::Mul) {
711 // We can propagate 'nsw' if we know that
712 // %Y = mul nsw i16 %X, C
713 // %Z = add nsw i16 %Y, %X
715 // %Z = mul nsw i16 %X, C+1
717 // iff C+1 isn't INT_MIN
719 if (match(V, m_APInt(CInt))) {
720 if (!CInt->isMinSignedValue())
721 BO->setHasNoSignedWrap(HasNSW);
724 // nuw can be propagated with any constant or nuw value.
725 BO->setHasNoUnsignedWrap(HasNUW);
730 return SimplifiedInst;
733 /// This tries to simplify binary operations which some other binary operation
734 /// distributes over either by factorizing out common terms
735 /// (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this results in
736 /// simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is a win).
737 /// Returns the simplified value, or null if it didn't simplify.
738 Value *InstCombinerImpl::SimplifyUsingDistributiveLaws(BinaryOperator &I) {
739 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
740 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
741 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
742 Instruction::BinaryOps TopLevelOpcode = I.getOpcode();
746 Value *A, *B, *C, *D;
747 Instruction::BinaryOps LHSOpcode, RHSOpcode;
749 LHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op0, A, B);
751 RHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op1, C, D);
753 // The instruction has the form "(A op' B) op (C op' D)". Try to factorize
755 if (Op0 && Op1 && LHSOpcode == RHSOpcode)
756 if (Value *V = tryFactorization(I, LHSOpcode, A, B, C, D))
759 // The instruction has the form "(A op' B) op (C)". Try to factorize common
762 if (Value *Ident = getIdentityValue(LHSOpcode, RHS))
763 if (Value *V = tryFactorization(I, LHSOpcode, A, B, RHS, Ident))
766 // The instruction has the form "(B) op (C op' D)". Try to factorize common
769 if (Value *Ident = getIdentityValue(RHSOpcode, LHS))
770 if (Value *V = tryFactorization(I, RHSOpcode, LHS, Ident, C, D))
775 if (Op0 && rightDistributesOverLeft(Op0->getOpcode(), TopLevelOpcode)) {
776 // The instruction has the form "(A op' B) op C". See if expanding it out
777 // to "(A op C) op' (B op C)" results in simplifications.
778 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
779 Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op'
781 // Disable the use of undef because it's not safe to distribute undef.
782 auto SQDistributive = SQ.getWithInstruction(&I).getWithoutUndef();
783 Value *L = SimplifyBinOp(TopLevelOpcode, A, C, SQDistributive);
784 Value *R = SimplifyBinOp(TopLevelOpcode, B, C, SQDistributive);
786 // Do "A op C" and "B op C" both simplify?
788 // They do! Return "L op' R".
790 C = Builder.CreateBinOp(InnerOpcode, L, R);
795 // Does "A op C" simplify to the identity value for the inner opcode?
796 if (L && L == ConstantExpr::getBinOpIdentity(InnerOpcode, L->getType())) {
797 // They do! Return "B op C".
799 C = Builder.CreateBinOp(TopLevelOpcode, B, C);
804 // Does "B op C" simplify to the identity value for the inner opcode?
805 if (R && R == ConstantExpr::getBinOpIdentity(InnerOpcode, R->getType())) {
806 // They do! Return "A op C".
808 C = Builder.CreateBinOp(TopLevelOpcode, A, C);
814 if (Op1 && leftDistributesOverRight(TopLevelOpcode, Op1->getOpcode())) {
815 // The instruction has the form "A op (B op' C)". See if expanding it out
816 // to "(A op B) op' (A op C)" results in simplifications.
817 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
818 Instruction::BinaryOps InnerOpcode = Op1->getOpcode(); // op'
820 // Disable the use of undef because it's not safe to distribute undef.
821 auto SQDistributive = SQ.getWithInstruction(&I).getWithoutUndef();
822 Value *L = SimplifyBinOp(TopLevelOpcode, A, B, SQDistributive);
823 Value *R = SimplifyBinOp(TopLevelOpcode, A, C, SQDistributive);
825 // Do "A op B" and "A op C" both simplify?
827 // They do! Return "L op' R".
829 A = Builder.CreateBinOp(InnerOpcode, L, R);
834 // Does "A op B" simplify to the identity value for the inner opcode?
835 if (L && L == ConstantExpr::getBinOpIdentity(InnerOpcode, L->getType())) {
836 // They do! Return "A op C".
838 A = Builder.CreateBinOp(TopLevelOpcode, A, C);
843 // Does "A op C" simplify to the identity value for the inner opcode?
844 if (R && R == ConstantExpr::getBinOpIdentity(InnerOpcode, R->getType())) {
845 // They do! Return "A op B".
847 A = Builder.CreateBinOp(TopLevelOpcode, A, B);
853 return SimplifySelectsFeedingBinaryOp(I, LHS, RHS);
856 Value *InstCombinerImpl::SimplifySelectsFeedingBinaryOp(BinaryOperator &I,
859 Value *A, *B, *C, *D, *E, *F;
860 bool LHSIsSelect = match(LHS, m_Select(m_Value(A), m_Value(B), m_Value(C)));
861 bool RHSIsSelect = match(RHS, m_Select(m_Value(D), m_Value(E), m_Value(F)));
862 if (!LHSIsSelect && !RHSIsSelect)
866 BuilderTy::FastMathFlagGuard Guard(Builder);
867 if (isa<FPMathOperator>(&I)) {
868 FMF = I.getFastMathFlags();
869 Builder.setFastMathFlags(FMF);
872 Instruction::BinaryOps Opcode = I.getOpcode();
873 SimplifyQuery Q = SQ.getWithInstruction(&I);
875 Value *Cond, *True = nullptr, *False = nullptr;
876 if (LHSIsSelect && RHSIsSelect && A == D) {
877 // (A ? B : C) op (A ? E : F) -> A ? (B op E) : (C op F)
879 True = SimplifyBinOp(Opcode, B, E, FMF, Q);
880 False = SimplifyBinOp(Opcode, C, F, FMF, Q);
882 if (LHS->hasOneUse() && RHS->hasOneUse()) {
884 True = Builder.CreateBinOp(Opcode, B, E);
885 else if (True && !False)
886 False = Builder.CreateBinOp(Opcode, C, F);
888 } else if (LHSIsSelect && LHS->hasOneUse()) {
889 // (A ? B : C) op Y -> A ? (B op Y) : (C op Y)
891 True = SimplifyBinOp(Opcode, B, RHS, FMF, Q);
892 False = SimplifyBinOp(Opcode, C, RHS, FMF, Q);
893 } else if (RHSIsSelect && RHS->hasOneUse()) {
894 // X op (D ? E : F) -> D ? (X op E) : (X op F)
896 True = SimplifyBinOp(Opcode, LHS, E, FMF, Q);
897 False = SimplifyBinOp(Opcode, LHS, F, FMF, Q);
903 Value *SI = Builder.CreateSelect(Cond, True, False);
908 /// Freely adapt every user of V as-if V was changed to !V.
909 /// WARNING: only if canFreelyInvertAllUsersOf() said this can be done.
910 void InstCombinerImpl::freelyInvertAllUsersOf(Value *I) {
911 for (User *U : I->users()) {
912 switch (cast<Instruction>(U)->getOpcode()) {
913 case Instruction::Select: {
914 auto *SI = cast<SelectInst>(U);
916 SI->swapProfMetadata();
919 case Instruction::Br:
920 cast<BranchInst>(U)->swapSuccessors(); // swaps prof metadata too
922 case Instruction::Xor:
923 replaceInstUsesWith(cast<Instruction>(*U), I);
926 llvm_unreachable("Got unexpected user - out of sync with "
927 "canFreelyInvertAllUsersOf() ?");
932 /// Given a 'sub' instruction, return the RHS of the instruction if the LHS is a
933 /// constant zero (which is the 'negate' form).
934 Value *InstCombinerImpl::dyn_castNegVal(Value *V) const {
936 if (match(V, m_Neg(m_Value(NegV))))
939 // Constants can be considered to be negated values if they can be folded.
940 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
941 return ConstantExpr::getNeg(C);
943 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
944 if (C->getType()->getElementType()->isIntegerTy())
945 return ConstantExpr::getNeg(C);
947 if (ConstantVector *CV = dyn_cast<ConstantVector>(V)) {
948 for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) {
949 Constant *Elt = CV->getAggregateElement(i);
953 if (isa<UndefValue>(Elt))
956 if (!isa<ConstantInt>(Elt))
959 return ConstantExpr::getNeg(CV);
962 // Negate integer vector splats.
963 if (auto *CV = dyn_cast<Constant>(V))
964 if (CV->getType()->isVectorTy() &&
965 CV->getType()->getScalarType()->isIntegerTy() && CV->getSplatValue())
966 return ConstantExpr::getNeg(CV);
971 /// A binop with a constant operand and a sign-extended boolean operand may be
972 /// converted into a select of constants by applying the binary operation to
973 /// the constant with the two possible values of the extended boolean (0 or -1).
974 Instruction *InstCombinerImpl::foldBinopOfSextBoolToSelect(BinaryOperator &BO) {
975 // TODO: Handle non-commutative binop (constant is operand 0).
976 // TODO: Handle zext.
977 // TODO: Peek through 'not' of cast.
978 Value *BO0 = BO.getOperand(0);
979 Value *BO1 = BO.getOperand(1);
982 if (!match(BO0, m_SExt(m_Value(X))) || !match(BO1, m_ImmConstant(C)) ||
983 !X->getType()->isIntOrIntVectorTy(1))
986 // bo (sext i1 X), C --> select X, (bo -1, C), (bo 0, C)
987 Constant *Ones = ConstantInt::getAllOnesValue(BO.getType());
988 Constant *Zero = ConstantInt::getNullValue(BO.getType());
989 Constant *TVal = ConstantExpr::get(BO.getOpcode(), Ones, C);
990 Constant *FVal = ConstantExpr::get(BO.getOpcode(), Zero, C);
991 return SelectInst::Create(X, TVal, FVal);
994 static Value *foldOperationIntoSelectOperand(Instruction &I, Value *SO,
995 InstCombiner::BuilderTy &Builder) {
996 if (auto *Cast = dyn_cast<CastInst>(&I))
997 return Builder.CreateCast(Cast->getOpcode(), SO, I.getType());
999 if (auto *II = dyn_cast<IntrinsicInst>(&I)) {
1000 assert(canConstantFoldCallTo(II, cast<Function>(II->getCalledOperand())) &&
1001 "Expected constant-foldable intrinsic");
1002 Intrinsic::ID IID = II->getIntrinsicID();
1003 if (II->arg_size() == 1)
1004 return Builder.CreateUnaryIntrinsic(IID, SO);
1006 // This works for real binary ops like min/max (where we always expect the
1007 // constant operand to be canonicalized as op1) and unary ops with a bonus
1008 // constant argument like ctlz/cttz.
1009 // TODO: Handle non-commutative binary intrinsics as below for binops.
1010 assert(II->arg_size() == 2 && "Expected binary intrinsic");
1011 assert(isa<Constant>(II->getArgOperand(1)) && "Expected constant operand");
1012 return Builder.CreateBinaryIntrinsic(IID, SO, II->getArgOperand(1));
1015 assert(I.isBinaryOp() && "Unexpected opcode for select folding");
1017 // Figure out if the constant is the left or the right argument.
1018 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1019 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1021 if (auto *SOC = dyn_cast<Constant>(SO)) {
1023 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1024 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1027 Value *Op0 = SO, *Op1 = ConstOperand;
1029 std::swap(Op0, Op1);
1031 Value *NewBO = Builder.CreateBinOp(cast<BinaryOperator>(&I)->getOpcode(), Op0,
1032 Op1, SO->getName() + ".op");
1033 if (auto *NewBOI = dyn_cast<Instruction>(NewBO))
1034 NewBOI->copyIRFlags(&I);
1038 Instruction *InstCombinerImpl::FoldOpIntoSelect(Instruction &Op,
1040 // Don't modify shared select instructions.
1041 if (!SI->hasOneUse())
1044 Value *TV = SI->getTrueValue();
1045 Value *FV = SI->getFalseValue();
1046 if (!(isa<Constant>(TV) || isa<Constant>(FV)))
1049 // Bool selects with constant operands can be folded to logical ops.
1050 if (SI->getType()->isIntOrIntVectorTy(1))
1053 // If it's a bitcast involving vectors, make sure it has the same number of
1054 // elements on both sides.
1055 if (auto *BC = dyn_cast<BitCastInst>(&Op)) {
1056 VectorType *DestTy = dyn_cast<VectorType>(BC->getDestTy());
1057 VectorType *SrcTy = dyn_cast<VectorType>(BC->getSrcTy());
1059 // Verify that either both or neither are vectors.
1060 if ((SrcTy == nullptr) != (DestTy == nullptr))
1063 // If vectors, verify that they have the same number of elements.
1064 if (SrcTy && SrcTy->getElementCount() != DestTy->getElementCount())
1068 // Test if a CmpInst instruction is used exclusively by a select as
1069 // part of a minimum or maximum operation. If so, refrain from doing
1070 // any other folding. This helps out other analyses which understand
1071 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
1072 // and CodeGen. And in this case, at least one of the comparison
1073 // operands has at least one user besides the compare (the select),
1074 // which would often largely negate the benefit of folding anyway.
1075 if (auto *CI = dyn_cast<CmpInst>(SI->getCondition())) {
1076 if (CI->hasOneUse()) {
1077 Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1);
1079 // FIXME: This is a hack to avoid infinite looping with min/max patterns.
1080 // We have to ensure that vector constants that only differ with
1081 // undef elements are treated as equivalent.
1082 auto areLooselyEqual = [](Value *A, Value *B) {
1086 // Test for vector constants.
1087 Constant *ConstA, *ConstB;
1088 if (!match(A, m_Constant(ConstA)) || !match(B, m_Constant(ConstB)))
1091 // TODO: Deal with FP constants?
1092 if (!A->getType()->isIntOrIntVectorTy() || A->getType() != B->getType())
1095 // Compare for equality including undefs as equal.
1096 auto *Cmp = ConstantExpr::getCompare(ICmpInst::ICMP_EQ, ConstA, ConstB);
1098 return match(Cmp, m_APIntAllowUndef(C)) && C->isOne();
1101 if ((areLooselyEqual(TV, Op0) && areLooselyEqual(FV, Op1)) ||
1102 (areLooselyEqual(FV, Op0) && areLooselyEqual(TV, Op1)))
1107 Value *NewTV = foldOperationIntoSelectOperand(Op, TV, Builder);
1108 Value *NewFV = foldOperationIntoSelectOperand(Op, FV, Builder);
1109 return SelectInst::Create(SI->getCondition(), NewTV, NewFV, "", nullptr, SI);
1112 static Value *foldOperationIntoPhiValue(BinaryOperator *I, Value *InV,
1113 InstCombiner::BuilderTy &Builder) {
1114 bool ConstIsRHS = isa<Constant>(I->getOperand(1));
1115 Constant *C = cast<Constant>(I->getOperand(ConstIsRHS));
1117 if (auto *InC = dyn_cast<Constant>(InV)) {
1119 return ConstantExpr::get(I->getOpcode(), InC, C);
1120 return ConstantExpr::get(I->getOpcode(), C, InC);
1123 Value *Op0 = InV, *Op1 = C;
1125 std::swap(Op0, Op1);
1127 Value *RI = Builder.CreateBinOp(I->getOpcode(), Op0, Op1, "phi.bo");
1128 auto *FPInst = dyn_cast<Instruction>(RI);
1129 if (FPInst && isa<FPMathOperator>(FPInst))
1130 FPInst->copyFastMathFlags(I);
1134 Instruction *InstCombinerImpl::foldOpIntoPhi(Instruction &I, PHINode *PN) {
1135 unsigned NumPHIValues = PN->getNumIncomingValues();
1136 if (NumPHIValues == 0)
1139 // We normally only transform phis with a single use. However, if a PHI has
1140 // multiple uses and they are all the same operation, we can fold *all* of the
1141 // uses into the PHI.
1142 if (!PN->hasOneUse()) {
1143 // Walk the use list for the instruction, comparing them to I.
1144 for (User *U : PN->users()) {
1145 Instruction *UI = cast<Instruction>(U);
1146 if (UI != &I && !I.isIdenticalTo(UI))
1149 // Otherwise, we can replace *all* users with the new PHI we form.
1152 // Check to see if all of the operands of the PHI are simple constants
1153 // (constantint/constantfp/undef). If there is one non-constant value,
1154 // remember the BB it is in. If there is more than one or if *it* is a PHI,
1155 // bail out. We don't do arbitrary constant expressions here because moving
1156 // their computation can be expensive without a cost model.
1157 BasicBlock *NonConstBB = nullptr;
1158 for (unsigned i = 0; i != NumPHIValues; ++i) {
1159 Value *InVal = PN->getIncomingValue(i);
1160 // For non-freeze, require constant operand
1161 // For freeze, require non-undef, non-poison operand
1162 if (!isa<FreezeInst>(I) && match(InVal, m_ImmConstant()))
1164 if (isa<FreezeInst>(I) && isGuaranteedNotToBeUndefOrPoison(InVal))
1167 if (isa<PHINode>(InVal)) return nullptr; // Itself a phi.
1168 if (NonConstBB) return nullptr; // More than one non-const value.
1170 NonConstBB = PN->getIncomingBlock(i);
1172 // If the InVal is an invoke at the end of the pred block, then we can't
1173 // insert a computation after it without breaking the edge.
1174 if (isa<InvokeInst>(InVal))
1175 if (cast<Instruction>(InVal)->getParent() == NonConstBB)
1178 // If the incoming non-constant value is in I's block, we will remove one
1179 // instruction, but insert another equivalent one, leading to infinite
1181 if (isPotentiallyReachable(I.getParent(), NonConstBB, nullptr, &DT, LI))
1185 // If there is exactly one non-constant value, we can insert a copy of the
1186 // operation in that block. However, if this is a critical edge, we would be
1187 // inserting the computation on some other paths (e.g. inside a loop). Only
1188 // do this if the pred block is unconditionally branching into the phi block.
1189 // Also, make sure that the pred block is not dead code.
1190 if (NonConstBB != nullptr) {
1191 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1192 if (!BI || !BI->isUnconditional() || !DT.isReachableFromEntry(NonConstBB))
1196 // Okay, we can do the transformation: create the new PHI node.
1197 PHINode *NewPN = PHINode::Create(I.getType(), PN->getNumIncomingValues());
1198 InsertNewInstBefore(NewPN, *PN);
1199 NewPN->takeName(PN);
1201 // If we are going to have to insert a new computation, do so right before the
1202 // predecessor's terminator.
1204 Builder.SetInsertPoint(NonConstBB->getTerminator());
1206 // Next, add all of the operands to the PHI.
1207 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) {
1208 // We only currently try to fold the condition of a select when it is a phi,
1209 // not the true/false values.
1210 Value *TrueV = SI->getTrueValue();
1211 Value *FalseV = SI->getFalseValue();
1212 BasicBlock *PhiTransBB = PN->getParent();
1213 for (unsigned i = 0; i != NumPHIValues; ++i) {
1214 BasicBlock *ThisBB = PN->getIncomingBlock(i);
1215 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB);
1216 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB);
1217 Value *InV = nullptr;
1218 // Beware of ConstantExpr: it may eventually evaluate to getNullValue,
1219 // even if currently isNullValue gives false.
1220 Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i));
1221 // For vector constants, we cannot use isNullValue to fold into
1222 // FalseVInPred versus TrueVInPred. When we have individual nonzero
1223 // elements in the vector, we will incorrectly fold InC to
1225 if (InC && isa<ConstantInt>(InC))
1226 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred;
1228 // Generate the select in the same block as PN's current incoming block.
1229 // Note: ThisBB need not be the NonConstBB because vector constants
1230 // which are constants by definition are handled here.
1231 // FIXME: This can lead to an increase in IR generation because we might
1232 // generate selects for vector constant phi operand, that could not be
1233 // folded to TrueVInPred or FalseVInPred as done for ConstantInt. For
1234 // non-vector phis, this transformation was always profitable because
1235 // the select would be generated exactly once in the NonConstBB.
1236 Builder.SetInsertPoint(ThisBB->getTerminator());
1237 InV = Builder.CreateSelect(PN->getIncomingValue(i), TrueVInPred,
1238 FalseVInPred, "phi.sel");
1240 NewPN->addIncoming(InV, ThisBB);
1242 } else if (CmpInst *CI = dyn_cast<CmpInst>(&I)) {
1243 Constant *C = cast<Constant>(I.getOperand(1));
1244 for (unsigned i = 0; i != NumPHIValues; ++i) {
1245 Value *InV = nullptr;
1246 if (auto *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
1247 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1249 InV = Builder.CreateCmp(CI->getPredicate(), PN->getIncomingValue(i),
1251 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1253 } else if (auto *BO = dyn_cast<BinaryOperator>(&I)) {
1254 for (unsigned i = 0; i != NumPHIValues; ++i) {
1255 Value *InV = foldOperationIntoPhiValue(BO, PN->getIncomingValue(i),
1257 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1259 } else if (isa<FreezeInst>(&I)) {
1260 for (unsigned i = 0; i != NumPHIValues; ++i) {
1262 if (NonConstBB == PN->getIncomingBlock(i))
1263 InV = Builder.CreateFreeze(PN->getIncomingValue(i), "phi.fr");
1265 InV = PN->getIncomingValue(i);
1266 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1269 CastInst *CI = cast<CastInst>(&I);
1270 Type *RetTy = CI->getType();
1271 for (unsigned i = 0; i != NumPHIValues; ++i) {
1273 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
1274 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1276 InV = Builder.CreateCast(CI->getOpcode(), PN->getIncomingValue(i),
1277 I.getType(), "phi.cast");
1278 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1282 for (User *U : make_early_inc_range(PN->users())) {
1283 Instruction *User = cast<Instruction>(U);
1284 if (User == &I) continue;
1285 replaceInstUsesWith(*User, NewPN);
1286 eraseInstFromFunction(*User);
1288 return replaceInstUsesWith(I, NewPN);
1291 Instruction *InstCombinerImpl::foldBinopWithPhiOperands(BinaryOperator &BO) {
1292 // TODO: This should be similar to the incoming values check in foldOpIntoPhi:
1293 // we are guarding against replicating the binop in >1 predecessor.
1294 // This could miss matching a phi with 2 constant incoming values.
1295 auto *Phi0 = dyn_cast<PHINode>(BO.getOperand(0));
1296 auto *Phi1 = dyn_cast<PHINode>(BO.getOperand(1));
1297 if (!Phi0 || !Phi1 || !Phi0->hasOneUse() || !Phi1->hasOneUse() ||
1298 Phi0->getNumOperands() != 2 || Phi1->getNumOperands() != 2)
1301 // TODO: Remove the restriction for binop being in the same block as the phis.
1302 if (BO.getParent() != Phi0->getParent() ||
1303 BO.getParent() != Phi1->getParent())
1306 // Match a pair of incoming constants for one of the predecessor blocks.
1307 BasicBlock *ConstBB, *OtherBB;
1309 if (match(Phi0->getIncomingValue(0), m_ImmConstant(C0))) {
1310 ConstBB = Phi0->getIncomingBlock(0);
1311 OtherBB = Phi0->getIncomingBlock(1);
1312 } else if (match(Phi0->getIncomingValue(1), m_ImmConstant(C0))) {
1313 ConstBB = Phi0->getIncomingBlock(1);
1314 OtherBB = Phi0->getIncomingBlock(0);
1318 if (!match(Phi1->getIncomingValueForBlock(ConstBB), m_ImmConstant(C1)))
1321 // The block that we are hoisting to must reach here unconditionally.
1322 // Otherwise, we could be speculatively executing an expensive or
1323 // non-speculative op.
1324 auto *PredBlockBranch = dyn_cast<BranchInst>(OtherBB->getTerminator());
1325 if (!PredBlockBranch || PredBlockBranch->isConditional() ||
1326 !DT.isReachableFromEntry(OtherBB))
1329 // TODO: This check could be tightened to only apply to binops (div/rem) that
1330 // are not safe to speculatively execute. But that could allow hoisting
1331 // potentially expensive instructions (fdiv for example).
1332 for (auto BBIter = BO.getParent()->begin(); &*BBIter != &BO; ++BBIter)
1333 if (!isGuaranteedToTransferExecutionToSuccessor(&*BBIter))
1336 // Make a new binop in the predecessor block with the non-constant incoming
1338 Builder.SetInsertPoint(PredBlockBranch);
1339 Value *NewBO = Builder.CreateBinOp(BO.getOpcode(),
1340 Phi0->getIncomingValueForBlock(OtherBB),
1341 Phi1->getIncomingValueForBlock(OtherBB));
1342 if (auto *NotFoldedNewBO = dyn_cast<BinaryOperator>(NewBO))
1343 NotFoldedNewBO->copyIRFlags(&BO);
1345 // Fold constants for the predecessor block with constant incoming values.
1346 Constant *NewC = ConstantExpr::get(BO.getOpcode(), C0, C1);
1348 // Replace the binop with a phi of the new values. The old phis are dead.
1349 PHINode *NewPhi = PHINode::Create(BO.getType(), 2);
1350 NewPhi->addIncoming(NewBO, OtherBB);
1351 NewPhi->addIncoming(NewC, ConstBB);
1355 Instruction *InstCombinerImpl::foldBinOpIntoSelectOrPhi(BinaryOperator &I) {
1356 if (!isa<Constant>(I.getOperand(1)))
1359 if (auto *Sel = dyn_cast<SelectInst>(I.getOperand(0))) {
1360 if (Instruction *NewSel = FoldOpIntoSelect(I, Sel))
1362 } else if (auto *PN = dyn_cast<PHINode>(I.getOperand(0))) {
1363 if (Instruction *NewPhi = foldOpIntoPhi(I, PN))
1369 /// Given a pointer type and a constant offset, determine whether or not there
1370 /// is a sequence of GEP indices into the pointed type that will land us at the
1371 /// specified offset. If so, fill them into NewIndices and return the resultant
1372 /// element type, otherwise return null.
1373 static Type *findElementAtOffset(PointerType *PtrTy, int64_t IntOffset,
1374 SmallVectorImpl<Value *> &NewIndices,
1375 const DataLayout &DL) {
1376 // Only used by visitGEPOfBitcast(), which is skipped for opaque pointers.
1377 Type *Ty = PtrTy->getNonOpaquePointerElementType();
1381 APInt Offset(DL.getIndexTypeSizeInBits(PtrTy), IntOffset);
1382 SmallVector<APInt> Indices = DL.getGEPIndicesForOffset(Ty, Offset);
1383 if (!Offset.isZero())
1386 for (const APInt &Index : Indices)
1387 NewIndices.push_back(ConstantInt::get(PtrTy->getContext(), Index));
1391 static bool shouldMergeGEPs(GEPOperator &GEP, GEPOperator &Src) {
1392 // If this GEP has only 0 indices, it is the same pointer as
1393 // Src. If Src is not a trivial GEP too, don't combine
1395 if (GEP.hasAllZeroIndices() && !Src.hasAllZeroIndices() &&
1401 /// Return a value X such that Val = X * Scale, or null if none.
1402 /// If the multiplication is known not to overflow, then NoSignedWrap is set.
1403 Value *InstCombinerImpl::Descale(Value *Val, APInt Scale, bool &NoSignedWrap) {
1404 assert(isa<IntegerType>(Val->getType()) && "Can only descale integers!");
1405 assert(cast<IntegerType>(Val->getType())->getBitWidth() ==
1406 Scale.getBitWidth() && "Scale not compatible with value!");
1408 // If Val is zero or Scale is one then Val = Val * Scale.
1409 if (match(Val, m_Zero()) || Scale == 1) {
1410 NoSignedWrap = true;
1414 // If Scale is zero then it does not divide Val.
1415 if (Scale.isMinValue())
1418 // Look through chains of multiplications, searching for a constant that is
1419 // divisible by Scale. For example, descaling X*(Y*(Z*4)) by a factor of 4
1420 // will find the constant factor 4 and produce X*(Y*Z). Descaling X*(Y*8) by
1421 // a factor of 4 will produce X*(Y*2). The principle of operation is to bore
1424 // Val = M1 * X || Analysis starts here and works down
1425 // M1 = M2 * Y || Doesn't descend into terms with more
1426 // M2 = Z * 4 \/ than one use
1428 // Then to modify a term at the bottom:
1431 // M1 = Z * Y || Replaced M2 with Z
1433 // Then to work back up correcting nsw flags.
1435 // Op - the term we are currently analyzing. Starts at Val then drills down.
1436 // Replaced with its descaled value before exiting from the drill down loop.
1439 // Parent - initially null, but after drilling down notes where Op came from.
1440 // In the example above, Parent is (Val, 0) when Op is M1, because M1 is the
1441 // 0'th operand of Val.
1442 std::pair<Instruction *, unsigned> Parent;
1444 // Set if the transform requires a descaling at deeper levels that doesn't
1446 bool RequireNoSignedWrap = false;
1448 // Log base 2 of the scale. Negative if not a power of 2.
1449 int32_t logScale = Scale.exactLogBase2();
1451 for (;; Op = Parent.first->getOperand(Parent.second)) { // Drill down
1452 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
1453 // If Op is a constant divisible by Scale then descale to the quotient.
1454 APInt Quotient(Scale), Remainder(Scale); // Init ensures right bitwidth.
1455 APInt::sdivrem(CI->getValue(), Scale, Quotient, Remainder);
1456 if (!Remainder.isMinValue())
1457 // Not divisible by Scale.
1459 // Replace with the quotient in the parent.
1460 Op = ConstantInt::get(CI->getType(), Quotient);
1461 NoSignedWrap = true;
1465 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op)) {
1466 if (BO->getOpcode() == Instruction::Mul) {
1468 NoSignedWrap = BO->hasNoSignedWrap();
1469 if (RequireNoSignedWrap && !NoSignedWrap)
1472 // There are three cases for multiplication: multiplication by exactly
1473 // the scale, multiplication by a constant different to the scale, and
1474 // multiplication by something else.
1475 Value *LHS = BO->getOperand(0);
1476 Value *RHS = BO->getOperand(1);
1478 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
1479 // Multiplication by a constant.
1480 if (CI->getValue() == Scale) {
1481 // Multiplication by exactly the scale, replace the multiplication
1482 // by its left-hand side in the parent.
1487 // Otherwise drill down into the constant.
1488 if (!Op->hasOneUse())
1491 Parent = std::make_pair(BO, 1);
1495 // Multiplication by something else. Drill down into the left-hand side
1496 // since that's where the reassociate pass puts the good stuff.
1497 if (!Op->hasOneUse())
1500 Parent = std::make_pair(BO, 0);
1504 if (logScale > 0 && BO->getOpcode() == Instruction::Shl &&
1505 isa<ConstantInt>(BO->getOperand(1))) {
1506 // Multiplication by a power of 2.
1507 NoSignedWrap = BO->hasNoSignedWrap();
1508 if (RequireNoSignedWrap && !NoSignedWrap)
1511 Value *LHS = BO->getOperand(0);
1512 int32_t Amt = cast<ConstantInt>(BO->getOperand(1))->
1513 getLimitedValue(Scale.getBitWidth());
1516 if (Amt == logScale) {
1517 // Multiplication by exactly the scale, replace the multiplication
1518 // by its left-hand side in the parent.
1522 if (Amt < logScale || !Op->hasOneUse())
1525 // Multiplication by more than the scale. Reduce the multiplying amount
1526 // by the scale in the parent.
1527 Parent = std::make_pair(BO, 1);
1528 Op = ConstantInt::get(BO->getType(), Amt - logScale);
1533 if (!Op->hasOneUse())
1536 if (CastInst *Cast = dyn_cast<CastInst>(Op)) {
1537 if (Cast->getOpcode() == Instruction::SExt) {
1538 // Op is sign-extended from a smaller type, descale in the smaller type.
1539 unsigned SmallSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1540 APInt SmallScale = Scale.trunc(SmallSize);
1541 // Suppose Op = sext X, and we descale X as Y * SmallScale. We want to
1542 // descale Op as (sext Y) * Scale. In order to have
1543 // sext (Y * SmallScale) = (sext Y) * Scale
1544 // some conditions need to hold however: SmallScale must sign-extend to
1545 // Scale and the multiplication Y * SmallScale should not overflow.
1546 if (SmallScale.sext(Scale.getBitWidth()) != Scale)
1547 // SmallScale does not sign-extend to Scale.
1549 assert(SmallScale.exactLogBase2() == logScale);
1550 // Require that Y * SmallScale must not overflow.
1551 RequireNoSignedWrap = true;
1553 // Drill down through the cast.
1554 Parent = std::make_pair(Cast, 0);
1559 if (Cast->getOpcode() == Instruction::Trunc) {
1560 // Op is truncated from a larger type, descale in the larger type.
1561 // Suppose Op = trunc X, and we descale X as Y * sext Scale. Then
1562 // trunc (Y * sext Scale) = (trunc Y) * Scale
1563 // always holds. However (trunc Y) * Scale may overflow even if
1564 // trunc (Y * sext Scale) does not, so nsw flags need to be cleared
1565 // from this point up in the expression (see later).
1566 if (RequireNoSignedWrap)
1569 // Drill down through the cast.
1570 unsigned LargeSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1571 Parent = std::make_pair(Cast, 0);
1572 Scale = Scale.sext(LargeSize);
1573 if (logScale + 1 == (int32_t)Cast->getType()->getPrimitiveSizeInBits())
1575 assert(Scale.exactLogBase2() == logScale);
1580 // Unsupported expression, bail out.
1584 // If Op is zero then Val = Op * Scale.
1585 if (match(Op, m_Zero())) {
1586 NoSignedWrap = true;
1590 // We know that we can successfully descale, so from here on we can safely
1591 // modify the IR. Op holds the descaled version of the deepest term in the
1592 // expression. NoSignedWrap is 'true' if multiplying Op by Scale is known
1596 // The expression only had one term.
1599 // Rewrite the parent using the descaled version of its operand.
1600 assert(Parent.first->hasOneUse() && "Drilled down when more than one use!");
1601 assert(Op != Parent.first->getOperand(Parent.second) &&
1602 "Descaling was a no-op?");
1603 replaceOperand(*Parent.first, Parent.second, Op);
1604 Worklist.push(Parent.first);
1606 // Now work back up the expression correcting nsw flags. The logic is based
1607 // on the following observation: if X * Y is known not to overflow as a signed
1608 // multiplication, and Y is replaced by a value Z with smaller absolute value,
1609 // then X * Z will not overflow as a signed multiplication either. As we work
1610 // our way up, having NoSignedWrap 'true' means that the descaled value at the
1611 // current level has strictly smaller absolute value than the original.
1612 Instruction *Ancestor = Parent.first;
1614 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Ancestor)) {
1615 // If the multiplication wasn't nsw then we can't say anything about the
1616 // value of the descaled multiplication, and we have to clear nsw flags
1617 // from this point on up.
1618 bool OpNoSignedWrap = BO->hasNoSignedWrap();
1619 NoSignedWrap &= OpNoSignedWrap;
1620 if (NoSignedWrap != OpNoSignedWrap) {
1621 BO->setHasNoSignedWrap(NoSignedWrap);
1622 Worklist.push(Ancestor);
1624 } else if (Ancestor->getOpcode() == Instruction::Trunc) {
1625 // The fact that the descaled input to the trunc has smaller absolute
1626 // value than the original input doesn't tell us anything useful about
1627 // the absolute values of the truncations.
1628 NoSignedWrap = false;
1630 assert((Ancestor->getOpcode() != Instruction::SExt || NoSignedWrap) &&
1631 "Failed to keep proper track of nsw flags while drilling down?");
1633 if (Ancestor == Val)
1634 // Got to the top, all done!
1637 // Move up one level in the expression.
1638 assert(Ancestor->hasOneUse() && "Drilled down when more than one use!");
1639 Ancestor = Ancestor->user_back();
1643 Instruction *InstCombinerImpl::foldVectorBinop(BinaryOperator &Inst) {
1644 if (!isa<VectorType>(Inst.getType()))
1647 BinaryOperator::BinaryOps Opcode = Inst.getOpcode();
1648 Value *LHS = Inst.getOperand(0), *RHS = Inst.getOperand(1);
1649 assert(cast<VectorType>(LHS->getType())->getElementCount() ==
1650 cast<VectorType>(Inst.getType())->getElementCount());
1651 assert(cast<VectorType>(RHS->getType())->getElementCount() ==
1652 cast<VectorType>(Inst.getType())->getElementCount());
1654 // If both operands of the binop are vector concatenations, then perform the
1655 // narrow binop on each pair of the source operands followed by concatenation
1657 Value *L0, *L1, *R0, *R1;
1659 if (match(LHS, m_Shuffle(m_Value(L0), m_Value(L1), m_Mask(Mask))) &&
1660 match(RHS, m_Shuffle(m_Value(R0), m_Value(R1), m_SpecificMask(Mask))) &&
1661 LHS->hasOneUse() && RHS->hasOneUse() &&
1662 cast<ShuffleVectorInst>(LHS)->isConcat() &&
1663 cast<ShuffleVectorInst>(RHS)->isConcat()) {
1664 // This transform does not have the speculative execution constraint as
1665 // below because the shuffle is a concatenation. The new binops are
1666 // operating on exactly the same elements as the existing binop.
1667 // TODO: We could ease the mask requirement to allow different undef lanes,
1668 // but that requires an analysis of the binop-with-undef output value.
1669 Value *NewBO0 = Builder.CreateBinOp(Opcode, L0, R0);
1670 if (auto *BO = dyn_cast<BinaryOperator>(NewBO0))
1671 BO->copyIRFlags(&Inst);
1672 Value *NewBO1 = Builder.CreateBinOp(Opcode, L1, R1);
1673 if (auto *BO = dyn_cast<BinaryOperator>(NewBO1))
1674 BO->copyIRFlags(&Inst);
1675 return new ShuffleVectorInst(NewBO0, NewBO1, Mask);
1678 // It may not be safe to reorder shuffles and things like div, urem, etc.
1679 // because we may trap when executing those ops on unknown vector elements.
1681 if (!isSafeToSpeculativelyExecute(&Inst))
1684 auto createBinOpShuffle = [&](Value *X, Value *Y, ArrayRef<int> M) {
1685 Value *XY = Builder.CreateBinOp(Opcode, X, Y);
1686 if (auto *BO = dyn_cast<BinaryOperator>(XY))
1687 BO->copyIRFlags(&Inst);
1688 return new ShuffleVectorInst(XY, M);
1691 // If both arguments of the binary operation are shuffles that use the same
1692 // mask and shuffle within a single vector, move the shuffle after the binop.
1694 if (match(LHS, m_Shuffle(m_Value(V1), m_Undef(), m_Mask(Mask))) &&
1695 match(RHS, m_Shuffle(m_Value(V2), m_Undef(), m_SpecificMask(Mask))) &&
1696 V1->getType() == V2->getType() &&
1697 (LHS->hasOneUse() || RHS->hasOneUse() || LHS == RHS)) {
1698 // Op(shuffle(V1, Mask), shuffle(V2, Mask)) -> shuffle(Op(V1, V2), Mask)
1699 return createBinOpShuffle(V1, V2, Mask);
1702 // If both arguments of a commutative binop are select-shuffles that use the
1703 // same mask with commuted operands, the shuffles are unnecessary.
1704 if (Inst.isCommutative() &&
1705 match(LHS, m_Shuffle(m_Value(V1), m_Value(V2), m_Mask(Mask))) &&
1707 m_Shuffle(m_Specific(V2), m_Specific(V1), m_SpecificMask(Mask)))) {
1708 auto *LShuf = cast<ShuffleVectorInst>(LHS);
1709 auto *RShuf = cast<ShuffleVectorInst>(RHS);
1710 // TODO: Allow shuffles that contain undefs in the mask?
1711 // That is legal, but it reduces undef knowledge.
1712 // TODO: Allow arbitrary shuffles by shuffling after binop?
1713 // That might be legal, but we have to deal with poison.
1714 if (LShuf->isSelect() &&
1715 !is_contained(LShuf->getShuffleMask(), UndefMaskElem) &&
1716 RShuf->isSelect() &&
1717 !is_contained(RShuf->getShuffleMask(), UndefMaskElem)) {
1719 // LHS = shuffle V1, V2, <0, 5, 6, 3>
1720 // RHS = shuffle V2, V1, <0, 5, 6, 3>
1721 // LHS + RHS --> (V10+V20, V21+V11, V22+V12, V13+V23) --> V1 + V2
1722 Instruction *NewBO = BinaryOperator::Create(Opcode, V1, V2);
1723 NewBO->copyIRFlags(&Inst);
1728 // If one argument is a shuffle within one vector and the other is a constant,
1729 // try moving the shuffle after the binary operation. This canonicalization
1730 // intends to move shuffles closer to other shuffles and binops closer to
1731 // other binops, so they can be folded. It may also enable demanded elements
1734 auto *InstVTy = dyn_cast<FixedVectorType>(Inst.getType());
1737 m_c_BinOp(m_OneUse(m_Shuffle(m_Value(V1), m_Undef(), m_Mask(Mask))),
1738 m_ImmConstant(C))) &&
1739 cast<FixedVectorType>(V1->getType())->getNumElements() <=
1740 InstVTy->getNumElements()) {
1741 assert(InstVTy->getScalarType() == V1->getType()->getScalarType() &&
1742 "Shuffle should not change scalar type");
1744 // Find constant NewC that has property:
1745 // shuffle(NewC, ShMask) = C
1746 // If such constant does not exist (example: ShMask=<0,0> and C=<1,2>)
1747 // reorder is not possible. A 1-to-1 mapping is not required. Example:
1748 // ShMask = <1,1,2,2> and C = <5,5,6,6> --> NewC = <undef,5,6,undef>
1749 bool ConstOp1 = isa<Constant>(RHS);
1750 ArrayRef<int> ShMask = Mask;
1751 unsigned SrcVecNumElts =
1752 cast<FixedVectorType>(V1->getType())->getNumElements();
1753 UndefValue *UndefScalar = UndefValue::get(C->getType()->getScalarType());
1754 SmallVector<Constant *, 16> NewVecC(SrcVecNumElts, UndefScalar);
1755 bool MayChange = true;
1756 unsigned NumElts = InstVTy->getNumElements();
1757 for (unsigned I = 0; I < NumElts; ++I) {
1758 Constant *CElt = C->getAggregateElement(I);
1759 if (ShMask[I] >= 0) {
1760 assert(ShMask[I] < (int)NumElts && "Not expecting narrowing shuffle");
1761 Constant *NewCElt = NewVecC[ShMask[I]];
1763 // 1. The constant vector contains a constant expression.
1764 // 2. The shuffle needs an element of the constant vector that can't
1765 // be mapped to a new constant vector.
1766 // 3. This is a widening shuffle that copies elements of V1 into the
1767 // extended elements (extending with undef is allowed).
1768 if (!CElt || (!isa<UndefValue>(NewCElt) && NewCElt != CElt) ||
1769 I >= SrcVecNumElts) {
1773 NewVecC[ShMask[I]] = CElt;
1775 // If this is a widening shuffle, we must be able to extend with undef
1776 // elements. If the original binop does not produce an undef in the high
1777 // lanes, then this transform is not safe.
1778 // Similarly for undef lanes due to the shuffle mask, we can only
1779 // transform binops that preserve undef.
1780 // TODO: We could shuffle those non-undef constant values into the
1781 // result by using a constant vector (rather than an undef vector)
1782 // as operand 1 of the new binop, but that might be too aggressive
1783 // for target-independent shuffle creation.
1784 if (I >= SrcVecNumElts || ShMask[I] < 0) {
1785 Constant *MaybeUndef =
1786 ConstOp1 ? ConstantExpr::get(Opcode, UndefScalar, CElt)
1787 : ConstantExpr::get(Opcode, CElt, UndefScalar);
1788 if (!match(MaybeUndef, m_Undef())) {
1795 Constant *NewC = ConstantVector::get(NewVecC);
1796 // It may not be safe to execute a binop on a vector with undef elements
1797 // because the entire instruction can be folded to undef or create poison
1798 // that did not exist in the original code.
1799 if (Inst.isIntDivRem() || (Inst.isShift() && ConstOp1))
1800 NewC = getSafeVectorConstantForBinop(Opcode, NewC, ConstOp1);
1802 // Op(shuffle(V1, Mask), C) -> shuffle(Op(V1, NewC), Mask)
1803 // Op(C, shuffle(V1, Mask)) -> shuffle(Op(NewC, V1), Mask)
1804 Value *NewLHS = ConstOp1 ? V1 : NewC;
1805 Value *NewRHS = ConstOp1 ? NewC : V1;
1806 return createBinOpShuffle(NewLHS, NewRHS, Mask);
1810 // Try to reassociate to sink a splat shuffle after a binary operation.
1811 if (Inst.isAssociative() && Inst.isCommutative()) {
1812 // Canonicalize shuffle operand as LHS.
1813 if (isa<ShuffleVectorInst>(RHS))
1814 std::swap(LHS, RHS);
1817 ArrayRef<int> MaskC;
1821 m_OneUse(m_Shuffle(m_Value(X), m_Undef(), m_Mask(MaskC)))) ||
1822 !match(MaskC, m_SplatOrUndefMask(SplatIndex)) ||
1823 X->getType() != Inst.getType() ||
1824 !match(RHS, m_OneUse(m_BinOp(Opcode, m_Value(Y), m_Value(OtherOp)))))
1827 // FIXME: This may not be safe if the analysis allows undef elements. By
1828 // moving 'Y' before the splat shuffle, we are implicitly assuming
1829 // that it is not undef/poison at the splat index.
1830 if (isSplatValue(OtherOp, SplatIndex)) {
1831 std::swap(Y, OtherOp);
1832 } else if (!isSplatValue(Y, SplatIndex)) {
1836 // X and Y are splatted values, so perform the binary operation on those
1837 // values followed by a splat followed by the 2nd binary operation:
1838 // bo (splat X), (bo Y, OtherOp) --> bo (splat (bo X, Y)), OtherOp
1839 Value *NewBO = Builder.CreateBinOp(Opcode, X, Y);
1840 SmallVector<int, 8> NewMask(MaskC.size(), SplatIndex);
1841 Value *NewSplat = Builder.CreateShuffleVector(NewBO, NewMask);
1842 Instruction *R = BinaryOperator::Create(Opcode, NewSplat, OtherOp);
1844 // Intersect FMF on both new binops. Other (poison-generating) flags are
1845 // dropped to be safe.
1846 if (isa<FPMathOperator>(R)) {
1847 R->copyFastMathFlags(&Inst);
1850 if (auto *NewInstBO = dyn_cast<BinaryOperator>(NewBO))
1851 NewInstBO->copyIRFlags(R);
1858 /// Try to narrow the width of a binop if at least 1 operand is an extend of
1859 /// of a value. This requires a potentially expensive known bits check to make
1860 /// sure the narrow op does not overflow.
1861 Instruction *InstCombinerImpl::narrowMathIfNoOverflow(BinaryOperator &BO) {
1862 // We need at least one extended operand.
1863 Value *Op0 = BO.getOperand(0), *Op1 = BO.getOperand(1);
1865 // If this is a sub, we swap the operands since we always want an extension
1866 // on the RHS. The LHS can be an extension or a constant.
1867 if (BO.getOpcode() == Instruction::Sub)
1868 std::swap(Op0, Op1);
1871 bool IsSext = match(Op0, m_SExt(m_Value(X)));
1872 if (!IsSext && !match(Op0, m_ZExt(m_Value(X))))
1875 // If both operands are the same extension from the same source type and we
1876 // can eliminate at least one (hasOneUse), this might work.
1877 CastInst::CastOps CastOpc = IsSext ? Instruction::SExt : Instruction::ZExt;
1879 if (!(match(Op1, m_ZExtOrSExt(m_Value(Y))) && X->getType() == Y->getType() &&
1880 cast<Operator>(Op1)->getOpcode() == CastOpc &&
1881 (Op0->hasOneUse() || Op1->hasOneUse()))) {
1882 // If that did not match, see if we have a suitable constant operand.
1883 // Truncating and extending must produce the same constant.
1885 if (!Op0->hasOneUse() || !match(Op1, m_Constant(WideC)))
1887 Constant *NarrowC = ConstantExpr::getTrunc(WideC, X->getType());
1888 if (ConstantExpr::getCast(CastOpc, NarrowC, BO.getType()) != WideC)
1893 // Swap back now that we found our operands.
1894 if (BO.getOpcode() == Instruction::Sub)
1897 // Both operands have narrow versions. Last step: the math must not overflow
1898 // in the narrow width.
1899 if (!willNotOverflow(BO.getOpcode(), X, Y, BO, IsSext))
1902 // bo (ext X), (ext Y) --> ext (bo X, Y)
1903 // bo (ext X), C --> ext (bo X, C')
1904 Value *NarrowBO = Builder.CreateBinOp(BO.getOpcode(), X, Y, "narrow");
1905 if (auto *NewBinOp = dyn_cast<BinaryOperator>(NarrowBO)) {
1907 NewBinOp->setHasNoSignedWrap();
1909 NewBinOp->setHasNoUnsignedWrap();
1911 return CastInst::Create(CastOpc, NarrowBO, BO.getType());
1914 static bool isMergedGEPInBounds(GEPOperator &GEP1, GEPOperator &GEP2) {
1915 // At least one GEP must be inbounds.
1916 if (!GEP1.isInBounds() && !GEP2.isInBounds())
1919 return (GEP1.isInBounds() || GEP1.hasAllZeroIndices()) &&
1920 (GEP2.isInBounds() || GEP2.hasAllZeroIndices());
1923 /// Thread a GEP operation with constant indices through the constant true/false
1924 /// arms of a select.
1925 static Instruction *foldSelectGEP(GetElementPtrInst &GEP,
1926 InstCombiner::BuilderTy &Builder) {
1927 if (!GEP.hasAllConstantIndices())
1932 Constant *TrueC, *FalseC;
1933 if (!match(GEP.getPointerOperand(), m_Instruction(Sel)) ||
1935 m_Select(m_Value(Cond), m_Constant(TrueC), m_Constant(FalseC))))
1938 // gep (select Cond, TrueC, FalseC), IndexC --> select Cond, TrueC', FalseC'
1939 // Propagate 'inbounds' and metadata from existing instructions.
1940 // Note: using IRBuilder to create the constants for efficiency.
1941 SmallVector<Value *, 4> IndexC(GEP.indices());
1942 bool IsInBounds = GEP.isInBounds();
1943 Type *Ty = GEP.getSourceElementType();
1944 Value *NewTrueC = IsInBounds ? Builder.CreateInBoundsGEP(Ty, TrueC, IndexC)
1945 : Builder.CreateGEP(Ty, TrueC, IndexC);
1946 Value *NewFalseC = IsInBounds ? Builder.CreateInBoundsGEP(Ty, FalseC, IndexC)
1947 : Builder.CreateGEP(Ty, FalseC, IndexC);
1948 return SelectInst::Create(Cond, NewTrueC, NewFalseC, "", nullptr, Sel);
1951 Instruction *InstCombinerImpl::visitGEPOfGEP(GetElementPtrInst &GEP,
1953 // Combine Indices - If the source pointer to this getelementptr instruction
1954 // is a getelementptr instruction with matching element type, combine the
1955 // indices of the two getelementptr instructions into a single instruction.
1956 if (Src->getResultElementType() != GEP.getSourceElementType())
1959 if (!shouldMergeGEPs(*cast<GEPOperator>(&GEP), *Src))
1962 if (Src->getNumOperands() == 2 && GEP.getNumOperands() == 2 &&
1964 Value *GO1 = GEP.getOperand(1);
1965 Value *SO1 = Src->getOperand(1);
1968 // Try to reassociate loop invariant GEP chains to enable LICM.
1969 if (Loop *L = LI->getLoopFor(GEP.getParent())) {
1970 // Reassociate the two GEPs if SO1 is variant in the loop and GO1 is
1971 // invariant: this breaks the dependence between GEPs and allows LICM
1972 // to hoist the invariant part out of the loop.
1973 if (L->isLoopInvariant(GO1) && !L->isLoopInvariant(SO1)) {
1974 // We have to be careful here.
1975 // We have something like:
1976 // %src = getelementptr <ty>, <ty>* %base, <ty> %idx
1977 // %gep = getelementptr <ty>, <ty>* %src, <ty> %idx2
1978 // If we just swap idx & idx2 then we could inadvertantly
1979 // change %src from a vector to a scalar, or vice versa.
1981 // 1) %base a scalar & idx a scalar & idx2 a vector
1982 // => Swapping idx & idx2 turns %src into a vector type.
1983 // 2) %base a scalar & idx a vector & idx2 a scalar
1984 // => Swapping idx & idx2 turns %src in a scalar type
1985 // 3) %base, %idx, and %idx2 are scalars
1986 // => %src & %gep are scalars
1987 // => swapping idx & idx2 is safe
1988 // 4) %base a vector
1989 // => %src is a vector
1990 // => swapping idx & idx2 is safe.
1991 auto *SO0 = Src->getOperand(0);
1992 auto *SO0Ty = SO0->getType();
1993 if (!isa<VectorType>(GEP.getType()) || // case 3
1994 isa<VectorType>(SO0Ty)) { // case 4
1995 Src->setOperand(1, GO1);
1996 GEP.setOperand(1, SO1);
2000 // -- have to recreate %src & %gep
2001 // put NewSrc at same location as %src
2002 Builder.SetInsertPoint(cast<Instruction>(Src));
2003 Value *NewSrc = Builder.CreateGEP(
2004 GEP.getSourceElementType(), SO0, GO1, Src->getName());
2005 // Propagate 'inbounds' if the new source was not constant-folded.
2006 if (auto *NewSrcGEPI = dyn_cast<GetElementPtrInst>(NewSrc))
2007 NewSrcGEPI->setIsInBounds(Src->isInBounds());
2008 GetElementPtrInst *NewGEP = GetElementPtrInst::Create(
2009 GEP.getSourceElementType(), NewSrc, {SO1});
2010 NewGEP->setIsInBounds(GEP.isInBounds());
2018 // Note that if our source is a gep chain itself then we wait for that
2019 // chain to be resolved before we perform this transformation. This
2020 // avoids us creating a TON of code in some cases.
2021 if (auto *SrcGEP = dyn_cast<GEPOperator>(Src->getOperand(0)))
2022 if (SrcGEP->getNumOperands() == 2 && shouldMergeGEPs(*Src, *SrcGEP))
2023 return nullptr; // Wait until our source is folded to completion.
2025 SmallVector<Value*, 8> Indices;
2027 // Find out whether the last index in the source GEP is a sequential idx.
2028 bool EndsWithSequential = false;
2029 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src);
2031 EndsWithSequential = I.isSequential();
2033 // Can we combine the two pointer arithmetics offsets?
2034 if (EndsWithSequential) {
2035 // Replace: gep (gep %P, long B), long A, ...
2036 // With: T = long A+B; gep %P, T, ...
2037 Value *SO1 = Src->getOperand(Src->getNumOperands()-1);
2038 Value *GO1 = GEP.getOperand(1);
2040 // If they aren't the same type, then the input hasn't been processed
2041 // by the loop above yet (which canonicalizes sequential index types to
2042 // intptr_t). Just avoid transforming this until the input has been
2044 if (SO1->getType() != GO1->getType())
2048 SimplifyAddInst(GO1, SO1, false, false, SQ.getWithInstruction(&GEP));
2049 // Only do the combine when we are sure the cost after the
2050 // merge is never more than that before the merge.
2054 // Update the GEP in place if possible.
2055 if (Src->getNumOperands() == 2) {
2056 GEP.setIsInBounds(isMergedGEPInBounds(*Src, *cast<GEPOperator>(&GEP)));
2057 replaceOperand(GEP, 0, Src->getOperand(0));
2058 replaceOperand(GEP, 1, Sum);
2061 Indices.append(Src->op_begin()+1, Src->op_end()-1);
2062 Indices.push_back(Sum);
2063 Indices.append(GEP.op_begin()+2, GEP.op_end());
2064 } else if (isa<Constant>(*GEP.idx_begin()) &&
2065 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
2066 Src->getNumOperands() != 1) {
2067 // Otherwise we can do the fold if the first index of the GEP is a zero
2068 Indices.append(Src->op_begin()+1, Src->op_end());
2069 Indices.append(GEP.idx_begin()+1, GEP.idx_end());
2072 if (!Indices.empty())
2073 return isMergedGEPInBounds(*Src, *cast<GEPOperator>(&GEP))
2074 ? GetElementPtrInst::CreateInBounds(
2075 Src->getSourceElementType(), Src->getOperand(0), Indices,
2077 : GetElementPtrInst::Create(Src->getSourceElementType(),
2078 Src->getOperand(0), Indices,
2084 // Note that we may have also stripped an address space cast in between.
2085 Instruction *InstCombinerImpl::visitGEPOfBitcast(BitCastInst *BCI,
2086 GetElementPtrInst &GEP) {
2087 // With opaque pointers, there is no pointer element type we can use to
2088 // adjust the GEP type.
2089 PointerType *SrcType = cast<PointerType>(BCI->getSrcTy());
2090 if (SrcType->isOpaque())
2093 Type *GEPEltType = GEP.getSourceElementType();
2094 Type *SrcEltType = SrcType->getNonOpaquePointerElementType();
2095 Value *SrcOp = BCI->getOperand(0);
2097 // GEP directly using the source operand if this GEP is accessing an element
2098 // of a bitcasted pointer to vector or array of the same dimensions:
2099 // gep (bitcast <c x ty>* X to [c x ty]*), Y, Z --> gep X, Y, Z
2100 // gep (bitcast [c x ty]* X to <c x ty>*), Y, Z --> gep X, Y, Z
2101 auto areMatchingArrayAndVecTypes = [](Type *ArrTy, Type *VecTy,
2102 const DataLayout &DL) {
2103 auto *VecVTy = cast<FixedVectorType>(VecTy);
2104 return ArrTy->getArrayElementType() == VecVTy->getElementType() &&
2105 ArrTy->getArrayNumElements() == VecVTy->getNumElements() &&
2106 DL.getTypeAllocSize(ArrTy) == DL.getTypeAllocSize(VecTy);
2108 if (GEP.getNumOperands() == 3 &&
2109 ((GEPEltType->isArrayTy() && isa<FixedVectorType>(SrcEltType) &&
2110 areMatchingArrayAndVecTypes(GEPEltType, SrcEltType, DL)) ||
2111 (isa<FixedVectorType>(GEPEltType) && SrcEltType->isArrayTy() &&
2112 areMatchingArrayAndVecTypes(SrcEltType, GEPEltType, DL)))) {
2114 // Create a new GEP here, as using `setOperand()` followed by
2115 // `setSourceElementType()` won't actually update the type of the
2116 // existing GEP Value. Causing issues if this Value is accessed when
2117 // constructing an AddrSpaceCastInst
2118 SmallVector<Value *, 8> Indices(GEP.indices());
2119 Value *NGEP = GEP.isInBounds()
2120 ? Builder.CreateInBoundsGEP(SrcEltType, SrcOp, Indices)
2121 : Builder.CreateGEP(SrcEltType, SrcOp, Indices);
2122 NGEP->takeName(&GEP);
2124 // Preserve GEP address space to satisfy users
2125 if (NGEP->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
2126 return new AddrSpaceCastInst(NGEP, GEP.getType());
2128 return replaceInstUsesWith(GEP, NGEP);
2131 // See if we can simplify:
2132 // X = bitcast A* to B*
2133 // Y = gep X, <...constant indices...>
2134 // into a gep of the original struct. This is important for SROA and alias
2135 // analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
2136 unsigned OffsetBits = DL.getIndexTypeSizeInBits(GEP.getType());
2137 APInt Offset(OffsetBits, 0);
2139 // If the bitcast argument is an allocation, The bitcast is for convertion
2140 // to actual type of allocation. Removing such bitcasts, results in having
2141 // GEPs with i8* base and pure byte offsets. That means GEP is not aware of
2142 // struct or array hierarchy.
2143 // By avoiding such GEPs, phi translation and MemoryDependencyAnalysis have
2144 // a better chance to succeed.
2145 if (!isa<BitCastInst>(SrcOp) && GEP.accumulateConstantOffset(DL, Offset) &&
2146 !isAllocationFn(SrcOp, &TLI)) {
2147 // If this GEP instruction doesn't move the pointer, just replace the GEP
2148 // with a bitcast of the real input to the dest type.
2150 // If the bitcast is of an allocation, and the allocation will be
2151 // converted to match the type of the cast, don't touch this.
2152 if (isa<AllocaInst>(SrcOp)) {
2153 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
2154 if (Instruction *I = visitBitCast(*BCI)) {
2157 BCI->getParent()->getInstList().insert(BCI->getIterator(), I);
2158 replaceInstUsesWith(*BCI, I);
2164 if (SrcType->getPointerAddressSpace() != GEP.getAddressSpace())
2165 return new AddrSpaceCastInst(SrcOp, GEP.getType());
2166 return new BitCastInst(SrcOp, GEP.getType());
2169 // Otherwise, if the offset is non-zero, we need to find out if there is a
2170 // field at Offset in 'A's type. If so, we can pull the cast through the
2172 SmallVector<Value*, 8> NewIndices;
2173 if (findElementAtOffset(SrcType, Offset.getSExtValue(), NewIndices, DL)) {
2176 ? Builder.CreateInBoundsGEP(SrcEltType, SrcOp, NewIndices)
2177 : Builder.CreateGEP(SrcEltType, SrcOp, NewIndices);
2179 if (NGEP->getType() == GEP.getType())
2180 return replaceInstUsesWith(GEP, NGEP);
2181 NGEP->takeName(&GEP);
2183 if (NGEP->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
2184 return new AddrSpaceCastInst(NGEP, GEP.getType());
2185 return new BitCastInst(NGEP, GEP.getType());
2192 Instruction *InstCombinerImpl::visitGetElementPtrInst(GetElementPtrInst &GEP) {
2193 Value *PtrOp = GEP.getOperand(0);
2194 SmallVector<Value *, 8> Indices(GEP.indices());
2195 Type *GEPType = GEP.getType();
2196 Type *GEPEltType = GEP.getSourceElementType();
2197 bool IsGEPSrcEleScalable = isa<ScalableVectorType>(GEPEltType);
2198 if (Value *V = SimplifyGEPInst(GEPEltType, PtrOp, Indices, GEP.isInBounds(),
2199 SQ.getWithInstruction(&GEP)))
2200 return replaceInstUsesWith(GEP, V);
2202 // For vector geps, use the generic demanded vector support.
2203 // Skip if GEP return type is scalable. The number of elements is unknown at
2205 if (auto *GEPFVTy = dyn_cast<FixedVectorType>(GEPType)) {
2206 auto VWidth = GEPFVTy->getNumElements();
2207 APInt UndefElts(VWidth, 0);
2208 APInt AllOnesEltMask(APInt::getAllOnes(VWidth));
2209 if (Value *V = SimplifyDemandedVectorElts(&GEP, AllOnesEltMask,
2212 return replaceInstUsesWith(GEP, V);
2216 // TODO: 1) Scalarize splat operands, 2) scalarize entire instruction if
2217 // possible (decide on canonical form for pointer broadcast), 3) exploit
2218 // undef elements to decrease demanded bits
2221 // Eliminate unneeded casts for indices, and replace indices which displace
2222 // by multiples of a zero size type with zero.
2223 bool MadeChange = false;
2225 // Index width may not be the same width as pointer width.
2226 // Data layout chooses the right type based on supported integer types.
2227 Type *NewScalarIndexTy =
2228 DL.getIndexType(GEP.getPointerOperandType()->getScalarType());
2230 gep_type_iterator GTI = gep_type_begin(GEP);
2231 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end(); I != E;
2233 // Skip indices into struct types.
2237 Type *IndexTy = (*I)->getType();
2238 Type *NewIndexType =
2239 IndexTy->isVectorTy()
2240 ? VectorType::get(NewScalarIndexTy,
2241 cast<VectorType>(IndexTy)->getElementCount())
2244 // If the element type has zero size then any index over it is equivalent
2245 // to an index of zero, so replace it with zero if it is not zero already.
2246 Type *EltTy = GTI.getIndexedType();
2247 if (EltTy->isSized() && DL.getTypeAllocSize(EltTy).isZero())
2248 if (!isa<Constant>(*I) || !match(I->get(), m_Zero())) {
2249 *I = Constant::getNullValue(NewIndexType);
2253 if (IndexTy != NewIndexType) {
2254 // If we are using a wider index than needed for this platform, shrink
2255 // it to what we need. If narrower, sign-extend it to what we need.
2256 // This explicit cast can make subsequent optimizations more obvious.
2257 *I = Builder.CreateIntCast(*I, NewIndexType, true);
2264 // Check to see if the inputs to the PHI node are getelementptr instructions.
2265 if (auto *PN = dyn_cast<PHINode>(PtrOp)) {
2266 auto *Op1 = dyn_cast<GetElementPtrInst>(PN->getOperand(0));
2270 // Don't fold a GEP into itself through a PHI node. This can only happen
2271 // through the back-edge of a loop. Folding a GEP into itself means that
2272 // the value of the previous iteration needs to be stored in the meantime,
2273 // thus requiring an additional register variable to be live, but not
2274 // actually achieving anything (the GEP still needs to be executed once per
2281 for (auto I = PN->op_begin()+1, E = PN->op_end(); I !=E; ++I) {
2282 auto *Op2 = dyn_cast<GetElementPtrInst>(*I);
2283 if (!Op2 || Op1->getNumOperands() != Op2->getNumOperands())
2286 // As for Op1 above, don't try to fold a GEP into itself.
2290 // Keep track of the type as we walk the GEP.
2291 Type *CurTy = nullptr;
2293 for (unsigned J = 0, F = Op1->getNumOperands(); J != F; ++J) {
2294 if (Op1->getOperand(J)->getType() != Op2->getOperand(J)->getType())
2297 if (Op1->getOperand(J) != Op2->getOperand(J)) {
2299 // We have not seen any differences yet in the GEPs feeding the
2300 // PHI yet, so we record this one if it is allowed to be a
2303 // The first two arguments can vary for any GEP, the rest have to be
2304 // static for struct slots
2306 assert(CurTy && "No current type?");
2307 if (CurTy->isStructTy())
2313 // The GEP is different by more than one input. While this could be
2314 // extended to support GEPs that vary by more than one variable it
2315 // doesn't make sense since it greatly increases the complexity and
2316 // would result in an R+R+R addressing mode which no backend
2317 // directly supports and would need to be broken into several
2318 // simpler instructions anyway.
2323 // Sink down a layer of the type for the next iteration.
2326 CurTy = Op1->getSourceElementType();
2329 GetElementPtrInst::getTypeAtIndex(CurTy, Op1->getOperand(J));
2335 // If not all GEPs are identical we'll have to create a new PHI node.
2336 // Check that the old PHI node has only one use so that it will get
2338 if (DI != -1 && !PN->hasOneUse())
2341 auto *NewGEP = cast<GetElementPtrInst>(Op1->clone());
2343 // All the GEPs feeding the PHI are identical. Clone one down into our
2344 // BB so that it can be merged with the current GEP.
2346 // All the GEPs feeding the PHI differ at a single offset. Clone a GEP
2347 // into the current block so it can be merged, and create a new PHI to
2351 IRBuilderBase::InsertPointGuard Guard(Builder);
2352 Builder.SetInsertPoint(PN);
2353 NewPN = Builder.CreatePHI(Op1->getOperand(DI)->getType(),
2354 PN->getNumOperands());
2357 for (auto &I : PN->operands())
2358 NewPN->addIncoming(cast<GEPOperator>(I)->getOperand(DI),
2359 PN->getIncomingBlock(I));
2361 NewGEP->setOperand(DI, NewPN);
2364 GEP.getParent()->getInstList().insert(
2365 GEP.getParent()->getFirstInsertionPt(), NewGEP);
2366 replaceOperand(GEP, 0, NewGEP);
2370 if (auto *Src = dyn_cast<GEPOperator>(PtrOp))
2371 if (Instruction *I = visitGEPOfGEP(GEP, Src))
2374 // Skip if GEP source element type is scalable. The type alloc size is unknown
2376 if (GEP.getNumIndices() == 1 && !IsGEPSrcEleScalable) {
2377 unsigned AS = GEP.getPointerAddressSpace();
2378 if (GEP.getOperand(1)->getType()->getScalarSizeInBits() ==
2379 DL.getIndexSizeInBits(AS)) {
2380 uint64_t TyAllocSize = DL.getTypeAllocSize(GEPEltType).getFixedSize();
2382 bool Matched = false;
2385 if (TyAllocSize == 1) {
2386 V = GEP.getOperand(1);
2388 } else if (match(GEP.getOperand(1),
2389 m_AShr(m_Value(V), m_ConstantInt(C)))) {
2390 if (TyAllocSize == 1ULL << C)
2392 } else if (match(GEP.getOperand(1),
2393 m_SDiv(m_Value(V), m_ConstantInt(C)))) {
2394 if (TyAllocSize == C)
2398 // Canonicalize (gep i8* X, (ptrtoint Y)-(ptrtoint X)) to (bitcast Y), but
2399 // only if both point to the same underlying object (otherwise provenance
2400 // is not necessarily retained).
2402 Value *X = GEP.getOperand(0);
2404 match(V, m_Sub(m_PtrToInt(m_Value(Y)), m_PtrToInt(m_Specific(X)))) &&
2405 getUnderlyingObject(X) == getUnderlyingObject(Y))
2406 return CastInst::CreatePointerBitCastOrAddrSpaceCast(Y, GEPType);
2410 // We do not handle pointer-vector geps here.
2411 if (GEPType->isVectorTy())
2414 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0).
2415 Value *StrippedPtr = PtrOp->stripPointerCasts();
2416 PointerType *StrippedPtrTy = cast<PointerType>(StrippedPtr->getType());
2418 // TODO: The basic approach of these folds is not compatible with opaque
2419 // pointers, because we can't use bitcasts as a hint for a desirable GEP
2420 // type. Instead, we should perform canonicalization directly on the GEP
2421 // type. For now, skip these.
2422 if (StrippedPtr != PtrOp && !StrippedPtrTy->isOpaque()) {
2423 bool HasZeroPointerIndex = false;
2424 Type *StrippedPtrEltTy = StrippedPtrTy->getNonOpaquePointerElementType();
2426 if (auto *C = dyn_cast<ConstantInt>(GEP.getOperand(1)))
2427 HasZeroPointerIndex = C->isZero();
2429 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
2430 // into : GEP [10 x i8]* X, i32 0, ...
2432 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
2433 // into : GEP i8* X, ...
2435 // This occurs when the program declares an array extern like "int X[];"
2436 if (HasZeroPointerIndex) {
2437 if (auto *CATy = dyn_cast<ArrayType>(GEPEltType)) {
2438 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
2439 if (CATy->getElementType() == StrippedPtrEltTy) {
2440 // -> GEP i8* X, ...
2441 SmallVector<Value *, 8> Idx(drop_begin(GEP.indices()));
2442 GetElementPtrInst *Res = GetElementPtrInst::Create(
2443 StrippedPtrEltTy, StrippedPtr, Idx, GEP.getName());
2444 Res->setIsInBounds(GEP.isInBounds());
2445 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace())
2447 // Insert Res, and create an addrspacecast.
2449 // GEP (addrspacecast i8 addrspace(1)* X to [0 x i8]*), i32 0, ...
2451 // %0 = GEP i8 addrspace(1)* X, ...
2452 // addrspacecast i8 addrspace(1)* %0 to i8*
2453 return new AddrSpaceCastInst(Builder.Insert(Res), GEPType);
2456 if (auto *XATy = dyn_cast<ArrayType>(StrippedPtrEltTy)) {
2457 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
2458 if (CATy->getElementType() == XATy->getElementType()) {
2459 // -> GEP [10 x i8]* X, i32 0, ...
2460 // At this point, we know that the cast source type is a pointer
2461 // to an array of the same type as the destination pointer
2462 // array. Because the array type is never stepped over (there
2463 // is a leading zero) we can fold the cast into this GEP.
2464 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace()) {
2465 GEP.setSourceElementType(XATy);
2466 return replaceOperand(GEP, 0, StrippedPtr);
2468 // Cannot replace the base pointer directly because StrippedPtr's
2469 // address space is different. Instead, create a new GEP followed by
2470 // an addrspacecast.
2472 // GEP (addrspacecast [10 x i8] addrspace(1)* X to [0 x i8]*),
2475 // %0 = GEP [10 x i8] addrspace(1)* X, ...
2476 // addrspacecast i8 addrspace(1)* %0 to i8*
2477 SmallVector<Value *, 8> Idx(GEP.indices());
2480 ? Builder.CreateInBoundsGEP(StrippedPtrEltTy, StrippedPtr,
2482 : Builder.CreateGEP(StrippedPtrEltTy, StrippedPtr, Idx,
2484 return new AddrSpaceCastInst(NewGEP, GEPType);
2488 } else if (GEP.getNumOperands() == 2 && !IsGEPSrcEleScalable) {
2489 // Skip if GEP source element type is scalable. The type alloc size is
2490 // unknown at compile-time.
2491 // Transform things like: %t = getelementptr i32*
2492 // bitcast ([2 x i32]* %str to i32*), i32 %V into: %t1 = getelementptr [2
2493 // x i32]* %str, i32 0, i32 %V; bitcast
2494 if (StrippedPtrEltTy->isArrayTy() &&
2495 DL.getTypeAllocSize(StrippedPtrEltTy->getArrayElementType()) ==
2496 DL.getTypeAllocSize(GEPEltType)) {
2497 Type *IdxType = DL.getIndexType(GEPType);
2498 Value *Idx[2] = { Constant::getNullValue(IdxType), GEP.getOperand(1) };
2501 ? Builder.CreateInBoundsGEP(StrippedPtrEltTy, StrippedPtr, Idx,
2503 : Builder.CreateGEP(StrippedPtrEltTy, StrippedPtr, Idx,
2506 // V and GEP are both pointer types --> BitCast
2507 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP, GEPType);
2510 // Transform things like:
2511 // %V = mul i64 %N, 4
2512 // %t = getelementptr i8* bitcast (i32* %arr to i8*), i32 %V
2513 // into: %t1 = getelementptr i32* %arr, i32 %N; bitcast
2514 if (GEPEltType->isSized() && StrippedPtrEltTy->isSized()) {
2515 // Check that changing the type amounts to dividing the index by a scale
2517 uint64_t ResSize = DL.getTypeAllocSize(GEPEltType).getFixedSize();
2518 uint64_t SrcSize = DL.getTypeAllocSize(StrippedPtrEltTy).getFixedSize();
2519 if (ResSize && SrcSize % ResSize == 0) {
2520 Value *Idx = GEP.getOperand(1);
2521 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
2522 uint64_t Scale = SrcSize / ResSize;
2524 // Earlier transforms ensure that the index has the right type
2525 // according to Data Layout, which considerably simplifies the
2526 // logic by eliminating implicit casts.
2527 assert(Idx->getType() == DL.getIndexType(GEPType) &&
2528 "Index type does not match the Data Layout preferences");
2531 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
2532 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
2533 // If the multiplication NewIdx * Scale may overflow then the new
2534 // GEP may not be "inbounds".
2536 GEP.isInBounds() && NSW
2537 ? Builder.CreateInBoundsGEP(StrippedPtrEltTy, StrippedPtr,
2538 NewIdx, GEP.getName())
2539 : Builder.CreateGEP(StrippedPtrEltTy, StrippedPtr, NewIdx,
2542 // The NewGEP must be pointer typed, so must the old one -> BitCast
2543 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
2549 // Similarly, transform things like:
2550 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
2551 // (where tmp = 8*tmp2) into:
2552 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
2553 if (GEPEltType->isSized() && StrippedPtrEltTy->isSized() &&
2554 StrippedPtrEltTy->isArrayTy()) {
2555 // Check that changing to the array element type amounts to dividing the
2556 // index by a scale factor.
2557 uint64_t ResSize = DL.getTypeAllocSize(GEPEltType).getFixedSize();
2558 uint64_t ArrayEltSize =
2559 DL.getTypeAllocSize(StrippedPtrEltTy->getArrayElementType())
2561 if (ResSize && ArrayEltSize % ResSize == 0) {
2562 Value *Idx = GEP.getOperand(1);
2563 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
2564 uint64_t Scale = ArrayEltSize / ResSize;
2566 // Earlier transforms ensure that the index has the right type
2567 // according to the Data Layout, which considerably simplifies
2568 // the logic by eliminating implicit casts.
2569 assert(Idx->getType() == DL.getIndexType(GEPType) &&
2570 "Index type does not match the Data Layout preferences");
2573 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
2574 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
2575 // If the multiplication NewIdx * Scale may overflow then the new
2576 // GEP may not be "inbounds".
2577 Type *IndTy = DL.getIndexType(GEPType);
2578 Value *Off[2] = {Constant::getNullValue(IndTy), NewIdx};
2581 GEP.isInBounds() && NSW
2582 ? Builder.CreateInBoundsGEP(StrippedPtrEltTy, StrippedPtr,
2584 : Builder.CreateGEP(StrippedPtrEltTy, StrippedPtr, Off,
2586 // The NewGEP must be pointer typed, so must the old one -> BitCast
2587 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
2595 // addrspacecast between types is canonicalized as a bitcast, then an
2596 // addrspacecast. To take advantage of the below bitcast + struct GEP, look
2597 // through the addrspacecast.
2598 Value *ASCStrippedPtrOp = PtrOp;
2599 if (auto *ASC = dyn_cast<AddrSpaceCastInst>(PtrOp)) {
2600 // X = bitcast A addrspace(1)* to B addrspace(1)*
2601 // Y = addrspacecast A addrspace(1)* to B addrspace(2)*
2602 // Z = gep Y, <...constant indices...>
2603 // Into an addrspacecasted GEP of the struct.
2604 if (auto *BC = dyn_cast<BitCastInst>(ASC->getOperand(0)))
2605 ASCStrippedPtrOp = BC;
2608 if (auto *BCI = dyn_cast<BitCastInst>(ASCStrippedPtrOp))
2609 if (Instruction *I = visitGEPOfBitcast(BCI, GEP))
2612 if (!GEP.isInBounds()) {
2614 DL.getIndexSizeInBits(PtrOp->getType()->getPointerAddressSpace());
2615 APInt BasePtrOffset(IdxWidth, 0);
2616 Value *UnderlyingPtrOp =
2617 PtrOp->stripAndAccumulateInBoundsConstantOffsets(DL,
2619 if (auto *AI = dyn_cast<AllocaInst>(UnderlyingPtrOp)) {
2620 if (GEP.accumulateConstantOffset(DL, BasePtrOffset) &&
2621 BasePtrOffset.isNonNegative()) {
2624 DL.getTypeAllocSize(AI->getAllocatedType()).getKnownMinSize());
2625 if (BasePtrOffset.ule(AllocSize)) {
2626 return GetElementPtrInst::CreateInBounds(
2627 GEP.getSourceElementType(), PtrOp, Indices, GEP.getName());
2633 if (Instruction *R = foldSelectGEP(GEP, Builder))
2639 static bool isNeverEqualToUnescapedAlloc(Value *V, const TargetLibraryInfo &TLI,
2641 if (isa<ConstantPointerNull>(V))
2643 if (auto *LI = dyn_cast<LoadInst>(V))
2644 return isa<GlobalVariable>(LI->getPointerOperand());
2645 // Two distinct allocations will never be equal.
2646 return isAllocLikeFn(V, &TLI) && V != AI;
2649 /// Given a call CB which uses an address UsedV, return true if we can prove the
2650 /// call's only possible effect is storing to V.
2651 static bool isRemovableWrite(CallBase &CB, Value *UsedV,
2652 const TargetLibraryInfo &TLI) {
2653 if (!CB.use_empty())
2654 // TODO: add recursion if returned attribute is present
2657 if (CB.isTerminator())
2658 // TODO: remove implementation restriction
2661 if (!CB.willReturn() || !CB.doesNotThrow())
2664 // If the only possible side effect of the call is writing to the alloca,
2665 // and the result isn't used, we can safely remove any reads implied by the
2666 // call including those which might read the alloca itself.
2667 Optional<MemoryLocation> Dest = MemoryLocation::getForDest(&CB, TLI);
2668 return Dest && Dest->Ptr == UsedV;
2671 static bool isAllocSiteRemovable(Instruction *AI,
2672 SmallVectorImpl<WeakTrackingVH> &Users,
2673 const TargetLibraryInfo &TLI) {
2674 SmallVector<Instruction*, 4> Worklist;
2675 Worklist.push_back(AI);
2678 Instruction *PI = Worklist.pop_back_val();
2679 for (User *U : PI->users()) {
2680 Instruction *I = cast<Instruction>(U);
2681 switch (I->getOpcode()) {
2683 // Give up the moment we see something we can't handle.
2686 case Instruction::AddrSpaceCast:
2687 case Instruction::BitCast:
2688 case Instruction::GetElementPtr:
2689 Users.emplace_back(I);
2690 Worklist.push_back(I);
2693 case Instruction::ICmp: {
2694 ICmpInst *ICI = cast<ICmpInst>(I);
2695 // We can fold eq/ne comparisons with null to false/true, respectively.
2696 // We also fold comparisons in some conditions provided the alloc has
2697 // not escaped (see isNeverEqualToUnescapedAlloc).
2698 if (!ICI->isEquality())
2700 unsigned OtherIndex = (ICI->getOperand(0) == PI) ? 1 : 0;
2701 if (!isNeverEqualToUnescapedAlloc(ICI->getOperand(OtherIndex), TLI, AI))
2703 Users.emplace_back(I);
2707 case Instruction::Call:
2708 // Ignore no-op and store intrinsics.
2709 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
2710 switch (II->getIntrinsicID()) {
2714 case Intrinsic::memmove:
2715 case Intrinsic::memcpy:
2716 case Intrinsic::memset: {
2717 MemIntrinsic *MI = cast<MemIntrinsic>(II);
2718 if (MI->isVolatile() || MI->getRawDest() != PI)
2722 case Intrinsic::assume:
2723 case Intrinsic::invariant_start:
2724 case Intrinsic::invariant_end:
2725 case Intrinsic::lifetime_start:
2726 case Intrinsic::lifetime_end:
2727 case Intrinsic::objectsize:
2728 Users.emplace_back(I);
2730 case Intrinsic::launder_invariant_group:
2731 case Intrinsic::strip_invariant_group:
2732 Users.emplace_back(I);
2733 Worklist.push_back(I);
2738 if (isRemovableWrite(*cast<CallBase>(I), PI, TLI)) {
2739 Users.emplace_back(I);
2743 if (isFreeCall(I, &TLI)) {
2744 Users.emplace_back(I);
2748 if (isReallocLikeFn(I, &TLI)) {
2749 Users.emplace_back(I);
2750 Worklist.push_back(I);
2756 case Instruction::Store: {
2757 StoreInst *SI = cast<StoreInst>(I);
2758 if (SI->isVolatile() || SI->getPointerOperand() != PI)
2760 Users.emplace_back(I);
2764 llvm_unreachable("missing a return?");
2766 } while (!Worklist.empty());
2770 Instruction *InstCombinerImpl::visitAllocSite(Instruction &MI) {
2771 assert(isa<AllocaInst>(MI) || isAllocRemovable(&cast<CallBase>(MI), &TLI));
2773 // If we have a malloc call which is only used in any amount of comparisons to
2774 // null and free calls, delete the calls and replace the comparisons with true
2775 // or false as appropriate.
2777 // This is based on the principle that we can substitute our own allocation
2778 // function (which will never return null) rather than knowledge of the
2779 // specific function being called. In some sense this can change the permitted
2780 // outputs of a program (when we convert a malloc to an alloca, the fact that
2781 // the allocation is now on the stack is potentially visible, for example),
2782 // but we believe in a permissible manner.
2783 SmallVector<WeakTrackingVH, 64> Users;
2785 // If we are removing an alloca with a dbg.declare, insert dbg.value calls
2786 // before each store.
2787 SmallVector<DbgVariableIntrinsic *, 8> DVIs;
2788 std::unique_ptr<DIBuilder> DIB;
2789 if (isa<AllocaInst>(MI)) {
2790 findDbgUsers(DVIs, &MI);
2791 DIB.reset(new DIBuilder(*MI.getModule(), /*AllowUnresolved=*/false));
2794 if (isAllocSiteRemovable(&MI, Users, TLI)) {
2795 for (unsigned i = 0, e = Users.size(); i != e; ++i) {
2796 // Lowering all @llvm.objectsize calls first because they may
2797 // use a bitcast/GEP of the alloca we are removing.
2801 Instruction *I = cast<Instruction>(&*Users[i]);
2803 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
2804 if (II->getIntrinsicID() == Intrinsic::objectsize) {
2806 lowerObjectSizeCall(II, DL, &TLI, /*MustSucceed=*/true);
2807 replaceInstUsesWith(*I, Result);
2808 eraseInstFromFunction(*I);
2809 Users[i] = nullptr; // Skip examining in the next loop.
2813 for (unsigned i = 0, e = Users.size(); i != e; ++i) {
2817 Instruction *I = cast<Instruction>(&*Users[i]);
2819 if (ICmpInst *C = dyn_cast<ICmpInst>(I)) {
2820 replaceInstUsesWith(*C,
2821 ConstantInt::get(Type::getInt1Ty(C->getContext()),
2822 C->isFalseWhenEqual()));
2823 } else if (auto *SI = dyn_cast<StoreInst>(I)) {
2824 for (auto *DVI : DVIs)
2825 if (DVI->isAddressOfVariable())
2826 ConvertDebugDeclareToDebugValue(DVI, SI, *DIB);
2828 // Casts, GEP, or anything else: we're about to delete this instruction,
2829 // so it can not have any valid uses.
2830 replaceInstUsesWith(*I, PoisonValue::get(I->getType()));
2832 eraseInstFromFunction(*I);
2835 if (InvokeInst *II = dyn_cast<InvokeInst>(&MI)) {
2836 // Replace invoke with a NOP intrinsic to maintain the original CFG
2837 Module *M = II->getModule();
2838 Function *F = Intrinsic::getDeclaration(M, Intrinsic::donothing);
2839 InvokeInst::Create(F, II->getNormalDest(), II->getUnwindDest(),
2840 None, "", II->getParent());
2843 // Remove debug intrinsics which describe the value contained within the
2844 // alloca. In addition to removing dbg.{declare,addr} which simply point to
2845 // the alloca, remove dbg.value(<alloca>, ..., DW_OP_deref)'s as well, e.g.:
2848 // define void @foo(i32 %0) {
2849 // %a = alloca i32 ; Deleted.
2850 // store i32 %0, i32* %a
2851 // dbg.value(i32 %0, "arg0") ; Not deleted.
2852 // dbg.value(i32* %a, "arg0", DW_OP_deref) ; Deleted.
2853 // call void @trivially_inlinable_no_op(i32* %a)
2858 // This may not be required if we stop describing the contents of allocas
2859 // using dbg.value(<alloca>, ..., DW_OP_deref), but we currently do this in
2860 // the LowerDbgDeclare utility.
2862 // If there is a dead store to `%a` in @trivially_inlinable_no_op, the
2863 // "arg0" dbg.value may be stale after the call. However, failing to remove
2864 // the DW_OP_deref dbg.value causes large gaps in location coverage.
2865 for (auto *DVI : DVIs)
2866 if (DVI->isAddressOfVariable() || DVI->getExpression()->startsWithDeref())
2867 DVI->eraseFromParent();
2869 return eraseInstFromFunction(MI);
2874 /// Move the call to free before a NULL test.
2876 /// Check if this free is accessed after its argument has been test
2877 /// against NULL (property 0).
2878 /// If yes, it is legal to move this call in its predecessor block.
2880 /// The move is performed only if the block containing the call to free
2881 /// will be removed, i.e.:
2882 /// 1. it has only one predecessor P, and P has two successors
2883 /// 2. it contains the call, noops, and an unconditional branch
2884 /// 3. its successor is the same as its predecessor's successor
2886 /// The profitability is out-of concern here and this function should
2887 /// be called only if the caller knows this transformation would be
2888 /// profitable (e.g., for code size).
2889 static Instruction *tryToMoveFreeBeforeNullTest(CallInst &FI,
2890 const DataLayout &DL) {
2891 Value *Op = FI.getArgOperand(0);
2892 BasicBlock *FreeInstrBB = FI.getParent();
2893 BasicBlock *PredBB = FreeInstrBB->getSinglePredecessor();
2895 // Validate part of constraint #1: Only one predecessor
2896 // FIXME: We can extend the number of predecessor, but in that case, we
2897 // would duplicate the call to free in each predecessor and it may
2898 // not be profitable even for code size.
2902 // Validate constraint #2: Does this block contains only the call to
2903 // free, noops, and an unconditional branch?
2905 Instruction *FreeInstrBBTerminator = FreeInstrBB->getTerminator();
2906 if (!match(FreeInstrBBTerminator, m_UnconditionalBr(SuccBB)))
2909 // If there are only 2 instructions in the block, at this point,
2910 // this is the call to free and unconditional.
2911 // If there are more than 2 instructions, check that they are noops
2912 // i.e., they won't hurt the performance of the generated code.
2913 if (FreeInstrBB->size() != 2) {
2914 for (const Instruction &Inst : FreeInstrBB->instructionsWithoutDebug()) {
2915 if (&Inst == &FI || &Inst == FreeInstrBBTerminator)
2917 auto *Cast = dyn_cast<CastInst>(&Inst);
2918 if (!Cast || !Cast->isNoopCast(DL))
2922 // Validate the rest of constraint #1 by matching on the pred branch.
2923 Instruction *TI = PredBB->getTerminator();
2924 BasicBlock *TrueBB, *FalseBB;
2925 ICmpInst::Predicate Pred;
2926 if (!match(TI, m_Br(m_ICmp(Pred,
2927 m_CombineOr(m_Specific(Op),
2928 m_Specific(Op->stripPointerCasts())),
2932 if (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE)
2935 // Validate constraint #3: Ensure the null case just falls through.
2936 if (SuccBB != (Pred == ICmpInst::ICMP_EQ ? TrueBB : FalseBB))
2938 assert(FreeInstrBB == (Pred == ICmpInst::ICMP_EQ ? FalseBB : TrueBB) &&
2939 "Broken CFG: missing edge from predecessor to successor");
2941 // At this point, we know that everything in FreeInstrBB can be moved
2943 for (Instruction &Instr : llvm::make_early_inc_range(*FreeInstrBB)) {
2944 if (&Instr == FreeInstrBBTerminator)
2946 Instr.moveBefore(TI);
2948 assert(FreeInstrBB->size() == 1 &&
2949 "Only the branch instruction should remain");
2951 // Now that we've moved the call to free before the NULL check, we have to
2952 // remove any attributes on its parameter that imply it's non-null, because
2953 // those attributes might have only been valid because of the NULL check, and
2954 // we can get miscompiles if we keep them. This is conservative if non-null is
2955 // also implied by something other than the NULL check, but it's guaranteed to
2956 // be correct, and the conservativeness won't matter in practice, since the
2957 // attributes are irrelevant for the call to free itself and the pointer
2958 // shouldn't be used after the call.
2959 AttributeList Attrs = FI.getAttributes();
2960 Attrs = Attrs.removeParamAttribute(FI.getContext(), 0, Attribute::NonNull);
2961 Attribute Dereferenceable = Attrs.getParamAttr(0, Attribute::Dereferenceable);
2962 if (Dereferenceable.isValid()) {
2963 uint64_t Bytes = Dereferenceable.getDereferenceableBytes();
2964 Attrs = Attrs.removeParamAttribute(FI.getContext(), 0,
2965 Attribute::Dereferenceable);
2966 Attrs = Attrs.addDereferenceableOrNullParamAttr(FI.getContext(), 0, Bytes);
2968 FI.setAttributes(Attrs);
2973 Instruction *InstCombinerImpl::visitFree(CallInst &FI) {
2974 Value *Op = FI.getArgOperand(0);
2976 // free undef -> unreachable.
2977 if (isa<UndefValue>(Op)) {
2978 // Leave a marker since we can't modify the CFG here.
2979 CreateNonTerminatorUnreachable(&FI);
2980 return eraseInstFromFunction(FI);
2983 // If we have 'free null' delete the instruction. This can happen in stl code
2984 // when lots of inlining happens.
2985 if (isa<ConstantPointerNull>(Op))
2986 return eraseInstFromFunction(FI);
2988 // If we had free(realloc(...)) with no intervening uses, then eliminate the
2989 // realloc() entirely.
2990 if (CallInst *CI = dyn_cast<CallInst>(Op)) {
2991 if (CI->hasOneUse() && isReallocLikeFn(CI, &TLI)) {
2992 return eraseInstFromFunction(
2993 *replaceInstUsesWith(*CI, CI->getOperand(0)));
2997 // If we optimize for code size, try to move the call to free before the null
2998 // test so that simplify cfg can remove the empty block and dead code
2999 // elimination the branch. I.e., helps to turn something like:
3000 // if (foo) free(foo);
3004 // Note that we can only do this for 'free' and not for any flavor of
3005 // 'operator delete'; there is no 'operator delete' symbol for which we are
3006 // permitted to invent a call, even if we're passing in a null pointer.
3009 if (TLI.getLibFunc(FI, Func) && TLI.has(Func) && Func == LibFunc_free)
3010 if (Instruction *I = tryToMoveFreeBeforeNullTest(FI, DL))
3017 static bool isMustTailCall(Value *V) {
3018 if (auto *CI = dyn_cast<CallInst>(V))
3019 return CI->isMustTailCall();
3023 Instruction *InstCombinerImpl::visitReturnInst(ReturnInst &RI) {
3024 if (RI.getNumOperands() == 0) // ret void
3027 Value *ResultOp = RI.getOperand(0);
3028 Type *VTy = ResultOp->getType();
3029 if (!VTy->isIntegerTy() || isa<Constant>(ResultOp))
3032 // Don't replace result of musttail calls.
3033 if (isMustTailCall(ResultOp))
3036 // There might be assume intrinsics dominating this return that completely
3037 // determine the value. If so, constant fold it.
3038 KnownBits Known = computeKnownBits(ResultOp, 0, &RI);
3039 if (Known.isConstant())
3040 return replaceOperand(RI, 0,
3041 Constant::getIntegerValue(VTy, Known.getConstant()));
3046 // WARNING: keep in sync with SimplifyCFGOpt::simplifyUnreachable()!
3047 Instruction *InstCombinerImpl::visitUnreachableInst(UnreachableInst &I) {
3048 // Try to remove the previous instruction if it must lead to unreachable.
3049 // This includes instructions like stores and "llvm.assume" that may not get
3050 // removed by simple dead code elimination.
3051 while (Instruction *Prev = I.getPrevNonDebugInstruction()) {
3052 // While we theoretically can erase EH, that would result in a block that
3053 // used to start with an EH no longer starting with EH, which is invalid.
3054 // To make it valid, we'd need to fixup predecessors to no longer refer to
3055 // this block, but that changes CFG, which is not allowed in InstCombine.
3056 if (Prev->isEHPad())
3057 return nullptr; // Can not drop any more instructions. We're done here.
3059 if (!isGuaranteedToTransferExecutionToSuccessor(Prev))
3060 return nullptr; // Can not drop any more instructions. We're done here.
3061 // Otherwise, this instruction can be freely erased,
3062 // even if it is not side-effect free.
3064 // A value may still have uses before we process it here (for example, in
3065 // another unreachable block), so convert those to poison.
3066 replaceInstUsesWith(*Prev, PoisonValue::get(Prev->getType()));
3067 eraseInstFromFunction(*Prev);
3069 assert(I.getParent()->sizeWithoutDebug() == 1 && "The block is now empty.");
3070 // FIXME: recurse into unconditional predecessors?
3074 Instruction *InstCombinerImpl::visitUnconditionalBranchInst(BranchInst &BI) {
3075 assert(BI.isUnconditional() && "Only for unconditional branches.");
3077 // If this store is the second-to-last instruction in the basic block
3078 // (excluding debug info and bitcasts of pointers) and if the block ends with
3079 // an unconditional branch, try to move the store to the successor block.
3081 auto GetLastSinkableStore = [](BasicBlock::iterator BBI) {
3082 auto IsNoopInstrForStoreMerging = [](BasicBlock::iterator BBI) {
3083 return BBI->isDebugOrPseudoInst() ||
3084 (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy());
3087 BasicBlock::iterator FirstInstr = BBI->getParent()->begin();
3089 if (BBI != FirstInstr)
3091 } while (BBI != FirstInstr && IsNoopInstrForStoreMerging(BBI));
3093 return dyn_cast<StoreInst>(BBI);
3096 if (StoreInst *SI = GetLastSinkableStore(BasicBlock::iterator(BI)))
3097 if (mergeStoreIntoSuccessor(*SI))
3103 Instruction *InstCombinerImpl::visitBranchInst(BranchInst &BI) {
3104 if (BI.isUnconditional())
3105 return visitUnconditionalBranchInst(BI);
3107 // Change br (not X), label True, label False to: br X, label False, True
3109 if (match(&BI, m_Br(m_Not(m_Value(X)), m_BasicBlock(), m_BasicBlock())) &&
3110 !isa<Constant>(X)) {
3111 // Swap Destinations and condition...
3112 BI.swapSuccessors();
3113 return replaceOperand(BI, 0, X);
3116 // If the condition is irrelevant, remove the use so that other
3117 // transforms on the condition become more effective.
3118 if (!isa<ConstantInt>(BI.getCondition()) &&
3119 BI.getSuccessor(0) == BI.getSuccessor(1))
3120 return replaceOperand(
3121 BI, 0, ConstantInt::getFalse(BI.getCondition()->getType()));
3123 // Canonicalize, for example, fcmp_one -> fcmp_oeq.
3124 CmpInst::Predicate Pred;
3125 if (match(&BI, m_Br(m_OneUse(m_FCmp(Pred, m_Value(), m_Value())),
3126 m_BasicBlock(), m_BasicBlock())) &&
3127 !isCanonicalPredicate(Pred)) {
3128 // Swap destinations and condition.
3129 CmpInst *Cond = cast<CmpInst>(BI.getCondition());
3130 Cond->setPredicate(CmpInst::getInversePredicate(Pred));
3131 BI.swapSuccessors();
3132 Worklist.push(Cond);
3139 Instruction *InstCombinerImpl::visitSwitchInst(SwitchInst &SI) {
3140 Value *Cond = SI.getCondition();
3142 ConstantInt *AddRHS;
3143 if (match(Cond, m_Add(m_Value(Op0), m_ConstantInt(AddRHS)))) {
3144 // Change 'switch (X+4) case 1:' into 'switch (X) case -3'.
3145 for (auto Case : SI.cases()) {
3146 Constant *NewCase = ConstantExpr::getSub(Case.getCaseValue(), AddRHS);
3147 assert(isa<ConstantInt>(NewCase) &&
3148 "Result of expression should be constant");
3149 Case.setValue(cast<ConstantInt>(NewCase));
3151 return replaceOperand(SI, 0, Op0);
3154 KnownBits Known = computeKnownBits(Cond, 0, &SI);
3155 unsigned LeadingKnownZeros = Known.countMinLeadingZeros();
3156 unsigned LeadingKnownOnes = Known.countMinLeadingOnes();
3158 // Compute the number of leading bits we can ignore.
3159 // TODO: A better way to determine this would use ComputeNumSignBits().
3160 for (auto &C : SI.cases()) {
3161 LeadingKnownZeros = std::min(
3162 LeadingKnownZeros, C.getCaseValue()->getValue().countLeadingZeros());
3163 LeadingKnownOnes = std::min(
3164 LeadingKnownOnes, C.getCaseValue()->getValue().countLeadingOnes());
3167 unsigned NewWidth = Known.getBitWidth() - std::max(LeadingKnownZeros, LeadingKnownOnes);
3169 // Shrink the condition operand if the new type is smaller than the old type.
3170 // But do not shrink to a non-standard type, because backend can't generate
3171 // good code for that yet.
3172 // TODO: We can make it aggressive again after fixing PR39569.
3173 if (NewWidth > 0 && NewWidth < Known.getBitWidth() &&
3174 shouldChangeType(Known.getBitWidth(), NewWidth)) {
3175 IntegerType *Ty = IntegerType::get(SI.getContext(), NewWidth);
3176 Builder.SetInsertPoint(&SI);
3177 Value *NewCond = Builder.CreateTrunc(Cond, Ty, "trunc");
3179 for (auto Case : SI.cases()) {
3180 APInt TruncatedCase = Case.getCaseValue()->getValue().trunc(NewWidth);
3181 Case.setValue(ConstantInt::get(SI.getContext(), TruncatedCase));
3183 return replaceOperand(SI, 0, NewCond);
3189 Instruction *InstCombinerImpl::visitExtractValueInst(ExtractValueInst &EV) {
3190 Value *Agg = EV.getAggregateOperand();
3192 if (!EV.hasIndices())
3193 return replaceInstUsesWith(EV, Agg);
3195 if (Value *V = SimplifyExtractValueInst(Agg, EV.getIndices(),
3196 SQ.getWithInstruction(&EV)))
3197 return replaceInstUsesWith(EV, V);
3199 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
3200 // We're extracting from an insertvalue instruction, compare the indices
3201 const unsigned *exti, *exte, *insi, *inse;
3202 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
3203 exte = EV.idx_end(), inse = IV->idx_end();
3204 exti != exte && insi != inse;
3207 // The insert and extract both reference distinctly different elements.
3208 // This means the extract is not influenced by the insert, and we can
3209 // replace the aggregate operand of the extract with the aggregate
3210 // operand of the insert. i.e., replace
3211 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
3212 // %E = extractvalue { i32, { i32 } } %I, 0
3214 // %E = extractvalue { i32, { i32 } } %A, 0
3215 return ExtractValueInst::Create(IV->getAggregateOperand(),
3218 if (exti == exte && insi == inse)
3219 // Both iterators are at the end: Index lists are identical. Replace
3220 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
3221 // %C = extractvalue { i32, { i32 } } %B, 1, 0
3223 return replaceInstUsesWith(EV, IV->getInsertedValueOperand());
3225 // The extract list is a prefix of the insert list. i.e. replace
3226 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
3227 // %E = extractvalue { i32, { i32 } } %I, 1
3229 // %X = extractvalue { i32, { i32 } } %A, 1
3230 // %E = insertvalue { i32 } %X, i32 42, 0
3231 // by switching the order of the insert and extract (though the
3232 // insertvalue should be left in, since it may have other uses).
3233 Value *NewEV = Builder.CreateExtractValue(IV->getAggregateOperand(),
3235 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
3236 makeArrayRef(insi, inse));
3239 // The insert list is a prefix of the extract list
3240 // We can simply remove the common indices from the extract and make it
3241 // operate on the inserted value instead of the insertvalue result.
3243 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
3244 // %E = extractvalue { i32, { i32 } } %I, 1, 0
3246 // %E extractvalue { i32 } { i32 42 }, 0
3247 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
3248 makeArrayRef(exti, exte));
3250 if (WithOverflowInst *WO = dyn_cast<WithOverflowInst>(Agg)) {
3251 // We're extracting from an overflow intrinsic, see if we're the only user,
3252 // which allows us to simplify multiple result intrinsics to simpler
3253 // things that just get one value.
3254 if (WO->hasOneUse()) {
3255 // Check if we're grabbing only the result of a 'with overflow' intrinsic
3256 // and replace it with a traditional binary instruction.
3257 if (*EV.idx_begin() == 0) {
3258 Instruction::BinaryOps BinOp = WO->getBinaryOp();
3259 Value *LHS = WO->getLHS(), *RHS = WO->getRHS();
3260 // Replace the old instruction's uses with poison.
3261 replaceInstUsesWith(*WO, PoisonValue::get(WO->getType()));
3262 eraseInstFromFunction(*WO);
3263 return BinaryOperator::Create(BinOp, LHS, RHS);
3266 assert(*EV.idx_begin() == 1 &&
3267 "unexpected extract index for overflow inst");
3269 // If only the overflow result is used, and the right hand side is a
3270 // constant (or constant splat), we can remove the intrinsic by directly
3271 // checking for overflow.
3273 if (match(WO->getRHS(), m_APInt(C))) {
3274 // Compute the no-wrap range for LHS given RHS=C, then construct an
3275 // equivalent icmp, potentially using an offset.
3277 ConstantRange::makeExactNoWrapRegion(WO->getBinaryOp(), *C,
3278 WO->getNoWrapKind());
3280 CmpInst::Predicate Pred;
3281 APInt NewRHSC, Offset;
3282 NWR.getEquivalentICmp(Pred, NewRHSC, Offset);
3283 auto *OpTy = WO->getRHS()->getType();
3284 auto *NewLHS = WO->getLHS();
3286 NewLHS = Builder.CreateAdd(NewLHS, ConstantInt::get(OpTy, Offset));
3287 return new ICmpInst(ICmpInst::getInversePredicate(Pred), NewLHS,
3288 ConstantInt::get(OpTy, NewRHSC));
3292 if (LoadInst *L = dyn_cast<LoadInst>(Agg))
3293 // If the (non-volatile) load only has one use, we can rewrite this to a
3294 // load from a GEP. This reduces the size of the load. If a load is used
3295 // only by extractvalue instructions then this either must have been
3296 // optimized before, or it is a struct with padding, in which case we
3297 // don't want to do the transformation as it loses padding knowledge.
3298 if (L->isSimple() && L->hasOneUse()) {
3299 // extractvalue has integer indices, getelementptr has Value*s. Convert.
3300 SmallVector<Value*, 4> Indices;
3301 // Prefix an i32 0 since we need the first element.
3302 Indices.push_back(Builder.getInt32(0));
3303 for (unsigned Idx : EV.indices())
3304 Indices.push_back(Builder.getInt32(Idx));
3306 // We need to insert these at the location of the old load, not at that of
3307 // the extractvalue.
3308 Builder.SetInsertPoint(L);
3309 Value *GEP = Builder.CreateInBoundsGEP(L->getType(),
3310 L->getPointerOperand(), Indices);
3311 Instruction *NL = Builder.CreateLoad(EV.getType(), GEP);
3312 // Whatever aliasing information we had for the orignal load must also
3313 // hold for the smaller load, so propagate the annotations.
3314 NL->setAAMetadata(L->getAAMetadata());
3315 // Returning the load directly will cause the main loop to insert it in
3316 // the wrong spot, so use replaceInstUsesWith().
3317 return replaceInstUsesWith(EV, NL);
3319 // We could simplify extracts from other values. Note that nested extracts may
3320 // already be simplified implicitly by the above: extract (extract (insert) )
3321 // will be translated into extract ( insert ( extract ) ) first and then just
3322 // the value inserted, if appropriate. Similarly for extracts from single-use
3323 // loads: extract (extract (load)) will be translated to extract (load (gep))
3324 // and if again single-use then via load (gep (gep)) to load (gep).
3325 // However, double extracts from e.g. function arguments or return values
3326 // aren't handled yet.
3330 /// Return 'true' if the given typeinfo will match anything.
3331 static bool isCatchAll(EHPersonality Personality, Constant *TypeInfo) {
3332 switch (Personality) {
3333 case EHPersonality::GNU_C:
3334 case EHPersonality::GNU_C_SjLj:
3335 case EHPersonality::Rust:
3336 // The GCC C EH and Rust personality only exists to support cleanups, so
3337 // it's not clear what the semantics of catch clauses are.
3339 case EHPersonality::Unknown:
3341 case EHPersonality::GNU_Ada:
3342 // While __gnat_all_others_value will match any Ada exception, it doesn't
3343 // match foreign exceptions (or didn't, before gcc-4.7).
3345 case EHPersonality::GNU_CXX:
3346 case EHPersonality::GNU_CXX_SjLj:
3347 case EHPersonality::GNU_ObjC:
3348 case EHPersonality::MSVC_X86SEH:
3349 case EHPersonality::MSVC_TableSEH:
3350 case EHPersonality::MSVC_CXX:
3351 case EHPersonality::CoreCLR:
3352 case EHPersonality::Wasm_CXX:
3353 case EHPersonality::XL_CXX:
3354 return TypeInfo->isNullValue();
3356 llvm_unreachable("invalid enum");
3359 static bool shorter_filter(const Value *LHS, const Value *RHS) {
3361 cast<ArrayType>(LHS->getType())->getNumElements()
3363 cast<ArrayType>(RHS->getType())->getNumElements();
3366 Instruction *InstCombinerImpl::visitLandingPadInst(LandingPadInst &LI) {
3367 // The logic here should be correct for any real-world personality function.
3368 // However if that turns out not to be true, the offending logic can always
3369 // be conditioned on the personality function, like the catch-all logic is.
3370 EHPersonality Personality =
3371 classifyEHPersonality(LI.getParent()->getParent()->getPersonalityFn());
3373 // Simplify the list of clauses, eg by removing repeated catch clauses
3374 // (these are often created by inlining).
3375 bool MakeNewInstruction = false; // If true, recreate using the following:
3376 SmallVector<Constant *, 16> NewClauses; // - Clauses for the new instruction;
3377 bool CleanupFlag = LI.isCleanup(); // - The new instruction is a cleanup.
3379 SmallPtrSet<Value *, 16> AlreadyCaught; // Typeinfos known caught already.
3380 for (unsigned i = 0, e = LI.getNumClauses(); i != e; ++i) {
3381 bool isLastClause = i + 1 == e;
3382 if (LI.isCatch(i)) {
3384 Constant *CatchClause = LI.getClause(i);
3385 Constant *TypeInfo = CatchClause->stripPointerCasts();
3387 // If we already saw this clause, there is no point in having a second
3389 if (AlreadyCaught.insert(TypeInfo).second) {
3390 // This catch clause was not already seen.
3391 NewClauses.push_back(CatchClause);
3393 // Repeated catch clause - drop the redundant copy.
3394 MakeNewInstruction = true;
3397 // If this is a catch-all then there is no point in keeping any following
3398 // clauses or marking the landingpad as having a cleanup.
3399 if (isCatchAll(Personality, TypeInfo)) {
3401 MakeNewInstruction = true;
3402 CleanupFlag = false;
3406 // A filter clause. If any of the filter elements were already caught
3407 // then they can be dropped from the filter. It is tempting to try to
3408 // exploit the filter further by saying that any typeinfo that does not
3409 // occur in the filter can't be caught later (and thus can be dropped).
3410 // However this would be wrong, since typeinfos can match without being
3411 // equal (for example if one represents a C++ class, and the other some
3412 // class derived from it).
3413 assert(LI.isFilter(i) && "Unsupported landingpad clause!");
3414 Constant *FilterClause = LI.getClause(i);
3415 ArrayType *FilterType = cast<ArrayType>(FilterClause->getType());
3416 unsigned NumTypeInfos = FilterType->getNumElements();
3418 // An empty filter catches everything, so there is no point in keeping any
3419 // following clauses or marking the landingpad as having a cleanup. By
3420 // dealing with this case here the following code is made a bit simpler.
3421 if (!NumTypeInfos) {
3422 NewClauses.push_back(FilterClause);
3424 MakeNewInstruction = true;
3425 CleanupFlag = false;
3429 bool MakeNewFilter = false; // If true, make a new filter.
3430 SmallVector<Constant *, 16> NewFilterElts; // New elements.
3431 if (isa<ConstantAggregateZero>(FilterClause)) {
3432 // Not an empty filter - it contains at least one null typeinfo.
3433 assert(NumTypeInfos > 0 && "Should have handled empty filter already!");
3434 Constant *TypeInfo =
3435 Constant::getNullValue(FilterType->getElementType());
3436 // If this typeinfo is a catch-all then the filter can never match.
3437 if (isCatchAll(Personality, TypeInfo)) {
3438 // Throw the filter away.
3439 MakeNewInstruction = true;
3443 // There is no point in having multiple copies of this typeinfo, so
3444 // discard all but the first copy if there is more than one.
3445 NewFilterElts.push_back(TypeInfo);
3446 if (NumTypeInfos > 1)
3447 MakeNewFilter = true;
3449 ConstantArray *Filter = cast<ConstantArray>(FilterClause);
3450 SmallPtrSet<Value *, 16> SeenInFilter; // For uniquing the elements.
3451 NewFilterElts.reserve(NumTypeInfos);
3453 // Remove any filter elements that were already caught or that already
3454 // occurred in the filter. While there, see if any of the elements are
3455 // catch-alls. If so, the filter can be discarded.
3456 bool SawCatchAll = false;
3457 for (unsigned j = 0; j != NumTypeInfos; ++j) {
3458 Constant *Elt = Filter->getOperand(j);
3459 Constant *TypeInfo = Elt->stripPointerCasts();
3460 if (isCatchAll(Personality, TypeInfo)) {
3461 // This element is a catch-all. Bail out, noting this fact.
3466 // Even if we've seen a type in a catch clause, we don't want to
3467 // remove it from the filter. An unexpected type handler may be
3468 // set up for a call site which throws an exception of the same
3469 // type caught. In order for the exception thrown by the unexpected
3470 // handler to propagate correctly, the filter must be correctly
3471 // described for the call site.
3475 // void unexpected() { throw 1;}
3476 // void foo() throw (int) {
3477 // std::set_unexpected(unexpected);
3480 // } catch (int i) {}
3483 // There is no point in having multiple copies of the same typeinfo in
3484 // a filter, so only add it if we didn't already.
3485 if (SeenInFilter.insert(TypeInfo).second)
3486 NewFilterElts.push_back(cast<Constant>(Elt));
3488 // A filter containing a catch-all cannot match anything by definition.
3490 // Throw the filter away.
3491 MakeNewInstruction = true;
3495 // If we dropped something from the filter, make a new one.
3496 if (NewFilterElts.size() < NumTypeInfos)
3497 MakeNewFilter = true;
3499 if (MakeNewFilter) {
3500 FilterType = ArrayType::get(FilterType->getElementType(),
3501 NewFilterElts.size());
3502 FilterClause = ConstantArray::get(FilterType, NewFilterElts);
3503 MakeNewInstruction = true;
3506 NewClauses.push_back(FilterClause);
3508 // If the new filter is empty then it will catch everything so there is
3509 // no point in keeping any following clauses or marking the landingpad
3510 // as having a cleanup. The case of the original filter being empty was
3511 // already handled above.
3512 if (MakeNewFilter && !NewFilterElts.size()) {
3513 assert(MakeNewInstruction && "New filter but not a new instruction!");
3514 CleanupFlag = false;
3520 // If several filters occur in a row then reorder them so that the shortest
3521 // filters come first (those with the smallest number of elements). This is
3522 // advantageous because shorter filters are more likely to match, speeding up
3523 // unwinding, but mostly because it increases the effectiveness of the other
3524 // filter optimizations below.
3525 for (unsigned i = 0, e = NewClauses.size(); i + 1 < e; ) {
3527 // Find the maximal 'j' s.t. the range [i, j) consists entirely of filters.
3528 for (j = i; j != e; ++j)
3529 if (!isa<ArrayType>(NewClauses[j]->getType()))
3532 // Check whether the filters are already sorted by length. We need to know
3533 // if sorting them is actually going to do anything so that we only make a
3534 // new landingpad instruction if it does.
3535 for (unsigned k = i; k + 1 < j; ++k)
3536 if (shorter_filter(NewClauses[k+1], NewClauses[k])) {
3537 // Not sorted, so sort the filters now. Doing an unstable sort would be
3538 // correct too but reordering filters pointlessly might confuse users.
3539 std::stable_sort(NewClauses.begin() + i, NewClauses.begin() + j,
3541 MakeNewInstruction = true;
3545 // Look for the next batch of filters.
3549 // If typeinfos matched if and only if equal, then the elements of a filter L
3550 // that occurs later than a filter F could be replaced by the intersection of
3551 // the elements of F and L. In reality two typeinfos can match without being
3552 // equal (for example if one represents a C++ class, and the other some class
3553 // derived from it) so it would be wrong to perform this transform in general.
3554 // However the transform is correct and useful if F is a subset of L. In that
3555 // case L can be replaced by F, and thus removed altogether since repeating a
3556 // filter is pointless. So here we look at all pairs of filters F and L where
3557 // L follows F in the list of clauses, and remove L if every element of F is
3558 // an element of L. This can occur when inlining C++ functions with exception
3560 for (unsigned i = 0; i + 1 < NewClauses.size(); ++i) {
3561 // Examine each filter in turn.
3562 Value *Filter = NewClauses[i];
3563 ArrayType *FTy = dyn_cast<ArrayType>(Filter->getType());
3565 // Not a filter - skip it.
3567 unsigned FElts = FTy->getNumElements();
3568 // Examine each filter following this one. Doing this backwards means that
3569 // we don't have to worry about filters disappearing under us when removed.
3570 for (unsigned j = NewClauses.size() - 1; j != i; --j) {
3571 Value *LFilter = NewClauses[j];
3572 ArrayType *LTy = dyn_cast<ArrayType>(LFilter->getType());
3574 // Not a filter - skip it.
3576 // If Filter is a subset of LFilter, i.e. every element of Filter is also
3577 // an element of LFilter, then discard LFilter.
3578 SmallVectorImpl<Constant *>::iterator J = NewClauses.begin() + j;
3579 // If Filter is empty then it is a subset of LFilter.
3582 NewClauses.erase(J);
3583 MakeNewInstruction = true;
3584 // Move on to the next filter.
3587 unsigned LElts = LTy->getNumElements();
3588 // If Filter is longer than LFilter then it cannot be a subset of it.
3590 // Move on to the next filter.
3592 // At this point we know that LFilter has at least one element.
3593 if (isa<ConstantAggregateZero>(LFilter)) { // LFilter only contains zeros.
3594 // Filter is a subset of LFilter iff Filter contains only zeros (as we
3595 // already know that Filter is not longer than LFilter).
3596 if (isa<ConstantAggregateZero>(Filter)) {
3597 assert(FElts <= LElts && "Should have handled this case earlier!");
3599 NewClauses.erase(J);
3600 MakeNewInstruction = true;
3602 // Move on to the next filter.
3605 ConstantArray *LArray = cast<ConstantArray>(LFilter);
3606 if (isa<ConstantAggregateZero>(Filter)) { // Filter only contains zeros.
3607 // Since Filter is non-empty and contains only zeros, it is a subset of
3608 // LFilter iff LFilter contains a zero.
3609 assert(FElts > 0 && "Should have eliminated the empty filter earlier!");
3610 for (unsigned l = 0; l != LElts; ++l)
3611 if (LArray->getOperand(l)->isNullValue()) {
3612 // LFilter contains a zero - discard it.
3613 NewClauses.erase(J);
3614 MakeNewInstruction = true;
3617 // Move on to the next filter.
3620 // At this point we know that both filters are ConstantArrays. Loop over
3621 // operands to see whether every element of Filter is also an element of
3622 // LFilter. Since filters tend to be short this is probably faster than
3623 // using a method that scales nicely.
3624 ConstantArray *FArray = cast<ConstantArray>(Filter);
3625 bool AllFound = true;
3626 for (unsigned f = 0; f != FElts; ++f) {
3627 Value *FTypeInfo = FArray->getOperand(f)->stripPointerCasts();
3629 for (unsigned l = 0; l != LElts; ++l) {
3630 Value *LTypeInfo = LArray->getOperand(l)->stripPointerCasts();
3631 if (LTypeInfo == FTypeInfo) {
3641 NewClauses.erase(J);
3642 MakeNewInstruction = true;
3644 // Move on to the next filter.
3648 // If we changed any of the clauses, replace the old landingpad instruction
3650 if (MakeNewInstruction) {
3651 LandingPadInst *NLI = LandingPadInst::Create(LI.getType(),
3653 for (unsigned i = 0, e = NewClauses.size(); i != e; ++i)
3654 NLI->addClause(NewClauses[i]);
3655 // A landing pad with no clauses must have the cleanup flag set. It is
3656 // theoretically possible, though highly unlikely, that we eliminated all
3657 // clauses. If so, force the cleanup flag to true.
3658 if (NewClauses.empty())
3660 NLI->setCleanup(CleanupFlag);
3664 // Even if none of the clauses changed, we may nonetheless have understood
3665 // that the cleanup flag is pointless. Clear it if so.
3666 if (LI.isCleanup() != CleanupFlag) {
3667 assert(!CleanupFlag && "Adding a cleanup, not removing one?!");
3668 LI.setCleanup(CleanupFlag);
3676 InstCombinerImpl::pushFreezeToPreventPoisonFromPropagating(FreezeInst &OrigFI) {
3677 // Try to push freeze through instructions that propagate but don't produce
3678 // poison as far as possible. If an operand of freeze follows three
3679 // conditions 1) one-use, 2) does not produce poison, and 3) has all but one
3680 // guaranteed-non-poison operands then push the freeze through to the one
3681 // operand that is not guaranteed non-poison. The actual transform is as
3683 // Op1 = ... ; Op1 can be posion
3684 // Op0 = Inst(Op1, NonPoisonOps...) ; Op0 has only one use and only have
3685 // ; single guaranteed-non-poison operands
3686 // ... = Freeze(Op0)
3689 // Op1.fr = Freeze(Op1)
3690 // ... = Inst(Op1.fr, NonPoisonOps...)
3691 auto *OrigOp = OrigFI.getOperand(0);
3692 auto *OrigOpInst = dyn_cast<Instruction>(OrigOp);
3694 // While we could change the other users of OrigOp to use freeze(OrigOp), that
3695 // potentially reduces their optimization potential, so let's only do this iff
3696 // the OrigOp is only used by the freeze.
3697 if (!OrigOpInst || !OrigOpInst->hasOneUse() || isa<PHINode>(OrigOp))
3700 // We can't push the freeze through an instruction which can itself create
3701 // poison. If the only source of new poison is flags, we can simply
3702 // strip them (since we know the only use is the freeze and nothing can
3703 // benefit from them.)
3704 if (canCreateUndefOrPoison(cast<Operator>(OrigOp), /*ConsiderFlags*/ false))
3707 // If operand is guaranteed not to be poison, there is no need to add freeze
3708 // to the operand. So we first find the operand that is not guaranteed to be
3710 Use *MaybePoisonOperand = nullptr;
3711 for (Use &U : OrigOpInst->operands()) {
3712 if (isGuaranteedNotToBeUndefOrPoison(U.get()))
3714 if (!MaybePoisonOperand)
3715 MaybePoisonOperand = &U;
3720 OrigOpInst->dropPoisonGeneratingFlags();
3722 // If all operands are guaranteed to be non-poison, we can drop freeze.
3723 if (!MaybePoisonOperand)
3726 auto *FrozenMaybePoisonOperand = new FreezeInst(
3727 MaybePoisonOperand->get(), MaybePoisonOperand->get()->getName() + ".fr");
3729 replaceUse(*MaybePoisonOperand, FrozenMaybePoisonOperand);
3730 FrozenMaybePoisonOperand->insertBefore(OrigOpInst);
3734 bool InstCombinerImpl::freezeDominatedUses(FreezeInst &FI) {
3735 Value *Op = FI.getOperand(0);
3737 if (isa<Constant>(Op))
3740 bool Changed = false;
3741 Op->replaceUsesWithIf(&FI, [&](Use &U) -> bool {
3742 bool Dominates = DT.dominates(&FI, U);
3743 Changed |= Dominates;
3750 Instruction *InstCombinerImpl::visitFreeze(FreezeInst &I) {
3751 Value *Op0 = I.getOperand(0);
3753 if (Value *V = SimplifyFreezeInst(Op0, SQ.getWithInstruction(&I)))
3754 return replaceInstUsesWith(I, V);
3756 // freeze (phi const, x) --> phi const, (freeze x)
3757 if (auto *PN = dyn_cast<PHINode>(Op0)) {
3758 if (Instruction *NV = foldOpIntoPhi(I, PN))
3762 if (Value *NI = pushFreezeToPreventPoisonFromPropagating(I))
3763 return replaceInstUsesWith(I, NI);
3765 if (match(Op0, m_Undef())) {
3766 // If I is freeze(undef), see its uses and fold it to the best constant.
3768 // - select's condition: pick the value that leads to choosing a constant
3769 // - other ops: pick 0
3770 Constant *BestValue = nullptr;
3771 Constant *NullValue = Constant::getNullValue(I.getType());
3772 for (const auto *U : I.users()) {
3773 Constant *C = NullValue;
3775 if (match(U, m_Or(m_Value(), m_Value())))
3776 C = Constant::getAllOnesValue(I.getType());
3777 else if (const auto *SI = dyn_cast<SelectInst>(U)) {
3778 if (SI->getCondition() == &I) {
3779 APInt CondVal(1, isa<Constant>(SI->getFalseValue()) ? 0 : 1);
3780 C = Constant::getIntegerValue(I.getType(), CondVal);
3786 else if (BestValue != C)
3787 BestValue = NullValue;
3790 return replaceInstUsesWith(I, BestValue);
3793 // Replace all dominated uses of Op to freeze(Op).
3794 if (freezeDominatedUses(I))
3800 /// Check for case where the call writes to an otherwise dead alloca. This
3801 /// shows up for unused out-params in idiomatic C/C++ code. Note that this
3802 /// helper *only* analyzes the write; doesn't check any other legality aspect.
3803 static bool SoleWriteToDeadLocal(Instruction *I, TargetLibraryInfo &TLI) {
3804 auto *CB = dyn_cast<CallBase>(I);
3806 // TODO: handle e.g. store to alloca here - only worth doing if we extend
3807 // to allow reload along used path as described below. Otherwise, this
3808 // is simply a store to a dead allocation which will be removed.
3810 Optional<MemoryLocation> Dest = MemoryLocation::getForDest(CB, TLI);
3813 auto *AI = dyn_cast<AllocaInst>(getUnderlyingObject(Dest->Ptr));
3815 // TODO: allow malloc?
3817 // TODO: allow memory access dominated by move point? Note that since AI
3818 // could have a reference to itself captured by the call, we would need to
3819 // account for cycles in doing so.
3820 SmallVector<const User *> AllocaUsers;
3821 SmallPtrSet<const User *, 4> Visited;
3822 auto pushUsers = [&](const Instruction &I) {
3823 for (const User *U : I.users()) {
3824 if (Visited.insert(U).second)
3825 AllocaUsers.push_back(U);
3829 while (!AllocaUsers.empty()) {
3830 auto *UserI = cast<Instruction>(AllocaUsers.pop_back_val());
3831 if (isa<BitCastInst>(UserI) || isa<GetElementPtrInst>(UserI) ||
3832 isa<AddrSpaceCastInst>(UserI)) {
3838 // TODO: support lifetime.start/end here
3844 /// Try to move the specified instruction from its current block into the
3845 /// beginning of DestBlock, which can only happen if it's safe to move the
3846 /// instruction past all of the instructions between it and the end of its
3848 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock,
3849 TargetLibraryInfo &TLI) {
3850 assert(I->getUniqueUndroppableUser() && "Invariants didn't hold!");
3851 BasicBlock *SrcBlock = I->getParent();
3853 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
3854 if (isa<PHINode>(I) || I->isEHPad() || I->mayThrow() || !I->willReturn() ||
3858 // Do not sink static or dynamic alloca instructions. Static allocas must
3859 // remain in the entry block, and dynamic allocas must not be sunk in between
3860 // a stacksave / stackrestore pair, which would incorrectly shorten its
3862 if (isa<AllocaInst>(I))
3865 // Do not sink into catchswitch blocks.
3866 if (isa<CatchSwitchInst>(DestBlock->getTerminator()))
3869 // Do not sink convergent call instructions.
3870 if (auto *CI = dyn_cast<CallInst>(I)) {
3871 if (CI->isConvergent())
3875 // Unless we can prove that the memory write isn't visibile except on the
3876 // path we're sinking to, we must bail.
3877 if (I->mayWriteToMemory()) {
3878 if (!SoleWriteToDeadLocal(I, TLI))
3882 // We can only sink load instructions if there is nothing between the load and
3883 // the end of block that could change the value.
3884 if (I->mayReadFromMemory()) {
3885 // We don't want to do any sophisticated alias analysis, so we only check
3886 // the instructions after I in I's parent block if we try to sink to its
3888 if (DestBlock->getUniquePredecessor() != I->getParent())
3890 for (BasicBlock::iterator Scan = std::next(I->getIterator()),
3891 E = I->getParent()->end();
3893 if (Scan->mayWriteToMemory())
3897 I->dropDroppableUses([DestBlock](const Use *U) {
3898 if (auto *I = dyn_cast<Instruction>(U->getUser()))
3899 return I->getParent() != DestBlock;
3902 /// FIXME: We could remove droppable uses that are not dominated by
3903 /// the new position.
3905 BasicBlock::iterator InsertPos = DestBlock->getFirstInsertionPt();
3906 I->moveBefore(&*InsertPos);
3909 // Also sink all related debug uses from the source basic block. Otherwise we
3910 // get debug use before the def. Attempt to salvage debug uses first, to
3911 // maximise the range variables have location for. If we cannot salvage, then
3912 // mark the location undef: we know it was supposed to receive a new location
3913 // here, but that computation has been sunk.
3914 SmallVector<DbgVariableIntrinsic *, 2> DbgUsers;
3915 findDbgUsers(DbgUsers, I);
3916 // Process the sinking DbgUsers in reverse order, as we only want to clone the
3917 // last appearing debug intrinsic for each given variable.
3918 SmallVector<DbgVariableIntrinsic *, 2> DbgUsersToSink;
3919 for (DbgVariableIntrinsic *DVI : DbgUsers)
3920 if (DVI->getParent() == SrcBlock)
3921 DbgUsersToSink.push_back(DVI);
3922 llvm::sort(DbgUsersToSink,
3923 [](auto *A, auto *B) { return B->comesBefore(A); });
3925 SmallVector<DbgVariableIntrinsic *, 2> DIIClones;
3926 SmallSet<DebugVariable, 4> SunkVariables;
3927 for (auto User : DbgUsersToSink) {
3928 // A dbg.declare instruction should not be cloned, since there can only be
3929 // one per variable fragment. It should be left in the original place
3930 // because the sunk instruction is not an alloca (otherwise we could not be
3932 if (isa<DbgDeclareInst>(User))
3935 DebugVariable DbgUserVariable =
3936 DebugVariable(User->getVariable(), User->getExpression(),
3937 User->getDebugLoc()->getInlinedAt());
3939 if (!SunkVariables.insert(DbgUserVariable).second)
3942 DIIClones.emplace_back(cast<DbgVariableIntrinsic>(User->clone()));
3943 if (isa<DbgDeclareInst>(User) && isa<CastInst>(I))
3944 DIIClones.back()->replaceVariableLocationOp(I, I->getOperand(0));
3945 LLVM_DEBUG(dbgs() << "CLONE: " << *DIIClones.back() << '\n');
3948 // Perform salvaging without the clones, then sink the clones.
3949 if (!DIIClones.empty()) {
3950 salvageDebugInfoForDbgValues(*I, DbgUsers);
3951 // The clones are in reverse order of original appearance, reverse again to
3952 // maintain the original order.
3953 for (auto &DIIClone : llvm::reverse(DIIClones)) {
3954 DIIClone->insertBefore(&*InsertPos);
3955 LLVM_DEBUG(dbgs() << "SINK: " << *DIIClone << '\n');
3962 bool InstCombinerImpl::run() {
3963 while (!Worklist.isEmpty()) {
3964 // Walk deferred instructions in reverse order, and push them to the
3965 // worklist, which means they'll end up popped from the worklist in-order.
3966 while (Instruction *I = Worklist.popDeferred()) {
3967 // Check to see if we can DCE the instruction. We do this already here to
3968 // reduce the number of uses and thus allow other folds to trigger.
3969 // Note that eraseInstFromFunction() may push additional instructions on
3970 // the deferred worklist, so this will DCE whole instruction chains.
3971 if (isInstructionTriviallyDead(I, &TLI)) {
3972 eraseInstFromFunction(*I);
3980 Instruction *I = Worklist.removeOne();
3981 if (I == nullptr) continue; // skip null values.
3983 // Check to see if we can DCE the instruction.
3984 if (isInstructionTriviallyDead(I, &TLI)) {
3985 eraseInstFromFunction(*I);
3990 if (!DebugCounter::shouldExecute(VisitCounter))
3993 // Instruction isn't dead, see if we can constant propagate it.
3994 if (!I->use_empty() &&
3995 (I->getNumOperands() == 0 || isa<Constant>(I->getOperand(0)))) {
3996 if (Constant *C = ConstantFoldInstruction(I, DL, &TLI)) {
3997 LLVM_DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " << *I
4000 // Add operands to the worklist.
4001 replaceInstUsesWith(*I, C);
4003 if (isInstructionTriviallyDead(I, &TLI))
4004 eraseInstFromFunction(*I);
4005 MadeIRChange = true;
4010 // See if we can trivially sink this instruction to its user if we can
4011 // prove that the successor is not executed more frequently than our block.
4012 // Return the UserBlock if successful.
4013 auto getOptionalSinkBlockForInst =
4014 [this](Instruction *I) -> Optional<BasicBlock *> {
4015 if (!EnableCodeSinking)
4017 auto *UserInst = cast_or_null<Instruction>(I->getUniqueUndroppableUser());
4021 BasicBlock *BB = I->getParent();
4022 BasicBlock *UserParent = nullptr;
4024 // Special handling for Phi nodes - get the block the use occurs in.
4025 if (PHINode *PN = dyn_cast<PHINode>(UserInst)) {
4026 for (unsigned i = 0; i < PN->getNumIncomingValues(); i++) {
4027 if (PN->getIncomingValue(i) == I) {
4028 // Bail out if we have uses in different blocks. We don't do any
4029 // sophisticated analysis (i.e finding NearestCommonDominator of these
4031 if (UserParent && UserParent != PN->getIncomingBlock(i))
4033 UserParent = PN->getIncomingBlock(i);
4036 assert(UserParent && "expected to find user block!");
4038 UserParent = UserInst->getParent();
4040 // Try sinking to another block. If that block is unreachable, then do
4041 // not bother. SimplifyCFG should handle it.
4042 if (UserParent == BB || !DT.isReachableFromEntry(UserParent))
4045 auto *Term = UserParent->getTerminator();
4046 // See if the user is one of our successors that has only one
4047 // predecessor, so that we don't have to split the critical edge.
4048 // Another option where we can sink is a block that ends with a
4049 // terminator that does not pass control to other block (such as
4050 // return or unreachable or resume). In this case:
4051 // - I dominates the User (by SSA form);
4052 // - the User will be executed at most once.
4053 // So sinking I down to User is always profitable or neutral.
4054 if (UserParent->getUniquePredecessor() == BB || succ_empty(Term)) {
4055 assert(DT.dominates(BB, UserParent) && "Dominance relation broken?");
4061 auto OptBB = getOptionalSinkBlockForInst(I);
4063 auto *UserParent = *OptBB;
4064 // Okay, the CFG is simple enough, try to sink this instruction.
4065 if (TryToSinkInstruction(I, UserParent, TLI)) {
4066 LLVM_DEBUG(dbgs() << "IC: Sink: " << *I << '\n');
4067 MadeIRChange = true;
4068 // We'll add uses of the sunk instruction below, but since
4069 // sinking can expose opportunities for it's *operands* add
4070 // them to the worklist
4071 for (Use &U : I->operands())
4072 if (Instruction *OpI = dyn_cast<Instruction>(U.get()))
4077 // Now that we have an instruction, try combining it to simplify it.
4078 Builder.SetInsertPoint(I);
4079 Builder.CollectMetadataToCopy(
4080 I, {LLVMContext::MD_dbg, LLVMContext::MD_annotation});
4085 LLVM_DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str(););
4086 LLVM_DEBUG(dbgs() << "IC: Visiting: " << OrigI << '\n');
4088 if (Instruction *Result = visit(*I)) {
4090 // Should we replace the old instruction with a new one?
4092 LLVM_DEBUG(dbgs() << "IC: Old = " << *I << '\n'
4093 << " New = " << *Result << '\n');
4095 Result->copyMetadata(*I,
4096 {LLVMContext::MD_dbg, LLVMContext::MD_annotation});
4097 // Everything uses the new instruction now.
4098 I->replaceAllUsesWith(Result);
4100 // Move the name to the new instruction first.
4101 Result->takeName(I);
4103 // Insert the new instruction into the basic block...
4104 BasicBlock *InstParent = I->getParent();
4105 BasicBlock::iterator InsertPos = I->getIterator();
4107 // Are we replace a PHI with something that isn't a PHI, or vice versa?
4108 if (isa<PHINode>(Result) != isa<PHINode>(I)) {
4109 // We need to fix up the insertion point.
4110 if (isa<PHINode>(I)) // PHI -> Non-PHI
4111 InsertPos = InstParent->getFirstInsertionPt();
4112 else // Non-PHI -> PHI
4113 InsertPos = InstParent->getFirstNonPHI()->getIterator();
4116 InstParent->getInstList().insert(InsertPos, Result);
4118 // Push the new instruction and any users onto the worklist.
4119 Worklist.pushUsersToWorkList(*Result);
4120 Worklist.push(Result);
4122 eraseInstFromFunction(*I);
4124 LLVM_DEBUG(dbgs() << "IC: Mod = " << OrigI << '\n'
4125 << " New = " << *I << '\n');
4127 // If the instruction was modified, it's possible that it is now dead.
4128 // if so, remove it.
4129 if (isInstructionTriviallyDead(I, &TLI)) {
4130 eraseInstFromFunction(*I);
4132 Worklist.pushUsersToWorkList(*I);
4136 MadeIRChange = true;
4141 return MadeIRChange;
4144 // Track the scopes used by !alias.scope and !noalias. In a function, a
4145 // @llvm.experimental.noalias.scope.decl is only useful if that scope is used
4146 // by both sets. If not, the declaration of the scope can be safely omitted.
4147 // The MDNode of the scope can be omitted as well for the instructions that are
4148 // part of this function. We do not do that at this point, as this might become
4149 // too time consuming to do.
4150 class AliasScopeTracker {
4151 SmallPtrSet<const MDNode *, 8> UsedAliasScopesAndLists;
4152 SmallPtrSet<const MDNode *, 8> UsedNoAliasScopesAndLists;
4155 void analyse(Instruction *I) {
4156 // This seems to be faster than checking 'mayReadOrWriteMemory()'.
4157 if (!I->hasMetadataOtherThanDebugLoc())
4160 auto Track = [](Metadata *ScopeList, auto &Container) {
4161 const auto *MDScopeList = dyn_cast_or_null<MDNode>(ScopeList);
4162 if (!MDScopeList || !Container.insert(MDScopeList).second)
4164 for (auto &MDOperand : MDScopeList->operands())
4165 if (auto *MDScope = dyn_cast<MDNode>(MDOperand))
4166 Container.insert(MDScope);
4169 Track(I->getMetadata(LLVMContext::MD_alias_scope), UsedAliasScopesAndLists);
4170 Track(I->getMetadata(LLVMContext::MD_noalias), UsedNoAliasScopesAndLists);
4173 bool isNoAliasScopeDeclDead(Instruction *Inst) {
4174 NoAliasScopeDeclInst *Decl = dyn_cast<NoAliasScopeDeclInst>(Inst);
4178 assert(Decl->use_empty() &&
4179 "llvm.experimental.noalias.scope.decl in use ?");
4180 const MDNode *MDSL = Decl->getScopeList();
4181 assert(MDSL->getNumOperands() == 1 &&
4182 "llvm.experimental.noalias.scope should refer to a single scope");
4183 auto &MDOperand = MDSL->getOperand(0);
4184 if (auto *MD = dyn_cast<MDNode>(MDOperand))
4185 return !UsedAliasScopesAndLists.contains(MD) ||
4186 !UsedNoAliasScopesAndLists.contains(MD);
4188 // Not an MDNode ? throw away.
4193 /// Populate the IC worklist from a function, by walking it in depth-first
4194 /// order and adding all reachable code to the worklist.
4196 /// This has a couple of tricks to make the code faster and more powerful. In
4197 /// particular, we constant fold and DCE instructions as we go, to avoid adding
4198 /// them to the worklist (this significantly speeds up instcombine on code where
4199 /// many instructions are dead or constant). Additionally, if we find a branch
4200 /// whose condition is a known constant, we only visit the reachable successors.
4201 static bool prepareICWorklistFromFunction(Function &F, const DataLayout &DL,
4202 const TargetLibraryInfo *TLI,
4203 InstructionWorklist &ICWorklist) {
4204 bool MadeIRChange = false;
4205 SmallPtrSet<BasicBlock *, 32> Visited;
4206 SmallVector<BasicBlock*, 256> Worklist;
4207 Worklist.push_back(&F.front());
4209 SmallVector<Instruction *, 128> InstrsForInstructionWorklist;
4210 DenseMap<Constant *, Constant *> FoldedConstants;
4211 AliasScopeTracker SeenAliasScopes;
4214 BasicBlock *BB = Worklist.pop_back_val();
4216 // We have now visited this block! If we've already been here, ignore it.
4217 if (!Visited.insert(BB).second)
4220 for (Instruction &Inst : llvm::make_early_inc_range(*BB)) {
4221 // ConstantProp instruction if trivially constant.
4222 if (!Inst.use_empty() &&
4223 (Inst.getNumOperands() == 0 || isa<Constant>(Inst.getOperand(0))))
4224 if (Constant *C = ConstantFoldInstruction(&Inst, DL, TLI)) {
4225 LLVM_DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " << Inst
4227 Inst.replaceAllUsesWith(C);
4229 if (isInstructionTriviallyDead(&Inst, TLI))
4230 Inst.eraseFromParent();
4231 MadeIRChange = true;
4235 // See if we can constant fold its operands.
4236 for (Use &U : Inst.operands()) {
4237 if (!isa<ConstantVector>(U) && !isa<ConstantExpr>(U))
4240 auto *C = cast<Constant>(U);
4241 Constant *&FoldRes = FoldedConstants[C];
4243 FoldRes = ConstantFoldConstant(C, DL, TLI);
4246 LLVM_DEBUG(dbgs() << "IC: ConstFold operand of: " << Inst
4247 << "\n Old = " << *C
4248 << "\n New = " << *FoldRes << '\n');
4250 MadeIRChange = true;
4254 // Skip processing debug and pseudo intrinsics in InstCombine. Processing
4255 // these call instructions consumes non-trivial amount of time and
4256 // provides no value for the optimization.
4257 if (!Inst.isDebugOrPseudoInst()) {
4258 InstrsForInstructionWorklist.push_back(&Inst);
4259 SeenAliasScopes.analyse(&Inst);
4263 // Recursively visit successors. If this is a branch or switch on a
4264 // constant, only visit the reachable successor.
4265 Instruction *TI = BB->getTerminator();
4266 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
4267 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
4268 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
4269 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
4270 Worklist.push_back(ReachableBB);
4273 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
4274 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
4275 Worklist.push_back(SI->findCaseValue(Cond)->getCaseSuccessor());
4280 append_range(Worklist, successors(TI));
4281 } while (!Worklist.empty());
4283 // Remove instructions inside unreachable blocks. This prevents the
4284 // instcombine code from having to deal with some bad special cases, and
4285 // reduces use counts of instructions.
4286 for (BasicBlock &BB : F) {
4287 if (Visited.count(&BB))
4290 unsigned NumDeadInstInBB;
4291 unsigned NumDeadDbgInstInBB;
4292 std::tie(NumDeadInstInBB, NumDeadDbgInstInBB) =
4293 removeAllNonTerminatorAndEHPadInstructions(&BB);
4295 MadeIRChange |= NumDeadInstInBB + NumDeadDbgInstInBB > 0;
4296 NumDeadInst += NumDeadInstInBB;
4299 // Once we've found all of the instructions to add to instcombine's worklist,
4300 // add them in reverse order. This way instcombine will visit from the top
4301 // of the function down. This jives well with the way that it adds all uses
4302 // of instructions to the worklist after doing a transformation, thus avoiding
4303 // some N^2 behavior in pathological cases.
4304 ICWorklist.reserve(InstrsForInstructionWorklist.size());
4305 for (Instruction *Inst : reverse(InstrsForInstructionWorklist)) {
4306 // DCE instruction if trivially dead. As we iterate in reverse program
4307 // order here, we will clean up whole chains of dead instructions.
4308 if (isInstructionTriviallyDead(Inst, TLI) ||
4309 SeenAliasScopes.isNoAliasScopeDeclDead(Inst)) {
4311 LLVM_DEBUG(dbgs() << "IC: DCE: " << *Inst << '\n');
4312 salvageDebugInfo(*Inst);
4313 Inst->eraseFromParent();
4314 MadeIRChange = true;
4318 ICWorklist.push(Inst);
4321 return MadeIRChange;
4324 static bool combineInstructionsOverFunction(
4325 Function &F, InstructionWorklist &Worklist, AliasAnalysis *AA,
4326 AssumptionCache &AC, TargetLibraryInfo &TLI, TargetTransformInfo &TTI,
4327 DominatorTree &DT, OptimizationRemarkEmitter &ORE, BlockFrequencyInfo *BFI,
4328 ProfileSummaryInfo *PSI, unsigned MaxIterations, LoopInfo *LI) {
4329 auto &DL = F.getParent()->getDataLayout();
4330 MaxIterations = std::min(MaxIterations, LimitMaxIterations.getValue());
4332 /// Builder - This is an IRBuilder that automatically inserts new
4333 /// instructions into the worklist when they are created.
4334 IRBuilder<TargetFolder, IRBuilderCallbackInserter> Builder(
4335 F.getContext(), TargetFolder(DL),
4336 IRBuilderCallbackInserter([&Worklist, &AC](Instruction *I) {
4338 if (auto *Assume = dyn_cast<AssumeInst>(I))
4339 AC.registerAssumption(Assume);
4342 // Lower dbg.declare intrinsics otherwise their value may be clobbered
4344 bool MadeIRChange = false;
4345 if (ShouldLowerDbgDeclare)
4346 MadeIRChange = LowerDbgDeclare(F);
4348 // Iterate while there is work to do.
4349 unsigned Iteration = 0;
4351 ++NumWorklistIterations;
4354 if (Iteration > InfiniteLoopDetectionThreshold) {
4356 "Instruction Combining seems stuck in an infinite loop after " +
4357 Twine(InfiniteLoopDetectionThreshold) + " iterations.");
4360 if (Iteration > MaxIterations) {
4361 LLVM_DEBUG(dbgs() << "\n\n[IC] Iteration limit #" << MaxIterations
4362 << " on " << F.getName()
4363 << " reached; stopping before reaching a fixpoint\n");
4367 LLVM_DEBUG(dbgs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
4368 << F.getName() << "\n");
4370 MadeIRChange |= prepareICWorklistFromFunction(F, DL, &TLI, Worklist);
4372 InstCombinerImpl IC(Worklist, Builder, F.hasMinSize(), AA, AC, TLI, TTI, DT,
4373 ORE, BFI, PSI, DL, LI);
4374 IC.MaxArraySizeForCombine = MaxArraySize;
4379 MadeIRChange = true;
4382 return MadeIRChange;
4385 InstCombinePass::InstCombinePass() : MaxIterations(LimitMaxIterations) {}
4387 InstCombinePass::InstCombinePass(unsigned MaxIterations)
4388 : MaxIterations(MaxIterations) {}
4390 PreservedAnalyses InstCombinePass::run(Function &F,
4391 FunctionAnalysisManager &AM) {
4392 auto &AC = AM.getResult<AssumptionAnalysis>(F);
4393 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
4394 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
4395 auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F);
4396 auto &TTI = AM.getResult<TargetIRAnalysis>(F);
4398 auto *LI = AM.getCachedResult<LoopAnalysis>(F);
4400 auto *AA = &AM.getResult<AAManager>(F);
4401 auto &MAMProxy = AM.getResult<ModuleAnalysisManagerFunctionProxy>(F);
4402 ProfileSummaryInfo *PSI =
4403 MAMProxy.getCachedResult<ProfileSummaryAnalysis>(*F.getParent());
4404 auto *BFI = (PSI && PSI->hasProfileSummary()) ?
4405 &AM.getResult<BlockFrequencyAnalysis>(F) : nullptr;
4407 if (!combineInstructionsOverFunction(F, Worklist, AA, AC, TLI, TTI, DT, ORE,
4408 BFI, PSI, MaxIterations, LI))
4409 // No changes, all analyses are preserved.
4410 return PreservedAnalyses::all();
4412 // Mark all the analyses that instcombine updates as preserved.
4413 PreservedAnalyses PA;
4414 PA.preserveSet<CFGAnalyses>();
4418 void InstructionCombiningPass::getAnalysisUsage(AnalysisUsage &AU) const {
4419 AU.setPreservesCFG();
4420 AU.addRequired<AAResultsWrapperPass>();
4421 AU.addRequired<AssumptionCacheTracker>();
4422 AU.addRequired<TargetLibraryInfoWrapperPass>();
4423 AU.addRequired<TargetTransformInfoWrapperPass>();
4424 AU.addRequired<DominatorTreeWrapperPass>();
4425 AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
4426 AU.addPreserved<DominatorTreeWrapperPass>();
4427 AU.addPreserved<AAResultsWrapperPass>();
4428 AU.addPreserved<BasicAAWrapperPass>();
4429 AU.addPreserved<GlobalsAAWrapperPass>();
4430 AU.addRequired<ProfileSummaryInfoWrapperPass>();
4431 LazyBlockFrequencyInfoPass::getLazyBFIAnalysisUsage(AU);
4434 bool InstructionCombiningPass::runOnFunction(Function &F) {
4435 if (skipFunction(F))
4438 // Required analyses.
4439 auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
4440 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
4441 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
4442 auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
4443 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
4444 auto &ORE = getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
4446 // Optional analyses.
4447 auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>();
4448 auto *LI = LIWP ? &LIWP->getLoopInfo() : nullptr;
4449 ProfileSummaryInfo *PSI =
4450 &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
4451 BlockFrequencyInfo *BFI =
4452 (PSI && PSI->hasProfileSummary()) ?
4453 &getAnalysis<LazyBlockFrequencyInfoPass>().getBFI() :
4456 return combineInstructionsOverFunction(F, Worklist, AA, AC, TLI, TTI, DT, ORE,
4457 BFI, PSI, MaxIterations, LI);
4460 char InstructionCombiningPass::ID = 0;
4462 InstructionCombiningPass::InstructionCombiningPass()
4463 : FunctionPass(ID), MaxIterations(InstCombineDefaultMaxIterations) {
4464 initializeInstructionCombiningPassPass(*PassRegistry::getPassRegistry());
4467 InstructionCombiningPass::InstructionCombiningPass(unsigned MaxIterations)
4468 : FunctionPass(ID), MaxIterations(MaxIterations) {
4469 initializeInstructionCombiningPassPass(*PassRegistry::getPassRegistry());
4472 INITIALIZE_PASS_BEGIN(InstructionCombiningPass, "instcombine",
4473 "Combine redundant instructions", false, false)
4474 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
4475 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
4476 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
4477 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
4478 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
4479 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
4480 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)
4481 INITIALIZE_PASS_DEPENDENCY(LazyBlockFrequencyInfoPass)
4482 INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)
4483 INITIALIZE_PASS_END(InstructionCombiningPass, "instcombine",
4484 "Combine redundant instructions", false, false)
4486 // Initialization Routines
4487 void llvm::initializeInstCombine(PassRegistry &Registry) {
4488 initializeInstructionCombiningPassPass(Registry);
4491 void LLVMInitializeInstCombine(LLVMPassRegistryRef R) {
4492 initializeInstructionCombiningPassPass(*unwrap(R));
4495 FunctionPass *llvm::createInstructionCombiningPass() {
4496 return new InstructionCombiningPass();
4499 FunctionPass *llvm::createInstructionCombiningPass(unsigned MaxIterations) {
4500 return new InstructionCombiningPass(MaxIterations);
4503 void LLVMAddInstructionCombiningPass(LLVMPassManagerRef PM) {
4504 unwrap(PM)->add(createInstructionCombiningPass());