InstCombine.h revision dce4a407a24b04eebc6a376f8e62b41aaa7b071f
1//===- InstCombine.h - Main InstCombine pass definition ---------*- C++ -*-===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9
10#ifndef INSTCOMBINE_INSTCOMBINE_H
11#define INSTCOMBINE_INSTCOMBINE_H
12
13#include "InstCombineWorklist.h"
14#include "llvm/Analysis/TargetFolder.h"
15#include "llvm/Analysis/ValueTracking.h"
16#include "llvm/IR/IRBuilder.h"
17#include "llvm/IR/InstVisitor.h"
18#include "llvm/IR/IntrinsicInst.h"
19#include "llvm/IR/Operator.h"
20#include "llvm/Pass.h"
21#include "llvm/Transforms/Utils/SimplifyLibCalls.h"
22
23#define DEBUG_TYPE "instcombine"
24
25namespace llvm {
26class CallSite;
27class DataLayout;
28class TargetLibraryInfo;
29class DbgDeclareInst;
30class MemIntrinsic;
31class MemSetInst;
32
33/// SelectPatternFlavor - We can match a variety of different patterns for
34/// select operations.
35enum SelectPatternFlavor {
36  SPF_UNKNOWN = 0,
37  SPF_SMIN,
38  SPF_UMIN,
39  SPF_SMAX,
40  SPF_UMAX
41  // SPF_ABS - TODO.
42};
43
44/// getComplexity:  Assign a complexity or rank value to LLVM Values...
45///   0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
46static inline unsigned getComplexity(Value *V) {
47  if (isa<Instruction>(V)) {
48    if (BinaryOperator::isNeg(V) || BinaryOperator::isFNeg(V) ||
49        BinaryOperator::isNot(V))
50      return 3;
51    return 4;
52  }
53  if (isa<Argument>(V))
54    return 3;
55  return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
56}
57
58/// AddOne - Add one to a Constant
59static inline Constant *AddOne(Constant *C) {
60  return ConstantExpr::getAdd(C, ConstantInt::get(C->getType(), 1));
61}
62/// SubOne - Subtract one from a Constant
63static inline Constant *SubOne(Constant *C) {
64  return ConstantExpr::getSub(C, ConstantInt::get(C->getType(), 1));
65}
66
67/// InstCombineIRInserter - This is an IRBuilder insertion helper that works
68/// just like the normal insertion helper, but also adds any new instructions
69/// to the instcombine worklist.
70class LLVM_LIBRARY_VISIBILITY InstCombineIRInserter
71    : public IRBuilderDefaultInserter<true> {
72  InstCombineWorklist &Worklist;
73
74public:
75  InstCombineIRInserter(InstCombineWorklist &WL) : Worklist(WL) {}
76
77  void InsertHelper(Instruction *I, const Twine &Name, BasicBlock *BB,
78                    BasicBlock::iterator InsertPt) const {
79    IRBuilderDefaultInserter<true>::InsertHelper(I, Name, BB, InsertPt);
80    Worklist.Add(I);
81  }
82};
83
84/// InstCombiner - The -instcombine pass.
85class LLVM_LIBRARY_VISIBILITY InstCombiner
86    : public FunctionPass,
87      public InstVisitor<InstCombiner, Instruction *> {
88  const DataLayout *DL;
89  TargetLibraryInfo *TLI;
90  bool MadeIRChange;
91  LibCallSimplifier *Simplifier;
92  bool MinimizeSize;
93
94public:
95  /// Worklist - All of the instructions that need to be simplified.
96  InstCombineWorklist Worklist;
97
98  /// Builder - This is an IRBuilder that automatically inserts new
99  /// instructions into the worklist when they are created.
100  typedef IRBuilder<true, TargetFolder, InstCombineIRInserter> BuilderTy;
101  BuilderTy *Builder;
102
103  static char ID; // Pass identification, replacement for typeid
104  InstCombiner() : FunctionPass(ID), DL(nullptr), Builder(nullptr) {
105    MinimizeSize = false;
106    initializeInstCombinerPass(*PassRegistry::getPassRegistry());
107  }
108
109public:
110  bool runOnFunction(Function &F) override;
111
112  bool DoOneIteration(Function &F, unsigned ItNum);
113
114  void getAnalysisUsage(AnalysisUsage &AU) const override;
115
116  const DataLayout *getDataLayout() const { return DL; }
117
118  TargetLibraryInfo *getTargetLibraryInfo() const { return TLI; }
119
120  // Visitation implementation - Implement instruction combining for different
121  // instruction types.  The semantics are as follows:
122  // Return Value:
123  //    null        - No change was made
124  //     I          - Change was made, I is still valid, I may be dead though
125  //   otherwise    - Change was made, replace I with returned instruction
126  //
127  Instruction *visitAdd(BinaryOperator &I);
128  Instruction *visitFAdd(BinaryOperator &I);
129  Value *OptimizePointerDifference(Value *LHS, Value *RHS, Type *Ty);
130  Instruction *visitSub(BinaryOperator &I);
131  Instruction *visitFSub(BinaryOperator &I);
132  Instruction *visitMul(BinaryOperator &I);
133  Value *foldFMulConst(Instruction *FMulOrDiv, Constant *C,
134                       Instruction *InsertBefore);
135  Instruction *visitFMul(BinaryOperator &I);
136  Instruction *visitURem(BinaryOperator &I);
137  Instruction *visitSRem(BinaryOperator &I);
138  Instruction *visitFRem(BinaryOperator &I);
139  bool SimplifyDivRemOfSelect(BinaryOperator &I);
140  Instruction *commonRemTransforms(BinaryOperator &I);
141  Instruction *commonIRemTransforms(BinaryOperator &I);
142  Instruction *commonDivTransforms(BinaryOperator &I);
143  Instruction *commonIDivTransforms(BinaryOperator &I);
144  Instruction *visitUDiv(BinaryOperator &I);
145  Instruction *visitSDiv(BinaryOperator &I);
146  Instruction *visitFDiv(BinaryOperator &I);
147  Value *FoldAndOfICmps(ICmpInst *LHS, ICmpInst *RHS);
148  Value *FoldAndOfFCmps(FCmpInst *LHS, FCmpInst *RHS);
149  Instruction *visitAnd(BinaryOperator &I);
150  Value *FoldOrOfICmps(ICmpInst *LHS, ICmpInst *RHS);
151  Value *FoldOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS);
152  Instruction *FoldOrWithConstants(BinaryOperator &I, Value *Op, Value *A,
153                                   Value *B, Value *C);
154  Instruction *visitOr(BinaryOperator &I);
155  Instruction *visitXor(BinaryOperator &I);
156  Instruction *visitShl(BinaryOperator &I);
157  Instruction *visitAShr(BinaryOperator &I);
158  Instruction *visitLShr(BinaryOperator &I);
159  Instruction *commonShiftTransforms(BinaryOperator &I);
160  Instruction *FoldFCmp_IntToFP_Cst(FCmpInst &I, Instruction *LHSI,
161                                    Constant *RHSC);
162  Instruction *FoldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP,
163                                            GlobalVariable *GV, CmpInst &ICI,
164                                            ConstantInt *AndCst = nullptr);
165  Instruction *visitFCmpInst(FCmpInst &I);
166  Instruction *visitICmpInst(ICmpInst &I);
167  Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
168  Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI, Instruction *LHS,
169                                              ConstantInt *RHS);
170  Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
171                              ConstantInt *DivRHS);
172  Instruction *FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *DivI,
173                              ConstantInt *DivRHS);
174  Instruction *FoldICmpAddOpCst(Instruction &ICI, Value *X, ConstantInt *CI,
175                                ICmpInst::Predicate Pred);
176  Instruction *FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
177                           ICmpInst::Predicate Cond, Instruction &I);
178  Instruction *FoldShiftByConstant(Value *Op0, Constant *Op1,
179                                   BinaryOperator &I);
180  Instruction *commonCastTransforms(CastInst &CI);
181  Instruction *commonPointerCastTransforms(CastInst &CI);
182  Instruction *visitTrunc(TruncInst &CI);
183  Instruction *visitZExt(ZExtInst &CI);
184  Instruction *visitSExt(SExtInst &CI);
185  Instruction *visitFPTrunc(FPTruncInst &CI);
186  Instruction *visitFPExt(CastInst &CI);
187  Instruction *visitFPToUI(FPToUIInst &FI);
188  Instruction *visitFPToSI(FPToSIInst &FI);
189  Instruction *visitUIToFP(CastInst &CI);
190  Instruction *visitSIToFP(CastInst &CI);
191  Instruction *visitPtrToInt(PtrToIntInst &CI);
192  Instruction *visitIntToPtr(IntToPtrInst &CI);
193  Instruction *visitBitCast(BitCastInst &CI);
194  Instruction *visitAddrSpaceCast(AddrSpaceCastInst &CI);
195  Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI, Instruction *FI);
196  Instruction *FoldSelectIntoOp(SelectInst &SI, Value *, Value *);
197  Instruction *FoldSPFofSPF(Instruction *Inner, SelectPatternFlavor SPF1,
198                            Value *A, Value *B, Instruction &Outer,
199                            SelectPatternFlavor SPF2, Value *C);
200  Instruction *visitSelectInst(SelectInst &SI);
201  Instruction *visitSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI);
202  Instruction *visitCallInst(CallInst &CI);
203  Instruction *visitInvokeInst(InvokeInst &II);
204
205  Instruction *SliceUpIllegalIntegerPHI(PHINode &PN);
206  Instruction *visitPHINode(PHINode &PN);
207  Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
208  Instruction *visitAllocaInst(AllocaInst &AI);
209  Instruction *visitAllocSite(Instruction &FI);
210  Instruction *visitFree(CallInst &FI);
211  Instruction *visitLoadInst(LoadInst &LI);
212  Instruction *visitStoreInst(StoreInst &SI);
213  Instruction *visitBranchInst(BranchInst &BI);
214  Instruction *visitSwitchInst(SwitchInst &SI);
215  Instruction *visitInsertValueInst(InsertValueInst &IV);
216  Instruction *visitInsertElementInst(InsertElementInst &IE);
217  Instruction *visitExtractElementInst(ExtractElementInst &EI);
218  Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
219  Instruction *visitExtractValueInst(ExtractValueInst &EV);
220  Instruction *visitLandingPadInst(LandingPadInst &LI);
221
222  // visitInstruction - Specify what to return for unhandled instructions...
223  Instruction *visitInstruction(Instruction &I) { return nullptr; }
224
225private:
226  bool ShouldChangeType(Type *From, Type *To) const;
227  Value *dyn_castNegVal(Value *V) const;
228  Value *dyn_castFNegVal(Value *V, bool NoSignedZero = false) const;
229  Type *FindElementAtOffset(Type *PtrTy, int64_t Offset,
230                            SmallVectorImpl<Value *> &NewIndices);
231  Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI);
232
233  /// ShouldOptimizeCast - Return true if the cast from "V to Ty" actually
234  /// results in any code being generated and is interesting to optimize out. If
235  /// the cast can be eliminated by some other simple transformation, we prefer
236  /// to do the simplification first.
237  bool ShouldOptimizeCast(Instruction::CastOps opcode, const Value *V,
238                          Type *Ty);
239
240  Instruction *visitCallSite(CallSite CS);
241  Instruction *tryOptimizeCall(CallInst *CI, const DataLayout *DL);
242  bool transformConstExprCastCall(CallSite CS);
243  Instruction *transformCallThroughTrampoline(CallSite CS,
244                                              IntrinsicInst *Tramp);
245  Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI,
246                                 bool DoXform = true);
247  Instruction *transformSExtICmp(ICmpInst *ICI, Instruction &CI);
248  bool WillNotOverflowSignedAdd(Value *LHS, Value *RHS);
249  Value *EmitGEPOffset(User *GEP);
250  Instruction *scalarizePHI(ExtractElementInst &EI, PHINode *PN);
251  Value *EvaluateInDifferentElementOrder(Value *V, ArrayRef<int> Mask);
252
253public:
254  // InsertNewInstBefore - insert an instruction New before instruction Old
255  // in the program.  Add the new instruction to the worklist.
256  //
257  Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
258    assert(New && !New->getParent() &&
259           "New instruction already inserted into a basic block!");
260    BasicBlock *BB = Old.getParent();
261    BB->getInstList().insert(&Old, New); // Insert inst
262    Worklist.Add(New);
263    return New;
264  }
265
266  // InsertNewInstWith - same as InsertNewInstBefore, but also sets the
267  // debug loc.
268  //
269  Instruction *InsertNewInstWith(Instruction *New, Instruction &Old) {
270    New->setDebugLoc(Old.getDebugLoc());
271    return InsertNewInstBefore(New, Old);
272  }
273
274  // ReplaceInstUsesWith - This method is to be used when an instruction is
275  // found to be dead, replacable with another preexisting expression.  Here
276  // we add all uses of I to the worklist, replace all uses of I with the new
277  // value, then return I, so that the inst combiner will know that I was
278  // modified.
279  //
280  Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
281    Worklist.AddUsersToWorkList(I); // Add all modified instrs to worklist.
282
283    // If we are replacing the instruction with itself, this must be in a
284    // segment of unreachable code, so just clobber the instruction.
285    if (&I == V)
286      V = UndefValue::get(I.getType());
287
288    DEBUG(dbgs() << "IC: Replacing " << I << "\n"
289                    "    with " << *V << '\n');
290
291    I.replaceAllUsesWith(V);
292    return &I;
293  }
294
295  // EraseInstFromFunction - When dealing with an instruction that has side
296  // effects or produces a void value, we can't rely on DCE to delete the
297  // instruction.  Instead, visit methods should return the value returned by
298  // this function.
299  Instruction *EraseInstFromFunction(Instruction &I) {
300    DEBUG(dbgs() << "IC: ERASE " << I << '\n');
301
302    assert(I.use_empty() && "Cannot erase instruction that is used!");
303    // Make sure that we reprocess all operands now that we reduced their
304    // use counts.
305    if (I.getNumOperands() < 8) {
306      for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
307        if (Instruction *Op = dyn_cast<Instruction>(*i))
308          Worklist.Add(Op);
309    }
310    Worklist.Remove(&I);
311    I.eraseFromParent();
312    MadeIRChange = true;
313    return nullptr; // Don't do anything with FI
314  }
315
316  void computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
317                        unsigned Depth = 0) const {
318    return llvm::computeKnownBits(V, KnownZero, KnownOne, DL, Depth);
319  }
320
321  bool MaskedValueIsZero(Value *V, const APInt &Mask,
322                         unsigned Depth = 0) const {
323    return llvm::MaskedValueIsZero(V, Mask, DL, Depth);
324  }
325  unsigned ComputeNumSignBits(Value *Op, unsigned Depth = 0) const {
326    return llvm::ComputeNumSignBits(Op, DL, Depth);
327  }
328
329private:
330  /// SimplifyAssociativeOrCommutative - This performs a few simplifications for
331  /// operators which are associative or commutative.
332  bool SimplifyAssociativeOrCommutative(BinaryOperator &I);
333
334  /// SimplifyUsingDistributiveLaws - This tries to simplify binary operations
335  /// which some other binary operation distributes over either by factorizing
336  /// out common terms (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this
337  /// results in simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is
338  /// a win).  Returns the simplified value, or null if it didn't simplify.
339  Value *SimplifyUsingDistributiveLaws(BinaryOperator &I);
340
341  /// SimplifyDemandedUseBits - Attempts to replace V with a simpler value
342  /// based on the demanded bits.
343  Value *SimplifyDemandedUseBits(Value *V, APInt DemandedMask, APInt &KnownZero,
344                                 APInt &KnownOne, unsigned Depth);
345  bool SimplifyDemandedBits(Use &U, APInt DemandedMask, APInt &KnownZero,
346                            APInt &KnownOne, unsigned Depth = 0);
347  /// Helper routine of SimplifyDemandedUseBits. It tries to simplify demanded
348  /// bit for "r1 = shr x, c1; r2 = shl r1, c2" instruction sequence.
349  Value *SimplifyShrShlDemandedBits(Instruction *Lsr, Instruction *Sftl,
350                                    APInt DemandedMask, APInt &KnownZero,
351                                    APInt &KnownOne);
352
353  /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
354  /// SimplifyDemandedBits knows about.  See if the instruction has any
355  /// properties that allow us to simplify its operands.
356  bool SimplifyDemandedInstructionBits(Instruction &Inst);
357
358  Value *SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
359                                    APInt &UndefElts, unsigned Depth = 0);
360
361  Value *SimplifyVectorOp(BinaryOperator &Inst);
362
363  // FoldOpIntoPhi - Given a binary operator, cast instruction, or select
364  // which has a PHI node as operand #0, see if we can fold the instruction
365  // into the PHI (which is only possible if all operands to the PHI are
366  // constants).
367  //
368  Instruction *FoldOpIntoPhi(Instruction &I);
369
370  // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
371  // operator and they all are only used by the PHI, PHI together their
372  // inputs, and do the operation once, to the result of the PHI.
373  Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
374  Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
375  Instruction *FoldPHIArgGEPIntoPHI(PHINode &PN);
376  Instruction *FoldPHIArgLoadIntoPHI(PHINode &PN);
377
378  Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
379                        ConstantInt *AndRHS, BinaryOperator &TheAnd);
380
381  Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
382                            bool isSub, Instruction &I);
383  Value *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi, bool isSigned,
384                         bool Inside);
385  Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocaInst &AI);
386  Instruction *MatchBSwap(BinaryOperator &I);
387  bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
388  Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
389  Instruction *SimplifyMemSet(MemSetInst *MI);
390
391  Value *EvaluateInDifferentType(Value *V, Type *Ty, bool isSigned);
392
393  /// Descale - Return a value X such that Val = X * Scale, or null if none.  If
394  /// the multiplication is known not to overflow then NoSignedWrap is set.
395  Value *Descale(Value *Val, APInt Scale, bool &NoSignedWrap);
396};
397
398} // end namespace llvm.
399
400#undef DEBUG_TYPE
401
402#endif
403