InstCombine.h revision 36b56886974eae4f9c5ebc96befd3e7bfe5de338 
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
23namespace llvm {
24  class CallSite;
25  class DataLayout;
26  class TargetLibraryInfo;
27  class DbgDeclareInst;
28  class MemIntrinsic;
29  class MemSetInst;
30
31/// SelectPatternFlavor - We can match a variety of different patterns for
32/// select operations.
33enum SelectPatternFlavor {
34  SPF_UNKNOWN = 0,
35  SPF_SMIN, SPF_UMIN,
36  SPF_SMAX, SPF_UMAX
37  //SPF_ABS - TODO.
38};
39
40/// getComplexity:  Assign a complexity or rank value to LLVM Values...
41///   0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
42static inline unsigned getComplexity(Value *V) {
43  if (isa<Instruction>(V)) {
44    if (BinaryOperator::isNeg(V) ||
45        BinaryOperator::isFNeg(V) ||
46        BinaryOperator::isNot(V))
47      return 3;
48    return 4;
49  }
50  if (isa<Argument>(V)) return 3;
51  return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
52}
53
54/// AddOne - Add one to a Constant
55static inline Constant *AddOne(Constant *C) {
56  return ConstantExpr::getAdd(C, ConstantInt::get(C->getType(), 1));
57}
58/// SubOne - Subtract one from a Constant
59static inline Constant *SubOne(Constant *C) {
60  return ConstantExpr::getSub(C, ConstantInt::get(C->getType(), 1));
61}
62
63
64/// InstCombineIRInserter - This is an IRBuilder insertion helper that works
65/// just like the normal insertion helper, but also adds any new instructions
66/// to the instcombine worklist.
67class LLVM_LIBRARY_VISIBILITY InstCombineIRInserter
68    : public IRBuilderDefaultInserter<true> {
69  InstCombineWorklist &Worklist;
70public:
71  InstCombineIRInserter(InstCombineWorklist &WL) : Worklist(WL) {}
72
73  void InsertHelper(Instruction *I, const Twine &Name,
74                    BasicBlock *BB, BasicBlock::iterator InsertPt) const {
75    IRBuilderDefaultInserter<true>::InsertHelper(I, Name, BB, InsertPt);
76    Worklist.Add(I);
77  }
78};
79
80/// InstCombiner - The -instcombine pass.
81class LLVM_LIBRARY_VISIBILITY InstCombiner
82                             : public FunctionPass,
83                               public InstVisitor<InstCombiner, Instruction*> {
84  const DataLayout *DL;
85  TargetLibraryInfo *TLI;
86  bool MadeIRChange;
87  LibCallSimplifier *Simplifier;
88  bool MinimizeSize;
89public:
90  /// Worklist - All of the instructions that need to be simplified.
91  InstCombineWorklist Worklist;
92
93  /// Builder - This is an IRBuilder that automatically inserts new
94  /// instructions into the worklist when they are created.
95  typedef IRBuilder<true, TargetFolder, InstCombineIRInserter> BuilderTy;
96  BuilderTy *Builder;
97
98  static char ID; // Pass identification, replacement for typeid
99  InstCombiner() : FunctionPass(ID), DL(0), Builder(0) {
100    MinimizeSize = false;
101    initializeInstCombinerPass(*PassRegistry::getPassRegistry());
102  }
103
104public:
105  bool runOnFunction(Function &F) override;
106
107  bool DoOneIteration(Function &F, unsigned ItNum);
108
109  void getAnalysisUsage(AnalysisUsage &AU) const override;
110
111  const DataLayout *getDataLayout() const { return DL; }
112
113  TargetLibraryInfo *getTargetLibraryInfo() const { return TLI; }
114
115  // Visitation implementation - Implement instruction combining for different
116  // instruction types.  The semantics are as follows:
117  // Return Value:
118  //    null        - No change was made
119  //     I          - Change was made, I is still valid, I may be dead though
120  //   otherwise    - Change was made, replace I with returned instruction
121  //
122  Instruction *visitAdd(BinaryOperator &I);
123  Instruction *visitFAdd(BinaryOperator &I);
124  Value *OptimizePointerDifference(Value *LHS, Value *RHS, Type *Ty);
125  Instruction *visitSub(BinaryOperator &I);
126  Instruction *visitFSub(BinaryOperator &I);
127  Instruction *visitMul(BinaryOperator &I);
128  Value *foldFMulConst(Instruction *FMulOrDiv, Constant *C,
129                       Instruction *InsertBefore);
130  Instruction *visitFMul(BinaryOperator &I);
131  Instruction *visitURem(BinaryOperator &I);
132  Instruction *visitSRem(BinaryOperator &I);
133  Instruction *visitFRem(BinaryOperator &I);
134  bool SimplifyDivRemOfSelect(BinaryOperator &I);
135  Instruction *commonRemTransforms(BinaryOperator &I);
136  Instruction *commonIRemTransforms(BinaryOperator &I);
137  Instruction *commonDivTransforms(BinaryOperator &I);
138  Instruction *commonIDivTransforms(BinaryOperator &I);
139  Instruction *visitUDiv(BinaryOperator &I);
140  Instruction *visitSDiv(BinaryOperator &I);
141  Instruction *visitFDiv(BinaryOperator &I);
142  Value *FoldAndOfICmps(ICmpInst *LHS, ICmpInst *RHS);
143  Value *FoldAndOfFCmps(FCmpInst *LHS, FCmpInst *RHS);
144  Instruction *visitAnd(BinaryOperator &I);
145  Value *FoldOrOfICmps(ICmpInst *LHS, ICmpInst *RHS);
146  Value *FoldOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS);
147  Instruction *FoldOrWithConstants(BinaryOperator &I, Value *Op,
148                                   Value *A, Value *B, Value *C);
149  Instruction *visitOr (BinaryOperator &I);
150  Instruction *visitXor(BinaryOperator &I);
151  Instruction *visitShl(BinaryOperator &I);
152  Instruction *visitAShr(BinaryOperator &I);
153  Instruction *visitLShr(BinaryOperator &I);
154  Instruction *commonShiftTransforms(BinaryOperator &I);
155  Instruction *FoldFCmp_IntToFP_Cst(FCmpInst &I, Instruction *LHSI,
156                                    Constant *RHSC);
157  Instruction *FoldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP,
158                                            GlobalVariable *GV, CmpInst &ICI,
159                                            ConstantInt *AndCst = 0);
160  Instruction *visitFCmpInst(FCmpInst &I);
161  Instruction *visitICmpInst(ICmpInst &I);
162  Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
163  Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
164                                              Instruction *LHS,
165                                              ConstantInt *RHS);
166  Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
167                              ConstantInt *DivRHS);
168  Instruction *FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *DivI,
169                              ConstantInt *DivRHS);
170  Instruction *FoldICmpAddOpCst(Instruction &ICI, Value *X, ConstantInt *CI,
171                                ICmpInst::Predicate Pred);
172  Instruction *FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
173                           ICmpInst::Predicate Cond, Instruction &I);
174  Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
175                                   BinaryOperator &I);
176  Instruction *commonCastTransforms(CastInst &CI);
177  Instruction *commonPointerCastTransforms(CastInst &CI);
178  Instruction *visitTrunc(TruncInst &CI);
179  Instruction *visitZExt(ZExtInst &CI);
180  Instruction *visitSExt(SExtInst &CI);
181  Instruction *visitFPTrunc(FPTruncInst &CI);
182  Instruction *visitFPExt(CastInst &CI);
183  Instruction *visitFPToUI(FPToUIInst &FI);
184  Instruction *visitFPToSI(FPToSIInst &FI);
185  Instruction *visitUIToFP(CastInst &CI);
186  Instruction *visitSIToFP(CastInst &CI);
187  Instruction *visitPtrToInt(PtrToIntInst &CI);
188  Instruction *visitIntToPtr(IntToPtrInst &CI);
189  Instruction *visitBitCast(BitCastInst &CI);
190  Instruction *visitAddrSpaceCast(AddrSpaceCastInst &CI);
191  Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
192                              Instruction *FI);
193  Instruction *FoldSelectIntoOp(SelectInst &SI, Value*, Value*);
194  Instruction *FoldSPFofSPF(Instruction *Inner, SelectPatternFlavor SPF1,
195                            Value *A, Value *B, Instruction &Outer,
196                            SelectPatternFlavor SPF2, Value *C);
197  Instruction *visitSelectInst(SelectInst &SI);
198  Instruction *visitSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI);
199  Instruction *visitCallInst(CallInst &CI);
200  Instruction *visitInvokeInst(InvokeInst &II);
201
202  Instruction *SliceUpIllegalIntegerPHI(PHINode &PN);
203  Instruction *visitPHINode(PHINode &PN);
204  Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
205  Instruction *visitAllocaInst(AllocaInst &AI);
206  Instruction *visitAllocSite(Instruction &FI);
207  Instruction *visitFree(CallInst &FI);
208  Instruction *visitLoadInst(LoadInst &LI);
209  Instruction *visitStoreInst(StoreInst &SI);
210  Instruction *visitBranchInst(BranchInst &BI);
211  Instruction *visitSwitchInst(SwitchInst &SI);
212  Instruction *visitInsertElementInst(InsertElementInst &IE);
213  Instruction *visitExtractElementInst(ExtractElementInst &EI);
214  Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
215  Instruction *visitExtractValueInst(ExtractValueInst &EV);
216  Instruction *visitLandingPadInst(LandingPadInst &LI);
217
218  // visitInstruction - Specify what to return for unhandled instructions...
219  Instruction *visitInstruction(Instruction &I) { return 0; }
220
221private:
222  bool ShouldChangeType(Type *From, Type *To) const;
223  Value *dyn_castNegVal(Value *V) const;
224  Value *dyn_castFNegVal(Value *V, bool NoSignedZero=false) const;
225  Type *FindElementAtOffset(Type *PtrTy, int64_t Offset,
226                            SmallVectorImpl<Value*> &NewIndices);
227  Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI);
228
229  /// ShouldOptimizeCast - Return true if the cast from "V to Ty" actually
230  /// results in any code being generated and is interesting to optimize out. If
231  /// the cast can be eliminated by some other simple transformation, we prefer
232  /// to do the simplification first.
233  bool ShouldOptimizeCast(Instruction::CastOps opcode,const Value *V,
234                          Type *Ty);
235
236  Instruction *visitCallSite(CallSite CS);
237  Instruction *tryOptimizeCall(CallInst *CI, const DataLayout *DL);
238  bool transformConstExprCastCall(CallSite CS);
239  Instruction *transformCallThroughTrampoline(CallSite CS,
240                                              IntrinsicInst *Tramp);
241  Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI,
242                                 bool DoXform = true);
243  Instruction *transformSExtICmp(ICmpInst *ICI, Instruction &CI);
244  bool WillNotOverflowSignedAdd(Value *LHS, Value *RHS);
245  Value *EmitGEPOffset(User *GEP);
246  Instruction *scalarizePHI(ExtractElementInst &EI, PHINode *PN);
247  Value *EvaluateInDifferentElementOrder(Value *V, ArrayRef<int> Mask);
248
249public:
250  // InsertNewInstBefore - insert an instruction New before instruction Old
251  // in the program.  Add the new instruction to the worklist.
252  //
253  Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
254    assert(New && New->getParent() == 0 &&
255           "New instruction already inserted into a basic block!");
256    BasicBlock *BB = Old.getParent();
257    BB->getInstList().insert(&Old, New);  // Insert inst
258    Worklist.Add(New);
259    return New;
260  }
261
262  // InsertNewInstWith - same as InsertNewInstBefore, but also sets the
263  // debug loc.
264  //
265  Instruction *InsertNewInstWith(Instruction *New, Instruction &Old) {
266    New->setDebugLoc(Old.getDebugLoc());
267    return InsertNewInstBefore(New, Old);
268  }
269
270  // ReplaceInstUsesWith - This method is to be used when an instruction is
271  // found to be dead, replacable with another preexisting expression.  Here
272  // we add all uses of I to the worklist, replace all uses of I with the new
273  // value, then return I, so that the inst combiner will know that I was
274  // modified.
275  //
276  Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
277    Worklist.AddUsersToWorkList(I);   // Add all modified instrs to worklist.
278
279    // If we are replacing the instruction with itself, this must be in a
280    // segment of unreachable code, so just clobber the instruction.
281    if (&I == V)
282      V = UndefValue::get(I.getType());
283
284    DEBUG(dbgs() << "IC: Replacing " << I << "\n"
285                    "    with " << *V << '\n');
286
287    I.replaceAllUsesWith(V);
288    return &I;
289  }
290
291  // EraseInstFromFunction - When dealing with an instruction that has side
292  // effects or produces a void value, we can't rely on DCE to delete the
293  // instruction.  Instead, visit methods should return the value returned by
294  // this function.
295  Instruction *EraseInstFromFunction(Instruction &I) {
296    DEBUG(dbgs() << "IC: ERASE " << I << '\n');
297
298    assert(I.use_empty() && "Cannot erase instruction that is used!");
299    // Make sure that we reprocess all operands now that we reduced their
300    // use counts.
301    if (I.getNumOperands() < 8) {
302      for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
303        if (Instruction *Op = dyn_cast<Instruction>(*i))
304          Worklist.Add(Op);
305    }
306    Worklist.Remove(&I);
307    I.eraseFromParent();
308    MadeIRChange = true;
309    return 0;  // Don't do anything with FI
310  }
311
312  void ComputeMaskedBits(Value *V, APInt &KnownZero,
313                         APInt &KnownOne, unsigned Depth = 0) const {
314    return llvm::ComputeMaskedBits(V, KnownZero, KnownOne, DL, Depth);
315  }
316
317  bool MaskedValueIsZero(Value *V, const APInt &Mask,
318                         unsigned Depth = 0) const {
319    return llvm::MaskedValueIsZero(V, Mask, DL, Depth);
320  }
321  unsigned ComputeNumSignBits(Value *Op, unsigned Depth = 0) const {
322    return llvm::ComputeNumSignBits(Op, DL, Depth);
323  }
324
325private:
326
327  /// SimplifyAssociativeOrCommutative - This performs a few simplifications for
328  /// operators which are associative or commutative.
329  bool SimplifyAssociativeOrCommutative(BinaryOperator &I);
330
331  /// SimplifyUsingDistributiveLaws - This tries to simplify binary operations
332  /// which some other binary operation distributes over either by factorizing
333  /// out common terms (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this
334  /// results in simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is
335  /// a win).  Returns the simplified value, or null if it didn't simplify.
336  Value *SimplifyUsingDistributiveLaws(BinaryOperator &I);
337
338  /// SimplifyDemandedUseBits - Attempts to replace V with a simpler value
339  /// based on the demanded bits.
340  Value *SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
341                                 APInt& KnownZero, APInt& KnownOne,
342                                 unsigned Depth);
343  bool SimplifyDemandedBits(Use &U, APInt DemandedMask,
344                            APInt& KnownZero, APInt& KnownOne,
345                            unsigned Depth=0);
346  /// Helper routine of SimplifyDemandedUseBits. It tries to simplify demanded
347  /// bit for "r1 = shr x, c1; r2 = shl r1, c2" instruction sequence.
348  Value *SimplifyShrShlDemandedBits(Instruction *Lsr, Instruction *Sftl,
349                                    APInt DemandedMask, APInt &KnownZero,
350                                    APInt &KnownOne);
351
352  /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
353  /// SimplifyDemandedBits knows about.  See if the instruction has any
354  /// properties that allow us to simplify its operands.
355  bool SimplifyDemandedInstructionBits(Instruction &Inst);
356
357  Value *SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
358                                    APInt& UndefElts, unsigned Depth = 0);
359
360  // FoldOpIntoPhi - Given a binary operator, cast instruction, or select
361  // which has a PHI node as operand #0, see if we can fold the instruction
362  // into the PHI (which is only possible if all operands to the PHI are
363  // constants).
364  //
365  Instruction *FoldOpIntoPhi(Instruction &I);
366
367  // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
368  // operator and they all are only used by the PHI, PHI together their
369  // inputs, and do the operation once, to the result of the PHI.
370  Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
371  Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
372  Instruction *FoldPHIArgGEPIntoPHI(PHINode &PN);
373  Instruction *FoldPHIArgLoadIntoPHI(PHINode &PN);
374
375
376  Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
377                        ConstantInt *AndRHS, BinaryOperator &TheAnd);
378
379  Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
380                            bool isSub, Instruction &I);
381  Value *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
382                         bool isSigned, bool Inside);
383  Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocaInst &AI);
384  Instruction *MatchBSwap(BinaryOperator &I);
385  bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
386  Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
387  Instruction *SimplifyMemSet(MemSetInst *MI);
388
389
390  Value *EvaluateInDifferentType(Value *V, Type *Ty, bool isSigned);
391
392  /// Descale - Return a value X such that Val = X * Scale, or null if none.  If
393  /// the multiplication is known not to overflow then NoSignedWrap is set.
394  Value *Descale(Value *Val, APInt Scale, bool &NoSignedWrap);
395};
396
397
398
399} // end namespace llvm.
400
401#endif
402