InstCombine.h revision 687140c818ba4b896329a83324714140b6580ef8 
1//===- InstCombine.h - Main InstCombine pass definition -------------------===//
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/Pass.h"
15#include "llvm/Analysis/ValueTracking.h"
16#include "llvm/Support/IRBuilder.h"
17#include "llvm/Support/InstVisitor.h"
18#include "llvm/Support/TargetFolder.h"
19
20namespace llvm {
21  class CallSite;
22  class TargetData;
23  class DbgDeclareInst;
24  class MemIntrinsic;
25  class MemSetInst;
26
27/// SelectPatternFlavor - We can match a variety of different patterns for
28/// select operations.
29enum SelectPatternFlavor {
30  SPF_UNKNOWN = 0,
31  SPF_SMIN, SPF_UMIN,
32  SPF_SMAX, SPF_UMAX
33  //SPF_ABS - TODO.
34};
35
36/// getComplexity:  Assign a complexity or rank value to LLVM Values...
37///   0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
38static inline unsigned getComplexity(Value *V) {
39  if (isa<Instruction>(V)) {
40    if (BinaryOperator::isNeg(V) ||
41        BinaryOperator::isFNeg(V) ||
42        BinaryOperator::isNot(V))
43      return 3;
44    return 4;
45  }
46  if (isa<Argument>(V)) return 3;
47  return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
48}
49
50
51/// InstCombineIRInserter - This is an IRBuilder insertion helper that works
52/// just like the normal insertion helper, but also adds any new instructions
53/// to the instcombine worklist.
54class LLVM_LIBRARY_VISIBILITY InstCombineIRInserter
55    : public IRBuilderDefaultInserter<true> {
56  InstCombineWorklist &Worklist;
57public:
58  InstCombineIRInserter(InstCombineWorklist &WL) : Worklist(WL) {}
59
60  void InsertHelper(Instruction *I, const Twine &Name,
61                    BasicBlock *BB, BasicBlock::iterator InsertPt) const {
62    IRBuilderDefaultInserter<true>::InsertHelper(I, Name, BB, InsertPt);
63    Worklist.Add(I);
64  }
65};
66
67/// InstCombiner - The -instcombine pass.
68class LLVM_LIBRARY_VISIBILITY InstCombiner
69                             : public FunctionPass,
70                               public InstVisitor<InstCombiner, Instruction*> {
71  TargetData *TD;
72  bool MustPreserveLCSSA;
73  bool MadeIRChange;
74public:
75  /// Worklist - All of the instructions that need to be simplified.
76  InstCombineWorklist Worklist;
77
78  /// Builder - This is an IRBuilder that automatically inserts new
79  /// instructions into the worklist when they are created.
80  typedef IRBuilder<true, TargetFolder, InstCombineIRInserter> BuilderTy;
81  BuilderTy *Builder;
82
83  static char ID; // Pass identification, replacement for typeid
84  InstCombiner() : FunctionPass(ID), TD(0), Builder(0) {
85    initializeInstCombinerPass(*PassRegistry::getPassRegistry());
86  }
87
88public:
89  virtual bool runOnFunction(Function &F);
90
91  bool DoOneIteration(Function &F, unsigned ItNum);
92
93  virtual void getAnalysisUsage(AnalysisUsage &AU) const;
94
95  TargetData *getTargetData() const { return TD; }
96
97  // Visitation implementation - Implement instruction combining for different
98  // instruction types.  The semantics are as follows:
99  // Return Value:
100  //    null        - No change was made
101  //     I          - Change was made, I is still valid, I may be dead though
102  //   otherwise    - Change was made, replace I with returned instruction
103  //
104  Instruction *visitAdd(BinaryOperator &I);
105  Instruction *visitFAdd(BinaryOperator &I);
106  Value *OptimizePointerDifference(Value *LHS, Value *RHS, const Type *Ty);
107  Instruction *visitSub(BinaryOperator &I);
108  Instruction *visitFSub(BinaryOperator &I);
109  Instruction *visitMul(BinaryOperator &I);
110  Instruction *visitFMul(BinaryOperator &I);
111  Instruction *visitURem(BinaryOperator &I);
112  Instruction *visitSRem(BinaryOperator &I);
113  Instruction *visitFRem(BinaryOperator &I);
114  bool SimplifyDivRemOfSelect(BinaryOperator &I);
115  Instruction *commonRemTransforms(BinaryOperator &I);
116  Instruction *commonIRemTransforms(BinaryOperator &I);
117  Instruction *commonDivTransforms(BinaryOperator &I);
118  Instruction *commonIDivTransforms(BinaryOperator &I);
119  Instruction *visitUDiv(BinaryOperator &I);
120  Instruction *visitSDiv(BinaryOperator &I);
121  Instruction *visitFDiv(BinaryOperator &I);
122  Value *FoldAndOfICmps(ICmpInst *LHS, ICmpInst *RHS);
123  Value *FoldAndOfFCmps(FCmpInst *LHS, FCmpInst *RHS);
124  Instruction *visitAnd(BinaryOperator &I);
125  Value *FoldOrOfICmps(ICmpInst *LHS, ICmpInst *RHS);
126  Value *FoldOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS);
127  Instruction *FoldOrWithConstants(BinaryOperator &I, Value *Op,
128                                   Value *A, Value *B, Value *C);
129  Instruction *visitOr (BinaryOperator &I);
130  Instruction *visitXor(BinaryOperator &I);
131  Instruction *visitShl(BinaryOperator &I);
132  Instruction *visitAShr(BinaryOperator &I);
133  Instruction *visitLShr(BinaryOperator &I);
134  Instruction *commonShiftTransforms(BinaryOperator &I);
135  Instruction *FoldFCmp_IntToFP_Cst(FCmpInst &I, Instruction *LHSI,
136                                    Constant *RHSC);
137  Instruction *FoldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP,
138                                            GlobalVariable *GV, CmpInst &ICI,
139                                            ConstantInt *AndCst = 0);
140  Instruction *visitFCmpInst(FCmpInst &I);
141  Instruction *visitICmpInst(ICmpInst &I);
142  Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
143  Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
144                                              Instruction *LHS,
145                                              ConstantInt *RHS);
146  Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
147                              ConstantInt *DivRHS);
148  Instruction *FoldICmpAddOpCst(ICmpInst &ICI, Value *X, ConstantInt *CI,
149                                ICmpInst::Predicate Pred, Value *TheAdd);
150  Instruction *FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
151                           ICmpInst::Predicate Cond, Instruction &I);
152  Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
153                                   BinaryOperator &I);
154  Instruction *commonCastTransforms(CastInst &CI);
155  Instruction *commonPointerCastTransforms(CastInst &CI);
156  Instruction *visitTrunc(TruncInst &CI);
157  Instruction *visitZExt(ZExtInst &CI);
158  Instruction *visitSExt(SExtInst &CI);
159  Instruction *visitFPTrunc(FPTruncInst &CI);
160  Instruction *visitFPExt(CastInst &CI);
161  Instruction *visitFPToUI(FPToUIInst &FI);
162  Instruction *visitFPToSI(FPToSIInst &FI);
163  Instruction *visitUIToFP(CastInst &CI);
164  Instruction *visitSIToFP(CastInst &CI);
165  Instruction *visitPtrToInt(PtrToIntInst &CI);
166  Instruction *visitIntToPtr(IntToPtrInst &CI);
167  Instruction *visitBitCast(BitCastInst &CI);
168  Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
169                              Instruction *FI);
170  Instruction *FoldSelectIntoOp(SelectInst &SI, Value*, Value*);
171  Instruction *FoldSPFofSPF(Instruction *Inner, SelectPatternFlavor SPF1,
172                            Value *A, Value *B, Instruction &Outer,
173                            SelectPatternFlavor SPF2, Value *C);
174  Instruction *visitSelectInst(SelectInst &SI);
175  Instruction *visitSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI);
176  Instruction *visitCallInst(CallInst &CI);
177  Instruction *visitInvokeInst(InvokeInst &II);
178
179  Instruction *SliceUpIllegalIntegerPHI(PHINode &PN);
180  Instruction *visitPHINode(PHINode &PN);
181  Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
182  Instruction *visitAllocaInst(AllocaInst &AI);
183  Instruction *visitMalloc(Instruction &FI);
184  Instruction *visitFree(CallInst &FI);
185  Instruction *visitLoadInst(LoadInst &LI);
186  Instruction *visitStoreInst(StoreInst &SI);
187  Instruction *visitBranchInst(BranchInst &BI);
188  Instruction *visitSwitchInst(SwitchInst &SI);
189  Instruction *visitInsertElementInst(InsertElementInst &IE);
190  Instruction *visitExtractElementInst(ExtractElementInst &EI);
191  Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
192  Instruction *visitExtractValueInst(ExtractValueInst &EV);
193
194  // visitInstruction - Specify what to return for unhandled instructions...
195  Instruction *visitInstruction(Instruction &I) { return 0; }
196
197private:
198  bool ShouldChangeType(const Type *From, const Type *To) const;
199  Value *dyn_castNegVal(Value *V) const;
200  Value *dyn_castFNegVal(Value *V) const;
201  const Type *FindElementAtOffset(const Type *Ty, int64_t Offset,
202                                  SmallVectorImpl<Value*> &NewIndices);
203  Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI);
204
205  /// ShouldOptimizeCast - Return true if the cast from "V to Ty" actually
206  /// results in any code being generated and is interesting to optimize out. If
207  /// the cast can be eliminated by some other simple transformation, we prefer
208  /// to do the simplification first.
209  bool ShouldOptimizeCast(Instruction::CastOps opcode,const Value *V,
210                          const Type *Ty);
211
212  Instruction *visitCallSite(CallSite CS);
213  Instruction *tryOptimizeCall(CallInst *CI, const TargetData *TD);
214  bool transformConstExprCastCall(CallSite CS);
215  Instruction *transformCallThroughTrampoline(CallSite CS);
216  Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI,
217                                 bool DoXform = true);
218  bool WillNotOverflowSignedAdd(Value *LHS, Value *RHS);
219  DbgDeclareInst *hasOneUsePlusDeclare(Value *V);
220  Value *EmitGEPOffset(User *GEP);
221
222public:
223  // InsertNewInstBefore - insert an instruction New before instruction Old
224  // in the program.  Add the new instruction to the worklist.
225  //
226  Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
227    assert(New && New->getParent() == 0 &&
228           "New instruction already inserted into a basic block!");
229    BasicBlock *BB = Old.getParent();
230    BB->getInstList().insert(&Old, New);  // Insert inst
231    Worklist.Add(New);
232    return New;
233  }
234
235  // ReplaceInstUsesWith - This method is to be used when an instruction is
236  // found to be dead, replacable with another preexisting expression.  Here
237  // we add all uses of I to the worklist, replace all uses of I with the new
238  // value, then return I, so that the inst combiner will know that I was
239  // modified.
240  //
241  Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
242    Worklist.AddUsersToWorkList(I);   // Add all modified instrs to worklist.
243
244    // If we are replacing the instruction with itself, this must be in a
245    // segment of unreachable code, so just clobber the instruction.
246    if (&I == V)
247      V = UndefValue::get(I.getType());
248
249    I.replaceAllUsesWith(V);
250    return &I;
251  }
252
253  // EraseInstFromFunction - When dealing with an instruction that has side
254  // effects or produces a void value, we can't rely on DCE to delete the
255  // instruction.  Instead, visit methods should return the value returned by
256  // this function.
257  Instruction *EraseInstFromFunction(Instruction &I) {
258    DEBUG(errs() << "IC: ERASE " << I << '\n');
259
260    assert(I.use_empty() && "Cannot erase instruction that is used!");
261    // Make sure that we reprocess all operands now that we reduced their
262    // use counts.
263    if (I.getNumOperands() < 8) {
264      for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
265        if (Instruction *Op = dyn_cast<Instruction>(*i))
266          Worklist.Add(Op);
267    }
268    Worklist.Remove(&I);
269    I.eraseFromParent();
270    MadeIRChange = true;
271    return 0;  // Don't do anything with FI
272  }
273
274  void ComputeMaskedBits(Value *V, const APInt &Mask, APInt &KnownZero,
275                         APInt &KnownOne, unsigned Depth = 0) const {
276    return llvm::ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
277  }
278
279  bool MaskedValueIsZero(Value *V, const APInt &Mask,
280                         unsigned Depth = 0) const {
281    return llvm::MaskedValueIsZero(V, Mask, TD, Depth);
282  }
283  unsigned ComputeNumSignBits(Value *Op, unsigned Depth = 0) const {
284    return llvm::ComputeNumSignBits(Op, TD, Depth);
285  }
286
287private:
288
289  /// SimplifyAssociativeOrCommutative - This performs a few simplifications for
290  /// operators which are associative or commutative.
291  bool SimplifyAssociativeOrCommutative(BinaryOperator &I);
292
293  /// SimplifyUsingDistributiveLaws - This tries to simplify binary operations
294  /// which some other binary operation distributes over either by factorizing
295  /// out common terms (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this
296  /// results in simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is
297  /// a win).  Returns the simplified value, or null if it didn't simplify.
298  Value *SimplifyUsingDistributiveLaws(BinaryOperator &I);
299
300  /// SimplifyDemandedUseBits - Attempts to replace V with a simpler value
301  /// based on the demanded bits.
302  Value *SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
303                                 APInt& KnownZero, APInt& KnownOne,
304                                 unsigned Depth);
305  bool SimplifyDemandedBits(Use &U, APInt DemandedMask,
306                            APInt& KnownZero, APInt& KnownOne,
307                            unsigned Depth=0);
308
309  /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
310  /// SimplifyDemandedBits knows about.  See if the instruction has any
311  /// properties that allow us to simplify its operands.
312  bool SimplifyDemandedInstructionBits(Instruction &Inst);
313
314  Value *SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
315                                    APInt& UndefElts, unsigned Depth = 0);
316
317  // FoldOpIntoPhi - Given a binary operator, cast instruction, or select
318  // which has a PHI node as operand #0, see if we can fold the instruction
319  // into the PHI (which is only possible if all operands to the PHI are
320  // constants).
321  //
322  // If AllowAggressive is true, FoldOpIntoPhi will allow certain transforms
323  // that would normally be unprofitable because they strongly encourage jump
324  // threading.
325  Instruction *FoldOpIntoPhi(Instruction &I, bool AllowAggressive = false);
326
327  // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
328  // operator and they all are only used by the PHI, PHI together their
329  // inputs, and do the operation once, to the result of the PHI.
330  Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
331  Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
332  Instruction *FoldPHIArgGEPIntoPHI(PHINode &PN);
333  Instruction *FoldPHIArgLoadIntoPHI(PHINode &PN);
334
335
336  Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
337                        ConstantInt *AndRHS, BinaryOperator &TheAnd);
338
339  Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
340                            bool isSub, Instruction &I);
341  Value *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
342                         bool isSigned, bool Inside);
343  Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocaInst &AI);
344  Instruction *MatchBSwap(BinaryOperator &I);
345  bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
346  Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
347  Instruction *SimplifyMemSet(MemSetInst *MI);
348
349
350  Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
351};
352
353
354
355} // end namespace llvm.
356
357#endif
358