InstCombine.h revision 16d8f8bd919b72866e687d99f3aa94a140137c59
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
86public:
87  virtual bool runOnFunction(Function &F);
88
89  bool DoOneIteration(Function &F, unsigned ItNum);
90
91  virtual void getAnalysisUsage(AnalysisUsage &AU) const;
92
93  TargetData *getTargetData() const { return TD; }
94
95  // Visitation implementation - Implement instruction combining for different
96  // instruction types.  The semantics are as follows:
97  // Return Value:
98  //    null        - No change was made
99  //     I          - Change was made, I is still valid, I may be dead though
100  //   otherwise    - Change was made, replace I with returned instruction
101  //
102  Instruction *visitAdd(BinaryOperator &I);
103  Instruction *visitFAdd(BinaryOperator &I);
104  Value *OptimizePointerDifference(Value *LHS, Value *RHS, const Type *Ty);
105  Instruction *visitSub(BinaryOperator &I);
106  Instruction *visitFSub(BinaryOperator &I);
107  Instruction *visitMul(BinaryOperator &I);
108  Instruction *visitFMul(BinaryOperator &I);
109  Instruction *visitURem(BinaryOperator &I);
110  Instruction *visitSRem(BinaryOperator &I);
111  Instruction *visitFRem(BinaryOperator &I);
112  bool SimplifyDivRemOfSelect(BinaryOperator &I);
113  Instruction *commonRemTransforms(BinaryOperator &I);
114  Instruction *commonIRemTransforms(BinaryOperator &I);
115  Instruction *commonDivTransforms(BinaryOperator &I);
116  Instruction *commonIDivTransforms(BinaryOperator &I);
117  Instruction *visitUDiv(BinaryOperator &I);
118  Instruction *visitSDiv(BinaryOperator &I);
119  Instruction *visitFDiv(BinaryOperator &I);
120  Value *FoldAndOfICmps(ICmpInst *LHS, ICmpInst *RHS);
121  Value *FoldAndOfFCmps(FCmpInst *LHS, FCmpInst *RHS);
122  Instruction *visitAnd(BinaryOperator &I);
123  Value *FoldOrOfICmps(ICmpInst *LHS, ICmpInst *RHS);
124  Value *FoldOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS);
125  Instruction *FoldOrWithConstants(BinaryOperator &I, Value *Op,
126                                   Value *A, Value *B, Value *C);
127  Instruction *visitOr (BinaryOperator &I);
128  Instruction *visitXor(BinaryOperator &I);
129  Instruction *visitShl(BinaryOperator &I);
130  Instruction *visitAShr(BinaryOperator &I);
131  Instruction *visitLShr(BinaryOperator &I);
132  Instruction *commonShiftTransforms(BinaryOperator &I);
133  Instruction *FoldFCmp_IntToFP_Cst(FCmpInst &I, Instruction *LHSI,
134                                    Constant *RHSC);
135  Instruction *FoldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP,
136                                            GlobalVariable *GV, CmpInst &ICI,
137                                            ConstantInt *AndCst = 0);
138  Instruction *visitFCmpInst(FCmpInst &I);
139  Instruction *visitICmpInst(ICmpInst &I);
140  Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
141  Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
142                                              Instruction *LHS,
143                                              ConstantInt *RHS);
144  Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
145                              ConstantInt *DivRHS);
146  Instruction *FoldICmpAddOpCst(ICmpInst &ICI, Value *X, ConstantInt *CI,
147                                ICmpInst::Predicate Pred, Value *TheAdd);
148  Instruction *FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
149                           ICmpInst::Predicate Cond, Instruction &I);
150  Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
151                                   BinaryOperator &I);
152  Instruction *commonCastTransforms(CastInst &CI);
153  Instruction *commonPointerCastTransforms(CastInst &CI);
154  Instruction *visitTrunc(TruncInst &CI);
155  Instruction *visitZExt(ZExtInst &CI);
156  Instruction *visitSExt(SExtInst &CI);
157  Instruction *visitFPTrunc(FPTruncInst &CI);
158  Instruction *visitFPExt(CastInst &CI);
159  Instruction *visitFPToUI(FPToUIInst &FI);
160  Instruction *visitFPToSI(FPToSIInst &FI);
161  Instruction *visitUIToFP(CastInst &CI);
162  Instruction *visitSIToFP(CastInst &CI);
163  Instruction *visitPtrToInt(PtrToIntInst &CI);
164  Instruction *visitIntToPtr(IntToPtrInst &CI);
165  Instruction *visitBitCast(BitCastInst &CI);
166  Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
167                              Instruction *FI);
168  Instruction *FoldSelectIntoOp(SelectInst &SI, Value*, Value*);
169  Instruction *FoldSPFofSPF(Instruction *Inner, SelectPatternFlavor SPF1,
170                            Value *A, Value *B, Instruction &Outer,
171                            SelectPatternFlavor SPF2, Value *C);
172  Instruction *visitSelectInst(SelectInst &SI);
173  Instruction *visitSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI);
174  Instruction *visitCallInst(CallInst &CI);
175  Instruction *visitInvokeInst(InvokeInst &II);
176
177  Instruction *SliceUpIllegalIntegerPHI(PHINode &PN);
178  Instruction *visitPHINode(PHINode &PN);
179  Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
180  Instruction *visitAllocaInst(AllocaInst &AI);
181  Instruction *visitFree(Instruction &FI);
182  Instruction *visitLoadInst(LoadInst &LI);
183  Instruction *visitStoreInst(StoreInst &SI);
184  Instruction *visitBranchInst(BranchInst &BI);
185  Instruction *visitSwitchInst(SwitchInst &SI);
186  Instruction *visitInsertElementInst(InsertElementInst &IE);
187  Instruction *visitExtractElementInst(ExtractElementInst &EI);
188  Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
189  Instruction *visitExtractValueInst(ExtractValueInst &EV);
190
191  // visitInstruction - Specify what to return for unhandled instructions...
192  Instruction *visitInstruction(Instruction &I) { return 0; }
193
194private:
195  bool ShouldChangeType(const Type *From, const Type *To) const;
196  Value *dyn_castNegVal(Value *V) const;
197  Value *dyn_castFNegVal(Value *V) const;
198  const Type *FindElementAtOffset(const Type *Ty, int64_t Offset,
199                                  SmallVectorImpl<Value*> &NewIndices);
200  Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI);
201
202  /// ShouldOptimizeCast - Return true if the cast from "V to Ty" actually
203  /// results in any code being generated and is interesting to optimize out. If
204  /// the cast can be eliminated by some other simple transformation, we prefer
205  /// to do the simplification first.
206  bool ShouldOptimizeCast(Instruction::CastOps opcode,const Value *V,
207                          const Type *Ty);
208
209  Instruction *visitCallSite(CallSite CS);
210  Instruction *tryOptimizeCall(CallInst *CI, const TargetData *TD);
211  bool transformConstExprCastCall(CallSite CS);
212  Instruction *transformCallThroughTrampoline(CallSite CS);
213  Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI,
214                                 bool DoXform = true);
215  bool WillNotOverflowSignedAdd(Value *LHS, Value *RHS);
216  DbgDeclareInst *hasOneUsePlusDeclare(Value *V);
217  Value *EmitGEPOffset(User *GEP);
218
219public:
220  // InsertNewInstBefore - insert an instruction New before instruction Old
221  // in the program.  Add the new instruction to the worklist.
222  //
223  Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
224    assert(New && New->getParent() == 0 &&
225           "New instruction already inserted into a basic block!");
226    BasicBlock *BB = Old.getParent();
227    BB->getInstList().insert(&Old, New);  // Insert inst
228    Worklist.Add(New);
229    return New;
230  }
231
232  // ReplaceInstUsesWith - This method is to be used when an instruction is
233  // found to be dead, replacable with another preexisting expression.  Here
234  // we add all uses of I to the worklist, replace all uses of I with the new
235  // value, then return I, so that the inst combiner will know that I was
236  // modified.
237  //
238  Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
239    Worklist.AddUsersToWorkList(I);   // Add all modified instrs to worklist.
240
241    // If we are replacing the instruction with itself, this must be in a
242    // segment of unreachable code, so just clobber the instruction.
243    if (&I == V)
244      V = UndefValue::get(I.getType());
245
246    I.replaceAllUsesWith(V);
247    return &I;
248  }
249
250  // EraseInstFromFunction - When dealing with an instruction that has side
251  // effects or produces a void value, we can't rely on DCE to delete the
252  // instruction.  Instead, visit methods should return the value returned by
253  // this function.
254  Instruction *EraseInstFromFunction(Instruction &I) {
255    DEBUG(errs() << "IC: ERASE " << I << '\n');
256
257    assert(I.use_empty() && "Cannot erase instruction that is used!");
258    // Make sure that we reprocess all operands now that we reduced their
259    // use counts.
260    if (I.getNumOperands() < 8) {
261      for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
262        if (Instruction *Op = dyn_cast<Instruction>(*i))
263          Worklist.Add(Op);
264    }
265    Worklist.Remove(&I);
266    I.eraseFromParent();
267    MadeIRChange = true;
268    return 0;  // Don't do anything with FI
269  }
270
271  void ComputeMaskedBits(Value *V, const APInt &Mask, APInt &KnownZero,
272                         APInt &KnownOne, unsigned Depth = 0) const {
273    return llvm::ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
274  }
275
276  bool MaskedValueIsZero(Value *V, const APInt &Mask,
277                         unsigned Depth = 0) const {
278    return llvm::MaskedValueIsZero(V, Mask, TD, Depth);
279  }
280  unsigned ComputeNumSignBits(Value *Op, unsigned Depth = 0) const {
281    return llvm::ComputeNumSignBits(Op, TD, Depth);
282  }
283
284private:
285
286  /// SimplifyCommutative - This performs a few simplifications for
287  /// commutative operators.
288  bool SimplifyCommutative(BinaryOperator &I);
289
290  /// SimplifyDemandedUseBits - Attempts to replace V with a simpler value
291  /// based on the demanded bits.
292  Value *SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
293                                 APInt& KnownZero, APInt& KnownOne,
294                                 unsigned Depth);
295  bool SimplifyDemandedBits(Use &U, APInt DemandedMask,
296                            APInt& KnownZero, APInt& KnownOne,
297                            unsigned Depth=0);
298
299  /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
300  /// SimplifyDemandedBits knows about.  See if the instruction has any
301  /// properties that allow us to simplify its operands.
302  bool SimplifyDemandedInstructionBits(Instruction &Inst);
303
304  Value *SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
305                                    APInt& UndefElts, unsigned Depth = 0);
306
307  // FoldOpIntoPhi - Given a binary operator, cast instruction, or select
308  // which has a PHI node as operand #0, see if we can fold the instruction
309  // into the PHI (which is only possible if all operands to the PHI are
310  // constants).
311  //
312  // If AllowAggressive is true, FoldOpIntoPhi will allow certain transforms
313  // that would normally be unprofitable because they strongly encourage jump
314  // threading.
315  Instruction *FoldOpIntoPhi(Instruction &I, bool AllowAggressive = false);
316
317  // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
318  // operator and they all are only used by the PHI, PHI together their
319  // inputs, and do the operation once, to the result of the PHI.
320  Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
321  Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
322  Instruction *FoldPHIArgGEPIntoPHI(PHINode &PN);
323  Instruction *FoldPHIArgLoadIntoPHI(PHINode &PN);
324
325
326  Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
327                        ConstantInt *AndRHS, BinaryOperator &TheAnd);
328
329  Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
330                            bool isSub, Instruction &I);
331  Value *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
332                         bool isSigned, bool Inside);
333  Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocaInst &AI);
334  Instruction *MatchBSwap(BinaryOperator &I);
335  bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
336  Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
337  Instruction *SimplifyMemSet(MemSetInst *MI);
338
339
340  Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
341
342  unsigned GetOrEnforceKnownAlignment(Value *V,
343                                      unsigned PrefAlign = 0);
344
345};
346
347
348
349} // end namespace llvm.
350
351#endif
352