InstCombine.h revision 6dd44d3b7f33c9984dfb40461630d50c4fed1234 
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/ValueTracking.h"
15#include "llvm/IR/IRBuilder.h"
16#include "llvm/IR/IntrinsicInst.h"
17#include "llvm/IR/Operator.h"
18#include "llvm/InstVisitor.h"
19#include "llvm/Pass.h"
20#include "llvm/Support/TargetFolder.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
55/// InstCombineIRInserter - This is an IRBuilder insertion helper that works
56/// just like the normal insertion helper, but also adds any new instructions
57/// to the instcombine worklist.
58class LLVM_LIBRARY_VISIBILITY InstCombineIRInserter
59    : public IRBuilderDefaultInserter<true> {
60  InstCombineWorklist &Worklist;
61public:
62  InstCombineIRInserter(InstCombineWorklist &WL) : Worklist(WL) {}
63
64  void InsertHelper(Instruction *I, const Twine &Name,
65                    BasicBlock *BB, BasicBlock::iterator InsertPt) const {
66    IRBuilderDefaultInserter<true>::InsertHelper(I, Name, BB, InsertPt);
67    Worklist.Add(I);
68  }
69};
70
71/// InstCombiner - The -instcombine pass.
72class LLVM_LIBRARY_VISIBILITY InstCombiner
73                             : public FunctionPass,
74                               public InstVisitor<InstCombiner, Instruction*> {
75  DataLayout *TD;
76  TargetLibraryInfo *TLI;
77  bool MadeIRChange;
78  LibCallSimplifier *Simplifier;
79  bool MinimizeSize;
80public:
81  /// Worklist - All of the instructions that need to be simplified.
82  InstCombineWorklist Worklist;
83
84  /// Builder - This is an IRBuilder that automatically inserts new
85  /// instructions into the worklist when they are created.
86  typedef IRBuilder<true, TargetFolder, InstCombineIRInserter> BuilderTy;
87  BuilderTy *Builder;
88
89  static char ID; // Pass identification, replacement for typeid
90  InstCombiner() : FunctionPass(ID), TD(0), Builder(0) {
91    MinimizeSize = false;
92    initializeInstCombinerPass(*PassRegistry::getPassRegistry());
93  }
94
95public:
96  virtual bool runOnFunction(Function &F);
97
98  bool DoOneIteration(Function &F, unsigned ItNum);
99
100  virtual void getAnalysisUsage(AnalysisUsage &AU) const;
101
102  DataLayout *getDataLayout() const { return TD; }
103
104  TargetLibraryInfo *getTargetLibraryInfo() const { return TLI; }
105
106  // Visitation implementation - Implement instruction combining for different
107  // instruction types.  The semantics are as follows:
108  // Return Value:
109  //    null        - No change was made
110  //     I          - Change was made, I is still valid, I may be dead though
111  //   otherwise    - Change was made, replace I with returned instruction
112  //
113  Instruction *visitAdd(BinaryOperator &I);
114  Instruction *visitFAdd(BinaryOperator &I);
115  Value *OptimizePointerDifference(Value *LHS, Value *RHS, Type *Ty);
116  Instruction *visitSub(BinaryOperator &I);
117  Instruction *visitFSub(BinaryOperator &I);
118  Instruction *visitMul(BinaryOperator &I);
119  Value *foldFMulConst(Instruction *FMulOrDiv, ConstantFP *C,
120                       Instruction *InsertBefore);
121  Instruction *visitFMul(BinaryOperator &I);
122  Instruction *visitURem(BinaryOperator &I);
123  Instruction *visitSRem(BinaryOperator &I);
124  Instruction *visitFRem(BinaryOperator &I);
125  bool SimplifyDivRemOfSelect(BinaryOperator &I);
126  Instruction *commonRemTransforms(BinaryOperator &I);
127  Instruction *commonIRemTransforms(BinaryOperator &I);
128  Instruction *commonDivTransforms(BinaryOperator &I);
129  Instruction *commonIDivTransforms(BinaryOperator &I);
130  Instruction *visitUDiv(BinaryOperator &I);
131  Instruction *visitSDiv(BinaryOperator &I);
132  Instruction *visitFDiv(BinaryOperator &I);
133  Value *FoldAndOfICmps(ICmpInst *LHS, ICmpInst *RHS);
134  Value *FoldAndOfFCmps(FCmpInst *LHS, FCmpInst *RHS);
135  Instruction *visitAnd(BinaryOperator &I);
136  Value *FoldOrOfICmps(ICmpInst *LHS, ICmpInst *RHS);
137  Value *FoldOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS);
138  Instruction *FoldOrWithConstants(BinaryOperator &I, Value *Op,
139                                   Value *A, Value *B, Value *C);
140  Instruction *visitOr (BinaryOperator &I);
141  Instruction *visitXor(BinaryOperator &I);
142  Instruction *visitShl(BinaryOperator &I);
143  Instruction *visitAShr(BinaryOperator &I);
144  Instruction *visitLShr(BinaryOperator &I);
145  Instruction *commonShiftTransforms(BinaryOperator &I);
146  Instruction *FoldFCmp_IntToFP_Cst(FCmpInst &I, Instruction *LHSI,
147                                    Constant *RHSC);
148  Instruction *FoldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP,
149                                            GlobalVariable *GV, CmpInst &ICI,
150                                            ConstantInt *AndCst = 0);
151  Instruction *visitFCmpInst(FCmpInst &I);
152  Instruction *visitICmpInst(ICmpInst &I);
153  Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
154  Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
155                                              Instruction *LHS,
156                                              ConstantInt *RHS);
157  Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
158                              ConstantInt *DivRHS);
159  Instruction *FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *DivI,
160                              ConstantInt *DivRHS);
161  Instruction *FoldICmpAddOpCst(Instruction &ICI, Value *X, ConstantInt *CI,
162                                ICmpInst::Predicate Pred);
163  Instruction *FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
164                           ICmpInst::Predicate Cond, Instruction &I);
165  Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
166                                   BinaryOperator &I);
167  Instruction *commonCastTransforms(CastInst &CI);
168  Instruction *commonPointerCastTransforms(CastInst &CI);
169  Instruction *visitTrunc(TruncInst &CI);
170  Instruction *visitZExt(ZExtInst &CI);
171  Instruction *visitSExt(SExtInst &CI);
172  Instruction *visitFPTrunc(FPTruncInst &CI);
173  Instruction *visitFPExt(CastInst &CI);
174  Instruction *visitFPToUI(FPToUIInst &FI);
175  Instruction *visitFPToSI(FPToSIInst &FI);
176  Instruction *visitUIToFP(CastInst &CI);
177  Instruction *visitSIToFP(CastInst &CI);
178  Instruction *visitPtrToInt(PtrToIntInst &CI);
179  Instruction *visitIntToPtr(IntToPtrInst &CI);
180  Instruction *visitBitCast(BitCastInst &CI);
181  Instruction *visitAddrSpaceCast(AddrSpaceCastInst &CI);
182  Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
183                              Instruction *FI);
184  Instruction *FoldSelectIntoOp(SelectInst &SI, Value*, Value*);
185  Instruction *FoldSPFofSPF(Instruction *Inner, SelectPatternFlavor SPF1,
186                            Value *A, Value *B, Instruction &Outer,
187                            SelectPatternFlavor SPF2, Value *C);
188  Instruction *visitSelectInst(SelectInst &SI);
189  Instruction *visitSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI);
190  Instruction *visitCallInst(CallInst &CI);
191  Instruction *visitInvokeInst(InvokeInst &II);
192
193  Instruction *SliceUpIllegalIntegerPHI(PHINode &PN);
194  Instruction *visitPHINode(PHINode &PN);
195  Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
196  Instruction *visitAllocaInst(AllocaInst &AI);
197  Instruction *visitAllocSite(Instruction &FI);
198  Instruction *visitFree(CallInst &FI);
199  Instruction *visitLoadInst(LoadInst &LI);
200  Instruction *visitStoreInst(StoreInst &SI);
201  Instruction *visitBranchInst(BranchInst &BI);
202  Instruction *visitSwitchInst(SwitchInst &SI);
203  Instruction *visitInsertElementInst(InsertElementInst &IE);
204  Instruction *visitExtractElementInst(ExtractElementInst &EI);
205  Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
206  Instruction *visitExtractValueInst(ExtractValueInst &EV);
207  Instruction *visitLandingPadInst(LandingPadInst &LI);
208
209  // visitInstruction - Specify what to return for unhandled instructions...
210  Instruction *visitInstruction(Instruction &I) { return 0; }
211
212private:
213  bool ShouldChangeType(Type *From, Type *To) const;
214  Value *dyn_castNegVal(Value *V) const;
215  Value *dyn_castFNegVal(Value *V, bool NoSignedZero=false) const;
216  Type *FindElementAtOffset(Type *PtrTy, int64_t Offset,
217                            SmallVectorImpl<Value*> &NewIndices);
218  Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI);
219
220  /// ShouldOptimizeCast - Return true if the cast from "V to Ty" actually
221  /// results in any code being generated and is interesting to optimize out. If
222  /// the cast can be eliminated by some other simple transformation, we prefer
223  /// to do the simplification first.
224  bool ShouldOptimizeCast(Instruction::CastOps opcode,const Value *V,
225                          Type *Ty);
226
227  Instruction *visitCallSite(CallSite CS);
228  Instruction *tryOptimizeCall(CallInst *CI, const DataLayout *TD);
229  bool transformConstExprCastCall(CallSite CS);
230  Instruction *transformCallThroughTrampoline(CallSite CS,
231                                              IntrinsicInst *Tramp);
232  Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI,
233                                 bool DoXform = true);
234  Instruction *transformSExtICmp(ICmpInst *ICI, Instruction &CI);
235  bool WillNotOverflowSignedAdd(Value *LHS, Value *RHS);
236  Value *EmitGEPOffset(User *GEP);
237  Instruction *scalarizePHI(ExtractElementInst &EI, PHINode *PN);
238  Value *EvaluateInDifferentElementOrder(Value *V, ArrayRef<int> Mask);
239
240public:
241  // InsertNewInstBefore - insert an instruction New before instruction Old
242  // in the program.  Add the new instruction to the worklist.
243  //
244  Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
245    assert(New && New->getParent() == 0 &&
246           "New instruction already inserted into a basic block!");
247    BasicBlock *BB = Old.getParent();
248    BB->getInstList().insert(&Old, New);  // Insert inst
249    Worklist.Add(New);
250    return New;
251  }
252
253  // InsertNewInstWith - same as InsertNewInstBefore, but also sets the
254  // debug loc.
255  //
256  Instruction *InsertNewInstWith(Instruction *New, Instruction &Old) {
257    New->setDebugLoc(Old.getDebugLoc());
258    return InsertNewInstBefore(New, Old);
259  }
260
261  // ReplaceInstUsesWith - This method is to be used when an instruction is
262  // found to be dead, replacable with another preexisting expression.  Here
263  // we add all uses of I to the worklist, replace all uses of I with the new
264  // value, then return I, so that the inst combiner will know that I was
265  // modified.
266  //
267  Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
268    Worklist.AddUsersToWorkList(I);   // Add all modified instrs to worklist.
269
270    // If we are replacing the instruction with itself, this must be in a
271    // segment of unreachable code, so just clobber the instruction.
272    if (&I == V)
273      V = UndefValue::get(I.getType());
274
275    DEBUG(dbgs() << "IC: Replacing " << I << "\n"
276                    "    with " << *V << '\n');
277
278    I.replaceAllUsesWith(V);
279    return &I;
280  }
281
282  // EraseInstFromFunction - When dealing with an instruction that has side
283  // effects or produces a void value, we can't rely on DCE to delete the
284  // instruction.  Instead, visit methods should return the value returned by
285  // this function.
286  Instruction *EraseInstFromFunction(Instruction &I) {
287    DEBUG(dbgs() << "IC: ERASE " << I << '\n');
288
289    assert(I.use_empty() && "Cannot erase instruction that is used!");
290    // Make sure that we reprocess all operands now that we reduced their
291    // use counts.
292    if (I.getNumOperands() < 8) {
293      for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
294        if (Instruction *Op = dyn_cast<Instruction>(*i))
295          Worklist.Add(Op);
296    }
297    Worklist.Remove(&I);
298    I.eraseFromParent();
299    MadeIRChange = true;
300    return 0;  // Don't do anything with FI
301  }
302
303  void ComputeMaskedBits(Value *V, APInt &KnownZero,
304                         APInt &KnownOne, unsigned Depth = 0) const {
305    return llvm::ComputeMaskedBits(V, KnownZero, KnownOne, TD, Depth);
306  }
307
308  bool MaskedValueIsZero(Value *V, const APInt &Mask,
309                         unsigned Depth = 0) const {
310    return llvm::MaskedValueIsZero(V, Mask, TD, Depth);
311  }
312  unsigned ComputeNumSignBits(Value *Op, unsigned Depth = 0) const {
313    return llvm::ComputeNumSignBits(Op, TD, Depth);
314  }
315
316private:
317
318  /// SimplifyAssociativeOrCommutative - This performs a few simplifications for
319  /// operators which are associative or commutative.
320  bool SimplifyAssociativeOrCommutative(BinaryOperator &I);
321
322  /// SimplifyUsingDistributiveLaws - This tries to simplify binary operations
323  /// which some other binary operation distributes over either by factorizing
324  /// out common terms (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this
325  /// results in simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is
326  /// a win).  Returns the simplified value, or null if it didn't simplify.
327  Value *SimplifyUsingDistributiveLaws(BinaryOperator &I);
328
329  /// SimplifyDemandedUseBits - Attempts to replace V with a simpler value
330  /// based on the demanded bits.
331  Value *SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
332                                 APInt& KnownZero, APInt& KnownOne,
333                                 unsigned Depth);
334  bool SimplifyDemandedBits(Use &U, APInt DemandedMask,
335                            APInt& KnownZero, APInt& KnownOne,
336                            unsigned Depth=0);
337  /// Helper routine of SimplifyDemandedUseBits. It tries to simplify demanded
338  /// bit for "r1 = shr x, c1; r2 = shl r1, c2" instruction sequence.
339  Value *SimplifyShrShlDemandedBits(Instruction *Lsr, Instruction *Sftl,
340                                    APInt DemandedMask, APInt &KnownZero,
341                                    APInt &KnownOne);
342
343  /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
344  /// SimplifyDemandedBits knows about.  See if the instruction has any
345  /// properties that allow us to simplify its operands.
346  bool SimplifyDemandedInstructionBits(Instruction &Inst);
347
348  Value *SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
349                                    APInt& UndefElts, unsigned Depth = 0);
350
351  // FoldOpIntoPhi - Given a binary operator, cast instruction, or select
352  // which has a PHI node as operand #0, see if we can fold the instruction
353  // into the PHI (which is only possible if all operands to the PHI are
354  // constants).
355  //
356  Instruction *FoldOpIntoPhi(Instruction &I);
357
358  // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
359  // operator and they all are only used by the PHI, PHI together their
360  // inputs, and do the operation once, to the result of the PHI.
361  Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
362  Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
363  Instruction *FoldPHIArgGEPIntoPHI(PHINode &PN);
364  Instruction *FoldPHIArgLoadIntoPHI(PHINode &PN);
365
366
367  Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
368                        ConstantInt *AndRHS, BinaryOperator &TheAnd);
369
370  Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
371                            bool isSub, Instruction &I);
372  Value *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
373                         bool isSigned, bool Inside);
374  Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocaInst &AI);
375  Instruction *MatchBSwap(BinaryOperator &I);
376  bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
377  Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
378  Instruction *SimplifyMemSet(MemSetInst *MI);
379
380
381  Value *EvaluateInDifferentType(Value *V, Type *Ty, bool isSigned);
382
383  /// Descale - Return a value X such that Val = X * Scale, or null if none.  If
384  /// the multiplication is known not to overflow then NoSignedWrap is set.
385  Value *Descale(Value *Val, APInt Scale, bool &NoSignedWrap);
386};
387
388
389
390} // end namespace llvm.
391
392#endif
393