GlobalOpt.cpp revision 081ce940e7351e90fff829320b7dc6738a6b3815
1//===- GlobalOpt.cpp - Optimize Global Variables --------------------------===//
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// This pass transforms simple global variables that never have their address
11// taken.  If obviously true, it marks read/write globals as constant, deletes
12// variables only stored to, etc.
13//
14//===----------------------------------------------------------------------===//
15
16#define DEBUG_TYPE "globalopt"
17#include "llvm/Transforms/IPO.h"
18#include "llvm/CallingConv.h"
19#include "llvm/Constants.h"
20#include "llvm/DerivedTypes.h"
21#include "llvm/Instructions.h"
22#include "llvm/IntrinsicInst.h"
23#include "llvm/Module.h"
24#include "llvm/Pass.h"
25#include "llvm/Analysis/ConstantFolding.h"
26#include "llvm/Target/TargetData.h"
27#include "llvm/Support/Compiler.h"
28#include "llvm/Support/Debug.h"
29#include "llvm/ADT/SmallPtrSet.h"
30#include "llvm/ADT/SmallVector.h"
31#include "llvm/ADT/Statistic.h"
32#include "llvm/ADT/StringExtras.h"
33#include <algorithm>
34#include <set>
35using namespace llvm;
36
37STATISTIC(NumMarked    , "Number of globals marked constant");
38STATISTIC(NumSRA       , "Number of aggregate globals broken into scalars");
39STATISTIC(NumHeapSRA   , "Number of heap objects SRA'd");
40STATISTIC(NumSubstitute,"Number of globals with initializers stored into them");
41STATISTIC(NumDeleted   , "Number of globals deleted");
42STATISTIC(NumFnDeleted , "Number of functions deleted");
43STATISTIC(NumGlobUses  , "Number of global uses devirtualized");
44STATISTIC(NumLocalized , "Number of globals localized");
45STATISTIC(NumShrunkToBool  , "Number of global vars shrunk to booleans");
46STATISTIC(NumFastCallFns   , "Number of functions converted to fastcc");
47STATISTIC(NumCtorsEvaluated, "Number of static ctors evaluated");
48
49namespace {
50  struct VISIBILITY_HIDDEN GlobalOpt : public ModulePass {
51    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
52      AU.addRequired<TargetData>();
53    }
54    static char ID; // Pass identification, replacement for typeid
55    GlobalOpt() : ModulePass((intptr_t)&ID) {}
56
57    bool runOnModule(Module &M);
58
59  private:
60    GlobalVariable *FindGlobalCtors(Module &M);
61    bool OptimizeFunctions(Module &M);
62    bool OptimizeGlobalVars(Module &M);
63    bool OptimizeGlobalCtorsList(GlobalVariable *&GCL);
64    bool ProcessInternalGlobal(GlobalVariable *GV,Module::global_iterator &GVI);
65  };
66
67  char GlobalOpt::ID = 0;
68  RegisterPass<GlobalOpt> X("globalopt", "Global Variable Optimizer");
69}
70
71ModulePass *llvm::createGlobalOptimizerPass() { return new GlobalOpt(); }
72
73/// GlobalStatus - As we analyze each global, keep track of some information
74/// about it.  If we find out that the address of the global is taken, none of
75/// this info will be accurate.
76struct VISIBILITY_HIDDEN GlobalStatus {
77  /// isLoaded - True if the global is ever loaded.  If the global isn't ever
78  /// loaded it can be deleted.
79  bool isLoaded;
80
81  /// StoredType - Keep track of what stores to the global look like.
82  ///
83  enum StoredType {
84    /// NotStored - There is no store to this global.  It can thus be marked
85    /// constant.
86    NotStored,
87
88    /// isInitializerStored - This global is stored to, but the only thing
89    /// stored is the constant it was initialized with.  This is only tracked
90    /// for scalar globals.
91    isInitializerStored,
92
93    /// isStoredOnce - This global is stored to, but only its initializer and
94    /// one other value is ever stored to it.  If this global isStoredOnce, we
95    /// track the value stored to it in StoredOnceValue below.  This is only
96    /// tracked for scalar globals.
97    isStoredOnce,
98
99    /// isStored - This global is stored to by multiple values or something else
100    /// that we cannot track.
101    isStored
102  } StoredType;
103
104  /// StoredOnceValue - If only one value (besides the initializer constant) is
105  /// ever stored to this global, keep track of what value it is.
106  Value *StoredOnceValue;
107
108  /// AccessingFunction/HasMultipleAccessingFunctions - These start out
109  /// null/false.  When the first accessing function is noticed, it is recorded.
110  /// When a second different accessing function is noticed,
111  /// HasMultipleAccessingFunctions is set to true.
112  Function *AccessingFunction;
113  bool HasMultipleAccessingFunctions;
114
115  /// HasNonInstructionUser - Set to true if this global has a user that is not
116  /// an instruction (e.g. a constant expr or GV initializer).
117  bool HasNonInstructionUser;
118
119  /// HasPHIUser - Set to true if this global has a user that is a PHI node.
120  bool HasPHIUser;
121
122  /// isNotSuitableForSRA - Keep track of whether any SRA preventing users of
123  /// the global exist.  Such users include GEP instruction with variable
124  /// indexes, and non-gep/load/store users like constant expr casts.
125  bool isNotSuitableForSRA;
126
127  GlobalStatus() : isLoaded(false), StoredType(NotStored), StoredOnceValue(0),
128                   AccessingFunction(0), HasMultipleAccessingFunctions(false),
129                   HasNonInstructionUser(false), HasPHIUser(false),
130                   isNotSuitableForSRA(false) {}
131};
132
133
134
135/// ConstantIsDead - Return true if the specified constant is (transitively)
136/// dead.  The constant may be used by other constants (e.g. constant arrays and
137/// constant exprs) as long as they are dead, but it cannot be used by anything
138/// else.
139static bool ConstantIsDead(Constant *C) {
140  if (isa<GlobalValue>(C)) return false;
141
142  for (Value::use_iterator UI = C->use_begin(), E = C->use_end(); UI != E; ++UI)
143    if (Constant *CU = dyn_cast<Constant>(*UI)) {
144      if (!ConstantIsDead(CU)) return false;
145    } else
146      return false;
147  return true;
148}
149
150
151/// AnalyzeGlobal - Look at all uses of the global and fill in the GlobalStatus
152/// structure.  If the global has its address taken, return true to indicate we
153/// can't do anything with it.
154///
155static bool AnalyzeGlobal(Value *V, GlobalStatus &GS,
156                          std::set<PHINode*> &PHIUsers) {
157  for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
158    if (ConstantExpr *CE = dyn_cast<ConstantExpr>(*UI)) {
159      GS.HasNonInstructionUser = true;
160
161      if (AnalyzeGlobal(CE, GS, PHIUsers)) return true;
162      if (CE->getOpcode() != Instruction::GetElementPtr)
163        GS.isNotSuitableForSRA = true;
164      else if (!GS.isNotSuitableForSRA) {
165        // Check to see if this ConstantExpr GEP is SRA'able.  In particular, we
166        // don't like < 3 operand CE's, and we don't like non-constant integer
167        // indices.
168        if (CE->getNumOperands() < 3 || !CE->getOperand(1)->isNullValue())
169          GS.isNotSuitableForSRA = true;
170        else {
171          for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
172            if (!isa<ConstantInt>(CE->getOperand(i))) {
173              GS.isNotSuitableForSRA = true;
174              break;
175            }
176        }
177      }
178
179    } else if (Instruction *I = dyn_cast<Instruction>(*UI)) {
180      if (!GS.HasMultipleAccessingFunctions) {
181        Function *F = I->getParent()->getParent();
182        if (GS.AccessingFunction == 0)
183          GS.AccessingFunction = F;
184        else if (GS.AccessingFunction != F)
185          GS.HasMultipleAccessingFunctions = true;
186      }
187      if (isa<LoadInst>(I)) {
188        GS.isLoaded = true;
189      } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
190        // Don't allow a store OF the address, only stores TO the address.
191        if (SI->getOperand(0) == V) return true;
192
193        // If this is a direct store to the global (i.e., the global is a scalar
194        // value, not an aggregate), keep more specific information about
195        // stores.
196        if (GS.StoredType != GlobalStatus::isStored)
197          if (GlobalVariable *GV = dyn_cast<GlobalVariable>(SI->getOperand(1))){
198            Value *StoredVal = SI->getOperand(0);
199            if (StoredVal == GV->getInitializer()) {
200              if (GS.StoredType < GlobalStatus::isInitializerStored)
201                GS.StoredType = GlobalStatus::isInitializerStored;
202            } else if (isa<LoadInst>(StoredVal) &&
203                       cast<LoadInst>(StoredVal)->getOperand(0) == GV) {
204              // G = G
205              if (GS.StoredType < GlobalStatus::isInitializerStored)
206                GS.StoredType = GlobalStatus::isInitializerStored;
207            } else if (GS.StoredType < GlobalStatus::isStoredOnce) {
208              GS.StoredType = GlobalStatus::isStoredOnce;
209              GS.StoredOnceValue = StoredVal;
210            } else if (GS.StoredType == GlobalStatus::isStoredOnce &&
211                       GS.StoredOnceValue == StoredVal) {
212              // noop.
213            } else {
214              GS.StoredType = GlobalStatus::isStored;
215            }
216          } else {
217            GS.StoredType = GlobalStatus::isStored;
218          }
219      } else if (isa<GetElementPtrInst>(I)) {
220        if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
221
222        // If the first two indices are constants, this can be SRA'd.
223        if (isa<GlobalVariable>(I->getOperand(0))) {
224          if (I->getNumOperands() < 3 || !isa<Constant>(I->getOperand(1)) ||
225              !cast<Constant>(I->getOperand(1))->isNullValue() ||
226              !isa<ConstantInt>(I->getOperand(2)))
227            GS.isNotSuitableForSRA = true;
228        } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(I->getOperand(0))){
229          if (CE->getOpcode() != Instruction::GetElementPtr ||
230              CE->getNumOperands() < 3 || I->getNumOperands() < 2 ||
231              !isa<Constant>(I->getOperand(0)) ||
232              !cast<Constant>(I->getOperand(0))->isNullValue())
233            GS.isNotSuitableForSRA = true;
234        } else {
235          GS.isNotSuitableForSRA = true;
236        }
237      } else if (isa<SelectInst>(I)) {
238        if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
239        GS.isNotSuitableForSRA = true;
240      } else if (PHINode *PN = dyn_cast<PHINode>(I)) {
241        // PHI nodes we can check just like select or GEP instructions, but we
242        // have to be careful about infinite recursion.
243        if (PHIUsers.insert(PN).second)  // Not already visited.
244          if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
245        GS.isNotSuitableForSRA = true;
246        GS.HasPHIUser = true;
247      } else if (isa<CmpInst>(I)) {
248        GS.isNotSuitableForSRA = true;
249      } else if (isa<MemCpyInst>(I) || isa<MemMoveInst>(I)) {
250        if (I->getOperand(1) == V)
251          GS.StoredType = GlobalStatus::isStored;
252        if (I->getOperand(2) == V)
253          GS.isLoaded = true;
254        GS.isNotSuitableForSRA = true;
255      } else if (isa<MemSetInst>(I)) {
256        assert(I->getOperand(1) == V && "Memset only takes one pointer!");
257        GS.StoredType = GlobalStatus::isStored;
258        GS.isNotSuitableForSRA = true;
259      } else {
260        return true;  // Any other non-load instruction might take address!
261      }
262    } else if (Constant *C = dyn_cast<Constant>(*UI)) {
263      GS.HasNonInstructionUser = true;
264      // We might have a dead and dangling constant hanging off of here.
265      if (!ConstantIsDead(C))
266        return true;
267    } else {
268      GS.HasNonInstructionUser = true;
269      // Otherwise must be some other user.
270      return true;
271    }
272
273  return false;
274}
275
276static Constant *getAggregateConstantElement(Constant *Agg, Constant *Idx) {
277  ConstantInt *CI = dyn_cast<ConstantInt>(Idx);
278  if (!CI) return 0;
279  unsigned IdxV = CI->getZExtValue();
280
281  if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg)) {
282    if (IdxV < CS->getNumOperands()) return CS->getOperand(IdxV);
283  } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg)) {
284    if (IdxV < CA->getNumOperands()) return CA->getOperand(IdxV);
285  } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Agg)) {
286    if (IdxV < CP->getNumOperands()) return CP->getOperand(IdxV);
287  } else if (isa<ConstantAggregateZero>(Agg)) {
288    if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
289      if (IdxV < STy->getNumElements())
290        return Constant::getNullValue(STy->getElementType(IdxV));
291    } else if (const SequentialType *STy =
292               dyn_cast<SequentialType>(Agg->getType())) {
293      return Constant::getNullValue(STy->getElementType());
294    }
295  } else if (isa<UndefValue>(Agg)) {
296    if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
297      if (IdxV < STy->getNumElements())
298        return UndefValue::get(STy->getElementType(IdxV));
299    } else if (const SequentialType *STy =
300               dyn_cast<SequentialType>(Agg->getType())) {
301      return UndefValue::get(STy->getElementType());
302    }
303  }
304  return 0;
305}
306
307
308/// CleanupConstantGlobalUsers - We just marked GV constant.  Loop over all
309/// users of the global, cleaning up the obvious ones.  This is largely just a
310/// quick scan over the use list to clean up the easy and obvious cruft.  This
311/// returns true if it made a change.
312static bool CleanupConstantGlobalUsers(Value *V, Constant *Init) {
313  bool Changed = false;
314  for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;) {
315    User *U = *UI++;
316
317    if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
318      if (Init) {
319        // Replace the load with the initializer.
320        LI->replaceAllUsesWith(Init);
321        LI->eraseFromParent();
322        Changed = true;
323      }
324    } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
325      // Store must be unreachable or storing Init into the global.
326      SI->eraseFromParent();
327      Changed = true;
328    } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
329      if (CE->getOpcode() == Instruction::GetElementPtr) {
330        Constant *SubInit = 0;
331        if (Init)
332          SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
333        Changed |= CleanupConstantGlobalUsers(CE, SubInit);
334      } else if (CE->getOpcode() == Instruction::BitCast &&
335                 isa<PointerType>(CE->getType())) {
336        // Pointer cast, delete any stores and memsets to the global.
337        Changed |= CleanupConstantGlobalUsers(CE, 0);
338      }
339
340      if (CE->use_empty()) {
341        CE->destroyConstant();
342        Changed = true;
343      }
344    } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
345      // Do not transform "gepinst (gep constexpr (GV))" here, because forming
346      // "gepconstexpr (gep constexpr (GV))" will cause the two gep's to fold
347      // and will invalidate our notion of what Init is.
348      Constant *SubInit = 0;
349      if (!isa<ConstantExpr>(GEP->getOperand(0))) {
350        ConstantExpr *CE =
351          dyn_cast_or_null<ConstantExpr>(ConstantFoldInstruction(GEP));
352        if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr)
353          SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
354      }
355      Changed |= CleanupConstantGlobalUsers(GEP, SubInit);
356
357      if (GEP->use_empty()) {
358        GEP->eraseFromParent();
359        Changed = true;
360      }
361    } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv
362      if (MI->getRawDest() == V) {
363        MI->eraseFromParent();
364        Changed = true;
365      }
366
367    } else if (Constant *C = dyn_cast<Constant>(U)) {
368      // If we have a chain of dead constantexprs or other things dangling from
369      // us, and if they are all dead, nuke them without remorse.
370      if (ConstantIsDead(C)) {
371        C->destroyConstant();
372        // This could have invalidated UI, start over from scratch.
373        CleanupConstantGlobalUsers(V, Init);
374        return true;
375      }
376    }
377  }
378  return Changed;
379}
380
381/// SRAGlobal - Perform scalar replacement of aggregates on the specified global
382/// variable.  This opens the door for other optimizations by exposing the
383/// behavior of the program in a more fine-grained way.  We have determined that
384/// this transformation is safe already.  We return the first global variable we
385/// insert so that the caller can reprocess it.
386static GlobalVariable *SRAGlobal(GlobalVariable *GV) {
387  assert(GV->hasInternalLinkage() && !GV->isConstant());
388  Constant *Init = GV->getInitializer();
389  const Type *Ty = Init->getType();
390
391  std::vector<GlobalVariable*> NewGlobals;
392  Module::GlobalListType &Globals = GV->getParent()->getGlobalList();
393
394  if (const StructType *STy = dyn_cast<StructType>(Ty)) {
395    NewGlobals.reserve(STy->getNumElements());
396    for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
397      Constant *In = getAggregateConstantElement(Init,
398                                            ConstantInt::get(Type::Int32Ty, i));
399      assert(In && "Couldn't get element of initializer?");
400      GlobalVariable *NGV = new GlobalVariable(STy->getElementType(i), false,
401                                               GlobalVariable::InternalLinkage,
402                                               In, GV->getName()+"."+utostr(i),
403                                               (Module *)NULL,
404                                               GV->isThreadLocal());
405      Globals.insert(GV, NGV);
406      NewGlobals.push_back(NGV);
407    }
408  } else if (const SequentialType *STy = dyn_cast<SequentialType>(Ty)) {
409    unsigned NumElements = 0;
410    if (const ArrayType *ATy = dyn_cast<ArrayType>(STy))
411      NumElements = ATy->getNumElements();
412    else if (const VectorType *PTy = dyn_cast<VectorType>(STy))
413      NumElements = PTy->getNumElements();
414    else
415      assert(0 && "Unknown aggregate sequential type!");
416
417    if (NumElements > 16 && GV->hasNUsesOrMore(16))
418      return 0; // It's not worth it.
419    NewGlobals.reserve(NumElements);
420    for (unsigned i = 0, e = NumElements; i != e; ++i) {
421      Constant *In = getAggregateConstantElement(Init,
422                                            ConstantInt::get(Type::Int32Ty, i));
423      assert(In && "Couldn't get element of initializer?");
424
425      GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false,
426                                               GlobalVariable::InternalLinkage,
427                                               In, GV->getName()+"."+utostr(i),
428                                               (Module *)NULL,
429                                               GV->isThreadLocal());
430      Globals.insert(GV, NGV);
431      NewGlobals.push_back(NGV);
432    }
433  }
434
435  if (NewGlobals.empty())
436    return 0;
437
438  DOUT << "PERFORMING GLOBAL SRA ON: " << *GV;
439
440  Constant *NullInt = Constant::getNullValue(Type::Int32Ty);
441
442  // Loop over all of the uses of the global, replacing the constantexpr geps,
443  // with smaller constantexpr geps or direct references.
444  while (!GV->use_empty()) {
445    User *GEP = GV->use_back();
446    assert(((isa<ConstantExpr>(GEP) &&
447             cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)||
448            isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!");
449
450    // Ignore the 1th operand, which has to be zero or else the program is quite
451    // broken (undefined).  Get the 2nd operand, which is the structure or array
452    // index.
453    unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
454    if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access.
455
456    Value *NewPtr = NewGlobals[Val];
457
458    // Form a shorter GEP if needed.
459    if (GEP->getNumOperands() > 3)
460      if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) {
461        SmallVector<Constant*, 8> Idxs;
462        Idxs.push_back(NullInt);
463        for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i)
464          Idxs.push_back(CE->getOperand(i));
465        NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr),
466                                                &Idxs[0], Idxs.size());
467      } else {
468        GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP);
469        SmallVector<Value*, 8> Idxs;
470        Idxs.push_back(NullInt);
471        for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i)
472          Idxs.push_back(GEPI->getOperand(i));
473        NewPtr = new GetElementPtrInst(NewPtr, Idxs.begin(), Idxs.end(),
474                                       GEPI->getName()+"."+utostr(Val), GEPI);
475      }
476    GEP->replaceAllUsesWith(NewPtr);
477
478    if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP))
479      GEPI->eraseFromParent();
480    else
481      cast<ConstantExpr>(GEP)->destroyConstant();
482  }
483
484  // Delete the old global, now that it is dead.
485  Globals.erase(GV);
486  ++NumSRA;
487
488  // Loop over the new globals array deleting any globals that are obviously
489  // dead.  This can arise due to scalarization of a structure or an array that
490  // has elements that are dead.
491  unsigned FirstGlobal = 0;
492  for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i)
493    if (NewGlobals[i]->use_empty()) {
494      Globals.erase(NewGlobals[i]);
495      if (FirstGlobal == i) ++FirstGlobal;
496    }
497
498  return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : 0;
499}
500
501/// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified
502/// value will trap if the value is dynamically null.  PHIs keeps track of any
503/// phi nodes we've seen to avoid reprocessing them.
504static bool AllUsesOfValueWillTrapIfNull(Value *V,
505                                         SmallPtrSet<PHINode*, 8> &PHIs) {
506  for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
507    if (isa<LoadInst>(*UI)) {
508      // Will trap.
509    } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
510      if (SI->getOperand(0) == V) {
511        //cerr << "NONTRAPPING USE: " << **UI;
512        return false;  // Storing the value.
513      }
514    } else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
515      if (CI->getOperand(0) != V) {
516        //cerr << "NONTRAPPING USE: " << **UI;
517        return false;  // Not calling the ptr
518      }
519    } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
520      if (II->getOperand(0) != V) {
521        //cerr << "NONTRAPPING USE: " << **UI;
522        return false;  // Not calling the ptr
523      }
524    } else if (BitCastInst *CI = dyn_cast<BitCastInst>(*UI)) {
525      if (!AllUsesOfValueWillTrapIfNull(CI, PHIs)) return false;
526    } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) {
527      if (!AllUsesOfValueWillTrapIfNull(GEPI, PHIs)) return false;
528    } else if (PHINode *PN = dyn_cast<PHINode>(*UI)) {
529      // If we've already seen this phi node, ignore it, it has already been
530      // checked.
531      if (PHIs.insert(PN))
532        return AllUsesOfValueWillTrapIfNull(PN, PHIs);
533    } else if (isa<ICmpInst>(*UI) &&
534               isa<ConstantPointerNull>(UI->getOperand(1))) {
535      // Ignore setcc X, null
536    } else {
537      //cerr << "NONTRAPPING USE: " << **UI;
538      return false;
539    }
540  return true;
541}
542
543/// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads
544/// from GV will trap if the loaded value is null.  Note that this also permits
545/// comparisons of the loaded value against null, as a special case.
546static bool AllUsesOfLoadedValueWillTrapIfNull(GlobalVariable *GV) {
547  for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI!=E; ++UI)
548    if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
549      SmallPtrSet<PHINode*, 8> PHIs;
550      if (!AllUsesOfValueWillTrapIfNull(LI, PHIs))
551        return false;
552    } else if (isa<StoreInst>(*UI)) {
553      // Ignore stores to the global.
554    } else {
555      // We don't know or understand this user, bail out.
556      //cerr << "UNKNOWN USER OF GLOBAL!: " << **UI;
557      return false;
558    }
559
560  return true;
561}
562
563static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) {
564  bool Changed = false;
565  for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) {
566    Instruction *I = cast<Instruction>(*UI++);
567    if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
568      LI->setOperand(0, NewV);
569      Changed = true;
570    } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
571      if (SI->getOperand(1) == V) {
572        SI->setOperand(1, NewV);
573        Changed = true;
574      }
575    } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
576      if (I->getOperand(0) == V) {
577        // Calling through the pointer!  Turn into a direct call, but be careful
578        // that the pointer is not also being passed as an argument.
579        I->setOperand(0, NewV);
580        Changed = true;
581        bool PassedAsArg = false;
582        for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i)
583          if (I->getOperand(i) == V) {
584            PassedAsArg = true;
585            I->setOperand(i, NewV);
586          }
587
588        if (PassedAsArg) {
589          // Being passed as an argument also.  Be careful to not invalidate UI!
590          UI = V->use_begin();
591        }
592      }
593    } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
594      Changed |= OptimizeAwayTrappingUsesOfValue(CI,
595                                ConstantExpr::getCast(CI->getOpcode(),
596                                                      NewV, CI->getType()));
597      if (CI->use_empty()) {
598        Changed = true;
599        CI->eraseFromParent();
600      }
601    } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
602      // Should handle GEP here.
603      SmallVector<Constant*, 8> Idxs;
604      Idxs.reserve(GEPI->getNumOperands()-1);
605      for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
606        if (Constant *C = dyn_cast<Constant>(GEPI->getOperand(i)))
607          Idxs.push_back(C);
608        else
609          break;
610      if (Idxs.size() == GEPI->getNumOperands()-1)
611        Changed |= OptimizeAwayTrappingUsesOfValue(GEPI,
612                                ConstantExpr::getGetElementPtr(NewV, &Idxs[0],
613                                                               Idxs.size()));
614      if (GEPI->use_empty()) {
615        Changed = true;
616        GEPI->eraseFromParent();
617      }
618    }
619  }
620
621  return Changed;
622}
623
624
625/// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null
626/// value stored into it.  If there are uses of the loaded value that would trap
627/// if the loaded value is dynamically null, then we know that they cannot be
628/// reachable with a null optimize away the load.
629static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV) {
630  std::vector<LoadInst*> Loads;
631  bool Changed = false;
632
633  // Replace all uses of loads with uses of uses of the stored value.
634  for (Value::use_iterator GUI = GV->use_begin(), E = GV->use_end();
635       GUI != E; ++GUI)
636    if (LoadInst *LI = dyn_cast<LoadInst>(*GUI)) {
637      Loads.push_back(LI);
638      Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV);
639    } else {
640      // If we get here we could have stores, selects, or phi nodes whose values
641      // are loaded.
642      assert((isa<StoreInst>(*GUI) || isa<PHINode>(*GUI) ||
643              isa<SelectInst>(*GUI)) &&
644             "Only expect load and stores!");
645    }
646
647  if (Changed) {
648    DOUT << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV;
649    ++NumGlobUses;
650  }
651
652  // Delete all of the loads we can, keeping track of whether we nuked them all!
653  bool AllLoadsGone = true;
654  while (!Loads.empty()) {
655    LoadInst *L = Loads.back();
656    if (L->use_empty()) {
657      L->eraseFromParent();
658      Changed = true;
659    } else {
660      AllLoadsGone = false;
661    }
662    Loads.pop_back();
663  }
664
665  // If we nuked all of the loads, then none of the stores are needed either,
666  // nor is the global.
667  if (AllLoadsGone) {
668    DOUT << "  *** GLOBAL NOW DEAD!\n";
669    CleanupConstantGlobalUsers(GV, 0);
670    if (GV->use_empty()) {
671      GV->eraseFromParent();
672      ++NumDeleted;
673    }
674    Changed = true;
675  }
676  return Changed;
677}
678
679/// ConstantPropUsersOf - Walk the use list of V, constant folding all of the
680/// instructions that are foldable.
681static void ConstantPropUsersOf(Value *V) {
682  for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; )
683    if (Instruction *I = dyn_cast<Instruction>(*UI++))
684      if (Constant *NewC = ConstantFoldInstruction(I)) {
685        I->replaceAllUsesWith(NewC);
686
687        // Advance UI to the next non-I use to avoid invalidating it!
688        // Instructions could multiply use V.
689        while (UI != E && *UI == I)
690          ++UI;
691        I->eraseFromParent();
692      }
693}
694
695/// OptimizeGlobalAddressOfMalloc - This function takes the specified global
696/// variable, and transforms the program as if it always contained the result of
697/// the specified malloc.  Because it is always the result of the specified
698/// malloc, there is no reason to actually DO the malloc.  Instead, turn the
699/// malloc into a global, and any loads of GV as uses of the new global.
700static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV,
701                                                     MallocInst *MI) {
702  DOUT << "PROMOTING MALLOC GLOBAL: " << *GV << "  MALLOC = " << *MI;
703  ConstantInt *NElements = cast<ConstantInt>(MI->getArraySize());
704
705  if (NElements->getZExtValue() != 1) {
706    // If we have an array allocation, transform it to a single element
707    // allocation to make the code below simpler.
708    Type *NewTy = ArrayType::get(MI->getAllocatedType(),
709                                 NElements->getZExtValue());
710    MallocInst *NewMI =
711      new MallocInst(NewTy, Constant::getNullValue(Type::Int32Ty),
712                     MI->getAlignment(), MI->getName(), MI);
713    Value* Indices[2];
714    Indices[0] = Indices[1] = Constant::getNullValue(Type::Int32Ty);
715    Value *NewGEP = new GetElementPtrInst(NewMI, Indices, Indices + 2,
716                                          NewMI->getName()+".el0", MI);
717    MI->replaceAllUsesWith(NewGEP);
718    MI->eraseFromParent();
719    MI = NewMI;
720  }
721
722  // Create the new global variable.  The contents of the malloc'd memory is
723  // undefined, so initialize with an undef value.
724  Constant *Init = UndefValue::get(MI->getAllocatedType());
725  GlobalVariable *NewGV = new GlobalVariable(MI->getAllocatedType(), false,
726                                             GlobalValue::InternalLinkage, Init,
727                                             GV->getName()+".body",
728                                             (Module *)NULL,
729                                             GV->isThreadLocal());
730  GV->getParent()->getGlobalList().insert(GV, NewGV);
731
732  // Anything that used the malloc now uses the global directly.
733  MI->replaceAllUsesWith(NewGV);
734
735  Constant *RepValue = NewGV;
736  if (NewGV->getType() != GV->getType()->getElementType())
737    RepValue = ConstantExpr::getBitCast(RepValue,
738                                        GV->getType()->getElementType());
739
740  // If there is a comparison against null, we will insert a global bool to
741  // keep track of whether the global was initialized yet or not.
742  GlobalVariable *InitBool =
743    new GlobalVariable(Type::Int1Ty, false, GlobalValue::InternalLinkage,
744                       ConstantInt::getFalse(), GV->getName()+".init",
745                       (Module *)NULL, GV->isThreadLocal());
746  bool InitBoolUsed = false;
747
748  // Loop over all uses of GV, processing them in turn.
749  std::vector<StoreInst*> Stores;
750  while (!GV->use_empty())
751    if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) {
752      while (!LI->use_empty()) {
753        Use &LoadUse = LI->use_begin().getUse();
754        if (!isa<ICmpInst>(LoadUse.getUser()))
755          LoadUse = RepValue;
756        else {
757          ICmpInst *CI = cast<ICmpInst>(LoadUse.getUser());
758          // Replace the cmp X, 0 with a use of the bool value.
759          Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", CI);
760          InitBoolUsed = true;
761          switch (CI->getPredicate()) {
762          default: assert(0 && "Unknown ICmp Predicate!");
763          case ICmpInst::ICMP_ULT:
764          case ICmpInst::ICMP_SLT:
765            LV = ConstantInt::getFalse();   // X < null -> always false
766            break;
767          case ICmpInst::ICMP_ULE:
768          case ICmpInst::ICMP_SLE:
769          case ICmpInst::ICMP_EQ:
770            LV = BinaryOperator::createNot(LV, "notinit", CI);
771            break;
772          case ICmpInst::ICMP_NE:
773          case ICmpInst::ICMP_UGE:
774          case ICmpInst::ICMP_SGE:
775          case ICmpInst::ICMP_UGT:
776          case ICmpInst::ICMP_SGT:
777            break;  // no change.
778          }
779          CI->replaceAllUsesWith(LV);
780          CI->eraseFromParent();
781        }
782      }
783      LI->eraseFromParent();
784    } else {
785      StoreInst *SI = cast<StoreInst>(GV->use_back());
786      // The global is initialized when the store to it occurs.
787      new StoreInst(ConstantInt::getTrue(), InitBool, SI);
788      SI->eraseFromParent();
789    }
790
791  // If the initialization boolean was used, insert it, otherwise delete it.
792  if (!InitBoolUsed) {
793    while (!InitBool->use_empty())  // Delete initializations
794      cast<Instruction>(InitBool->use_back())->eraseFromParent();
795    delete InitBool;
796  } else
797    GV->getParent()->getGlobalList().insert(GV, InitBool);
798
799
800  // Now the GV is dead, nuke it and the malloc.
801  GV->eraseFromParent();
802  MI->eraseFromParent();
803
804  // To further other optimizations, loop over all users of NewGV and try to
805  // constant prop them.  This will promote GEP instructions with constant
806  // indices into GEP constant-exprs, which will allow global-opt to hack on it.
807  ConstantPropUsersOf(NewGV);
808  if (RepValue != NewGV)
809    ConstantPropUsersOf(RepValue);
810
811  return NewGV;
812}
813
814/// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking
815/// to make sure that there are no complex uses of V.  We permit simple things
816/// like dereferencing the pointer, but not storing through the address, unless
817/// it is to the specified global.
818static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Instruction *V,
819                                                      GlobalVariable *GV,
820                                              SmallPtrSet<PHINode*, 8> &PHIs) {
821  for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
822    if (isa<LoadInst>(*UI) || isa<CmpInst>(*UI)) {
823      // Fine, ignore.
824    } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
825      if (SI->getOperand(0) == V && SI->getOperand(1) != GV)
826        return false;  // Storing the pointer itself... bad.
827      // Otherwise, storing through it, or storing into GV... fine.
828    } else if (isa<GetElementPtrInst>(*UI)) {
829      if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(cast<Instruction>(*UI),
830                                                     GV, PHIs))
831        return false;
832    } else if (PHINode *PN = dyn_cast<PHINode>(*UI)) {
833      // PHIs are ok if all uses are ok.  Don't infinitely recurse through PHI
834      // cycles.
835      if (PHIs.insert(PN))
836        if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(PN, GV, PHIs))
837          return false;
838    } else {
839      return false;
840    }
841  return true;
842}
843
844/// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV
845/// somewhere.  Transform all uses of the allocation into loads from the
846/// global and uses of the resultant pointer.  Further, delete the store into
847/// GV.  This assumes that these value pass the
848/// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate.
849static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc,
850                                          GlobalVariable *GV) {
851  while (!Alloc->use_empty()) {
852    Instruction *U = cast<Instruction>(*Alloc->use_begin());
853    Instruction *InsertPt = U;
854    if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
855      // If this is the store of the allocation into the global, remove it.
856      if (SI->getOperand(1) == GV) {
857        SI->eraseFromParent();
858        continue;
859      }
860    } else if (PHINode *PN = dyn_cast<PHINode>(U)) {
861      // Insert the load in the corresponding predecessor, not right before the
862      // PHI.
863      unsigned PredNo = Alloc->use_begin().getOperandNo()/2;
864      InsertPt = PN->getIncomingBlock(PredNo)->getTerminator();
865    }
866
867    // Insert a load from the global, and use it instead of the malloc.
868    Value *NL = new LoadInst(GV, GV->getName()+".val", InsertPt);
869    U->replaceUsesOfWith(Alloc, NL);
870  }
871}
872
873/// GlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from
874/// GV are simple enough to perform HeapSRA, return true.
875static bool GlobalLoadUsesSimpleEnoughForHeapSRA(GlobalVariable *GV,
876                                                 MallocInst *MI) {
877  for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;
878       ++UI)
879    if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
880      // We permit two users of the load: setcc comparing against the null
881      // pointer, and a getelementptr of a specific form.
882      for (Value::use_iterator UI = LI->use_begin(), E = LI->use_end(); UI != E;
883           ++UI) {
884        // Comparison against null is ok.
885        if (ICmpInst *ICI = dyn_cast<ICmpInst>(*UI)) {
886          if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
887            return false;
888          continue;
889        }
890
891        // getelementptr is also ok, but only a simple form.
892        if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) {
893          // Must index into the array and into the struct.
894          if (GEPI->getNumOperands() < 3)
895            return false;
896
897          // Otherwise the GEP is ok.
898          continue;
899        }
900
901        if (PHINode *PN = dyn_cast<PHINode>(*UI)) {
902          // We have a phi of a load from the global.  We can only handle this
903          // if the other PHI'd values are actually the same.  In this case,
904          // the rewriter will just drop the phi entirely.
905          for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
906            Value *IV = PN->getIncomingValue(i);
907            if (IV == LI) continue;  // Trivial the same.
908
909            // If the phi'd value is from the malloc that initializes the value,
910            // we can xform it.
911            if (IV == MI) continue;
912
913            // Otherwise, we don't know what it is.
914            return false;
915          }
916          return true;
917        }
918
919        // Otherwise we don't know what this is, not ok.
920        return false;
921      }
922    }
923  return true;
924}
925
926/// GetHeapSROALoad - Return the load for the specified field of the HeapSROA'd
927/// value, lazily creating it on demand.
928static Value *GetHeapSROALoad(Instruction *Load, unsigned FieldNo,
929                              const std::vector<GlobalVariable*> &FieldGlobals,
930                              std::vector<Value *> &InsertedLoadsForPtr) {
931  if (InsertedLoadsForPtr.size() <= FieldNo)
932    InsertedLoadsForPtr.resize(FieldNo+1);
933  if (InsertedLoadsForPtr[FieldNo] == 0)
934    InsertedLoadsForPtr[FieldNo] = new LoadInst(FieldGlobals[FieldNo],
935                                                Load->getName()+".f" +
936                                                utostr(FieldNo), Load);
937  return InsertedLoadsForPtr[FieldNo];
938}
939
940/// RewriteHeapSROALoadUser - Given a load instruction and a value derived from
941/// the load, rewrite the derived value to use the HeapSRoA'd load.
942static void RewriteHeapSROALoadUser(LoadInst *Load, Instruction *LoadUser,
943                               const std::vector<GlobalVariable*> &FieldGlobals,
944                                    std::vector<Value *> &InsertedLoadsForPtr) {
945  // If this is a comparison against null, handle it.
946  if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) {
947    assert(isa<ConstantPointerNull>(SCI->getOperand(1)));
948    // If we have a setcc of the loaded pointer, we can use a setcc of any
949    // field.
950    Value *NPtr;
951    if (InsertedLoadsForPtr.empty()) {
952      NPtr = GetHeapSROALoad(Load, 0, FieldGlobals, InsertedLoadsForPtr);
953    } else {
954      NPtr = InsertedLoadsForPtr.back();
955    }
956
957    Value *New = new ICmpInst(SCI->getPredicate(), NPtr,
958                              Constant::getNullValue(NPtr->getType()),
959                              SCI->getName(), SCI);
960    SCI->replaceAllUsesWith(New);
961    SCI->eraseFromParent();
962    return;
963  }
964
965  // Handle 'getelementptr Ptr, Idx, uint FieldNo ...'
966  if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) {
967    assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2))
968           && "Unexpected GEPI!");
969
970    // Load the pointer for this field.
971    unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
972    Value *NewPtr = GetHeapSROALoad(Load, FieldNo,
973                                    FieldGlobals, InsertedLoadsForPtr);
974
975    // Create the new GEP idx vector.
976    SmallVector<Value*, 8> GEPIdx;
977    GEPIdx.push_back(GEPI->getOperand(1));
978    GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end());
979
980    Value *NGEPI = new GetElementPtrInst(NewPtr, GEPIdx.begin(), GEPIdx.end(),
981                                         GEPI->getName(), GEPI);
982    GEPI->replaceAllUsesWith(NGEPI);
983    GEPI->eraseFromParent();
984    return;
985  }
986
987  // Handle PHI nodes.  PHI nodes must be merging in the same values, plus
988  // potentially the original malloc.  Insert phi nodes for each field, then
989  // process uses of the PHI.
990  PHINode *PN = cast<PHINode>(LoadUser);
991  std::vector<Value *> PHIsForField;
992  PHIsForField.resize(FieldGlobals.size());
993  for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
994    Value *LoadV = GetHeapSROALoad(Load, i, FieldGlobals, InsertedLoadsForPtr);
995
996    PHINode *FieldPN = new PHINode(LoadV->getType(),
997                                   PN->getName()+"."+utostr(i), PN);
998    // Fill in the predecessor values.
999    for (unsigned pred = 0, e = PN->getNumIncomingValues(); pred != e; ++pred) {
1000      // Each predecessor either uses the load or the original malloc.
1001      Value *InVal = PN->getIncomingValue(pred);
1002      BasicBlock *BB = PN->getIncomingBlock(pred);
1003      Value *NewVal;
1004      if (isa<MallocInst>(InVal)) {
1005        // Insert a reload from the global in the predecessor.
1006        NewVal = GetHeapSROALoad(BB->getTerminator(), i, FieldGlobals,
1007                                 PHIsForField);
1008      } else {
1009        NewVal = InsertedLoadsForPtr[i];
1010      }
1011      FieldPN->addIncoming(NewVal, BB);
1012    }
1013    PHIsForField[i] = FieldPN;
1014  }
1015
1016  // Since PHIsForField specifies a phi for every input value, the lazy inserter
1017  // will never insert a load.
1018  while (!PN->use_empty())
1019    RewriteHeapSROALoadUser(Load, PN->use_back(), FieldGlobals, PHIsForField);
1020  PN->eraseFromParent();
1021}
1022
1023/// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global.  Ptr
1024/// is a value loaded from the global.  Eliminate all uses of Ptr, making them
1025/// use FieldGlobals instead.  All uses of loaded values satisfy
1026/// GlobalLoadUsesSimpleEnoughForHeapSRA.
1027static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load,
1028                             const std::vector<GlobalVariable*> &FieldGlobals) {
1029  std::vector<Value *> InsertedLoadsForPtr;
1030  //InsertedLoadsForPtr.resize(FieldGlobals.size());
1031  while (!Load->use_empty())
1032    RewriteHeapSROALoadUser(Load, Load->use_back(),
1033                            FieldGlobals, InsertedLoadsForPtr);
1034}
1035
1036/// PerformHeapAllocSRoA - MI is an allocation of an array of structures.  Break
1037/// it up into multiple allocations of arrays of the fields.
1038static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, MallocInst *MI){
1039  DOUT << "SROA HEAP ALLOC: " << *GV << "  MALLOC = " << *MI;
1040  const StructType *STy = cast<StructType>(MI->getAllocatedType());
1041
1042  // There is guaranteed to be at least one use of the malloc (storing
1043  // it into GV).  If there are other uses, change them to be uses of
1044  // the global to simplify later code.  This also deletes the store
1045  // into GV.
1046  ReplaceUsesOfMallocWithGlobal(MI, GV);
1047
1048  // Okay, at this point, there are no users of the malloc.  Insert N
1049  // new mallocs at the same place as MI, and N globals.
1050  std::vector<GlobalVariable*> FieldGlobals;
1051  std::vector<MallocInst*> FieldMallocs;
1052
1053  for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){
1054    const Type *FieldTy = STy->getElementType(FieldNo);
1055    const Type *PFieldTy = PointerType::getUnqual(FieldTy);
1056
1057    GlobalVariable *NGV =
1058      new GlobalVariable(PFieldTy, false, GlobalValue::InternalLinkage,
1059                         Constant::getNullValue(PFieldTy),
1060                         GV->getName() + ".f" + utostr(FieldNo), GV,
1061                         GV->isThreadLocal());
1062    FieldGlobals.push_back(NGV);
1063
1064    MallocInst *NMI = new MallocInst(FieldTy, MI->getArraySize(),
1065                                     MI->getName() + ".f" + utostr(FieldNo),MI);
1066    FieldMallocs.push_back(NMI);
1067    new StoreInst(NMI, NGV, MI);
1068  }
1069
1070  // The tricky aspect of this transformation is handling the case when malloc
1071  // fails.  In the original code, malloc failing would set the result pointer
1072  // of malloc to null.  In this case, some mallocs could succeed and others
1073  // could fail.  As such, we emit code that looks like this:
1074  //    F0 = malloc(field0)
1075  //    F1 = malloc(field1)
1076  //    F2 = malloc(field2)
1077  //    if (F0 == 0 || F1 == 0 || F2 == 0) {
1078  //      if (F0) { free(F0); F0 = 0; }
1079  //      if (F1) { free(F1); F1 = 0; }
1080  //      if (F2) { free(F2); F2 = 0; }
1081  //    }
1082  Value *RunningOr = 0;
1083  for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) {
1084    Value *Cond = new ICmpInst(ICmpInst::ICMP_EQ, FieldMallocs[i],
1085                             Constant::getNullValue(FieldMallocs[i]->getType()),
1086                                  "isnull", MI);
1087    if (!RunningOr)
1088      RunningOr = Cond;   // First seteq
1089    else
1090      RunningOr = BinaryOperator::createOr(RunningOr, Cond, "tmp", MI);
1091  }
1092
1093  // Split the basic block at the old malloc.
1094  BasicBlock *OrigBB = MI->getParent();
1095  BasicBlock *ContBB = OrigBB->splitBasicBlock(MI, "malloc_cont");
1096
1097  // Create the block to check the first condition.  Put all these blocks at the
1098  // end of the function as they are unlikely to be executed.
1099  BasicBlock *NullPtrBlock = new BasicBlock("malloc_ret_null",
1100                                            OrigBB->getParent());
1101
1102  // Remove the uncond branch from OrigBB to ContBB, turning it into a cond
1103  // branch on RunningOr.
1104  OrigBB->getTerminator()->eraseFromParent();
1105  new BranchInst(NullPtrBlock, ContBB, RunningOr, OrigBB);
1106
1107  // Within the NullPtrBlock, we need to emit a comparison and branch for each
1108  // pointer, because some may be null while others are not.
1109  for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1110    Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock);
1111    Value *Cmp = new ICmpInst(ICmpInst::ICMP_NE, GVVal,
1112                              Constant::getNullValue(GVVal->getType()),
1113                              "tmp", NullPtrBlock);
1114    BasicBlock *FreeBlock = new BasicBlock("free_it", OrigBB->getParent());
1115    BasicBlock *NextBlock = new BasicBlock("next", OrigBB->getParent());
1116    new BranchInst(FreeBlock, NextBlock, Cmp, NullPtrBlock);
1117
1118    // Fill in FreeBlock.
1119    new FreeInst(GVVal, FreeBlock);
1120    new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i],
1121                  FreeBlock);
1122    new BranchInst(NextBlock, FreeBlock);
1123
1124    NullPtrBlock = NextBlock;
1125  }
1126
1127  new BranchInst(ContBB, NullPtrBlock);
1128
1129
1130  // MI is no longer needed, remove it.
1131  MI->eraseFromParent();
1132
1133
1134  // Okay, the malloc site is completely handled.  All of the uses of GV are now
1135  // loads, and all uses of those loads are simple.  Rewrite them to use loads
1136  // of the per-field globals instead.
1137  while (!GV->use_empty()) {
1138    if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) {
1139      RewriteUsesOfLoadForHeapSRoA(LI, FieldGlobals);
1140      LI->eraseFromParent();
1141    } else {
1142      // Must be a store of null.
1143      StoreInst *SI = cast<StoreInst>(GV->use_back());
1144      assert(isa<Constant>(SI->getOperand(0)) &&
1145             cast<Constant>(SI->getOperand(0))->isNullValue() &&
1146             "Unexpected heap-sra user!");
1147
1148      // Insert a store of null into each global.
1149      for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1150        Constant *Null =
1151          Constant::getNullValue(FieldGlobals[i]->getType()->getElementType());
1152        new StoreInst(Null, FieldGlobals[i], SI);
1153      }
1154      // Erase the original store.
1155      SI->eraseFromParent();
1156    }
1157  }
1158
1159  // The old global is now dead, remove it.
1160  GV->eraseFromParent();
1161
1162  ++NumHeapSRA;
1163  return FieldGlobals[0];
1164}
1165
1166
1167// OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge
1168// that only one value (besides its initializer) is ever stored to the global.
1169static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal,
1170                                     Module::global_iterator &GVI,
1171                                     TargetData &TD) {
1172  if (CastInst *CI = dyn_cast<CastInst>(StoredOnceVal))
1173    StoredOnceVal = CI->getOperand(0);
1174  else if (GetElementPtrInst *GEPI =dyn_cast<GetElementPtrInst>(StoredOnceVal)){
1175    // "getelementptr Ptr, 0, 0, 0" is really just a cast.
1176    bool IsJustACast = true;
1177    for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
1178      if (!isa<Constant>(GEPI->getOperand(i)) ||
1179          !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
1180        IsJustACast = false;
1181        break;
1182      }
1183    if (IsJustACast)
1184      StoredOnceVal = GEPI->getOperand(0);
1185  }
1186
1187  // If we are dealing with a pointer global that is initialized to null and
1188  // only has one (non-null) value stored into it, then we can optimize any
1189  // users of the loaded value (often calls and loads) that would trap if the
1190  // value was null.
1191  if (isa<PointerType>(GV->getInitializer()->getType()) &&
1192      GV->getInitializer()->isNullValue()) {
1193    if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) {
1194      if (GV->getInitializer()->getType() != SOVC->getType())
1195        SOVC = ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType());
1196
1197      // Optimize away any trapping uses of the loaded value.
1198      if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC))
1199        return true;
1200    } else if (MallocInst *MI = dyn_cast<MallocInst>(StoredOnceVal)) {
1201      // If this is a malloc of an abstract type, don't touch it.
1202      if (!MI->getAllocatedType()->isSized())
1203        return false;
1204
1205      // We can't optimize this global unless all uses of it are *known* to be
1206      // of the malloc value, not of the null initializer value (consider a use
1207      // that compares the global's value against zero to see if the malloc has
1208      // been reached).  To do this, we check to see if all uses of the global
1209      // would trap if the global were null: this proves that they must all
1210      // happen after the malloc.
1211      if (!AllUsesOfLoadedValueWillTrapIfNull(GV))
1212        return false;
1213
1214      // We can't optimize this if the malloc itself is used in a complex way,
1215      // for example, being stored into multiple globals.  This allows the
1216      // malloc to be stored into the specified global, loaded setcc'd, and
1217      // GEP'd.  These are all things we could transform to using the global
1218      // for.
1219      {
1220        SmallPtrSet<PHINode*, 8> PHIs;
1221        if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(MI, GV, PHIs))
1222          return false;
1223      }
1224
1225
1226      // If we have a global that is only initialized with a fixed size malloc,
1227      // transform the program to use global memory instead of malloc'd memory.
1228      // This eliminates dynamic allocation, avoids an indirection accessing the
1229      // data, and exposes the resultant global to further GlobalOpt.
1230      if (ConstantInt *NElements = dyn_cast<ConstantInt>(MI->getArraySize())) {
1231        // Restrict this transformation to only working on small allocations
1232        // (2048 bytes currently), as we don't want to introduce a 16M global or
1233        // something.
1234        if (NElements->getZExtValue()*
1235                     TD.getABITypeSize(MI->getAllocatedType()) < 2048) {
1236          GVI = OptimizeGlobalAddressOfMalloc(GV, MI);
1237          return true;
1238        }
1239      }
1240
1241      // If the allocation is an array of structures, consider transforming this
1242      // into multiple malloc'd arrays, one for each field.  This is basically
1243      // SRoA for malloc'd memory.
1244      if (const StructType *AllocTy =
1245                  dyn_cast<StructType>(MI->getAllocatedType())) {
1246        // This the structure has an unreasonable number of fields, leave it
1247        // alone.
1248        if (AllocTy->getNumElements() <= 16 && AllocTy->getNumElements() > 0 &&
1249            GlobalLoadUsesSimpleEnoughForHeapSRA(GV, MI)) {
1250          GVI = PerformHeapAllocSRoA(GV, MI);
1251          return true;
1252        }
1253      }
1254    }
1255  }
1256
1257  return false;
1258}
1259
1260/// ShrinkGlobalToBoolean - At this point, we have learned that the only two
1261/// values ever stored into GV are its initializer and OtherVal.
1262static void ShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) {
1263  // Create the new global, initializing it to false.
1264  GlobalVariable *NewGV = new GlobalVariable(Type::Int1Ty, false,
1265         GlobalValue::InternalLinkage, ConstantInt::getFalse(),
1266                                             GV->getName()+".b",
1267                                             (Module *)NULL,
1268                                             GV->isThreadLocal());
1269  GV->getParent()->getGlobalList().insert(GV, NewGV);
1270
1271  Constant *InitVal = GV->getInitializer();
1272  assert(InitVal->getType() != Type::Int1Ty && "No reason to shrink to bool!");
1273
1274  // If initialized to zero and storing one into the global, we can use a cast
1275  // instead of a select to synthesize the desired value.
1276  bool IsOneZero = false;
1277  if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal))
1278    IsOneZero = InitVal->isNullValue() && CI->isOne();
1279
1280  while (!GV->use_empty()) {
1281    Instruction *UI = cast<Instruction>(GV->use_back());
1282    if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
1283      // Change the store into a boolean store.
1284      bool StoringOther = SI->getOperand(0) == OtherVal;
1285      // Only do this if we weren't storing a loaded value.
1286      Value *StoreVal;
1287      if (StoringOther || SI->getOperand(0) == InitVal)
1288        StoreVal = ConstantInt::get(Type::Int1Ty, StoringOther);
1289      else {
1290        // Otherwise, we are storing a previously loaded copy.  To do this,
1291        // change the copy from copying the original value to just copying the
1292        // bool.
1293        Instruction *StoredVal = cast<Instruction>(SI->getOperand(0));
1294
1295        // If we're already replaced the input, StoredVal will be a cast or
1296        // select instruction.  If not, it will be a load of the original
1297        // global.
1298        if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
1299          assert(LI->getOperand(0) == GV && "Not a copy!");
1300          // Insert a new load, to preserve the saved value.
1301          StoreVal = new LoadInst(NewGV, LI->getName()+".b", LI);
1302        } else {
1303          assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) &&
1304                 "This is not a form that we understand!");
1305          StoreVal = StoredVal->getOperand(0);
1306          assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!");
1307        }
1308      }
1309      new StoreInst(StoreVal, NewGV, SI);
1310    } else if (!UI->use_empty()) {
1311      // Change the load into a load of bool then a select.
1312      LoadInst *LI = cast<LoadInst>(UI);
1313      LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", LI);
1314      Value *NSI;
1315      if (IsOneZero)
1316        NSI = new ZExtInst(NLI, LI->getType(), "", LI);
1317      else
1318        NSI = new SelectInst(NLI, OtherVal, InitVal, "", LI);
1319      NSI->takeName(LI);
1320      LI->replaceAllUsesWith(NSI);
1321    }
1322    UI->eraseFromParent();
1323  }
1324
1325  GV->eraseFromParent();
1326}
1327
1328
1329/// ProcessInternalGlobal - Analyze the specified global variable and optimize
1330/// it if possible.  If we make a change, return true.
1331bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV,
1332                                      Module::global_iterator &GVI) {
1333  std::set<PHINode*> PHIUsers;
1334  GlobalStatus GS;
1335  GV->removeDeadConstantUsers();
1336
1337  if (GV->use_empty()) {
1338    DOUT << "GLOBAL DEAD: " << *GV;
1339    GV->eraseFromParent();
1340    ++NumDeleted;
1341    return true;
1342  }
1343
1344  if (!AnalyzeGlobal(GV, GS, PHIUsers)) {
1345#if 0
1346    cerr << "Global: " << *GV;
1347    cerr << "  isLoaded = " << GS.isLoaded << "\n";
1348    cerr << "  StoredType = ";
1349    switch (GS.StoredType) {
1350    case GlobalStatus::NotStored: cerr << "NEVER STORED\n"; break;
1351    case GlobalStatus::isInitializerStored: cerr << "INIT STORED\n"; break;
1352    case GlobalStatus::isStoredOnce: cerr << "STORED ONCE\n"; break;
1353    case GlobalStatus::isStored: cerr << "stored\n"; break;
1354    }
1355    if (GS.StoredType == GlobalStatus::isStoredOnce && GS.StoredOnceValue)
1356      cerr << "  StoredOnceValue = " << *GS.StoredOnceValue << "\n";
1357    if (GS.AccessingFunction && !GS.HasMultipleAccessingFunctions)
1358      cerr << "  AccessingFunction = " << GS.AccessingFunction->getName()
1359                << "\n";
1360    cerr << "  HasMultipleAccessingFunctions =  "
1361              << GS.HasMultipleAccessingFunctions << "\n";
1362    cerr << "  HasNonInstructionUser = " << GS.HasNonInstructionUser<<"\n";
1363    cerr << "  isNotSuitableForSRA = " << GS.isNotSuitableForSRA << "\n";
1364    cerr << "\n";
1365#endif
1366
1367    // If this is a first class global and has only one accessing function
1368    // and this function is main (which we know is not recursive we can make
1369    // this global a local variable) we replace the global with a local alloca
1370    // in this function.
1371    //
1372    // NOTE: It doesn't make sense to promote non first class types since we
1373    // are just replacing static memory to stack memory.
1374    if (!GS.HasMultipleAccessingFunctions &&
1375        GS.AccessingFunction && !GS.HasNonInstructionUser &&
1376        GV->getType()->getElementType()->isFirstClassType() &&
1377        GS.AccessingFunction->getName() == "main" &&
1378        GS.AccessingFunction->hasExternalLinkage()) {
1379      DOUT << "LOCALIZING GLOBAL: " << *GV;
1380      Instruction* FirstI = GS.AccessingFunction->getEntryBlock().begin();
1381      const Type* ElemTy = GV->getType()->getElementType();
1382      // FIXME: Pass Global's alignment when globals have alignment
1383      AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), FirstI);
1384      if (!isa<UndefValue>(GV->getInitializer()))
1385        new StoreInst(GV->getInitializer(), Alloca, FirstI);
1386
1387      GV->replaceAllUsesWith(Alloca);
1388      GV->eraseFromParent();
1389      ++NumLocalized;
1390      return true;
1391    }
1392
1393    // If the global is never loaded (but may be stored to), it is dead.
1394    // Delete it now.
1395    if (!GS.isLoaded) {
1396      DOUT << "GLOBAL NEVER LOADED: " << *GV;
1397
1398      // Delete any stores we can find to the global.  We may not be able to
1399      // make it completely dead though.
1400      bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer());
1401
1402      // If the global is dead now, delete it.
1403      if (GV->use_empty()) {
1404        GV->eraseFromParent();
1405        ++NumDeleted;
1406        Changed = true;
1407      }
1408      return Changed;
1409
1410    } else if (GS.StoredType <= GlobalStatus::isInitializerStored) {
1411      DOUT << "MARKING CONSTANT: " << *GV;
1412      GV->setConstant(true);
1413
1414      // Clean up any obviously simplifiable users now.
1415      CleanupConstantGlobalUsers(GV, GV->getInitializer());
1416
1417      // If the global is dead now, just nuke it.
1418      if (GV->use_empty()) {
1419        DOUT << "   *** Marking constant allowed us to simplify "
1420             << "all users and delete global!\n";
1421        GV->eraseFromParent();
1422        ++NumDeleted;
1423      }
1424
1425      ++NumMarked;
1426      return true;
1427    } else if (!GS.isNotSuitableForSRA &&
1428               !GV->getInitializer()->getType()->isFirstClassType()) {
1429      if (GlobalVariable *FirstNewGV = SRAGlobal(GV)) {
1430        GVI = FirstNewGV;  // Don't skip the newly produced globals!
1431        return true;
1432      }
1433    } else if (GS.StoredType == GlobalStatus::isStoredOnce) {
1434      // If the initial value for the global was an undef value, and if only
1435      // one other value was stored into it, we can just change the
1436      // initializer to be an undef value, then delete all stores to the
1437      // global.  This allows us to mark it constant.
1438      if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1439        if (isa<UndefValue>(GV->getInitializer())) {
1440          // Change the initial value here.
1441          GV->setInitializer(SOVConstant);
1442
1443          // Clean up any obviously simplifiable users now.
1444          CleanupConstantGlobalUsers(GV, GV->getInitializer());
1445
1446          if (GV->use_empty()) {
1447            DOUT << "   *** Substituting initializer allowed us to "
1448                 << "simplify all users and delete global!\n";
1449            GV->eraseFromParent();
1450            ++NumDeleted;
1451          } else {
1452            GVI = GV;
1453          }
1454          ++NumSubstitute;
1455          return true;
1456        }
1457
1458      // Try to optimize globals based on the knowledge that only one value
1459      // (besides its initializer) is ever stored to the global.
1460      if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI,
1461                                   getAnalysis<TargetData>()))
1462        return true;
1463
1464      // Otherwise, if the global was not a boolean, we can shrink it to be a
1465      // boolean.
1466      if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1467        if (GV->getType()->getElementType() != Type::Int1Ty &&
1468            !GV->getType()->getElementType()->isFloatingPoint() &&
1469            !isa<VectorType>(GV->getType()->getElementType()) &&
1470            !GS.HasPHIUser && !GS.isNotSuitableForSRA) {
1471          DOUT << "   *** SHRINKING TO BOOL: " << *GV;
1472          ShrinkGlobalToBoolean(GV, SOVConstant);
1473          ++NumShrunkToBool;
1474          return true;
1475        }
1476    }
1477  }
1478  return false;
1479}
1480
1481/// OnlyCalledDirectly - Return true if the specified function is only called
1482/// directly.  In other words, its address is never taken.
1483static bool OnlyCalledDirectly(Function *F) {
1484  for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1485    Instruction *User = dyn_cast<Instruction>(*UI);
1486    if (!User) return false;
1487    if (!isa<CallInst>(User) && !isa<InvokeInst>(User)) return false;
1488
1489    // See if the function address is passed as an argument.
1490    for (unsigned i = 1, e = User->getNumOperands(); i != e; ++i)
1491      if (User->getOperand(i) == F) return false;
1492  }
1493  return true;
1494}
1495
1496/// ChangeCalleesToFastCall - Walk all of the direct calls of the specified
1497/// function, changing them to FastCC.
1498static void ChangeCalleesToFastCall(Function *F) {
1499  for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1500    Instruction *User = cast<Instruction>(*UI);
1501    if (CallInst *CI = dyn_cast<CallInst>(User))
1502      CI->setCallingConv(CallingConv::Fast);
1503    else
1504      cast<InvokeInst>(User)->setCallingConv(CallingConv::Fast);
1505  }
1506}
1507
1508bool GlobalOpt::OptimizeFunctions(Module &M) {
1509  bool Changed = false;
1510  // Optimize functions.
1511  for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) {
1512    Function *F = FI++;
1513    F->removeDeadConstantUsers();
1514    if (F->use_empty() && (F->hasInternalLinkage() ||
1515                           F->hasLinkOnceLinkage())) {
1516      M.getFunctionList().erase(F);
1517      Changed = true;
1518      ++NumFnDeleted;
1519    } else if (F->hasInternalLinkage() &&
1520               F->getCallingConv() == CallingConv::C &&  !F->isVarArg() &&
1521               OnlyCalledDirectly(F)) {
1522      // If this function has C calling conventions, is not a varargs
1523      // function, and is only called directly, promote it to use the Fast
1524      // calling convention.
1525      F->setCallingConv(CallingConv::Fast);
1526      ChangeCalleesToFastCall(F);
1527      ++NumFastCallFns;
1528      Changed = true;
1529    }
1530  }
1531  return Changed;
1532}
1533
1534bool GlobalOpt::OptimizeGlobalVars(Module &M) {
1535  bool Changed = false;
1536  for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
1537       GVI != E; ) {
1538    GlobalVariable *GV = GVI++;
1539    if (!GV->isConstant() && GV->hasInternalLinkage() &&
1540        GV->hasInitializer())
1541      Changed |= ProcessInternalGlobal(GV, GVI);
1542  }
1543  return Changed;
1544}
1545
1546/// FindGlobalCtors - Find the llvm.globalctors list, verifying that all
1547/// initializers have an init priority of 65535.
1548GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) {
1549  for (Module::global_iterator I = M.global_begin(), E = M.global_end();
1550       I != E; ++I)
1551    if (I->getName() == "llvm.global_ctors") {
1552      // Found it, verify it's an array of { int, void()* }.
1553      const ArrayType *ATy =dyn_cast<ArrayType>(I->getType()->getElementType());
1554      if (!ATy) return 0;
1555      const StructType *STy = dyn_cast<StructType>(ATy->getElementType());
1556      if (!STy || STy->getNumElements() != 2 ||
1557          STy->getElementType(0) != Type::Int32Ty) return 0;
1558      const PointerType *PFTy = dyn_cast<PointerType>(STy->getElementType(1));
1559      if (!PFTy) return 0;
1560      const FunctionType *FTy = dyn_cast<FunctionType>(PFTy->getElementType());
1561      if (!FTy || FTy->getReturnType() != Type::VoidTy || FTy->isVarArg() ||
1562          FTy->getNumParams() != 0)
1563        return 0;
1564
1565      // Verify that the initializer is simple enough for us to handle.
1566      if (!I->hasInitializer()) return 0;
1567      ConstantArray *CA = dyn_cast<ConstantArray>(I->getInitializer());
1568      if (!CA) return 0;
1569      for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i)
1570        if (ConstantStruct *CS = dyn_cast<ConstantStruct>(CA->getOperand(i))) {
1571          if (isa<ConstantPointerNull>(CS->getOperand(1)))
1572            continue;
1573
1574          // Must have a function or null ptr.
1575          if (!isa<Function>(CS->getOperand(1)))
1576            return 0;
1577
1578          // Init priority must be standard.
1579          ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0));
1580          if (!CI || CI->getZExtValue() != 65535)
1581            return 0;
1582        } else {
1583          return 0;
1584        }
1585
1586      return I;
1587    }
1588  return 0;
1589}
1590
1591/// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand,
1592/// return a list of the functions and null terminator as a vector.
1593static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) {
1594  ConstantArray *CA = cast<ConstantArray>(GV->getInitializer());
1595  std::vector<Function*> Result;
1596  Result.reserve(CA->getNumOperands());
1597  for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i) {
1598    ConstantStruct *CS = cast<ConstantStruct>(CA->getOperand(i));
1599    Result.push_back(dyn_cast<Function>(CS->getOperand(1)));
1600  }
1601  return Result;
1602}
1603
1604/// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the
1605/// specified array, returning the new global to use.
1606static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL,
1607                                          const std::vector<Function*> &Ctors) {
1608  // If we made a change, reassemble the initializer list.
1609  std::vector<Constant*> CSVals;
1610  CSVals.push_back(ConstantInt::get(Type::Int32Ty, 65535));
1611  CSVals.push_back(0);
1612
1613  // Create the new init list.
1614  std::vector<Constant*> CAList;
1615  for (unsigned i = 0, e = Ctors.size(); i != e; ++i) {
1616    if (Ctors[i]) {
1617      CSVals[1] = Ctors[i];
1618    } else {
1619      const Type *FTy = FunctionType::get(Type::VoidTy,
1620                                          std::vector<const Type*>(), false);
1621      const PointerType *PFTy = PointerType::getUnqual(FTy);
1622      CSVals[1] = Constant::getNullValue(PFTy);
1623      CSVals[0] = ConstantInt::get(Type::Int32Ty, 2147483647);
1624    }
1625    CAList.push_back(ConstantStruct::get(CSVals));
1626  }
1627
1628  // Create the array initializer.
1629  const Type *StructTy =
1630    cast<ArrayType>(GCL->getType()->getElementType())->getElementType();
1631  Constant *CA = ConstantArray::get(ArrayType::get(StructTy, CAList.size()),
1632                                    CAList);
1633
1634  // If we didn't change the number of elements, don't create a new GV.
1635  if (CA->getType() == GCL->getInitializer()->getType()) {
1636    GCL->setInitializer(CA);
1637    return GCL;
1638  }
1639
1640  // Create the new global and insert it next to the existing list.
1641  GlobalVariable *NGV = new GlobalVariable(CA->getType(), GCL->isConstant(),
1642                                           GCL->getLinkage(), CA, "",
1643                                           (Module *)NULL,
1644                                           GCL->isThreadLocal());
1645  GCL->getParent()->getGlobalList().insert(GCL, NGV);
1646  NGV->takeName(GCL);
1647
1648  // Nuke the old list, replacing any uses with the new one.
1649  if (!GCL->use_empty()) {
1650    Constant *V = NGV;
1651    if (V->getType() != GCL->getType())
1652      V = ConstantExpr::getBitCast(V, GCL->getType());
1653    GCL->replaceAllUsesWith(V);
1654  }
1655  GCL->eraseFromParent();
1656
1657  if (Ctors.size())
1658    return NGV;
1659  else
1660    return 0;
1661}
1662
1663
1664static Constant *getVal(std::map<Value*, Constant*> &ComputedValues,
1665                        Value *V) {
1666  if (Constant *CV = dyn_cast<Constant>(V)) return CV;
1667  Constant *R = ComputedValues[V];
1668  assert(R && "Reference to an uncomputed value!");
1669  return R;
1670}
1671
1672/// isSimpleEnoughPointerToCommit - Return true if this constant is simple
1673/// enough for us to understand.  In particular, if it is a cast of something,
1674/// we punt.  We basically just support direct accesses to globals and GEP's of
1675/// globals.  This should be kept up to date with CommitValueTo.
1676static bool isSimpleEnoughPointerToCommit(Constant *C) {
1677  if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) {
1678    if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage())
1679      return false;  // do not allow weak/linkonce/dllimport/dllexport linkage.
1680    return !GV->isDeclaration();  // reject external globals.
1681  }
1682  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
1683    // Handle a constantexpr gep.
1684    if (CE->getOpcode() == Instruction::GetElementPtr &&
1685        isa<GlobalVariable>(CE->getOperand(0))) {
1686      GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1687      if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage())
1688        return false;  // do not allow weak/linkonce/dllimport/dllexport linkage.
1689      return GV->hasInitializer() &&
1690             ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
1691    }
1692  return false;
1693}
1694
1695/// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global
1696/// initializer.  This returns 'Init' modified to reflect 'Val' stored into it.
1697/// At this point, the GEP operands of Addr [0, OpNo) have been stepped into.
1698static Constant *EvaluateStoreInto(Constant *Init, Constant *Val,
1699                                   ConstantExpr *Addr, unsigned OpNo) {
1700  // Base case of the recursion.
1701  if (OpNo == Addr->getNumOperands()) {
1702    assert(Val->getType() == Init->getType() && "Type mismatch!");
1703    return Val;
1704  }
1705
1706  if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
1707    std::vector<Constant*> Elts;
1708
1709    // Break up the constant into its elements.
1710    if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
1711      for (unsigned i = 0, e = CS->getNumOperands(); i != e; ++i)
1712        Elts.push_back(CS->getOperand(i));
1713    } else if (isa<ConstantAggregateZero>(Init)) {
1714      for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1715        Elts.push_back(Constant::getNullValue(STy->getElementType(i)));
1716    } else if (isa<UndefValue>(Init)) {
1717      for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1718        Elts.push_back(UndefValue::get(STy->getElementType(i)));
1719    } else {
1720      assert(0 && "This code is out of sync with "
1721             " ConstantFoldLoadThroughGEPConstantExpr");
1722    }
1723
1724    // Replace the element that we are supposed to.
1725    ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo));
1726    unsigned Idx = CU->getZExtValue();
1727    assert(Idx < STy->getNumElements() && "Struct index out of range!");
1728    Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1);
1729
1730    // Return the modified struct.
1731    return ConstantStruct::get(&Elts[0], Elts.size(), STy->isPacked());
1732  } else {
1733    ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo));
1734    const ArrayType *ATy = cast<ArrayType>(Init->getType());
1735
1736    // Break up the array into elements.
1737    std::vector<Constant*> Elts;
1738    if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
1739      for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i)
1740        Elts.push_back(CA->getOperand(i));
1741    } else if (isa<ConstantAggregateZero>(Init)) {
1742      Constant *Elt = Constant::getNullValue(ATy->getElementType());
1743      Elts.assign(ATy->getNumElements(), Elt);
1744    } else if (isa<UndefValue>(Init)) {
1745      Constant *Elt = UndefValue::get(ATy->getElementType());
1746      Elts.assign(ATy->getNumElements(), Elt);
1747    } else {
1748      assert(0 && "This code is out of sync with "
1749             " ConstantFoldLoadThroughGEPConstantExpr");
1750    }
1751
1752    assert(CI->getZExtValue() < ATy->getNumElements());
1753    Elts[CI->getZExtValue()] =
1754      EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1);
1755    return ConstantArray::get(ATy, Elts);
1756  }
1757}
1758
1759/// CommitValueTo - We have decided that Addr (which satisfies the predicate
1760/// isSimpleEnoughPointerToCommit) should get Val as its value.  Make it happen.
1761static void CommitValueTo(Constant *Val, Constant *Addr) {
1762  if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
1763    assert(GV->hasInitializer());
1764    GV->setInitializer(Val);
1765    return;
1766  }
1767
1768  ConstantExpr *CE = cast<ConstantExpr>(Addr);
1769  GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1770
1771  Constant *Init = GV->getInitializer();
1772  Init = EvaluateStoreInto(Init, Val, CE, 2);
1773  GV->setInitializer(Init);
1774}
1775
1776/// ComputeLoadResult - Return the value that would be computed by a load from
1777/// P after the stores reflected by 'memory' have been performed.  If we can't
1778/// decide, return null.
1779static Constant *ComputeLoadResult(Constant *P,
1780                                const std::map<Constant*, Constant*> &Memory) {
1781  // If this memory location has been recently stored, use the stored value: it
1782  // is the most up-to-date.
1783  std::map<Constant*, Constant*>::const_iterator I = Memory.find(P);
1784  if (I != Memory.end()) return I->second;
1785
1786  // Access it.
1787  if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
1788    if (GV->hasInitializer())
1789      return GV->getInitializer();
1790    return 0;
1791  }
1792
1793  // Handle a constantexpr getelementptr.
1794  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P))
1795    if (CE->getOpcode() == Instruction::GetElementPtr &&
1796        isa<GlobalVariable>(CE->getOperand(0))) {
1797      GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1798      if (GV->hasInitializer())
1799        return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
1800    }
1801
1802  return 0;  // don't know how to evaluate.
1803}
1804
1805/// EvaluateFunction - Evaluate a call to function F, returning true if
1806/// successful, false if we can't evaluate it.  ActualArgs contains the formal
1807/// arguments for the function.
1808static bool EvaluateFunction(Function *F, Constant *&RetVal,
1809                             const std::vector<Constant*> &ActualArgs,
1810                             std::vector<Function*> &CallStack,
1811                             std::map<Constant*, Constant*> &MutatedMemory,
1812                             std::vector<GlobalVariable*> &AllocaTmps) {
1813  // Check to see if this function is already executing (recursion).  If so,
1814  // bail out.  TODO: we might want to accept limited recursion.
1815  if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end())
1816    return false;
1817
1818  CallStack.push_back(F);
1819
1820  /// Values - As we compute SSA register values, we store their contents here.
1821  std::map<Value*, Constant*> Values;
1822
1823  // Initialize arguments to the incoming values specified.
1824  unsigned ArgNo = 0;
1825  for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E;
1826       ++AI, ++ArgNo)
1827    Values[AI] = ActualArgs[ArgNo];
1828
1829  /// ExecutedBlocks - We only handle non-looping, non-recursive code.  As such,
1830  /// we can only evaluate any one basic block at most once.  This set keeps
1831  /// track of what we have executed so we can detect recursive cases etc.
1832  std::set<BasicBlock*> ExecutedBlocks;
1833
1834  // CurInst - The current instruction we're evaluating.
1835  BasicBlock::iterator CurInst = F->begin()->begin();
1836
1837  // This is the main evaluation loop.
1838  while (1) {
1839    Constant *InstResult = 0;
1840
1841    if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) {
1842      if (SI->isVolatile()) return false;  // no volatile accesses.
1843      Constant *Ptr = getVal(Values, SI->getOperand(1));
1844      if (!isSimpleEnoughPointerToCommit(Ptr))
1845        // If this is too complex for us to commit, reject it.
1846        return false;
1847      Constant *Val = getVal(Values, SI->getOperand(0));
1848      MutatedMemory[Ptr] = Val;
1849    } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) {
1850      InstResult = ConstantExpr::get(BO->getOpcode(),
1851                                     getVal(Values, BO->getOperand(0)),
1852                                     getVal(Values, BO->getOperand(1)));
1853    } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) {
1854      InstResult = ConstantExpr::getCompare(CI->getPredicate(),
1855                                            getVal(Values, CI->getOperand(0)),
1856                                            getVal(Values, CI->getOperand(1)));
1857    } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) {
1858      InstResult = ConstantExpr::getCast(CI->getOpcode(),
1859                                         getVal(Values, CI->getOperand(0)),
1860                                         CI->getType());
1861    } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) {
1862      InstResult = ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)),
1863                                           getVal(Values, SI->getOperand(1)),
1864                                           getVal(Values, SI->getOperand(2)));
1865    } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) {
1866      Constant *P = getVal(Values, GEP->getOperand(0));
1867      SmallVector<Constant*, 8> GEPOps;
1868      for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
1869        GEPOps.push_back(getVal(Values, GEP->getOperand(i)));
1870      InstResult = ConstantExpr::getGetElementPtr(P, &GEPOps[0], GEPOps.size());
1871    } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) {
1872      if (LI->isVolatile()) return false;  // no volatile accesses.
1873      InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)),
1874                                     MutatedMemory);
1875      if (InstResult == 0) return false; // Could not evaluate load.
1876    } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) {
1877      if (AI->isArrayAllocation()) return false;  // Cannot handle array allocs.
1878      const Type *Ty = AI->getType()->getElementType();
1879      AllocaTmps.push_back(new GlobalVariable(Ty, false,
1880                                              GlobalValue::InternalLinkage,
1881                                              UndefValue::get(Ty),
1882                                              AI->getName()));
1883      InstResult = AllocaTmps.back();
1884    } else if (CallInst *CI = dyn_cast<CallInst>(CurInst)) {
1885      // Cannot handle inline asm.
1886      if (isa<InlineAsm>(CI->getOperand(0))) return false;
1887
1888      // Resolve function pointers.
1889      Function *Callee = dyn_cast<Function>(getVal(Values, CI->getOperand(0)));
1890      if (!Callee) return false;  // Cannot resolve.
1891
1892      std::vector<Constant*> Formals;
1893      for (unsigned i = 1, e = CI->getNumOperands(); i != e; ++i)
1894        Formals.push_back(getVal(Values, CI->getOperand(i)));
1895
1896      if (Callee->isDeclaration()) {
1897        // If this is a function we can constant fold, do it.
1898        if (Constant *C = ConstantFoldCall(Callee, &Formals[0],
1899                                           Formals.size())) {
1900          InstResult = C;
1901        } else {
1902          return false;
1903        }
1904      } else {
1905        if (Callee->getFunctionType()->isVarArg())
1906          return false;
1907
1908        Constant *RetVal;
1909
1910        // Execute the call, if successful, use the return value.
1911        if (!EvaluateFunction(Callee, RetVal, Formals, CallStack,
1912                              MutatedMemory, AllocaTmps))
1913          return false;
1914        InstResult = RetVal;
1915      }
1916    } else if (isa<TerminatorInst>(CurInst)) {
1917      BasicBlock *NewBB = 0;
1918      if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) {
1919        if (BI->isUnconditional()) {
1920          NewBB = BI->getSuccessor(0);
1921        } else {
1922          ConstantInt *Cond =
1923            dyn_cast<ConstantInt>(getVal(Values, BI->getCondition()));
1924          if (!Cond) return false;  // Cannot determine.
1925
1926          NewBB = BI->getSuccessor(!Cond->getZExtValue());
1927        }
1928      } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) {
1929        ConstantInt *Val =
1930          dyn_cast<ConstantInt>(getVal(Values, SI->getCondition()));
1931        if (!Val) return false;  // Cannot determine.
1932        NewBB = SI->getSuccessor(SI->findCaseValue(Val));
1933      } else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) {
1934        if (RI->getNumOperands())
1935          RetVal = getVal(Values, RI->getOperand(0));
1936
1937        CallStack.pop_back();  // return from fn.
1938        return true;  // We succeeded at evaluating this ctor!
1939      } else {
1940        // invoke, unwind, unreachable.
1941        return false;  // Cannot handle this terminator.
1942      }
1943
1944      // Okay, we succeeded in evaluating this control flow.  See if we have
1945      // executed the new block before.  If so, we have a looping function,
1946      // which we cannot evaluate in reasonable time.
1947      if (!ExecutedBlocks.insert(NewBB).second)
1948        return false;  // looped!
1949
1950      // Okay, we have never been in this block before.  Check to see if there
1951      // are any PHI nodes.  If so, evaluate them with information about where
1952      // we came from.
1953      BasicBlock *OldBB = CurInst->getParent();
1954      CurInst = NewBB->begin();
1955      PHINode *PN;
1956      for (; (PN = dyn_cast<PHINode>(CurInst)); ++CurInst)
1957        Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB));
1958
1959      // Do NOT increment CurInst.  We know that the terminator had no value.
1960      continue;
1961    } else {
1962      // Did not know how to evaluate this!
1963      return false;
1964    }
1965
1966    if (!CurInst->use_empty())
1967      Values[CurInst] = InstResult;
1968
1969    // Advance program counter.
1970    ++CurInst;
1971  }
1972}
1973
1974/// EvaluateStaticConstructor - Evaluate static constructors in the function, if
1975/// we can.  Return true if we can, false otherwise.
1976static bool EvaluateStaticConstructor(Function *F) {
1977  /// MutatedMemory - For each store we execute, we update this map.  Loads
1978  /// check this to get the most up-to-date value.  If evaluation is successful,
1979  /// this state is committed to the process.
1980  std::map<Constant*, Constant*> MutatedMemory;
1981
1982  /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable
1983  /// to represent its body.  This vector is needed so we can delete the
1984  /// temporary globals when we are done.
1985  std::vector<GlobalVariable*> AllocaTmps;
1986
1987  /// CallStack - This is used to detect recursion.  In pathological situations
1988  /// we could hit exponential behavior, but at least there is nothing
1989  /// unbounded.
1990  std::vector<Function*> CallStack;
1991
1992  // Call the function.
1993  Constant *RetValDummy;
1994  bool EvalSuccess = EvaluateFunction(F, RetValDummy, std::vector<Constant*>(),
1995                                       CallStack, MutatedMemory, AllocaTmps);
1996  if (EvalSuccess) {
1997    // We succeeded at evaluation: commit the result.
1998    DOUT << "FULLY EVALUATED GLOBAL CTOR FUNCTION '"
1999         << F->getName() << "' to " << MutatedMemory.size()
2000         << " stores.\n";
2001    for (std::map<Constant*, Constant*>::iterator I = MutatedMemory.begin(),
2002         E = MutatedMemory.end(); I != E; ++I)
2003      CommitValueTo(I->second, I->first);
2004  }
2005
2006  // At this point, we are done interpreting.  If we created any 'alloca'
2007  // temporaries, release them now.
2008  while (!AllocaTmps.empty()) {
2009    GlobalVariable *Tmp = AllocaTmps.back();
2010    AllocaTmps.pop_back();
2011
2012    // If there are still users of the alloca, the program is doing something
2013    // silly, e.g. storing the address of the alloca somewhere and using it
2014    // later.  Since this is undefined, we'll just make it be null.
2015    if (!Tmp->use_empty())
2016      Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType()));
2017    delete Tmp;
2018  }
2019
2020  return EvalSuccess;
2021}
2022
2023
2024
2025/// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible.
2026/// Return true if anything changed.
2027bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) {
2028  std::vector<Function*> Ctors = ParseGlobalCtors(GCL);
2029  bool MadeChange = false;
2030  if (Ctors.empty()) return false;
2031
2032  // Loop over global ctors, optimizing them when we can.
2033  for (unsigned i = 0; i != Ctors.size(); ++i) {
2034    Function *F = Ctors[i];
2035    // Found a null terminator in the middle of the list, prune off the rest of
2036    // the list.
2037    if (F == 0) {
2038      if (i != Ctors.size()-1) {
2039        Ctors.resize(i+1);
2040        MadeChange = true;
2041      }
2042      break;
2043    }
2044
2045    // We cannot simplify external ctor functions.
2046    if (F->empty()) continue;
2047
2048    // If we can evaluate the ctor at compile time, do.
2049    if (EvaluateStaticConstructor(F)) {
2050      Ctors.erase(Ctors.begin()+i);
2051      MadeChange = true;
2052      --i;
2053      ++NumCtorsEvaluated;
2054      continue;
2055    }
2056  }
2057
2058  if (!MadeChange) return false;
2059
2060  GCL = InstallGlobalCtors(GCL, Ctors);
2061  return true;
2062}
2063
2064
2065bool GlobalOpt::runOnModule(Module &M) {
2066  bool Changed = false;
2067
2068  // Try to find the llvm.globalctors list.
2069  GlobalVariable *GlobalCtors = FindGlobalCtors(M);
2070
2071  bool LocalChange = true;
2072  while (LocalChange) {
2073    LocalChange = false;
2074
2075    // Delete functions that are trivially dead, ccc -> fastcc
2076    LocalChange |= OptimizeFunctions(M);
2077
2078    // Optimize global_ctors list.
2079    if (GlobalCtors)
2080      LocalChange |= OptimizeGlobalCtorsList(GlobalCtors);
2081
2082    // Optimize non-address-taken globals.
2083    LocalChange |= OptimizeGlobalVars(M);
2084    Changed |= LocalChange;
2085  }
2086
2087  // TODO: Move all global ctors functions to the end of the module for code
2088  // layout.
2089
2090  return Changed;
2091}
2092