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/Operator.h"
25#include "llvm/Pass.h"
26#include "llvm/Analysis/ConstantFolding.h"
27#include "llvm/Analysis/MemoryBuiltins.h"
28#include "llvm/Target/TargetData.h"
29#include "llvm/Target/TargetLibraryInfo.h"
30#include "llvm/Support/CallSite.h"
31#include "llvm/Support/Debug.h"
32#include "llvm/Support/ErrorHandling.h"
33#include "llvm/Support/GetElementPtrTypeIterator.h"
34#include "llvm/Support/MathExtras.h"
35#include "llvm/Support/raw_ostream.h"
36#include "llvm/ADT/DenseMap.h"
37#include "llvm/ADT/SmallPtrSet.h"
38#include "llvm/ADT/SmallVector.h"
39#include "llvm/ADT/Statistic.h"
40#include "llvm/ADT/STLExtras.h"
41#include <algorithm>
42using namespace llvm;
43
44STATISTIC(NumMarked    , "Number of globals marked constant");
45STATISTIC(NumUnnamed   , "Number of globals marked unnamed_addr");
46STATISTIC(NumSRA       , "Number of aggregate globals broken into scalars");
47STATISTIC(NumHeapSRA   , "Number of heap objects SRA'd");
48STATISTIC(NumSubstitute,"Number of globals with initializers stored into them");
49STATISTIC(NumDeleted   , "Number of globals deleted");
50STATISTIC(NumFnDeleted , "Number of functions deleted");
51STATISTIC(NumGlobUses  , "Number of global uses devirtualized");
52STATISTIC(NumLocalized , "Number of globals localized");
53STATISTIC(NumShrunkToBool  , "Number of global vars shrunk to booleans");
54STATISTIC(NumFastCallFns   , "Number of functions converted to fastcc");
55STATISTIC(NumCtorsEvaluated, "Number of static ctors evaluated");
56STATISTIC(NumNestRemoved   , "Number of nest attributes removed");
57STATISTIC(NumAliasesResolved, "Number of global aliases resolved");
58STATISTIC(NumAliasesRemoved, "Number of global aliases eliminated");
59STATISTIC(NumCXXDtorsRemoved, "Number of global C++ destructors removed");
60
61namespace {
62  struct GlobalStatus;
63  struct GlobalOpt : public ModulePass {
64    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
65      AU.addRequired<TargetLibraryInfo>();
66    }
67    static char ID; // Pass identification, replacement for typeid
68    GlobalOpt() : ModulePass(ID) {
69      initializeGlobalOptPass(*PassRegistry::getPassRegistry());
70    }
71
72    bool runOnModule(Module &M);
73
74  private:
75    GlobalVariable *FindGlobalCtors(Module &M);
76    bool OptimizeFunctions(Module &M);
77    bool OptimizeGlobalVars(Module &M);
78    bool OptimizeGlobalAliases(Module &M);
79    bool OptimizeGlobalCtorsList(GlobalVariable *&GCL);
80    bool ProcessGlobal(GlobalVariable *GV,Module::global_iterator &GVI);
81    bool ProcessInternalGlobal(GlobalVariable *GV,Module::global_iterator &GVI,
82                               const SmallPtrSet<const PHINode*, 16> &PHIUsers,
83                               const GlobalStatus &GS);
84    bool OptimizeEmptyGlobalCXXDtors(Function *CXAAtExitFn);
85
86    TargetData *TD;
87    TargetLibraryInfo *TLI;
88  };
89}
90
91char GlobalOpt::ID = 0;
92INITIALIZE_PASS_BEGIN(GlobalOpt, "globalopt",
93                "Global Variable Optimizer", false, false)
94INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
95INITIALIZE_PASS_END(GlobalOpt, "globalopt",
96                "Global Variable Optimizer", false, false)
97
98ModulePass *llvm::createGlobalOptimizerPass() { return new GlobalOpt(); }
99
100namespace {
101
102/// GlobalStatus - As we analyze each global, keep track of some information
103/// about it.  If we find out that the address of the global is taken, none of
104/// this info will be accurate.
105struct GlobalStatus {
106  /// isCompared - True if the global's address is used in a comparison.
107  bool isCompared;
108
109  /// isLoaded - True if the global is ever loaded.  If the global isn't ever
110  /// loaded it can be deleted.
111  bool isLoaded;
112
113  /// StoredType - Keep track of what stores to the global look like.
114  ///
115  enum StoredType {
116    /// NotStored - There is no store to this global.  It can thus be marked
117    /// constant.
118    NotStored,
119
120    /// isInitializerStored - This global is stored to, but the only thing
121    /// stored is the constant it was initialized with.  This is only tracked
122    /// for scalar globals.
123    isInitializerStored,
124
125    /// isStoredOnce - This global is stored to, but only its initializer and
126    /// one other value is ever stored to it.  If this global isStoredOnce, we
127    /// track the value stored to it in StoredOnceValue below.  This is only
128    /// tracked for scalar globals.
129    isStoredOnce,
130
131    /// isStored - This global is stored to by multiple values or something else
132    /// that we cannot track.
133    isStored
134  } StoredType;
135
136  /// StoredOnceValue - If only one value (besides the initializer constant) is
137  /// ever stored to this global, keep track of what value it is.
138  Value *StoredOnceValue;
139
140  /// AccessingFunction/HasMultipleAccessingFunctions - These start out
141  /// null/false.  When the first accessing function is noticed, it is recorded.
142  /// When a second different accessing function is noticed,
143  /// HasMultipleAccessingFunctions is set to true.
144  const Function *AccessingFunction;
145  bool HasMultipleAccessingFunctions;
146
147  /// HasNonInstructionUser - Set to true if this global has a user that is not
148  /// an instruction (e.g. a constant expr or GV initializer).
149  bool HasNonInstructionUser;
150
151  /// HasPHIUser - Set to true if this global has a user that is a PHI node.
152  bool HasPHIUser;
153
154  /// AtomicOrdering - Set to the strongest atomic ordering requirement.
155  AtomicOrdering Ordering;
156
157  GlobalStatus() : isCompared(false), isLoaded(false), StoredType(NotStored),
158                   StoredOnceValue(0), AccessingFunction(0),
159                   HasMultipleAccessingFunctions(false),
160                   HasNonInstructionUser(false), HasPHIUser(false),
161                   Ordering(NotAtomic) {}
162};
163
164}
165
166/// StrongerOrdering - Return the stronger of the two ordering. If the two
167/// orderings are acquire and release, then return AcquireRelease.
168///
169static AtomicOrdering StrongerOrdering(AtomicOrdering X, AtomicOrdering Y) {
170  if (X == Acquire && Y == Release) return AcquireRelease;
171  if (Y == Acquire && X == Release) return AcquireRelease;
172  return (AtomicOrdering)std::max(X, Y);
173}
174
175/// SafeToDestroyConstant - It is safe to destroy a constant iff it is only used
176/// by constants itself.  Note that constants cannot be cyclic, so this test is
177/// pretty easy to implement recursively.
178///
179static bool SafeToDestroyConstant(const Constant *C) {
180  if (isa<GlobalValue>(C)) return false;
181
182  for (Value::const_use_iterator UI = C->use_begin(), E = C->use_end(); UI != E;
183       ++UI)
184    if (const Constant *CU = dyn_cast<Constant>(*UI)) {
185      if (!SafeToDestroyConstant(CU)) return false;
186    } else
187      return false;
188  return true;
189}
190
191
192/// AnalyzeGlobal - Look at all uses of the global and fill in the GlobalStatus
193/// structure.  If the global has its address taken, return true to indicate we
194/// can't do anything with it.
195///
196static bool AnalyzeGlobal(const Value *V, GlobalStatus &GS,
197                          SmallPtrSet<const PHINode*, 16> &PHIUsers) {
198  for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;
199       ++UI) {
200    const User *U = *UI;
201    if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
202      GS.HasNonInstructionUser = true;
203
204      // If the result of the constantexpr isn't pointer type, then we won't
205      // know to expect it in various places.  Just reject early.
206      if (!isa<PointerType>(CE->getType())) return true;
207
208      if (AnalyzeGlobal(CE, GS, PHIUsers)) return true;
209    } else if (const Instruction *I = dyn_cast<Instruction>(U)) {
210      if (!GS.HasMultipleAccessingFunctions) {
211        const Function *F = I->getParent()->getParent();
212        if (GS.AccessingFunction == 0)
213          GS.AccessingFunction = F;
214        else if (GS.AccessingFunction != F)
215          GS.HasMultipleAccessingFunctions = true;
216      }
217      if (const LoadInst *LI = dyn_cast<LoadInst>(I)) {
218        GS.isLoaded = true;
219        // Don't hack on volatile loads.
220        if (LI->isVolatile()) return true;
221        GS.Ordering = StrongerOrdering(GS.Ordering, LI->getOrdering());
222      } else if (const StoreInst *SI = dyn_cast<StoreInst>(I)) {
223        // Don't allow a store OF the address, only stores TO the address.
224        if (SI->getOperand(0) == V) return true;
225
226        // Don't hack on volatile stores.
227        if (SI->isVolatile()) return true;
228        GS.Ordering = StrongerOrdering(GS.Ordering, SI->getOrdering());
229
230        // If this is a direct store to the global (i.e., the global is a scalar
231        // value, not an aggregate), keep more specific information about
232        // stores.
233        if (GS.StoredType != GlobalStatus::isStored) {
234          if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(
235                                                           SI->getOperand(1))) {
236            Value *StoredVal = SI->getOperand(0);
237            if (StoredVal == GV->getInitializer()) {
238              if (GS.StoredType < GlobalStatus::isInitializerStored)
239                GS.StoredType = GlobalStatus::isInitializerStored;
240            } else if (isa<LoadInst>(StoredVal) &&
241                       cast<LoadInst>(StoredVal)->getOperand(0) == GV) {
242              if (GS.StoredType < GlobalStatus::isInitializerStored)
243                GS.StoredType = GlobalStatus::isInitializerStored;
244            } else if (GS.StoredType < GlobalStatus::isStoredOnce) {
245              GS.StoredType = GlobalStatus::isStoredOnce;
246              GS.StoredOnceValue = StoredVal;
247            } else if (GS.StoredType == GlobalStatus::isStoredOnce &&
248                       GS.StoredOnceValue == StoredVal) {
249              // noop.
250            } else {
251              GS.StoredType = GlobalStatus::isStored;
252            }
253          } else {
254            GS.StoredType = GlobalStatus::isStored;
255          }
256        }
257      } else if (isa<BitCastInst>(I)) {
258        if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
259      } else if (isa<GetElementPtrInst>(I)) {
260        if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
261      } else if (isa<SelectInst>(I)) {
262        if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
263      } else if (const PHINode *PN = dyn_cast<PHINode>(I)) {
264        // PHI nodes we can check just like select or GEP instructions, but we
265        // have to be careful about infinite recursion.
266        if (PHIUsers.insert(PN))  // Not already visited.
267          if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
268        GS.HasPHIUser = true;
269      } else if (isa<CmpInst>(I)) {
270        GS.isCompared = true;
271      } else if (const MemTransferInst *MTI = dyn_cast<MemTransferInst>(I)) {
272        if (MTI->isVolatile()) return true;
273        if (MTI->getArgOperand(0) == V)
274          GS.StoredType = GlobalStatus::isStored;
275        if (MTI->getArgOperand(1) == V)
276          GS.isLoaded = true;
277      } else if (const MemSetInst *MSI = dyn_cast<MemSetInst>(I)) {
278        assert(MSI->getArgOperand(0) == V && "Memset only takes one pointer!");
279        if (MSI->isVolatile()) return true;
280        GS.StoredType = GlobalStatus::isStored;
281      } else {
282        return true;  // Any other non-load instruction might take address!
283      }
284    } else if (const Constant *C = dyn_cast<Constant>(U)) {
285      GS.HasNonInstructionUser = true;
286      // We might have a dead and dangling constant hanging off of here.
287      if (!SafeToDestroyConstant(C))
288        return true;
289    } else {
290      GS.HasNonInstructionUser = true;
291      // Otherwise must be some other user.
292      return true;
293    }
294  }
295
296  return false;
297}
298
299/// isLeakCheckerRoot - Is this global variable possibly used by a leak checker
300/// as a root?  If so, we might not really want to eliminate the stores to it.
301static bool isLeakCheckerRoot(GlobalVariable *GV) {
302  // A global variable is a root if it is a pointer, or could plausibly contain
303  // a pointer.  There are two challenges; one is that we could have a struct
304  // the has an inner member which is a pointer.  We recurse through the type to
305  // detect these (up to a point).  The other is that we may actually be a union
306  // of a pointer and another type, and so our LLVM type is an integer which
307  // gets converted into a pointer, or our type is an [i8 x #] with a pointer
308  // potentially contained here.
309
310  if (GV->hasPrivateLinkage())
311    return false;
312
313  SmallVector<Type *, 4> Types;
314  Types.push_back(cast<PointerType>(GV->getType())->getElementType());
315
316  unsigned Limit = 20;
317  do {
318    Type *Ty = Types.pop_back_val();
319    switch (Ty->getTypeID()) {
320      default: break;
321      case Type::PointerTyID: return true;
322      case Type::ArrayTyID:
323      case Type::VectorTyID: {
324        SequentialType *STy = cast<SequentialType>(Ty);
325        Types.push_back(STy->getElementType());
326        break;
327      }
328      case Type::StructTyID: {
329        StructType *STy = cast<StructType>(Ty);
330        if (STy->isOpaque()) return true;
331        for (StructType::element_iterator I = STy->element_begin(),
332                 E = STy->element_end(); I != E; ++I) {
333          Type *InnerTy = *I;
334          if (isa<PointerType>(InnerTy)) return true;
335          if (isa<CompositeType>(InnerTy))
336            Types.push_back(InnerTy);
337        }
338        break;
339      }
340    }
341    if (--Limit == 0) return true;
342  } while (!Types.empty());
343  return false;
344}
345
346/// Given a value that is stored to a global but never read, determine whether
347/// it's safe to remove the store and the chain of computation that feeds the
348/// store.
349static bool IsSafeComputationToRemove(Value *V, const TargetLibraryInfo *TLI) {
350  do {
351    if (isa<Constant>(V))
352      return true;
353    if (!V->hasOneUse())
354      return false;
355    if (isa<LoadInst>(V) || isa<InvokeInst>(V) || isa<Argument>(V) ||
356        isa<GlobalValue>(V))
357      return false;
358    if (isAllocationFn(V, TLI))
359      return true;
360
361    Instruction *I = cast<Instruction>(V);
362    if (I->mayHaveSideEffects())
363      return false;
364    if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
365      if (!GEP->hasAllConstantIndices())
366        return false;
367    } else if (I->getNumOperands() != 1) {
368      return false;
369    }
370
371    V = I->getOperand(0);
372  } while (1);
373}
374
375/// CleanupPointerRootUsers - This GV is a pointer root.  Loop over all users
376/// of the global and clean up any that obviously don't assign the global a
377/// value that isn't dynamically allocated.
378///
379static bool CleanupPointerRootUsers(GlobalVariable *GV,
380                                    const TargetLibraryInfo *TLI) {
381  // A brief explanation of leak checkers.  The goal is to find bugs where
382  // pointers are forgotten, causing an accumulating growth in memory
383  // usage over time.  The common strategy for leak checkers is to whitelist the
384  // memory pointed to by globals at exit.  This is popular because it also
385  // solves another problem where the main thread of a C++ program may shut down
386  // before other threads that are still expecting to use those globals.  To
387  // handle that case, we expect the program may create a singleton and never
388  // destroy it.
389
390  bool Changed = false;
391
392  // If Dead[n].first is the only use of a malloc result, we can delete its
393  // chain of computation and the store to the global in Dead[n].second.
394  SmallVector<std::pair<Instruction *, Instruction *>, 32> Dead;
395
396  // Constants can't be pointers to dynamically allocated memory.
397  for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end();
398       UI != E;) {
399    User *U = *UI++;
400    if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
401      Value *V = SI->getValueOperand();
402      if (isa<Constant>(V)) {
403        Changed = true;
404        SI->eraseFromParent();
405      } else if (Instruction *I = dyn_cast<Instruction>(V)) {
406        if (I->hasOneUse())
407          Dead.push_back(std::make_pair(I, SI));
408      }
409    } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(U)) {
410      if (isa<Constant>(MSI->getValue())) {
411        Changed = true;
412        MSI->eraseFromParent();
413      } else if (Instruction *I = dyn_cast<Instruction>(MSI->getValue())) {
414        if (I->hasOneUse())
415          Dead.push_back(std::make_pair(I, MSI));
416      }
417    } else if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(U)) {
418      GlobalVariable *MemSrc = dyn_cast<GlobalVariable>(MTI->getSource());
419      if (MemSrc && MemSrc->isConstant()) {
420        Changed = true;
421        MTI->eraseFromParent();
422      } else if (Instruction *I = dyn_cast<Instruction>(MemSrc)) {
423        if (I->hasOneUse())
424          Dead.push_back(std::make_pair(I, MTI));
425      }
426    } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
427      if (CE->use_empty()) {
428        CE->destroyConstant();
429        Changed = true;
430      }
431    } else if (Constant *C = dyn_cast<Constant>(U)) {
432      if (SafeToDestroyConstant(C)) {
433        C->destroyConstant();
434        // This could have invalidated UI, start over from scratch.
435        Dead.clear();
436        CleanupPointerRootUsers(GV, TLI);
437        return true;
438      }
439    }
440  }
441
442  for (int i = 0, e = Dead.size(); i != e; ++i) {
443    if (IsSafeComputationToRemove(Dead[i].first, TLI)) {
444      Dead[i].second->eraseFromParent();
445      Instruction *I = Dead[i].first;
446      do {
447	if (isAllocationFn(I, TLI))
448	  break;
449        Instruction *J = dyn_cast<Instruction>(I->getOperand(0));
450        if (!J)
451          break;
452        I->eraseFromParent();
453        I = J;
454      } while (1);
455      I->eraseFromParent();
456    }
457  }
458
459  return Changed;
460}
461
462/// CleanupConstantGlobalUsers - We just marked GV constant.  Loop over all
463/// users of the global, cleaning up the obvious ones.  This is largely just a
464/// quick scan over the use list to clean up the easy and obvious cruft.  This
465/// returns true if it made a change.
466static bool CleanupConstantGlobalUsers(Value *V, Constant *Init,
467                                       TargetData *TD, TargetLibraryInfo *TLI) {
468  bool Changed = false;
469  for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;) {
470    User *U = *UI++;
471
472    if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
473      if (Init) {
474        // Replace the load with the initializer.
475        LI->replaceAllUsesWith(Init);
476        LI->eraseFromParent();
477        Changed = true;
478      }
479    } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
480      // Store must be unreachable or storing Init into the global.
481      SI->eraseFromParent();
482      Changed = true;
483    } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
484      if (CE->getOpcode() == Instruction::GetElementPtr) {
485        Constant *SubInit = 0;
486        if (Init)
487          SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
488        Changed |= CleanupConstantGlobalUsers(CE, SubInit, TD, TLI);
489      } else if (CE->getOpcode() == Instruction::BitCast &&
490                 CE->getType()->isPointerTy()) {
491        // Pointer cast, delete any stores and memsets to the global.
492        Changed |= CleanupConstantGlobalUsers(CE, 0, TD, TLI);
493      }
494
495      if (CE->use_empty()) {
496        CE->destroyConstant();
497        Changed = true;
498      }
499    } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
500      // Do not transform "gepinst (gep constexpr (GV))" here, because forming
501      // "gepconstexpr (gep constexpr (GV))" will cause the two gep's to fold
502      // and will invalidate our notion of what Init is.
503      Constant *SubInit = 0;
504      if (!isa<ConstantExpr>(GEP->getOperand(0))) {
505        ConstantExpr *CE =
506          dyn_cast_or_null<ConstantExpr>(ConstantFoldInstruction(GEP, TD, TLI));
507        if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr)
508          SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
509
510        // If the initializer is an all-null value and we have an inbounds GEP,
511        // we already know what the result of any load from that GEP is.
512        // TODO: Handle splats.
513        if (Init && isa<ConstantAggregateZero>(Init) && GEP->isInBounds())
514          SubInit = Constant::getNullValue(GEP->getType()->getElementType());
515      }
516      Changed |= CleanupConstantGlobalUsers(GEP, SubInit, TD, TLI);
517
518      if (GEP->use_empty()) {
519        GEP->eraseFromParent();
520        Changed = true;
521      }
522    } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv
523      if (MI->getRawDest() == V) {
524        MI->eraseFromParent();
525        Changed = true;
526      }
527
528    } else if (Constant *C = dyn_cast<Constant>(U)) {
529      // If we have a chain of dead constantexprs or other things dangling from
530      // us, and if they are all dead, nuke them without remorse.
531      if (SafeToDestroyConstant(C)) {
532        C->destroyConstant();
533        // This could have invalidated UI, start over from scratch.
534        CleanupConstantGlobalUsers(V, Init, TD, TLI);
535        return true;
536      }
537    }
538  }
539  return Changed;
540}
541
542/// isSafeSROAElementUse - Return true if the specified instruction is a safe
543/// user of a derived expression from a global that we want to SROA.
544static bool isSafeSROAElementUse(Value *V) {
545  // We might have a dead and dangling constant hanging off of here.
546  if (Constant *C = dyn_cast<Constant>(V))
547    return SafeToDestroyConstant(C);
548
549  Instruction *I = dyn_cast<Instruction>(V);
550  if (!I) return false;
551
552  // Loads are ok.
553  if (isa<LoadInst>(I)) return true;
554
555  // Stores *to* the pointer are ok.
556  if (StoreInst *SI = dyn_cast<StoreInst>(I))
557    return SI->getOperand(0) != V;
558
559  // Otherwise, it must be a GEP.
560  GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I);
561  if (GEPI == 0) return false;
562
563  if (GEPI->getNumOperands() < 3 || !isa<Constant>(GEPI->getOperand(1)) ||
564      !cast<Constant>(GEPI->getOperand(1))->isNullValue())
565    return false;
566
567  for (Value::use_iterator I = GEPI->use_begin(), E = GEPI->use_end();
568       I != E; ++I)
569    if (!isSafeSROAElementUse(*I))
570      return false;
571  return true;
572}
573
574
575/// IsUserOfGlobalSafeForSRA - U is a direct user of the specified global value.
576/// Look at it and its uses and decide whether it is safe to SROA this global.
577///
578static bool IsUserOfGlobalSafeForSRA(User *U, GlobalValue *GV) {
579  // The user of the global must be a GEP Inst or a ConstantExpr GEP.
580  if (!isa<GetElementPtrInst>(U) &&
581      (!isa<ConstantExpr>(U) ||
582       cast<ConstantExpr>(U)->getOpcode() != Instruction::GetElementPtr))
583    return false;
584
585  // Check to see if this ConstantExpr GEP is SRA'able.  In particular, we
586  // don't like < 3 operand CE's, and we don't like non-constant integer
587  // indices.  This enforces that all uses are 'gep GV, 0, C, ...' for some
588  // value of C.
589  if (U->getNumOperands() < 3 || !isa<Constant>(U->getOperand(1)) ||
590      !cast<Constant>(U->getOperand(1))->isNullValue() ||
591      !isa<ConstantInt>(U->getOperand(2)))
592    return false;
593
594  gep_type_iterator GEPI = gep_type_begin(U), E = gep_type_end(U);
595  ++GEPI;  // Skip over the pointer index.
596
597  // If this is a use of an array allocation, do a bit more checking for sanity.
598  if (ArrayType *AT = dyn_cast<ArrayType>(*GEPI)) {
599    uint64_t NumElements = AT->getNumElements();
600    ConstantInt *Idx = cast<ConstantInt>(U->getOperand(2));
601
602    // Check to make sure that index falls within the array.  If not,
603    // something funny is going on, so we won't do the optimization.
604    //
605    if (Idx->getZExtValue() >= NumElements)
606      return false;
607
608    // We cannot scalar repl this level of the array unless any array
609    // sub-indices are in-range constants.  In particular, consider:
610    // A[0][i].  We cannot know that the user isn't doing invalid things like
611    // allowing i to index an out-of-range subscript that accesses A[1].
612    //
613    // Scalar replacing *just* the outer index of the array is probably not
614    // going to be a win anyway, so just give up.
615    for (++GEPI; // Skip array index.
616         GEPI != E;
617         ++GEPI) {
618      uint64_t NumElements;
619      if (ArrayType *SubArrayTy = dyn_cast<ArrayType>(*GEPI))
620        NumElements = SubArrayTy->getNumElements();
621      else if (VectorType *SubVectorTy = dyn_cast<VectorType>(*GEPI))
622        NumElements = SubVectorTy->getNumElements();
623      else {
624        assert((*GEPI)->isStructTy() &&
625               "Indexed GEP type is not array, vector, or struct!");
626        continue;
627      }
628
629      ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPI.getOperand());
630      if (!IdxVal || IdxVal->getZExtValue() >= NumElements)
631        return false;
632    }
633  }
634
635  for (Value::use_iterator I = U->use_begin(), E = U->use_end(); I != E; ++I)
636    if (!isSafeSROAElementUse(*I))
637      return false;
638  return true;
639}
640
641/// GlobalUsersSafeToSRA - Look at all uses of the global and decide whether it
642/// is safe for us to perform this transformation.
643///
644static bool GlobalUsersSafeToSRA(GlobalValue *GV) {
645  for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end();
646       UI != E; ++UI) {
647    if (!IsUserOfGlobalSafeForSRA(*UI, GV))
648      return false;
649  }
650  return true;
651}
652
653
654/// SRAGlobal - Perform scalar replacement of aggregates on the specified global
655/// variable.  This opens the door for other optimizations by exposing the
656/// behavior of the program in a more fine-grained way.  We have determined that
657/// this transformation is safe already.  We return the first global variable we
658/// insert so that the caller can reprocess it.
659static GlobalVariable *SRAGlobal(GlobalVariable *GV, const TargetData &TD) {
660  // Make sure this global only has simple uses that we can SRA.
661  if (!GlobalUsersSafeToSRA(GV))
662    return 0;
663
664  assert(GV->hasLocalLinkage() && !GV->isConstant());
665  Constant *Init = GV->getInitializer();
666  Type *Ty = Init->getType();
667
668  std::vector<GlobalVariable*> NewGlobals;
669  Module::GlobalListType &Globals = GV->getParent()->getGlobalList();
670
671  // Get the alignment of the global, either explicit or target-specific.
672  unsigned StartAlignment = GV->getAlignment();
673  if (StartAlignment == 0)
674    StartAlignment = TD.getABITypeAlignment(GV->getType());
675
676  if (StructType *STy = dyn_cast<StructType>(Ty)) {
677    NewGlobals.reserve(STy->getNumElements());
678    const StructLayout &Layout = *TD.getStructLayout(STy);
679    for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
680      Constant *In = Init->getAggregateElement(i);
681      assert(In && "Couldn't get element of initializer?");
682      GlobalVariable *NGV = new GlobalVariable(STy->getElementType(i), false,
683                                               GlobalVariable::InternalLinkage,
684                                               In, GV->getName()+"."+Twine(i),
685                                               GV->getThreadLocalMode(),
686                                              GV->getType()->getAddressSpace());
687      Globals.insert(GV, NGV);
688      NewGlobals.push_back(NGV);
689
690      // Calculate the known alignment of the field.  If the original aggregate
691      // had 256 byte alignment for example, something might depend on that:
692      // propagate info to each field.
693      uint64_t FieldOffset = Layout.getElementOffset(i);
694      unsigned NewAlign = (unsigned)MinAlign(StartAlignment, FieldOffset);
695      if (NewAlign > TD.getABITypeAlignment(STy->getElementType(i)))
696        NGV->setAlignment(NewAlign);
697    }
698  } else if (SequentialType *STy = dyn_cast<SequentialType>(Ty)) {
699    unsigned NumElements = 0;
700    if (ArrayType *ATy = dyn_cast<ArrayType>(STy))
701      NumElements = ATy->getNumElements();
702    else
703      NumElements = cast<VectorType>(STy)->getNumElements();
704
705    if (NumElements > 16 && GV->hasNUsesOrMore(16))
706      return 0; // It's not worth it.
707    NewGlobals.reserve(NumElements);
708
709    uint64_t EltSize = TD.getTypeAllocSize(STy->getElementType());
710    unsigned EltAlign = TD.getABITypeAlignment(STy->getElementType());
711    for (unsigned i = 0, e = NumElements; i != e; ++i) {
712      Constant *In = Init->getAggregateElement(i);
713      assert(In && "Couldn't get element of initializer?");
714
715      GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false,
716                                               GlobalVariable::InternalLinkage,
717                                               In, GV->getName()+"."+Twine(i),
718                                               GV->getThreadLocalMode(),
719                                              GV->getType()->getAddressSpace());
720      Globals.insert(GV, NGV);
721      NewGlobals.push_back(NGV);
722
723      // Calculate the known alignment of the field.  If the original aggregate
724      // had 256 byte alignment for example, something might depend on that:
725      // propagate info to each field.
726      unsigned NewAlign = (unsigned)MinAlign(StartAlignment, EltSize*i);
727      if (NewAlign > EltAlign)
728        NGV->setAlignment(NewAlign);
729    }
730  }
731
732  if (NewGlobals.empty())
733    return 0;
734
735  DEBUG(dbgs() << "PERFORMING GLOBAL SRA ON: " << *GV);
736
737  Constant *NullInt =Constant::getNullValue(Type::getInt32Ty(GV->getContext()));
738
739  // Loop over all of the uses of the global, replacing the constantexpr geps,
740  // with smaller constantexpr geps or direct references.
741  while (!GV->use_empty()) {
742    User *GEP = GV->use_back();
743    assert(((isa<ConstantExpr>(GEP) &&
744             cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)||
745            isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!");
746
747    // Ignore the 1th operand, which has to be zero or else the program is quite
748    // broken (undefined).  Get the 2nd operand, which is the structure or array
749    // index.
750    unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
751    if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access.
752
753    Value *NewPtr = NewGlobals[Val];
754
755    // Form a shorter GEP if needed.
756    if (GEP->getNumOperands() > 3) {
757      if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) {
758        SmallVector<Constant*, 8> Idxs;
759        Idxs.push_back(NullInt);
760        for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i)
761          Idxs.push_back(CE->getOperand(i));
762        NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr), Idxs);
763      } else {
764        GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP);
765        SmallVector<Value*, 8> Idxs;
766        Idxs.push_back(NullInt);
767        for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i)
768          Idxs.push_back(GEPI->getOperand(i));
769        NewPtr = GetElementPtrInst::Create(NewPtr, Idxs,
770                                           GEPI->getName()+"."+Twine(Val),GEPI);
771      }
772    }
773    GEP->replaceAllUsesWith(NewPtr);
774
775    if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP))
776      GEPI->eraseFromParent();
777    else
778      cast<ConstantExpr>(GEP)->destroyConstant();
779  }
780
781  // Delete the old global, now that it is dead.
782  Globals.erase(GV);
783  ++NumSRA;
784
785  // Loop over the new globals array deleting any globals that are obviously
786  // dead.  This can arise due to scalarization of a structure or an array that
787  // has elements that are dead.
788  unsigned FirstGlobal = 0;
789  for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i)
790    if (NewGlobals[i]->use_empty()) {
791      Globals.erase(NewGlobals[i]);
792      if (FirstGlobal == i) ++FirstGlobal;
793    }
794
795  return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : 0;
796}
797
798/// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified
799/// value will trap if the value is dynamically null.  PHIs keeps track of any
800/// phi nodes we've seen to avoid reprocessing them.
801static bool AllUsesOfValueWillTrapIfNull(const Value *V,
802                                         SmallPtrSet<const PHINode*, 8> &PHIs) {
803  for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;
804       ++UI) {
805    const User *U = *UI;
806
807    if (isa<LoadInst>(U)) {
808      // Will trap.
809    } else if (const StoreInst *SI = dyn_cast<StoreInst>(U)) {
810      if (SI->getOperand(0) == V) {
811        //cerr << "NONTRAPPING USE: " << *U;
812        return false;  // Storing the value.
813      }
814    } else if (const CallInst *CI = dyn_cast<CallInst>(U)) {
815      if (CI->getCalledValue() != V) {
816        //cerr << "NONTRAPPING USE: " << *U;
817        return false;  // Not calling the ptr
818      }
819    } else if (const InvokeInst *II = dyn_cast<InvokeInst>(U)) {
820      if (II->getCalledValue() != V) {
821        //cerr << "NONTRAPPING USE: " << *U;
822        return false;  // Not calling the ptr
823      }
824    } else if (const BitCastInst *CI = dyn_cast<BitCastInst>(U)) {
825      if (!AllUsesOfValueWillTrapIfNull(CI, PHIs)) return false;
826    } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
827      if (!AllUsesOfValueWillTrapIfNull(GEPI, PHIs)) return false;
828    } else if (const PHINode *PN = dyn_cast<PHINode>(U)) {
829      // If we've already seen this phi node, ignore it, it has already been
830      // checked.
831      if (PHIs.insert(PN) && !AllUsesOfValueWillTrapIfNull(PN, PHIs))
832        return false;
833    } else if (isa<ICmpInst>(U) &&
834               isa<ConstantPointerNull>(UI->getOperand(1))) {
835      // Ignore icmp X, null
836    } else {
837      //cerr << "NONTRAPPING USE: " << *U;
838      return false;
839    }
840  }
841  return true;
842}
843
844/// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads
845/// from GV will trap if the loaded value is null.  Note that this also permits
846/// comparisons of the loaded value against null, as a special case.
847static bool AllUsesOfLoadedValueWillTrapIfNull(const GlobalVariable *GV) {
848  for (Value::const_use_iterator UI = GV->use_begin(), E = GV->use_end();
849       UI != E; ++UI) {
850    const User *U = *UI;
851
852    if (const LoadInst *LI = dyn_cast<LoadInst>(U)) {
853      SmallPtrSet<const PHINode*, 8> PHIs;
854      if (!AllUsesOfValueWillTrapIfNull(LI, PHIs))
855        return false;
856    } else if (isa<StoreInst>(U)) {
857      // Ignore stores to the global.
858    } else {
859      // We don't know or understand this user, bail out.
860      //cerr << "UNKNOWN USER OF GLOBAL!: " << *U;
861      return false;
862    }
863  }
864  return true;
865}
866
867static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) {
868  bool Changed = false;
869  for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) {
870    Instruction *I = cast<Instruction>(*UI++);
871    if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
872      LI->setOperand(0, NewV);
873      Changed = true;
874    } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
875      if (SI->getOperand(1) == V) {
876        SI->setOperand(1, NewV);
877        Changed = true;
878      }
879    } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
880      CallSite CS(I);
881      if (CS.getCalledValue() == V) {
882        // Calling through the pointer!  Turn into a direct call, but be careful
883        // that the pointer is not also being passed as an argument.
884        CS.setCalledFunction(NewV);
885        Changed = true;
886        bool PassedAsArg = false;
887        for (unsigned i = 0, e = CS.arg_size(); i != e; ++i)
888          if (CS.getArgument(i) == V) {
889            PassedAsArg = true;
890            CS.setArgument(i, NewV);
891          }
892
893        if (PassedAsArg) {
894          // Being passed as an argument also.  Be careful to not invalidate UI!
895          UI = V->use_begin();
896        }
897      }
898    } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
899      Changed |= OptimizeAwayTrappingUsesOfValue(CI,
900                                ConstantExpr::getCast(CI->getOpcode(),
901                                                      NewV, CI->getType()));
902      if (CI->use_empty()) {
903        Changed = true;
904        CI->eraseFromParent();
905      }
906    } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
907      // Should handle GEP here.
908      SmallVector<Constant*, 8> Idxs;
909      Idxs.reserve(GEPI->getNumOperands()-1);
910      for (User::op_iterator i = GEPI->op_begin() + 1, e = GEPI->op_end();
911           i != e; ++i)
912        if (Constant *C = dyn_cast<Constant>(*i))
913          Idxs.push_back(C);
914        else
915          break;
916      if (Idxs.size() == GEPI->getNumOperands()-1)
917        Changed |= OptimizeAwayTrappingUsesOfValue(GEPI,
918                          ConstantExpr::getGetElementPtr(NewV, Idxs));
919      if (GEPI->use_empty()) {
920        Changed = true;
921        GEPI->eraseFromParent();
922      }
923    }
924  }
925
926  return Changed;
927}
928
929
930/// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null
931/// value stored into it.  If there are uses of the loaded value that would trap
932/// if the loaded value is dynamically null, then we know that they cannot be
933/// reachable with a null optimize away the load.
934static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV,
935                                            TargetData *TD,
936                                            TargetLibraryInfo *TLI) {
937  bool Changed = false;
938
939  // Keep track of whether we are able to remove all the uses of the global
940  // other than the store that defines it.
941  bool AllNonStoreUsesGone = true;
942
943  // Replace all uses of loads with uses of uses of the stored value.
944  for (Value::use_iterator GUI = GV->use_begin(), E = GV->use_end(); GUI != E;){
945    User *GlobalUser = *GUI++;
946    if (LoadInst *LI = dyn_cast<LoadInst>(GlobalUser)) {
947      Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV);
948      // If we were able to delete all uses of the loads
949      if (LI->use_empty()) {
950        LI->eraseFromParent();
951        Changed = true;
952      } else {
953        AllNonStoreUsesGone = false;
954      }
955    } else if (isa<StoreInst>(GlobalUser)) {
956      // Ignore the store that stores "LV" to the global.
957      assert(GlobalUser->getOperand(1) == GV &&
958             "Must be storing *to* the global");
959    } else {
960      AllNonStoreUsesGone = false;
961
962      // If we get here we could have other crazy uses that are transitively
963      // loaded.
964      assert((isa<PHINode>(GlobalUser) || isa<SelectInst>(GlobalUser) ||
965              isa<ConstantExpr>(GlobalUser) || isa<CmpInst>(GlobalUser)) &&
966             "Only expect load and stores!");
967    }
968  }
969
970  if (Changed) {
971    DEBUG(dbgs() << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV);
972    ++NumGlobUses;
973  }
974
975  // If we nuked all of the loads, then none of the stores are needed either,
976  // nor is the global.
977  if (AllNonStoreUsesGone) {
978    if (isLeakCheckerRoot(GV)) {
979      Changed |= CleanupPointerRootUsers(GV, TLI);
980    } else {
981      Changed = true;
982      CleanupConstantGlobalUsers(GV, 0, TD, TLI);
983    }
984    if (GV->use_empty()) {
985      DEBUG(dbgs() << "  *** GLOBAL NOW DEAD!\n");
986      Changed = true;
987      GV->eraseFromParent();
988      ++NumDeleted;
989    }
990  }
991  return Changed;
992}
993
994/// ConstantPropUsersOf - Walk the use list of V, constant folding all of the
995/// instructions that are foldable.
996static void ConstantPropUsersOf(Value *V,
997                                TargetData *TD, TargetLibraryInfo *TLI) {
998  for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; )
999    if (Instruction *I = dyn_cast<Instruction>(*UI++))
1000      if (Constant *NewC = ConstantFoldInstruction(I, TD, TLI)) {
1001        I->replaceAllUsesWith(NewC);
1002
1003        // Advance UI to the next non-I use to avoid invalidating it!
1004        // Instructions could multiply use V.
1005        while (UI != E && *UI == I)
1006          ++UI;
1007        I->eraseFromParent();
1008      }
1009}
1010
1011/// OptimizeGlobalAddressOfMalloc - This function takes the specified global
1012/// variable, and transforms the program as if it always contained the result of
1013/// the specified malloc.  Because it is always the result of the specified
1014/// malloc, there is no reason to actually DO the malloc.  Instead, turn the
1015/// malloc into a global, and any loads of GV as uses of the new global.
1016static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV,
1017                                                     CallInst *CI,
1018                                                     Type *AllocTy,
1019                                                     ConstantInt *NElements,
1020                                                     TargetData *TD,
1021                                                     TargetLibraryInfo *TLI) {
1022  DEBUG(errs() << "PROMOTING GLOBAL: " << *GV << "  CALL = " << *CI << '\n');
1023
1024  Type *GlobalType;
1025  if (NElements->getZExtValue() == 1)
1026    GlobalType = AllocTy;
1027  else
1028    // If we have an array allocation, the global variable is of an array.
1029    GlobalType = ArrayType::get(AllocTy, NElements->getZExtValue());
1030
1031  // Create the new global variable.  The contents of the malloc'd memory is
1032  // undefined, so initialize with an undef value.
1033  GlobalVariable *NewGV = new GlobalVariable(*GV->getParent(),
1034                                             GlobalType, false,
1035                                             GlobalValue::InternalLinkage,
1036                                             UndefValue::get(GlobalType),
1037                                             GV->getName()+".body",
1038                                             GV,
1039                                             GV->getThreadLocalMode());
1040
1041  // If there are bitcast users of the malloc (which is typical, usually we have
1042  // a malloc + bitcast) then replace them with uses of the new global.  Update
1043  // other users to use the global as well.
1044  BitCastInst *TheBC = 0;
1045  while (!CI->use_empty()) {
1046    Instruction *User = cast<Instruction>(CI->use_back());
1047    if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
1048      if (BCI->getType() == NewGV->getType()) {
1049        BCI->replaceAllUsesWith(NewGV);
1050        BCI->eraseFromParent();
1051      } else {
1052        BCI->setOperand(0, NewGV);
1053      }
1054    } else {
1055      if (TheBC == 0)
1056        TheBC = new BitCastInst(NewGV, CI->getType(), "newgv", CI);
1057      User->replaceUsesOfWith(CI, TheBC);
1058    }
1059  }
1060
1061  Constant *RepValue = NewGV;
1062  if (NewGV->getType() != GV->getType()->getElementType())
1063    RepValue = ConstantExpr::getBitCast(RepValue,
1064                                        GV->getType()->getElementType());
1065
1066  // If there is a comparison against null, we will insert a global bool to
1067  // keep track of whether the global was initialized yet or not.
1068  GlobalVariable *InitBool =
1069    new GlobalVariable(Type::getInt1Ty(GV->getContext()), false,
1070                       GlobalValue::InternalLinkage,
1071                       ConstantInt::getFalse(GV->getContext()),
1072                       GV->getName()+".init", GV->getThreadLocalMode());
1073  bool InitBoolUsed = false;
1074
1075  // Loop over all uses of GV, processing them in turn.
1076  while (!GV->use_empty()) {
1077    if (StoreInst *SI = dyn_cast<StoreInst>(GV->use_back())) {
1078      // The global is initialized when the store to it occurs.
1079      new StoreInst(ConstantInt::getTrue(GV->getContext()), InitBool, false, 0,
1080                    SI->getOrdering(), SI->getSynchScope(), SI);
1081      SI->eraseFromParent();
1082      continue;
1083    }
1084
1085    LoadInst *LI = cast<LoadInst>(GV->use_back());
1086    while (!LI->use_empty()) {
1087      Use &LoadUse = LI->use_begin().getUse();
1088      if (!isa<ICmpInst>(LoadUse.getUser())) {
1089        LoadUse = RepValue;
1090        continue;
1091      }
1092
1093      ICmpInst *ICI = cast<ICmpInst>(LoadUse.getUser());
1094      // Replace the cmp X, 0 with a use of the bool value.
1095      // Sink the load to where the compare was, if atomic rules allow us to.
1096      Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", false, 0,
1097                               LI->getOrdering(), LI->getSynchScope(),
1098                               LI->isUnordered() ? (Instruction*)ICI : LI);
1099      InitBoolUsed = true;
1100      switch (ICI->getPredicate()) {
1101      default: llvm_unreachable("Unknown ICmp Predicate!");
1102      case ICmpInst::ICMP_ULT:
1103      case ICmpInst::ICMP_SLT:   // X < null -> always false
1104        LV = ConstantInt::getFalse(GV->getContext());
1105        break;
1106      case ICmpInst::ICMP_ULE:
1107      case ICmpInst::ICMP_SLE:
1108      case ICmpInst::ICMP_EQ:
1109        LV = BinaryOperator::CreateNot(LV, "notinit", ICI);
1110        break;
1111      case ICmpInst::ICMP_NE:
1112      case ICmpInst::ICMP_UGE:
1113      case ICmpInst::ICMP_SGE:
1114      case ICmpInst::ICMP_UGT:
1115      case ICmpInst::ICMP_SGT:
1116        break;  // no change.
1117      }
1118      ICI->replaceAllUsesWith(LV);
1119      ICI->eraseFromParent();
1120    }
1121    LI->eraseFromParent();
1122  }
1123
1124  // If the initialization boolean was used, insert it, otherwise delete it.
1125  if (!InitBoolUsed) {
1126    while (!InitBool->use_empty())  // Delete initializations
1127      cast<StoreInst>(InitBool->use_back())->eraseFromParent();
1128    delete InitBool;
1129  } else
1130    GV->getParent()->getGlobalList().insert(GV, InitBool);
1131
1132  // Now the GV is dead, nuke it and the malloc..
1133  GV->eraseFromParent();
1134  CI->eraseFromParent();
1135
1136  // To further other optimizations, loop over all users of NewGV and try to
1137  // constant prop them.  This will promote GEP instructions with constant
1138  // indices into GEP constant-exprs, which will allow global-opt to hack on it.
1139  ConstantPropUsersOf(NewGV, TD, TLI);
1140  if (RepValue != NewGV)
1141    ConstantPropUsersOf(RepValue, TD, TLI);
1142
1143  return NewGV;
1144}
1145
1146/// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking
1147/// to make sure that there are no complex uses of V.  We permit simple things
1148/// like dereferencing the pointer, but not storing through the address, unless
1149/// it is to the specified global.
1150static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(const Instruction *V,
1151                                                      const GlobalVariable *GV,
1152                                         SmallPtrSet<const PHINode*, 8> &PHIs) {
1153  for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end();
1154       UI != E; ++UI) {
1155    const Instruction *Inst = cast<Instruction>(*UI);
1156
1157    if (isa<LoadInst>(Inst) || isa<CmpInst>(Inst)) {
1158      continue; // Fine, ignore.
1159    }
1160
1161    if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
1162      if (SI->getOperand(0) == V && SI->getOperand(1) != GV)
1163        return false;  // Storing the pointer itself... bad.
1164      continue; // Otherwise, storing through it, or storing into GV... fine.
1165    }
1166
1167    // Must index into the array and into the struct.
1168    if (isa<GetElementPtrInst>(Inst) && Inst->getNumOperands() >= 3) {
1169      if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Inst, GV, PHIs))
1170        return false;
1171      continue;
1172    }
1173
1174    if (const PHINode *PN = dyn_cast<PHINode>(Inst)) {
1175      // PHIs are ok if all uses are ok.  Don't infinitely recurse through PHI
1176      // cycles.
1177      if (PHIs.insert(PN))
1178        if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(PN, GV, PHIs))
1179          return false;
1180      continue;
1181    }
1182
1183    if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Inst)) {
1184      if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(BCI, GV, PHIs))
1185        return false;
1186      continue;
1187    }
1188
1189    return false;
1190  }
1191  return true;
1192}
1193
1194/// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV
1195/// somewhere.  Transform all uses of the allocation into loads from the
1196/// global and uses of the resultant pointer.  Further, delete the store into
1197/// GV.  This assumes that these value pass the
1198/// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate.
1199static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc,
1200                                          GlobalVariable *GV) {
1201  while (!Alloc->use_empty()) {
1202    Instruction *U = cast<Instruction>(*Alloc->use_begin());
1203    Instruction *InsertPt = U;
1204    if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
1205      // If this is the store of the allocation into the global, remove it.
1206      if (SI->getOperand(1) == GV) {
1207        SI->eraseFromParent();
1208        continue;
1209      }
1210    } else if (PHINode *PN = dyn_cast<PHINode>(U)) {
1211      // Insert the load in the corresponding predecessor, not right before the
1212      // PHI.
1213      InsertPt = PN->getIncomingBlock(Alloc->use_begin())->getTerminator();
1214    } else if (isa<BitCastInst>(U)) {
1215      // Must be bitcast between the malloc and store to initialize the global.
1216      ReplaceUsesOfMallocWithGlobal(U, GV);
1217      U->eraseFromParent();
1218      continue;
1219    } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
1220      // If this is a "GEP bitcast" and the user is a store to the global, then
1221      // just process it as a bitcast.
1222      if (GEPI->hasAllZeroIndices() && GEPI->hasOneUse())
1223        if (StoreInst *SI = dyn_cast<StoreInst>(GEPI->use_back()))
1224          if (SI->getOperand(1) == GV) {
1225            // Must be bitcast GEP between the malloc and store to initialize
1226            // the global.
1227            ReplaceUsesOfMallocWithGlobal(GEPI, GV);
1228            GEPI->eraseFromParent();
1229            continue;
1230          }
1231    }
1232
1233    // Insert a load from the global, and use it instead of the malloc.
1234    Value *NL = new LoadInst(GV, GV->getName()+".val", InsertPt);
1235    U->replaceUsesOfWith(Alloc, NL);
1236  }
1237}
1238
1239/// LoadUsesSimpleEnoughForHeapSRA - Verify that all uses of V (a load, or a phi
1240/// of a load) are simple enough to perform heap SRA on.  This permits GEP's
1241/// that index through the array and struct field, icmps of null, and PHIs.
1242static bool LoadUsesSimpleEnoughForHeapSRA(const Value *V,
1243                        SmallPtrSet<const PHINode*, 32> &LoadUsingPHIs,
1244                        SmallPtrSet<const PHINode*, 32> &LoadUsingPHIsPerLoad) {
1245  // We permit two users of the load: setcc comparing against the null
1246  // pointer, and a getelementptr of a specific form.
1247  for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;
1248       ++UI) {
1249    const Instruction *User = cast<Instruction>(*UI);
1250
1251    // Comparison against null is ok.
1252    if (const ICmpInst *ICI = dyn_cast<ICmpInst>(User)) {
1253      if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
1254        return false;
1255      continue;
1256    }
1257
1258    // getelementptr is also ok, but only a simple form.
1259    if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1260      // Must index into the array and into the struct.
1261      if (GEPI->getNumOperands() < 3)
1262        return false;
1263
1264      // Otherwise the GEP is ok.
1265      continue;
1266    }
1267
1268    if (const PHINode *PN = dyn_cast<PHINode>(User)) {
1269      if (!LoadUsingPHIsPerLoad.insert(PN))
1270        // This means some phi nodes are dependent on each other.
1271        // Avoid infinite looping!
1272        return false;
1273      if (!LoadUsingPHIs.insert(PN))
1274        // If we have already analyzed this PHI, then it is safe.
1275        continue;
1276
1277      // Make sure all uses of the PHI are simple enough to transform.
1278      if (!LoadUsesSimpleEnoughForHeapSRA(PN,
1279                                          LoadUsingPHIs, LoadUsingPHIsPerLoad))
1280        return false;
1281
1282      continue;
1283    }
1284
1285    // Otherwise we don't know what this is, not ok.
1286    return false;
1287  }
1288
1289  return true;
1290}
1291
1292
1293/// AllGlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from
1294/// GV are simple enough to perform HeapSRA, return true.
1295static bool AllGlobalLoadUsesSimpleEnoughForHeapSRA(const GlobalVariable *GV,
1296                                                    Instruction *StoredVal) {
1297  SmallPtrSet<const PHINode*, 32> LoadUsingPHIs;
1298  SmallPtrSet<const PHINode*, 32> LoadUsingPHIsPerLoad;
1299  for (Value::const_use_iterator UI = GV->use_begin(), E = GV->use_end();
1300       UI != E; ++UI)
1301    if (const LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
1302      if (!LoadUsesSimpleEnoughForHeapSRA(LI, LoadUsingPHIs,
1303                                          LoadUsingPHIsPerLoad))
1304        return false;
1305      LoadUsingPHIsPerLoad.clear();
1306    }
1307
1308  // If we reach here, we know that all uses of the loads and transitive uses
1309  // (through PHI nodes) are simple enough to transform.  However, we don't know
1310  // that all inputs the to the PHI nodes are in the same equivalence sets.
1311  // Check to verify that all operands of the PHIs are either PHIS that can be
1312  // transformed, loads from GV, or MI itself.
1313  for (SmallPtrSet<const PHINode*, 32>::const_iterator I = LoadUsingPHIs.begin()
1314       , E = LoadUsingPHIs.end(); I != E; ++I) {
1315    const PHINode *PN = *I;
1316    for (unsigned op = 0, e = PN->getNumIncomingValues(); op != e; ++op) {
1317      Value *InVal = PN->getIncomingValue(op);
1318
1319      // PHI of the stored value itself is ok.
1320      if (InVal == StoredVal) continue;
1321
1322      if (const PHINode *InPN = dyn_cast<PHINode>(InVal)) {
1323        // One of the PHIs in our set is (optimistically) ok.
1324        if (LoadUsingPHIs.count(InPN))
1325          continue;
1326        return false;
1327      }
1328
1329      // Load from GV is ok.
1330      if (const LoadInst *LI = dyn_cast<LoadInst>(InVal))
1331        if (LI->getOperand(0) == GV)
1332          continue;
1333
1334      // UNDEF? NULL?
1335
1336      // Anything else is rejected.
1337      return false;
1338    }
1339  }
1340
1341  return true;
1342}
1343
1344static Value *GetHeapSROAValue(Value *V, unsigned FieldNo,
1345               DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1346                   std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1347  std::vector<Value*> &FieldVals = InsertedScalarizedValues[V];
1348
1349  if (FieldNo >= FieldVals.size())
1350    FieldVals.resize(FieldNo+1);
1351
1352  // If we already have this value, just reuse the previously scalarized
1353  // version.
1354  if (Value *FieldVal = FieldVals[FieldNo])
1355    return FieldVal;
1356
1357  // Depending on what instruction this is, we have several cases.
1358  Value *Result;
1359  if (LoadInst *LI = dyn_cast<LoadInst>(V)) {
1360    // This is a scalarized version of the load from the global.  Just create
1361    // a new Load of the scalarized global.
1362    Result = new LoadInst(GetHeapSROAValue(LI->getOperand(0), FieldNo,
1363                                           InsertedScalarizedValues,
1364                                           PHIsToRewrite),
1365                          LI->getName()+".f"+Twine(FieldNo), LI);
1366  } else if (PHINode *PN = dyn_cast<PHINode>(V)) {
1367    // PN's type is pointer to struct.  Make a new PHI of pointer to struct
1368    // field.
1369    StructType *ST =
1370      cast<StructType>(cast<PointerType>(PN->getType())->getElementType());
1371
1372    PHINode *NewPN =
1373     PHINode::Create(PointerType::getUnqual(ST->getElementType(FieldNo)),
1374                     PN->getNumIncomingValues(),
1375                     PN->getName()+".f"+Twine(FieldNo), PN);
1376    Result = NewPN;
1377    PHIsToRewrite.push_back(std::make_pair(PN, FieldNo));
1378  } else {
1379    llvm_unreachable("Unknown usable value");
1380  }
1381
1382  return FieldVals[FieldNo] = Result;
1383}
1384
1385/// RewriteHeapSROALoadUser - Given a load instruction and a value derived from
1386/// the load, rewrite the derived value to use the HeapSRoA'd load.
1387static void RewriteHeapSROALoadUser(Instruction *LoadUser,
1388             DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1389                   std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1390  // If this is a comparison against null, handle it.
1391  if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) {
1392    assert(isa<ConstantPointerNull>(SCI->getOperand(1)));
1393    // If we have a setcc of the loaded pointer, we can use a setcc of any
1394    // field.
1395    Value *NPtr = GetHeapSROAValue(SCI->getOperand(0), 0,
1396                                   InsertedScalarizedValues, PHIsToRewrite);
1397
1398    Value *New = new ICmpInst(SCI, SCI->getPredicate(), NPtr,
1399                              Constant::getNullValue(NPtr->getType()),
1400                              SCI->getName());
1401    SCI->replaceAllUsesWith(New);
1402    SCI->eraseFromParent();
1403    return;
1404  }
1405
1406  // Handle 'getelementptr Ptr, Idx, i32 FieldNo ...'
1407  if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) {
1408    assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2))
1409           && "Unexpected GEPI!");
1410
1411    // Load the pointer for this field.
1412    unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
1413    Value *NewPtr = GetHeapSROAValue(GEPI->getOperand(0), FieldNo,
1414                                     InsertedScalarizedValues, PHIsToRewrite);
1415
1416    // Create the new GEP idx vector.
1417    SmallVector<Value*, 8> GEPIdx;
1418    GEPIdx.push_back(GEPI->getOperand(1));
1419    GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end());
1420
1421    Value *NGEPI = GetElementPtrInst::Create(NewPtr, GEPIdx,
1422                                             GEPI->getName(), GEPI);
1423    GEPI->replaceAllUsesWith(NGEPI);
1424    GEPI->eraseFromParent();
1425    return;
1426  }
1427
1428  // Recursively transform the users of PHI nodes.  This will lazily create the
1429  // PHIs that are needed for individual elements.  Keep track of what PHIs we
1430  // see in InsertedScalarizedValues so that we don't get infinite loops (very
1431  // antisocial).  If the PHI is already in InsertedScalarizedValues, it has
1432  // already been seen first by another load, so its uses have already been
1433  // processed.
1434  PHINode *PN = cast<PHINode>(LoadUser);
1435  if (!InsertedScalarizedValues.insert(std::make_pair(PN,
1436                                              std::vector<Value*>())).second)
1437    return;
1438
1439  // If this is the first time we've seen this PHI, recursively process all
1440  // users.
1441  for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end(); UI != E; ) {
1442    Instruction *User = cast<Instruction>(*UI++);
1443    RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite);
1444  }
1445}
1446
1447/// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global.  Ptr
1448/// is a value loaded from the global.  Eliminate all uses of Ptr, making them
1449/// use FieldGlobals instead.  All uses of loaded values satisfy
1450/// AllGlobalLoadUsesSimpleEnoughForHeapSRA.
1451static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load,
1452               DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1453                   std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1454  for (Value::use_iterator UI = Load->use_begin(), E = Load->use_end();
1455       UI != E; ) {
1456    Instruction *User = cast<Instruction>(*UI++);
1457    RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite);
1458  }
1459
1460  if (Load->use_empty()) {
1461    Load->eraseFromParent();
1462    InsertedScalarizedValues.erase(Load);
1463  }
1464}
1465
1466/// PerformHeapAllocSRoA - CI is an allocation of an array of structures.  Break
1467/// it up into multiple allocations of arrays of the fields.
1468static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, CallInst *CI,
1469                                            Value *NElems, TargetData *TD,
1470                                            const TargetLibraryInfo *TLI) {
1471  DEBUG(dbgs() << "SROA HEAP ALLOC: " << *GV << "  MALLOC = " << *CI << '\n');
1472  Type *MAT = getMallocAllocatedType(CI, TLI);
1473  StructType *STy = cast<StructType>(MAT);
1474
1475  // There is guaranteed to be at least one use of the malloc (storing
1476  // it into GV).  If there are other uses, change them to be uses of
1477  // the global to simplify later code.  This also deletes the store
1478  // into GV.
1479  ReplaceUsesOfMallocWithGlobal(CI, GV);
1480
1481  // Okay, at this point, there are no users of the malloc.  Insert N
1482  // new mallocs at the same place as CI, and N globals.
1483  std::vector<Value*> FieldGlobals;
1484  std::vector<Value*> FieldMallocs;
1485
1486  for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){
1487    Type *FieldTy = STy->getElementType(FieldNo);
1488    PointerType *PFieldTy = PointerType::getUnqual(FieldTy);
1489
1490    GlobalVariable *NGV =
1491      new GlobalVariable(*GV->getParent(),
1492                         PFieldTy, false, GlobalValue::InternalLinkage,
1493                         Constant::getNullValue(PFieldTy),
1494                         GV->getName() + ".f" + Twine(FieldNo), GV,
1495                         GV->getThreadLocalMode());
1496    FieldGlobals.push_back(NGV);
1497
1498    unsigned TypeSize = TD->getTypeAllocSize(FieldTy);
1499    if (StructType *ST = dyn_cast<StructType>(FieldTy))
1500      TypeSize = TD->getStructLayout(ST)->getSizeInBytes();
1501    Type *IntPtrTy = TD->getIntPtrType(CI->getContext());
1502    Value *NMI = CallInst::CreateMalloc(CI, IntPtrTy, FieldTy,
1503                                        ConstantInt::get(IntPtrTy, TypeSize),
1504                                        NElems, 0,
1505                                        CI->getName() + ".f" + Twine(FieldNo));
1506    FieldMallocs.push_back(NMI);
1507    new StoreInst(NMI, NGV, CI);
1508  }
1509
1510  // The tricky aspect of this transformation is handling the case when malloc
1511  // fails.  In the original code, malloc failing would set the result pointer
1512  // of malloc to null.  In this case, some mallocs could succeed and others
1513  // could fail.  As such, we emit code that looks like this:
1514  //    F0 = malloc(field0)
1515  //    F1 = malloc(field1)
1516  //    F2 = malloc(field2)
1517  //    if (F0 == 0 || F1 == 0 || F2 == 0) {
1518  //      if (F0) { free(F0); F0 = 0; }
1519  //      if (F1) { free(F1); F1 = 0; }
1520  //      if (F2) { free(F2); F2 = 0; }
1521  //    }
1522  // The malloc can also fail if its argument is too large.
1523  Constant *ConstantZero = ConstantInt::get(CI->getArgOperand(0)->getType(), 0);
1524  Value *RunningOr = new ICmpInst(CI, ICmpInst::ICMP_SLT, CI->getArgOperand(0),
1525                                  ConstantZero, "isneg");
1526  for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) {
1527    Value *Cond = new ICmpInst(CI, ICmpInst::ICMP_EQ, FieldMallocs[i],
1528                             Constant::getNullValue(FieldMallocs[i]->getType()),
1529                               "isnull");
1530    RunningOr = BinaryOperator::CreateOr(RunningOr, Cond, "tmp", CI);
1531  }
1532
1533  // Split the basic block at the old malloc.
1534  BasicBlock *OrigBB = CI->getParent();
1535  BasicBlock *ContBB = OrigBB->splitBasicBlock(CI, "malloc_cont");
1536
1537  // Create the block to check the first condition.  Put all these blocks at the
1538  // end of the function as they are unlikely to be executed.
1539  BasicBlock *NullPtrBlock = BasicBlock::Create(OrigBB->getContext(),
1540                                                "malloc_ret_null",
1541                                                OrigBB->getParent());
1542
1543  // Remove the uncond branch from OrigBB to ContBB, turning it into a cond
1544  // branch on RunningOr.
1545  OrigBB->getTerminator()->eraseFromParent();
1546  BranchInst::Create(NullPtrBlock, ContBB, RunningOr, OrigBB);
1547
1548  // Within the NullPtrBlock, we need to emit a comparison and branch for each
1549  // pointer, because some may be null while others are not.
1550  for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1551    Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock);
1552    Value *Cmp = new ICmpInst(*NullPtrBlock, ICmpInst::ICMP_NE, GVVal,
1553                              Constant::getNullValue(GVVal->getType()));
1554    BasicBlock *FreeBlock = BasicBlock::Create(Cmp->getContext(), "free_it",
1555                                               OrigBB->getParent());
1556    BasicBlock *NextBlock = BasicBlock::Create(Cmp->getContext(), "next",
1557                                               OrigBB->getParent());
1558    Instruction *BI = BranchInst::Create(FreeBlock, NextBlock,
1559                                         Cmp, NullPtrBlock);
1560
1561    // Fill in FreeBlock.
1562    CallInst::CreateFree(GVVal, BI);
1563    new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i],
1564                  FreeBlock);
1565    BranchInst::Create(NextBlock, FreeBlock);
1566
1567    NullPtrBlock = NextBlock;
1568  }
1569
1570  BranchInst::Create(ContBB, NullPtrBlock);
1571
1572  // CI is no longer needed, remove it.
1573  CI->eraseFromParent();
1574
1575  /// InsertedScalarizedLoads - As we process loads, if we can't immediately
1576  /// update all uses of the load, keep track of what scalarized loads are
1577  /// inserted for a given load.
1578  DenseMap<Value*, std::vector<Value*> > InsertedScalarizedValues;
1579  InsertedScalarizedValues[GV] = FieldGlobals;
1580
1581  std::vector<std::pair<PHINode*, unsigned> > PHIsToRewrite;
1582
1583  // Okay, the malloc site is completely handled.  All of the uses of GV are now
1584  // loads, and all uses of those loads are simple.  Rewrite them to use loads
1585  // of the per-field globals instead.
1586  for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;) {
1587    Instruction *User = cast<Instruction>(*UI++);
1588
1589    if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1590      RewriteUsesOfLoadForHeapSRoA(LI, InsertedScalarizedValues, PHIsToRewrite);
1591      continue;
1592    }
1593
1594    // Must be a store of null.
1595    StoreInst *SI = cast<StoreInst>(User);
1596    assert(isa<ConstantPointerNull>(SI->getOperand(0)) &&
1597           "Unexpected heap-sra user!");
1598
1599    // Insert a store of null into each global.
1600    for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1601      PointerType *PT = cast<PointerType>(FieldGlobals[i]->getType());
1602      Constant *Null = Constant::getNullValue(PT->getElementType());
1603      new StoreInst(Null, FieldGlobals[i], SI);
1604    }
1605    // Erase the original store.
1606    SI->eraseFromParent();
1607  }
1608
1609  // While we have PHIs that are interesting to rewrite, do it.
1610  while (!PHIsToRewrite.empty()) {
1611    PHINode *PN = PHIsToRewrite.back().first;
1612    unsigned FieldNo = PHIsToRewrite.back().second;
1613    PHIsToRewrite.pop_back();
1614    PHINode *FieldPN = cast<PHINode>(InsertedScalarizedValues[PN][FieldNo]);
1615    assert(FieldPN->getNumIncomingValues() == 0 &&"Already processed this phi");
1616
1617    // Add all the incoming values.  This can materialize more phis.
1618    for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1619      Value *InVal = PN->getIncomingValue(i);
1620      InVal = GetHeapSROAValue(InVal, FieldNo, InsertedScalarizedValues,
1621                               PHIsToRewrite);
1622      FieldPN->addIncoming(InVal, PN->getIncomingBlock(i));
1623    }
1624  }
1625
1626  // Drop all inter-phi links and any loads that made it this far.
1627  for (DenseMap<Value*, std::vector<Value*> >::iterator
1628       I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1629       I != E; ++I) {
1630    if (PHINode *PN = dyn_cast<PHINode>(I->first))
1631      PN->dropAllReferences();
1632    else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1633      LI->dropAllReferences();
1634  }
1635
1636  // Delete all the phis and loads now that inter-references are dead.
1637  for (DenseMap<Value*, std::vector<Value*> >::iterator
1638       I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1639       I != E; ++I) {
1640    if (PHINode *PN = dyn_cast<PHINode>(I->first))
1641      PN->eraseFromParent();
1642    else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1643      LI->eraseFromParent();
1644  }
1645
1646  // The old global is now dead, remove it.
1647  GV->eraseFromParent();
1648
1649  ++NumHeapSRA;
1650  return cast<GlobalVariable>(FieldGlobals[0]);
1651}
1652
1653/// TryToOptimizeStoreOfMallocToGlobal - This function is called when we see a
1654/// pointer global variable with a single value stored it that is a malloc or
1655/// cast of malloc.
1656static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV,
1657                                               CallInst *CI,
1658                                               Type *AllocTy,
1659                                               AtomicOrdering Ordering,
1660                                               Module::global_iterator &GVI,
1661                                               TargetData *TD,
1662                                               TargetLibraryInfo *TLI) {
1663  if (!TD)
1664    return false;
1665
1666  // If this is a malloc of an abstract type, don't touch it.
1667  if (!AllocTy->isSized())
1668    return false;
1669
1670  // We can't optimize this global unless all uses of it are *known* to be
1671  // of the malloc value, not of the null initializer value (consider a use
1672  // that compares the global's value against zero to see if the malloc has
1673  // been reached).  To do this, we check to see if all uses of the global
1674  // would trap if the global were null: this proves that they must all
1675  // happen after the malloc.
1676  if (!AllUsesOfLoadedValueWillTrapIfNull(GV))
1677    return false;
1678
1679  // We can't optimize this if the malloc itself is used in a complex way,
1680  // for example, being stored into multiple globals.  This allows the
1681  // malloc to be stored into the specified global, loaded icmp'd, and
1682  // GEP'd.  These are all things we could transform to using the global
1683  // for.
1684  SmallPtrSet<const PHINode*, 8> PHIs;
1685  if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(CI, GV, PHIs))
1686    return false;
1687
1688  // If we have a global that is only initialized with a fixed size malloc,
1689  // transform the program to use global memory instead of malloc'd memory.
1690  // This eliminates dynamic allocation, avoids an indirection accessing the
1691  // data, and exposes the resultant global to further GlobalOpt.
1692  // We cannot optimize the malloc if we cannot determine malloc array size.
1693  Value *NElems = getMallocArraySize(CI, TD, TLI, true);
1694  if (!NElems)
1695    return false;
1696
1697  if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems))
1698    // Restrict this transformation to only working on small allocations
1699    // (2048 bytes currently), as we don't want to introduce a 16M global or
1700    // something.
1701    if (NElements->getZExtValue() * TD->getTypeAllocSize(AllocTy) < 2048) {
1702      GVI = OptimizeGlobalAddressOfMalloc(GV, CI, AllocTy, NElements, TD, TLI);
1703      return true;
1704    }
1705
1706  // If the allocation is an array of structures, consider transforming this
1707  // into multiple malloc'd arrays, one for each field.  This is basically
1708  // SRoA for malloc'd memory.
1709
1710  if (Ordering != NotAtomic)
1711    return false;
1712
1713  // If this is an allocation of a fixed size array of structs, analyze as a
1714  // variable size array.  malloc [100 x struct],1 -> malloc struct, 100
1715  if (NElems == ConstantInt::get(CI->getArgOperand(0)->getType(), 1))
1716    if (ArrayType *AT = dyn_cast<ArrayType>(AllocTy))
1717      AllocTy = AT->getElementType();
1718
1719  StructType *AllocSTy = dyn_cast<StructType>(AllocTy);
1720  if (!AllocSTy)
1721    return false;
1722
1723  // This the structure has an unreasonable number of fields, leave it
1724  // alone.
1725  if (AllocSTy->getNumElements() <= 16 && AllocSTy->getNumElements() != 0 &&
1726      AllGlobalLoadUsesSimpleEnoughForHeapSRA(GV, CI)) {
1727
1728    // If this is a fixed size array, transform the Malloc to be an alloc of
1729    // structs.  malloc [100 x struct],1 -> malloc struct, 100
1730    if (ArrayType *AT = dyn_cast<ArrayType>(getMallocAllocatedType(CI, TLI))) {
1731      Type *IntPtrTy = TD->getIntPtrType(CI->getContext());
1732      unsigned TypeSize = TD->getStructLayout(AllocSTy)->getSizeInBytes();
1733      Value *AllocSize = ConstantInt::get(IntPtrTy, TypeSize);
1734      Value *NumElements = ConstantInt::get(IntPtrTy, AT->getNumElements());
1735      Instruction *Malloc = CallInst::CreateMalloc(CI, IntPtrTy, AllocSTy,
1736                                                   AllocSize, NumElements,
1737                                                   0, CI->getName());
1738      Instruction *Cast = new BitCastInst(Malloc, CI->getType(), "tmp", CI);
1739      CI->replaceAllUsesWith(Cast);
1740      CI->eraseFromParent();
1741      if (BitCastInst *BCI = dyn_cast<BitCastInst>(Malloc))
1742        CI = cast<CallInst>(BCI->getOperand(0));
1743      else
1744        CI = cast<CallInst>(Malloc);
1745    }
1746
1747    GVI = PerformHeapAllocSRoA(GV, CI, getMallocArraySize(CI, TD, TLI, true),
1748                               TD, TLI);
1749    return true;
1750  }
1751
1752  return false;
1753}
1754
1755// OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge
1756// that only one value (besides its initializer) is ever stored to the global.
1757static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal,
1758                                     AtomicOrdering Ordering,
1759                                     Module::global_iterator &GVI,
1760                                     TargetData *TD, TargetLibraryInfo *TLI) {
1761  // Ignore no-op GEPs and bitcasts.
1762  StoredOnceVal = StoredOnceVal->stripPointerCasts();
1763
1764  // If we are dealing with a pointer global that is initialized to null and
1765  // only has one (non-null) value stored into it, then we can optimize any
1766  // users of the loaded value (often calls and loads) that would trap if the
1767  // value was null.
1768  if (GV->getInitializer()->getType()->isPointerTy() &&
1769      GV->getInitializer()->isNullValue()) {
1770    if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) {
1771      if (GV->getInitializer()->getType() != SOVC->getType())
1772        SOVC = ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType());
1773
1774      // Optimize away any trapping uses of the loaded value.
1775      if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC, TD, TLI))
1776        return true;
1777    } else if (CallInst *CI = extractMallocCall(StoredOnceVal, TLI)) {
1778      Type *MallocType = getMallocAllocatedType(CI, TLI);
1779      if (MallocType &&
1780          TryToOptimizeStoreOfMallocToGlobal(GV, CI, MallocType, Ordering, GVI,
1781                                             TD, TLI))
1782        return true;
1783    }
1784  }
1785
1786  return false;
1787}
1788
1789/// TryToShrinkGlobalToBoolean - At this point, we have learned that the only
1790/// two values ever stored into GV are its initializer and OtherVal.  See if we
1791/// can shrink the global into a boolean and select between the two values
1792/// whenever it is used.  This exposes the values to other scalar optimizations.
1793static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) {
1794  Type *GVElType = GV->getType()->getElementType();
1795
1796  // If GVElType is already i1, it is already shrunk.  If the type of the GV is
1797  // an FP value, pointer or vector, don't do this optimization because a select
1798  // between them is very expensive and unlikely to lead to later
1799  // simplification.  In these cases, we typically end up with "cond ? v1 : v2"
1800  // where v1 and v2 both require constant pool loads, a big loss.
1801  if (GVElType == Type::getInt1Ty(GV->getContext()) ||
1802      GVElType->isFloatingPointTy() ||
1803      GVElType->isPointerTy() || GVElType->isVectorTy())
1804    return false;
1805
1806  // Walk the use list of the global seeing if all the uses are load or store.
1807  // If there is anything else, bail out.
1808  for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I){
1809    User *U = *I;
1810    if (!isa<LoadInst>(U) && !isa<StoreInst>(U))
1811      return false;
1812  }
1813
1814  DEBUG(dbgs() << "   *** SHRINKING TO BOOL: " << *GV);
1815
1816  // Create the new global, initializing it to false.
1817  GlobalVariable *NewGV = new GlobalVariable(Type::getInt1Ty(GV->getContext()),
1818                                             false,
1819                                             GlobalValue::InternalLinkage,
1820                                        ConstantInt::getFalse(GV->getContext()),
1821                                             GV->getName()+".b",
1822                                             GV->getThreadLocalMode());
1823  GV->getParent()->getGlobalList().insert(GV, NewGV);
1824
1825  Constant *InitVal = GV->getInitializer();
1826  assert(InitVal->getType() != Type::getInt1Ty(GV->getContext()) &&
1827         "No reason to shrink to bool!");
1828
1829  // If initialized to zero and storing one into the global, we can use a cast
1830  // instead of a select to synthesize the desired value.
1831  bool IsOneZero = false;
1832  if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal))
1833    IsOneZero = InitVal->isNullValue() && CI->isOne();
1834
1835  while (!GV->use_empty()) {
1836    Instruction *UI = cast<Instruction>(GV->use_back());
1837    if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
1838      // Change the store into a boolean store.
1839      bool StoringOther = SI->getOperand(0) == OtherVal;
1840      // Only do this if we weren't storing a loaded value.
1841      Value *StoreVal;
1842      if (StoringOther || SI->getOperand(0) == InitVal)
1843        StoreVal = ConstantInt::get(Type::getInt1Ty(GV->getContext()),
1844                                    StoringOther);
1845      else {
1846        // Otherwise, we are storing a previously loaded copy.  To do this,
1847        // change the copy from copying the original value to just copying the
1848        // bool.
1849        Instruction *StoredVal = cast<Instruction>(SI->getOperand(0));
1850
1851        // If we've already replaced the input, StoredVal will be a cast or
1852        // select instruction.  If not, it will be a load of the original
1853        // global.
1854        if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
1855          assert(LI->getOperand(0) == GV && "Not a copy!");
1856          // Insert a new load, to preserve the saved value.
1857          StoreVal = new LoadInst(NewGV, LI->getName()+".b", false, 0,
1858                                  LI->getOrdering(), LI->getSynchScope(), LI);
1859        } else {
1860          assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) &&
1861                 "This is not a form that we understand!");
1862          StoreVal = StoredVal->getOperand(0);
1863          assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!");
1864        }
1865      }
1866      new StoreInst(StoreVal, NewGV, false, 0,
1867                    SI->getOrdering(), SI->getSynchScope(), SI);
1868    } else {
1869      // Change the load into a load of bool then a select.
1870      LoadInst *LI = cast<LoadInst>(UI);
1871      LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", false, 0,
1872                                   LI->getOrdering(), LI->getSynchScope(), LI);
1873      Value *NSI;
1874      if (IsOneZero)
1875        NSI = new ZExtInst(NLI, LI->getType(), "", LI);
1876      else
1877        NSI = SelectInst::Create(NLI, OtherVal, InitVal, "", LI);
1878      NSI->takeName(LI);
1879      LI->replaceAllUsesWith(NSI);
1880    }
1881    UI->eraseFromParent();
1882  }
1883
1884  GV->eraseFromParent();
1885  return true;
1886}
1887
1888
1889/// ProcessGlobal - Analyze the specified global variable and optimize it if
1890/// possible.  If we make a change, return true.
1891bool GlobalOpt::ProcessGlobal(GlobalVariable *GV,
1892                              Module::global_iterator &GVI) {
1893  if (!GV->isDiscardableIfUnused())
1894    return false;
1895
1896  // Do more involved optimizations if the global is internal.
1897  GV->removeDeadConstantUsers();
1898
1899  if (GV->use_empty()) {
1900    DEBUG(dbgs() << "GLOBAL DEAD: " << *GV);
1901    GV->eraseFromParent();
1902    ++NumDeleted;
1903    return true;
1904  }
1905
1906  if (!GV->hasLocalLinkage())
1907    return false;
1908
1909  SmallPtrSet<const PHINode*, 16> PHIUsers;
1910  GlobalStatus GS;
1911
1912  if (AnalyzeGlobal(GV, GS, PHIUsers))
1913    return false;
1914
1915  if (!GS.isCompared && !GV->hasUnnamedAddr()) {
1916    GV->setUnnamedAddr(true);
1917    NumUnnamed++;
1918  }
1919
1920  if (GV->isConstant() || !GV->hasInitializer())
1921    return false;
1922
1923  return ProcessInternalGlobal(GV, GVI, PHIUsers, GS);
1924}
1925
1926/// ProcessInternalGlobal - Analyze the specified global variable and optimize
1927/// it if possible.  If we make a change, return true.
1928bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV,
1929                                      Module::global_iterator &GVI,
1930                                const SmallPtrSet<const PHINode*, 16> &PHIUsers,
1931                                      const GlobalStatus &GS) {
1932  // If this is a first class global and has only one accessing function
1933  // and this function is main (which we know is not recursive we can make
1934  // this global a local variable) we replace the global with a local alloca
1935  // in this function.
1936  //
1937  // NOTE: It doesn't make sense to promote non single-value types since we
1938  // are just replacing static memory to stack memory.
1939  //
1940  // If the global is in different address space, don't bring it to stack.
1941  if (!GS.HasMultipleAccessingFunctions &&
1942      GS.AccessingFunction && !GS.HasNonInstructionUser &&
1943      GV->getType()->getElementType()->isSingleValueType() &&
1944      GS.AccessingFunction->getName() == "main" &&
1945      GS.AccessingFunction->hasExternalLinkage() &&
1946      GV->getType()->getAddressSpace() == 0) {
1947    DEBUG(dbgs() << "LOCALIZING GLOBAL: " << *GV);
1948    Instruction &FirstI = const_cast<Instruction&>(*GS.AccessingFunction
1949                                                   ->getEntryBlock().begin());
1950    Type *ElemTy = GV->getType()->getElementType();
1951    // FIXME: Pass Global's alignment when globals have alignment
1952    AllocaInst *Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), &FirstI);
1953    if (!isa<UndefValue>(GV->getInitializer()))
1954      new StoreInst(GV->getInitializer(), Alloca, &FirstI);
1955
1956    GV->replaceAllUsesWith(Alloca);
1957    GV->eraseFromParent();
1958    ++NumLocalized;
1959    return true;
1960  }
1961
1962  // If the global is never loaded (but may be stored to), it is dead.
1963  // Delete it now.
1964  if (!GS.isLoaded) {
1965    DEBUG(dbgs() << "GLOBAL NEVER LOADED: " << *GV);
1966
1967    bool Changed;
1968    if (isLeakCheckerRoot(GV)) {
1969      // Delete any constant stores to the global.
1970      Changed = CleanupPointerRootUsers(GV, TLI);
1971    } else {
1972      // Delete any stores we can find to the global.  We may not be able to
1973      // make it completely dead though.
1974      Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer(), TD, TLI);
1975    }
1976
1977    // If the global is dead now, delete it.
1978    if (GV->use_empty()) {
1979      GV->eraseFromParent();
1980      ++NumDeleted;
1981      Changed = true;
1982    }
1983    return Changed;
1984
1985  } else if (GS.StoredType <= GlobalStatus::isInitializerStored) {
1986    DEBUG(dbgs() << "MARKING CONSTANT: " << *GV);
1987    GV->setConstant(true);
1988
1989    // Clean up any obviously simplifiable users now.
1990    CleanupConstantGlobalUsers(GV, GV->getInitializer(), TD, TLI);
1991
1992    // If the global is dead now, just nuke it.
1993    if (GV->use_empty()) {
1994      DEBUG(dbgs() << "   *** Marking constant allowed us to simplify "
1995            << "all users and delete global!\n");
1996      GV->eraseFromParent();
1997      ++NumDeleted;
1998    }
1999
2000    ++NumMarked;
2001    return true;
2002  } else if (!GV->getInitializer()->getType()->isSingleValueType()) {
2003    if (TargetData *TD = getAnalysisIfAvailable<TargetData>())
2004      if (GlobalVariable *FirstNewGV = SRAGlobal(GV, *TD)) {
2005        GVI = FirstNewGV;  // Don't skip the newly produced globals!
2006        return true;
2007      }
2008  } else if (GS.StoredType == GlobalStatus::isStoredOnce) {
2009    // If the initial value for the global was an undef value, and if only
2010    // one other value was stored into it, we can just change the
2011    // initializer to be the stored value, then delete all stores to the
2012    // global.  This allows us to mark it constant.
2013    if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
2014      if (isa<UndefValue>(GV->getInitializer())) {
2015        // Change the initial value here.
2016        GV->setInitializer(SOVConstant);
2017
2018        // Clean up any obviously simplifiable users now.
2019        CleanupConstantGlobalUsers(GV, GV->getInitializer(), TD, TLI);
2020
2021        if (GV->use_empty()) {
2022          DEBUG(dbgs() << "   *** Substituting initializer allowed us to "
2023                       << "simplify all users and delete global!\n");
2024          GV->eraseFromParent();
2025          ++NumDeleted;
2026        } else {
2027          GVI = GV;
2028        }
2029        ++NumSubstitute;
2030        return true;
2031      }
2032
2033    // Try to optimize globals based on the knowledge that only one value
2034    // (besides its initializer) is ever stored to the global.
2035    if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GS.Ordering, GVI,
2036                                 TD, TLI))
2037      return true;
2038
2039    // Otherwise, if the global was not a boolean, we can shrink it to be a
2040    // boolean.
2041    if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
2042      if (TryToShrinkGlobalToBoolean(GV, SOVConstant)) {
2043        ++NumShrunkToBool;
2044        return true;
2045      }
2046  }
2047
2048  return false;
2049}
2050
2051/// ChangeCalleesToFastCall - Walk all of the direct calls of the specified
2052/// function, changing them to FastCC.
2053static void ChangeCalleesToFastCall(Function *F) {
2054  for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
2055    if (isa<BlockAddress>(*UI))
2056      continue;
2057    CallSite User(cast<Instruction>(*UI));
2058    User.setCallingConv(CallingConv::Fast);
2059  }
2060}
2061
2062static AttrListPtr StripNest(const AttrListPtr &Attrs) {
2063  for (unsigned i = 0, e = Attrs.getNumSlots(); i != e; ++i) {
2064    if ((Attrs.getSlot(i).Attrs & Attribute::Nest) == 0)
2065      continue;
2066
2067    // There can be only one.
2068    return Attrs.removeAttr(Attrs.getSlot(i).Index, Attribute::Nest);
2069  }
2070
2071  return Attrs;
2072}
2073
2074static void RemoveNestAttribute(Function *F) {
2075  F->setAttributes(StripNest(F->getAttributes()));
2076  for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
2077    if (isa<BlockAddress>(*UI))
2078      continue;
2079    CallSite User(cast<Instruction>(*UI));
2080    User.setAttributes(StripNest(User.getAttributes()));
2081  }
2082}
2083
2084bool GlobalOpt::OptimizeFunctions(Module &M) {
2085  bool Changed = false;
2086  // Optimize functions.
2087  for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) {
2088    Function *F = FI++;
2089    // Functions without names cannot be referenced outside this module.
2090    if (!F->hasName() && !F->isDeclaration())
2091      F->setLinkage(GlobalValue::InternalLinkage);
2092    F->removeDeadConstantUsers();
2093    if (F->isDefTriviallyDead()) {
2094      F->eraseFromParent();
2095      Changed = true;
2096      ++NumFnDeleted;
2097    } else if (F->hasLocalLinkage()) {
2098      if (F->getCallingConv() == CallingConv::C && !F->isVarArg() &&
2099          !F->hasAddressTaken()) {
2100        // If this function has C calling conventions, is not a varargs
2101        // function, and is only called directly, promote it to use the Fast
2102        // calling convention.
2103        F->setCallingConv(CallingConv::Fast);
2104        ChangeCalleesToFastCall(F);
2105        ++NumFastCallFns;
2106        Changed = true;
2107      }
2108
2109      if (F->getAttributes().hasAttrSomewhere(Attribute::Nest) &&
2110          !F->hasAddressTaken()) {
2111        // The function is not used by a trampoline intrinsic, so it is safe
2112        // to remove the 'nest' attribute.
2113        RemoveNestAttribute(F);
2114        ++NumNestRemoved;
2115        Changed = true;
2116      }
2117    }
2118  }
2119  return Changed;
2120}
2121
2122bool GlobalOpt::OptimizeGlobalVars(Module &M) {
2123  bool Changed = false;
2124  for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
2125       GVI != E; ) {
2126    GlobalVariable *GV = GVI++;
2127    // Global variables without names cannot be referenced outside this module.
2128    if (!GV->hasName() && !GV->isDeclaration())
2129      GV->setLinkage(GlobalValue::InternalLinkage);
2130    // Simplify the initializer.
2131    if (GV->hasInitializer())
2132      if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GV->getInitializer())) {
2133        Constant *New = ConstantFoldConstantExpression(CE, TD, TLI);
2134        if (New && New != CE)
2135          GV->setInitializer(New);
2136      }
2137
2138    Changed |= ProcessGlobal(GV, GVI);
2139  }
2140  return Changed;
2141}
2142
2143/// FindGlobalCtors - Find the llvm.global_ctors list, verifying that all
2144/// initializers have an init priority of 65535.
2145GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) {
2146  GlobalVariable *GV = M.getGlobalVariable("llvm.global_ctors");
2147  if (GV == 0) return 0;
2148
2149  // Verify that the initializer is simple enough for us to handle. We are
2150  // only allowed to optimize the initializer if it is unique.
2151  if (!GV->hasUniqueInitializer()) return 0;
2152
2153  if (isa<ConstantAggregateZero>(GV->getInitializer()))
2154    return GV;
2155  ConstantArray *CA = cast<ConstantArray>(GV->getInitializer());
2156
2157  for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) {
2158    if (isa<ConstantAggregateZero>(*i))
2159      continue;
2160    ConstantStruct *CS = cast<ConstantStruct>(*i);
2161    if (isa<ConstantPointerNull>(CS->getOperand(1)))
2162      continue;
2163
2164    // Must have a function or null ptr.
2165    if (!isa<Function>(CS->getOperand(1)))
2166      return 0;
2167
2168    // Init priority must be standard.
2169    ConstantInt *CI = cast<ConstantInt>(CS->getOperand(0));
2170    if (CI->getZExtValue() != 65535)
2171      return 0;
2172  }
2173
2174  return GV;
2175}
2176
2177/// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand,
2178/// return a list of the functions and null terminator as a vector.
2179static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) {
2180  if (GV->getInitializer()->isNullValue())
2181    return std::vector<Function*>();
2182  ConstantArray *CA = cast<ConstantArray>(GV->getInitializer());
2183  std::vector<Function*> Result;
2184  Result.reserve(CA->getNumOperands());
2185  for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) {
2186    ConstantStruct *CS = cast<ConstantStruct>(*i);
2187    Result.push_back(dyn_cast<Function>(CS->getOperand(1)));
2188  }
2189  return Result;
2190}
2191
2192/// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the
2193/// specified array, returning the new global to use.
2194static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL,
2195                                          const std::vector<Function*> &Ctors) {
2196  // If we made a change, reassemble the initializer list.
2197  Constant *CSVals[2];
2198  CSVals[0] = ConstantInt::get(Type::getInt32Ty(GCL->getContext()), 65535);
2199  CSVals[1] = 0;
2200
2201  StructType *StructTy =
2202    cast <StructType>(
2203    cast<ArrayType>(GCL->getType()->getElementType())->getElementType());
2204
2205  // Create the new init list.
2206  std::vector<Constant*> CAList;
2207  for (unsigned i = 0, e = Ctors.size(); i != e; ++i) {
2208    if (Ctors[i]) {
2209      CSVals[1] = Ctors[i];
2210    } else {
2211      Type *FTy = FunctionType::get(Type::getVoidTy(GCL->getContext()),
2212                                          false);
2213      PointerType *PFTy = PointerType::getUnqual(FTy);
2214      CSVals[1] = Constant::getNullValue(PFTy);
2215      CSVals[0] = ConstantInt::get(Type::getInt32Ty(GCL->getContext()),
2216                                   0x7fffffff);
2217    }
2218    CAList.push_back(ConstantStruct::get(StructTy, CSVals));
2219  }
2220
2221  // Create the array initializer.
2222  Constant *CA = ConstantArray::get(ArrayType::get(StructTy,
2223                                                   CAList.size()), CAList);
2224
2225  // If we didn't change the number of elements, don't create a new GV.
2226  if (CA->getType() == GCL->getInitializer()->getType()) {
2227    GCL->setInitializer(CA);
2228    return GCL;
2229  }
2230
2231  // Create the new global and insert it next to the existing list.
2232  GlobalVariable *NGV = new GlobalVariable(CA->getType(), GCL->isConstant(),
2233                                           GCL->getLinkage(), CA, "",
2234                                           GCL->getThreadLocalMode());
2235  GCL->getParent()->getGlobalList().insert(GCL, NGV);
2236  NGV->takeName(GCL);
2237
2238  // Nuke the old list, replacing any uses with the new one.
2239  if (!GCL->use_empty()) {
2240    Constant *V = NGV;
2241    if (V->getType() != GCL->getType())
2242      V = ConstantExpr::getBitCast(V, GCL->getType());
2243    GCL->replaceAllUsesWith(V);
2244  }
2245  GCL->eraseFromParent();
2246
2247  if (Ctors.size())
2248    return NGV;
2249  else
2250    return 0;
2251}
2252
2253
2254static inline bool
2255isSimpleEnoughValueToCommit(Constant *C,
2256                            SmallPtrSet<Constant*, 8> &SimpleConstants,
2257                            const TargetData *TD);
2258
2259
2260/// isSimpleEnoughValueToCommit - Return true if the specified constant can be
2261/// handled by the code generator.  We don't want to generate something like:
2262///   void *X = &X/42;
2263/// because the code generator doesn't have a relocation that can handle that.
2264///
2265/// This function should be called if C was not found (but just got inserted)
2266/// in SimpleConstants to avoid having to rescan the same constants all the
2267/// time.
2268static bool isSimpleEnoughValueToCommitHelper(Constant *C,
2269                                   SmallPtrSet<Constant*, 8> &SimpleConstants,
2270                                   const TargetData *TD) {
2271  // Simple integer, undef, constant aggregate zero, global addresses, etc are
2272  // all supported.
2273  if (C->getNumOperands() == 0 || isa<BlockAddress>(C) ||
2274      isa<GlobalValue>(C))
2275    return true;
2276
2277  // Aggregate values are safe if all their elements are.
2278  if (isa<ConstantArray>(C) || isa<ConstantStruct>(C) ||
2279      isa<ConstantVector>(C)) {
2280    for (unsigned i = 0, e = C->getNumOperands(); i != e; ++i) {
2281      Constant *Op = cast<Constant>(C->getOperand(i));
2282      if (!isSimpleEnoughValueToCommit(Op, SimpleConstants, TD))
2283        return false;
2284    }
2285    return true;
2286  }
2287
2288  // We don't know exactly what relocations are allowed in constant expressions,
2289  // so we allow &global+constantoffset, which is safe and uniformly supported
2290  // across targets.
2291  ConstantExpr *CE = cast<ConstantExpr>(C);
2292  switch (CE->getOpcode()) {
2293  case Instruction::BitCast:
2294    // Bitcast is fine if the casted value is fine.
2295    return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, TD);
2296
2297  case Instruction::IntToPtr:
2298  case Instruction::PtrToInt:
2299    // int <=> ptr is fine if the int type is the same size as the
2300    // pointer type.
2301    if (!TD || TD->getTypeSizeInBits(CE->getType()) !=
2302               TD->getTypeSizeInBits(CE->getOperand(0)->getType()))
2303      return false;
2304    return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, TD);
2305
2306  // GEP is fine if it is simple + constant offset.
2307  case Instruction::GetElementPtr:
2308    for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
2309      if (!isa<ConstantInt>(CE->getOperand(i)))
2310        return false;
2311    return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, TD);
2312
2313  case Instruction::Add:
2314    // We allow simple+cst.
2315    if (!isa<ConstantInt>(CE->getOperand(1)))
2316      return false;
2317    return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, TD);
2318  }
2319  return false;
2320}
2321
2322static inline bool
2323isSimpleEnoughValueToCommit(Constant *C,
2324                            SmallPtrSet<Constant*, 8> &SimpleConstants,
2325                            const TargetData *TD) {
2326  // If we already checked this constant, we win.
2327  if (!SimpleConstants.insert(C)) return true;
2328  // Check the constant.
2329  return isSimpleEnoughValueToCommitHelper(C, SimpleConstants, TD);
2330}
2331
2332
2333/// isSimpleEnoughPointerToCommit - Return true if this constant is simple
2334/// enough for us to understand.  In particular, if it is a cast to anything
2335/// other than from one pointer type to another pointer type, we punt.
2336/// We basically just support direct accesses to globals and GEP's of
2337/// globals.  This should be kept up to date with CommitValueTo.
2338static bool isSimpleEnoughPointerToCommit(Constant *C) {
2339  // Conservatively, avoid aggregate types. This is because we don't
2340  // want to worry about them partially overlapping other stores.
2341  if (!cast<PointerType>(C->getType())->getElementType()->isSingleValueType())
2342    return false;
2343
2344  if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
2345    // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or
2346    // external globals.
2347    return GV->hasUniqueInitializer();
2348
2349  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
2350    // Handle a constantexpr gep.
2351    if (CE->getOpcode() == Instruction::GetElementPtr &&
2352        isa<GlobalVariable>(CE->getOperand(0)) &&
2353        cast<GEPOperator>(CE)->isInBounds()) {
2354      GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2355      // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or
2356      // external globals.
2357      if (!GV->hasUniqueInitializer())
2358        return false;
2359
2360      // The first index must be zero.
2361      ConstantInt *CI = dyn_cast<ConstantInt>(*llvm::next(CE->op_begin()));
2362      if (!CI || !CI->isZero()) return false;
2363
2364      // The remaining indices must be compile-time known integers within the
2365      // notional bounds of the corresponding static array types.
2366      if (!CE->isGEPWithNoNotionalOverIndexing())
2367        return false;
2368
2369      return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
2370
2371    // A constantexpr bitcast from a pointer to another pointer is a no-op,
2372    // and we know how to evaluate it by moving the bitcast from the pointer
2373    // operand to the value operand.
2374    } else if (CE->getOpcode() == Instruction::BitCast &&
2375               isa<GlobalVariable>(CE->getOperand(0))) {
2376      // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or
2377      // external globals.
2378      return cast<GlobalVariable>(CE->getOperand(0))->hasUniqueInitializer();
2379    }
2380  }
2381
2382  return false;
2383}
2384
2385/// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global
2386/// initializer.  This returns 'Init' modified to reflect 'Val' stored into it.
2387/// At this point, the GEP operands of Addr [0, OpNo) have been stepped into.
2388static Constant *EvaluateStoreInto(Constant *Init, Constant *Val,
2389                                   ConstantExpr *Addr, unsigned OpNo) {
2390  // Base case of the recursion.
2391  if (OpNo == Addr->getNumOperands()) {
2392    assert(Val->getType() == Init->getType() && "Type mismatch!");
2393    return Val;
2394  }
2395
2396  SmallVector<Constant*, 32> Elts;
2397  if (StructType *STy = dyn_cast<StructType>(Init->getType())) {
2398    // Break up the constant into its elements.
2399    for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
2400      Elts.push_back(Init->getAggregateElement(i));
2401
2402    // Replace the element that we are supposed to.
2403    ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo));
2404    unsigned Idx = CU->getZExtValue();
2405    assert(Idx < STy->getNumElements() && "Struct index out of range!");
2406    Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1);
2407
2408    // Return the modified struct.
2409    return ConstantStruct::get(STy, Elts);
2410  }
2411
2412  ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo));
2413  SequentialType *InitTy = cast<SequentialType>(Init->getType());
2414
2415  uint64_t NumElts;
2416  if (ArrayType *ATy = dyn_cast<ArrayType>(InitTy))
2417    NumElts = ATy->getNumElements();
2418  else
2419    NumElts = InitTy->getVectorNumElements();
2420
2421  // Break up the array into elements.
2422  for (uint64_t i = 0, e = NumElts; i != e; ++i)
2423    Elts.push_back(Init->getAggregateElement(i));
2424
2425  assert(CI->getZExtValue() < NumElts);
2426  Elts[CI->getZExtValue()] =
2427    EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1);
2428
2429  if (Init->getType()->isArrayTy())
2430    return ConstantArray::get(cast<ArrayType>(InitTy), Elts);
2431  return ConstantVector::get(Elts);
2432}
2433
2434/// CommitValueTo - We have decided that Addr (which satisfies the predicate
2435/// isSimpleEnoughPointerToCommit) should get Val as its value.  Make it happen.
2436static void CommitValueTo(Constant *Val, Constant *Addr) {
2437  if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
2438    assert(GV->hasInitializer());
2439    GV->setInitializer(Val);
2440    return;
2441  }
2442
2443  ConstantExpr *CE = cast<ConstantExpr>(Addr);
2444  GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2445  GV->setInitializer(EvaluateStoreInto(GV->getInitializer(), Val, CE, 2));
2446}
2447
2448namespace {
2449
2450/// Evaluator - This class evaluates LLVM IR, producing the Constant
2451/// representing each SSA instruction.  Changes to global variables are stored
2452/// in a mapping that can be iterated over after the evaluation is complete.
2453/// Once an evaluation call fails, the evaluation object should not be reused.
2454class Evaluator {
2455public:
2456  Evaluator(const TargetData *TD, const TargetLibraryInfo *TLI)
2457    : TD(TD), TLI(TLI) {
2458    ValueStack.push_back(new DenseMap<Value*, Constant*>);
2459  }
2460
2461  ~Evaluator() {
2462    DeleteContainerPointers(ValueStack);
2463    while (!AllocaTmps.empty()) {
2464      GlobalVariable *Tmp = AllocaTmps.back();
2465      AllocaTmps.pop_back();
2466
2467      // If there are still users of the alloca, the program is doing something
2468      // silly, e.g. storing the address of the alloca somewhere and using it
2469      // later.  Since this is undefined, we'll just make it be null.
2470      if (!Tmp->use_empty())
2471        Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType()));
2472      delete Tmp;
2473    }
2474  }
2475
2476  /// EvaluateFunction - Evaluate a call to function F, returning true if
2477  /// successful, false if we can't evaluate it.  ActualArgs contains the formal
2478  /// arguments for the function.
2479  bool EvaluateFunction(Function *F, Constant *&RetVal,
2480                        const SmallVectorImpl<Constant*> &ActualArgs);
2481
2482  /// EvaluateBlock - Evaluate all instructions in block BB, returning true if
2483  /// successful, false if we can't evaluate it.  NewBB returns the next BB that
2484  /// control flows into, or null upon return.
2485  bool EvaluateBlock(BasicBlock::iterator CurInst, BasicBlock *&NextBB);
2486
2487  Constant *getVal(Value *V) {
2488    if (Constant *CV = dyn_cast<Constant>(V)) return CV;
2489    Constant *R = ValueStack.back()->lookup(V);
2490    assert(R && "Reference to an uncomputed value!");
2491    return R;
2492  }
2493
2494  void setVal(Value *V, Constant *C) {
2495    ValueStack.back()->operator[](V) = C;
2496  }
2497
2498  const DenseMap<Constant*, Constant*> &getMutatedMemory() const {
2499    return MutatedMemory;
2500  }
2501
2502  const SmallPtrSet<GlobalVariable*, 8> &getInvariants() const {
2503    return Invariants;
2504  }
2505
2506private:
2507  Constant *ComputeLoadResult(Constant *P);
2508
2509  /// ValueStack - As we compute SSA register values, we store their contents
2510  /// here. The back of the vector contains the current function and the stack
2511  /// contains the values in the calling frames.
2512  SmallVector<DenseMap<Value*, Constant*>*, 4> ValueStack;
2513
2514  /// CallStack - This is used to detect recursion.  In pathological situations
2515  /// we could hit exponential behavior, but at least there is nothing
2516  /// unbounded.
2517  SmallVector<Function*, 4> CallStack;
2518
2519  /// MutatedMemory - For each store we execute, we update this map.  Loads
2520  /// check this to get the most up-to-date value.  If evaluation is successful,
2521  /// this state is committed to the process.
2522  DenseMap<Constant*, Constant*> MutatedMemory;
2523
2524  /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable
2525  /// to represent its body.  This vector is needed so we can delete the
2526  /// temporary globals when we are done.
2527  SmallVector<GlobalVariable*, 32> AllocaTmps;
2528
2529  /// Invariants - These global variables have been marked invariant by the
2530  /// static constructor.
2531  SmallPtrSet<GlobalVariable*, 8> Invariants;
2532
2533  /// SimpleConstants - These are constants we have checked and know to be
2534  /// simple enough to live in a static initializer of a global.
2535  SmallPtrSet<Constant*, 8> SimpleConstants;
2536
2537  const TargetData *TD;
2538  const TargetLibraryInfo *TLI;
2539};
2540
2541}  // anonymous namespace
2542
2543/// ComputeLoadResult - Return the value that would be computed by a load from
2544/// P after the stores reflected by 'memory' have been performed.  If we can't
2545/// decide, return null.
2546Constant *Evaluator::ComputeLoadResult(Constant *P) {
2547  // If this memory location has been recently stored, use the stored value: it
2548  // is the most up-to-date.
2549  DenseMap<Constant*, Constant*>::const_iterator I = MutatedMemory.find(P);
2550  if (I != MutatedMemory.end()) return I->second;
2551
2552  // Access it.
2553  if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
2554    if (GV->hasDefinitiveInitializer())
2555      return GV->getInitializer();
2556    return 0;
2557  }
2558
2559  // Handle a constantexpr getelementptr.
2560  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P))
2561    if (CE->getOpcode() == Instruction::GetElementPtr &&
2562        isa<GlobalVariable>(CE->getOperand(0))) {
2563      GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2564      if (GV->hasDefinitiveInitializer())
2565        return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
2566    }
2567
2568  return 0;  // don't know how to evaluate.
2569}
2570
2571/// EvaluateBlock - Evaluate all instructions in block BB, returning true if
2572/// successful, false if we can't evaluate it.  NewBB returns the next BB that
2573/// control flows into, or null upon return.
2574bool Evaluator::EvaluateBlock(BasicBlock::iterator CurInst,
2575                              BasicBlock *&NextBB) {
2576  // This is the main evaluation loop.
2577  while (1) {
2578    Constant *InstResult = 0;
2579
2580    if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) {
2581      if (!SI->isSimple()) return false;  // no volatile/atomic accesses.
2582      Constant *Ptr = getVal(SI->getOperand(1));
2583      if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
2584        Ptr = ConstantFoldConstantExpression(CE, TD, TLI);
2585      if (!isSimpleEnoughPointerToCommit(Ptr))
2586        // If this is too complex for us to commit, reject it.
2587        return false;
2588
2589      Constant *Val = getVal(SI->getOperand(0));
2590
2591      // If this might be too difficult for the backend to handle (e.g. the addr
2592      // of one global variable divided by another) then we can't commit it.
2593      if (!isSimpleEnoughValueToCommit(Val, SimpleConstants, TD))
2594        return false;
2595
2596      if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
2597        if (CE->getOpcode() == Instruction::BitCast) {
2598          // If we're evaluating a store through a bitcast, then we need
2599          // to pull the bitcast off the pointer type and push it onto the
2600          // stored value.
2601          Ptr = CE->getOperand(0);
2602
2603          Type *NewTy = cast<PointerType>(Ptr->getType())->getElementType();
2604
2605          // In order to push the bitcast onto the stored value, a bitcast
2606          // from NewTy to Val's type must be legal.  If it's not, we can try
2607          // introspecting NewTy to find a legal conversion.
2608          while (!Val->getType()->canLosslesslyBitCastTo(NewTy)) {
2609            // If NewTy is a struct, we can convert the pointer to the struct
2610            // into a pointer to its first member.
2611            // FIXME: This could be extended to support arrays as well.
2612            if (StructType *STy = dyn_cast<StructType>(NewTy)) {
2613              NewTy = STy->getTypeAtIndex(0U);
2614
2615              IntegerType *IdxTy = IntegerType::get(NewTy->getContext(), 32);
2616              Constant *IdxZero = ConstantInt::get(IdxTy, 0, false);
2617              Constant * const IdxList[] = {IdxZero, IdxZero};
2618
2619              Ptr = ConstantExpr::getGetElementPtr(Ptr, IdxList);
2620              if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
2621                Ptr = ConstantFoldConstantExpression(CE, TD, TLI);
2622
2623            // If we can't improve the situation by introspecting NewTy,
2624            // we have to give up.
2625            } else {
2626              return false;
2627            }
2628          }
2629
2630          // If we found compatible types, go ahead and push the bitcast
2631          // onto the stored value.
2632          Val = ConstantExpr::getBitCast(Val, NewTy);
2633        }
2634
2635      MutatedMemory[Ptr] = Val;
2636    } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) {
2637      InstResult = ConstantExpr::get(BO->getOpcode(),
2638                                     getVal(BO->getOperand(0)),
2639                                     getVal(BO->getOperand(1)));
2640    } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) {
2641      InstResult = ConstantExpr::getCompare(CI->getPredicate(),
2642                                            getVal(CI->getOperand(0)),
2643                                            getVal(CI->getOperand(1)));
2644    } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) {
2645      InstResult = ConstantExpr::getCast(CI->getOpcode(),
2646                                         getVal(CI->getOperand(0)),
2647                                         CI->getType());
2648    } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) {
2649      InstResult = ConstantExpr::getSelect(getVal(SI->getOperand(0)),
2650                                           getVal(SI->getOperand(1)),
2651                                           getVal(SI->getOperand(2)));
2652    } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) {
2653      Constant *P = getVal(GEP->getOperand(0));
2654      SmallVector<Constant*, 8> GEPOps;
2655      for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end();
2656           i != e; ++i)
2657        GEPOps.push_back(getVal(*i));
2658      InstResult =
2659        ConstantExpr::getGetElementPtr(P, GEPOps,
2660                                       cast<GEPOperator>(GEP)->isInBounds());
2661    } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) {
2662      if (!LI->isSimple()) return false;  // no volatile/atomic accesses.
2663      Constant *Ptr = getVal(LI->getOperand(0));
2664      if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
2665        Ptr = ConstantFoldConstantExpression(CE, TD, TLI);
2666      InstResult = ComputeLoadResult(Ptr);
2667      if (InstResult == 0) return false; // Could not evaluate load.
2668    } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) {
2669      if (AI->isArrayAllocation()) return false;  // Cannot handle array allocs.
2670      Type *Ty = AI->getType()->getElementType();
2671      AllocaTmps.push_back(new GlobalVariable(Ty, false,
2672                                              GlobalValue::InternalLinkage,
2673                                              UndefValue::get(Ty),
2674                                              AI->getName()));
2675      InstResult = AllocaTmps.back();
2676    } else if (isa<CallInst>(CurInst) || isa<InvokeInst>(CurInst)) {
2677      CallSite CS(CurInst);
2678
2679      // Debug info can safely be ignored here.
2680      if (isa<DbgInfoIntrinsic>(CS.getInstruction())) {
2681        ++CurInst;
2682        continue;
2683      }
2684
2685      // Cannot handle inline asm.
2686      if (isa<InlineAsm>(CS.getCalledValue())) return false;
2687
2688      if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction())) {
2689        if (MemSetInst *MSI = dyn_cast<MemSetInst>(II)) {
2690          if (MSI->isVolatile()) return false;
2691          Constant *Ptr = getVal(MSI->getDest());
2692          Constant *Val = getVal(MSI->getValue());
2693          Constant *DestVal = ComputeLoadResult(getVal(Ptr));
2694          if (Val->isNullValue() && DestVal && DestVal->isNullValue()) {
2695            // This memset is a no-op.
2696            ++CurInst;
2697            continue;
2698          }
2699        }
2700
2701        if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
2702            II->getIntrinsicID() == Intrinsic::lifetime_end) {
2703          ++CurInst;
2704          continue;
2705        }
2706
2707        if (II->getIntrinsicID() == Intrinsic::invariant_start) {
2708          // We don't insert an entry into Values, as it doesn't have a
2709          // meaningful return value.
2710          if (!II->use_empty())
2711            return false;
2712          ConstantInt *Size = cast<ConstantInt>(II->getArgOperand(0));
2713          Value *PtrArg = getVal(II->getArgOperand(1));
2714          Value *Ptr = PtrArg->stripPointerCasts();
2715          if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Ptr)) {
2716            Type *ElemTy = cast<PointerType>(GV->getType())->getElementType();
2717            if (!Size->isAllOnesValue() &&
2718                Size->getValue().getLimitedValue() >=
2719                TD->getTypeStoreSize(ElemTy))
2720              Invariants.insert(GV);
2721          }
2722          // Continue even if we do nothing.
2723          ++CurInst;
2724          continue;
2725        }
2726        return false;
2727      }
2728
2729      // Resolve function pointers.
2730      Function *Callee = dyn_cast<Function>(getVal(CS.getCalledValue()));
2731      if (!Callee || Callee->mayBeOverridden())
2732        return false;  // Cannot resolve.
2733
2734      SmallVector<Constant*, 8> Formals;
2735      for (User::op_iterator i = CS.arg_begin(), e = CS.arg_end(); i != e; ++i)
2736        Formals.push_back(getVal(*i));
2737
2738      if (Callee->isDeclaration()) {
2739        // If this is a function we can constant fold, do it.
2740        if (Constant *C = ConstantFoldCall(Callee, Formals, TLI)) {
2741          InstResult = C;
2742        } else {
2743          return false;
2744        }
2745      } else {
2746        if (Callee->getFunctionType()->isVarArg())
2747          return false;
2748
2749        Constant *RetVal;
2750        // Execute the call, if successful, use the return value.
2751        ValueStack.push_back(new DenseMap<Value*, Constant*>);
2752        if (!EvaluateFunction(Callee, RetVal, Formals))
2753          return false;
2754        delete ValueStack.pop_back_val();
2755        InstResult = RetVal;
2756      }
2757    } else if (isa<TerminatorInst>(CurInst)) {
2758      if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) {
2759        if (BI->isUnconditional()) {
2760          NextBB = BI->getSuccessor(0);
2761        } else {
2762          ConstantInt *Cond =
2763            dyn_cast<ConstantInt>(getVal(BI->getCondition()));
2764          if (!Cond) return false;  // Cannot determine.
2765
2766          NextBB = BI->getSuccessor(!Cond->getZExtValue());
2767        }
2768      } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) {
2769        ConstantInt *Val =
2770          dyn_cast<ConstantInt>(getVal(SI->getCondition()));
2771        if (!Val) return false;  // Cannot determine.
2772        NextBB = SI->findCaseValue(Val).getCaseSuccessor();
2773      } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(CurInst)) {
2774        Value *Val = getVal(IBI->getAddress())->stripPointerCasts();
2775        if (BlockAddress *BA = dyn_cast<BlockAddress>(Val))
2776          NextBB = BA->getBasicBlock();
2777        else
2778          return false;  // Cannot determine.
2779      } else if (isa<ReturnInst>(CurInst)) {
2780        NextBB = 0;
2781      } else {
2782        // invoke, unwind, resume, unreachable.
2783        return false;  // Cannot handle this terminator.
2784      }
2785
2786      // We succeeded at evaluating this block!
2787      return true;
2788    } else {
2789      // Did not know how to evaluate this!
2790      return false;
2791    }
2792
2793    if (!CurInst->use_empty()) {
2794      if (ConstantExpr *CE = dyn_cast<ConstantExpr>(InstResult))
2795        InstResult = ConstantFoldConstantExpression(CE, TD, TLI);
2796
2797      setVal(CurInst, InstResult);
2798    }
2799
2800    // If we just processed an invoke, we finished evaluating the block.
2801    if (InvokeInst *II = dyn_cast<InvokeInst>(CurInst)) {
2802      NextBB = II->getNormalDest();
2803      return true;
2804    }
2805
2806    // Advance program counter.
2807    ++CurInst;
2808  }
2809}
2810
2811/// EvaluateFunction - Evaluate a call to function F, returning true if
2812/// successful, false if we can't evaluate it.  ActualArgs contains the formal
2813/// arguments for the function.
2814bool Evaluator::EvaluateFunction(Function *F, Constant *&RetVal,
2815                                 const SmallVectorImpl<Constant*> &ActualArgs) {
2816  // Check to see if this function is already executing (recursion).  If so,
2817  // bail out.  TODO: we might want to accept limited recursion.
2818  if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end())
2819    return false;
2820
2821  CallStack.push_back(F);
2822
2823  // Initialize arguments to the incoming values specified.
2824  unsigned ArgNo = 0;
2825  for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E;
2826       ++AI, ++ArgNo)
2827    setVal(AI, ActualArgs[ArgNo]);
2828
2829  // ExecutedBlocks - We only handle non-looping, non-recursive code.  As such,
2830  // we can only evaluate any one basic block at most once.  This set keeps
2831  // track of what we have executed so we can detect recursive cases etc.
2832  SmallPtrSet<BasicBlock*, 32> ExecutedBlocks;
2833
2834  // CurBB - The current basic block we're evaluating.
2835  BasicBlock *CurBB = F->begin();
2836
2837  BasicBlock::iterator CurInst = CurBB->begin();
2838
2839  while (1) {
2840    BasicBlock *NextBB = 0; // Initialized to avoid compiler warnings.
2841    if (!EvaluateBlock(CurInst, NextBB))
2842      return false;
2843
2844    if (NextBB == 0) {
2845      // Successfully running until there's no next block means that we found
2846      // the return.  Fill it the return value and pop the call stack.
2847      ReturnInst *RI = cast<ReturnInst>(CurBB->getTerminator());
2848      if (RI->getNumOperands())
2849        RetVal = getVal(RI->getOperand(0));
2850      CallStack.pop_back();
2851      return true;
2852    }
2853
2854    // Okay, we succeeded in evaluating this control flow.  See if we have
2855    // executed the new block before.  If so, we have a looping function,
2856    // which we cannot evaluate in reasonable time.
2857    if (!ExecutedBlocks.insert(NextBB))
2858      return false;  // looped!
2859
2860    // Okay, we have never been in this block before.  Check to see if there
2861    // are any PHI nodes.  If so, evaluate them with information about where
2862    // we came from.
2863    PHINode *PN = 0;
2864    for (CurInst = NextBB->begin();
2865         (PN = dyn_cast<PHINode>(CurInst)); ++CurInst)
2866      setVal(PN, getVal(PN->getIncomingValueForBlock(CurBB)));
2867
2868    // Advance to the next block.
2869    CurBB = NextBB;
2870  }
2871}
2872
2873/// EvaluateStaticConstructor - Evaluate static constructors in the function, if
2874/// we can.  Return true if we can, false otherwise.
2875static bool EvaluateStaticConstructor(Function *F, const TargetData *TD,
2876                                      const TargetLibraryInfo *TLI) {
2877  // Call the function.
2878  Evaluator Eval(TD, TLI);
2879  Constant *RetValDummy;
2880  bool EvalSuccess = Eval.EvaluateFunction(F, RetValDummy,
2881                                           SmallVector<Constant*, 0>());
2882
2883  if (EvalSuccess) {
2884    // We succeeded at evaluation: commit the result.
2885    DEBUG(dbgs() << "FULLY EVALUATED GLOBAL CTOR FUNCTION '"
2886          << F->getName() << "' to " << Eval.getMutatedMemory().size()
2887          << " stores.\n");
2888    for (DenseMap<Constant*, Constant*>::const_iterator I =
2889           Eval.getMutatedMemory().begin(), E = Eval.getMutatedMemory().end();
2890         I != E; ++I)
2891      CommitValueTo(I->second, I->first);
2892    for (SmallPtrSet<GlobalVariable*, 8>::const_iterator I =
2893           Eval.getInvariants().begin(), E = Eval.getInvariants().end();
2894         I != E; ++I)
2895      (*I)->setConstant(true);
2896  }
2897
2898  return EvalSuccess;
2899}
2900
2901/// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible.
2902/// Return true if anything changed.
2903bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) {
2904  std::vector<Function*> Ctors = ParseGlobalCtors(GCL);
2905  bool MadeChange = false;
2906  if (Ctors.empty()) return false;
2907
2908  // Loop over global ctors, optimizing them when we can.
2909  for (unsigned i = 0; i != Ctors.size(); ++i) {
2910    Function *F = Ctors[i];
2911    // Found a null terminator in the middle of the list, prune off the rest of
2912    // the list.
2913    if (F == 0) {
2914      if (i != Ctors.size()-1) {
2915        Ctors.resize(i+1);
2916        MadeChange = true;
2917      }
2918      break;
2919    }
2920
2921    // We cannot simplify external ctor functions.
2922    if (F->empty()) continue;
2923
2924    // If we can evaluate the ctor at compile time, do.
2925    if (EvaluateStaticConstructor(F, TD, TLI)) {
2926      Ctors.erase(Ctors.begin()+i);
2927      MadeChange = true;
2928      --i;
2929      ++NumCtorsEvaluated;
2930      continue;
2931    }
2932  }
2933
2934  if (!MadeChange) return false;
2935
2936  GCL = InstallGlobalCtors(GCL, Ctors);
2937  return true;
2938}
2939
2940bool GlobalOpt::OptimizeGlobalAliases(Module &M) {
2941  bool Changed = false;
2942
2943  for (Module::alias_iterator I = M.alias_begin(), E = M.alias_end();
2944       I != E;) {
2945    Module::alias_iterator J = I++;
2946    // Aliases without names cannot be referenced outside this module.
2947    if (!J->hasName() && !J->isDeclaration())
2948      J->setLinkage(GlobalValue::InternalLinkage);
2949    // If the aliasee may change at link time, nothing can be done - bail out.
2950    if (J->mayBeOverridden())
2951      continue;
2952
2953    Constant *Aliasee = J->getAliasee();
2954    GlobalValue *Target = cast<GlobalValue>(Aliasee->stripPointerCasts());
2955    Target->removeDeadConstantUsers();
2956    bool hasOneUse = Target->hasOneUse() && Aliasee->hasOneUse();
2957
2958    // Make all users of the alias use the aliasee instead.
2959    if (!J->use_empty()) {
2960      J->replaceAllUsesWith(Aliasee);
2961      ++NumAliasesResolved;
2962      Changed = true;
2963    }
2964
2965    // If the alias is externally visible, we may still be able to simplify it.
2966    if (!J->hasLocalLinkage()) {
2967      // If the aliasee has internal linkage, give it the name and linkage
2968      // of the alias, and delete the alias.  This turns:
2969      //   define internal ... @f(...)
2970      //   @a = alias ... @f
2971      // into:
2972      //   define ... @a(...)
2973      if (!Target->hasLocalLinkage())
2974        continue;
2975
2976      // Do not perform the transform if multiple aliases potentially target the
2977      // aliasee. This check also ensures that it is safe to replace the section
2978      // and other attributes of the aliasee with those of the alias.
2979      if (!hasOneUse)
2980        continue;
2981
2982      // Give the aliasee the name, linkage and other attributes of the alias.
2983      Target->takeName(J);
2984      Target->setLinkage(J->getLinkage());
2985      Target->GlobalValue::copyAttributesFrom(J);
2986    }
2987
2988    // Delete the alias.
2989    M.getAliasList().erase(J);
2990    ++NumAliasesRemoved;
2991    Changed = true;
2992  }
2993
2994  return Changed;
2995}
2996
2997static Function *FindCXAAtExit(Module &M, TargetLibraryInfo *TLI) {
2998  if (!TLI->has(LibFunc::cxa_atexit))
2999    return 0;
3000
3001  Function *Fn = M.getFunction(TLI->getName(LibFunc::cxa_atexit));
3002
3003  if (!Fn)
3004    return 0;
3005
3006  FunctionType *FTy = Fn->getFunctionType();
3007
3008  // Checking that the function has the right return type, the right number of
3009  // parameters and that they all have pointer types should be enough.
3010  if (!FTy->getReturnType()->isIntegerTy() ||
3011      FTy->getNumParams() != 3 ||
3012      !FTy->getParamType(0)->isPointerTy() ||
3013      !FTy->getParamType(1)->isPointerTy() ||
3014      !FTy->getParamType(2)->isPointerTy())
3015    return 0;
3016
3017  return Fn;
3018}
3019
3020/// cxxDtorIsEmpty - Returns whether the given function is an empty C++
3021/// destructor and can therefore be eliminated.
3022/// Note that we assume that other optimization passes have already simplified
3023/// the code so we only look for a function with a single basic block, where
3024/// the only allowed instructions are 'ret', 'call' to an empty C++ dtor and
3025/// other side-effect free instructions.
3026static bool cxxDtorIsEmpty(const Function &Fn,
3027                           SmallPtrSet<const Function *, 8> &CalledFunctions) {
3028  // FIXME: We could eliminate C++ destructors if they're readonly/readnone and
3029  // nounwind, but that doesn't seem worth doing.
3030  if (Fn.isDeclaration())
3031    return false;
3032
3033  if (++Fn.begin() != Fn.end())
3034    return false;
3035
3036  const BasicBlock &EntryBlock = Fn.getEntryBlock();
3037  for (BasicBlock::const_iterator I = EntryBlock.begin(), E = EntryBlock.end();
3038       I != E; ++I) {
3039    if (const CallInst *CI = dyn_cast<CallInst>(I)) {
3040      // Ignore debug intrinsics.
3041      if (isa<DbgInfoIntrinsic>(CI))
3042        continue;
3043
3044      const Function *CalledFn = CI->getCalledFunction();
3045
3046      if (!CalledFn)
3047        return false;
3048
3049      SmallPtrSet<const Function *, 8> NewCalledFunctions(CalledFunctions);
3050
3051      // Don't treat recursive functions as empty.
3052      if (!NewCalledFunctions.insert(CalledFn))
3053        return false;
3054
3055      if (!cxxDtorIsEmpty(*CalledFn, NewCalledFunctions))
3056        return false;
3057    } else if (isa<ReturnInst>(*I))
3058      return true; // We're done.
3059    else if (I->mayHaveSideEffects())
3060      return false; // Destructor with side effects, bail.
3061  }
3062
3063  return false;
3064}
3065
3066bool GlobalOpt::OptimizeEmptyGlobalCXXDtors(Function *CXAAtExitFn) {
3067  /// Itanium C++ ABI p3.3.5:
3068  ///
3069  ///   After constructing a global (or local static) object, that will require
3070  ///   destruction on exit, a termination function is registered as follows:
3071  ///
3072  ///   extern "C" int __cxa_atexit ( void (*f)(void *), void *p, void *d );
3073  ///
3074  ///   This registration, e.g. __cxa_atexit(f,p,d), is intended to cause the
3075  ///   call f(p) when DSO d is unloaded, before all such termination calls
3076  ///   registered before this one. It returns zero if registration is
3077  ///   successful, nonzero on failure.
3078
3079  // This pass will look for calls to __cxa_atexit where the function is trivial
3080  // and remove them.
3081  bool Changed = false;
3082
3083  for (Function::use_iterator I = CXAAtExitFn->use_begin(),
3084       E = CXAAtExitFn->use_end(); I != E;) {
3085    // We're only interested in calls. Theoretically, we could handle invoke
3086    // instructions as well, but neither llvm-gcc nor clang generate invokes
3087    // to __cxa_atexit.
3088    CallInst *CI = dyn_cast<CallInst>(*I++);
3089    if (!CI)
3090      continue;
3091
3092    Function *DtorFn =
3093      dyn_cast<Function>(CI->getArgOperand(0)->stripPointerCasts());
3094    if (!DtorFn)
3095      continue;
3096
3097    SmallPtrSet<const Function *, 8> CalledFunctions;
3098    if (!cxxDtorIsEmpty(*DtorFn, CalledFunctions))
3099      continue;
3100
3101    // Just remove the call.
3102    CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
3103    CI->eraseFromParent();
3104
3105    ++NumCXXDtorsRemoved;
3106
3107    Changed |= true;
3108  }
3109
3110  return Changed;
3111}
3112
3113bool GlobalOpt::runOnModule(Module &M) {
3114  bool Changed = false;
3115
3116  TD = getAnalysisIfAvailable<TargetData>();
3117  TLI = &getAnalysis<TargetLibraryInfo>();
3118
3119  // Try to find the llvm.globalctors list.
3120  GlobalVariable *GlobalCtors = FindGlobalCtors(M);
3121
3122  Function *CXAAtExitFn = FindCXAAtExit(M, TLI);
3123
3124  bool LocalChange = true;
3125  while (LocalChange) {
3126    LocalChange = false;
3127
3128    // Delete functions that are trivially dead, ccc -> fastcc
3129    LocalChange |= OptimizeFunctions(M);
3130
3131    // Optimize global_ctors list.
3132    if (GlobalCtors)
3133      LocalChange |= OptimizeGlobalCtorsList(GlobalCtors);
3134
3135    // Optimize non-address-taken globals.
3136    LocalChange |= OptimizeGlobalVars(M);
3137
3138    // Resolve aliases, when possible.
3139    LocalChange |= OptimizeGlobalAliases(M);
3140
3141    // Try to remove trivial global destructors.
3142    if (CXAAtExitFn)
3143      LocalChange |= OptimizeEmptyGlobalCXXDtors(CXAAtExitFn);
3144
3145    Changed |= LocalChange;
3146  }
3147
3148  // TODO: Move all global ctors functions to the end of the module for code
3149  // layout.
3150
3151  return Changed;
3152}
3153