GlobalOpt.cpp revision 183b0cfd9491234de2fb8fbfd97eeaa977512e8f
1//===- GlobalOpt.cpp - Optimize Global Variables --------------------------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This pass transforms simple global variables that never have their address
11// taken.  If obviously true, it marks read/write globals as constant, deletes
12// variables only stored to, etc.
13//
14//===----------------------------------------------------------------------===//
15
16#define DEBUG_TYPE "globalopt"
17#include "llvm/Transforms/IPO.h"
18#include "llvm/CallingConv.h"
19#include "llvm/Constants.h"
20#include "llvm/DerivedTypes.h"
21#include "llvm/Instructions.h"
22#include "llvm/IntrinsicInst.h"
23#include "llvm/Module.h"
24#include "llvm/Pass.h"
25#include "llvm/Analysis/ConstantFolding.h"
26#include "llvm/Target/TargetData.h"
27#include "llvm/Support/CallSite.h"
28#include "llvm/Support/Compiler.h"
29#include "llvm/Support/Debug.h"
30#include "llvm/Support/GetElementPtrTypeIterator.h"
31#include "llvm/Support/MathExtras.h"
32#include "llvm/ADT/SmallPtrSet.h"
33#include "llvm/ADT/SmallVector.h"
34#include "llvm/ADT/Statistic.h"
35#include "llvm/ADT/StringExtras.h"
36#include <algorithm>
37#include <map>
38#include <set>
39using namespace llvm;
40
41STATISTIC(NumMarked    , "Number of globals marked constant");
42STATISTIC(NumSRA       , "Number of aggregate globals broken into scalars");
43STATISTIC(NumHeapSRA   , "Number of heap objects SRA'd");
44STATISTIC(NumSubstitute,"Number of globals with initializers stored into them");
45STATISTIC(NumDeleted   , "Number of globals deleted");
46STATISTIC(NumFnDeleted , "Number of functions deleted");
47STATISTIC(NumGlobUses  , "Number of global uses devirtualized");
48STATISTIC(NumLocalized , "Number of globals localized");
49STATISTIC(NumShrunkToBool  , "Number of global vars shrunk to booleans");
50STATISTIC(NumFastCallFns   , "Number of functions converted to fastcc");
51STATISTIC(NumCtorsEvaluated, "Number of static ctors evaluated");
52STATISTIC(NumNestRemoved   , "Number of nest attributes removed");
53
54namespace {
55  struct VISIBILITY_HIDDEN GlobalOpt : public ModulePass {
56    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
57      AU.addRequired<TargetData>();
58    }
59    static char ID; // Pass identification, replacement for typeid
60    GlobalOpt() : ModulePass(&ID) {}
61
62    bool runOnModule(Module &M);
63
64  private:
65    GlobalVariable *FindGlobalCtors(Module &M);
66    bool OptimizeFunctions(Module &M);
67    bool OptimizeGlobalVars(Module &M);
68    bool ResolveAliases(Module &M);
69    bool OptimizeGlobalCtorsList(GlobalVariable *&GCL);
70    bool ProcessInternalGlobal(GlobalVariable *GV,Module::global_iterator &GVI);
71  };
72}
73
74char GlobalOpt::ID = 0;
75static RegisterPass<GlobalOpt> X("globalopt", "Global Variable Optimizer");
76
77ModulePass *llvm::createGlobalOptimizerPass() { return new GlobalOpt(); }
78
79namespace {
80
81/// GlobalStatus - As we analyze each global, keep track of some information
82/// about it.  If we find out that the address of the global is taken, none of
83/// this info will be accurate.
84struct VISIBILITY_HIDDEN GlobalStatus {
85  /// isLoaded - True if the global is ever loaded.  If the global isn't ever
86  /// loaded it can be deleted.
87  bool isLoaded;
88
89  /// StoredType - Keep track of what stores to the global look like.
90  ///
91  enum StoredType {
92    /// NotStored - There is no store to this global.  It can thus be marked
93    /// constant.
94    NotStored,
95
96    /// isInitializerStored - This global is stored to, but the only thing
97    /// stored is the constant it was initialized with.  This is only tracked
98    /// for scalar globals.
99    isInitializerStored,
100
101    /// isStoredOnce - This global is stored to, but only its initializer and
102    /// one other value is ever stored to it.  If this global isStoredOnce, we
103    /// track the value stored to it in StoredOnceValue below.  This is only
104    /// tracked for scalar globals.
105    isStoredOnce,
106
107    /// isStored - This global is stored to by multiple values or something else
108    /// that we cannot track.
109    isStored
110  } StoredType;
111
112  /// StoredOnceValue - If only one value (besides the initializer constant) is
113  /// ever stored to this global, keep track of what value it is.
114  Value *StoredOnceValue;
115
116  /// AccessingFunction/HasMultipleAccessingFunctions - These start out
117  /// null/false.  When the first accessing function is noticed, it is recorded.
118  /// When a second different accessing function is noticed,
119  /// HasMultipleAccessingFunctions is set to true.
120  Function *AccessingFunction;
121  bool HasMultipleAccessingFunctions;
122
123  /// HasNonInstructionUser - Set to true if this global has a user that is not
124  /// an instruction (e.g. a constant expr or GV initializer).
125  bool HasNonInstructionUser;
126
127  /// HasPHIUser - Set to true if this global has a user that is a PHI node.
128  bool HasPHIUser;
129
130  GlobalStatus() : isLoaded(false), StoredType(NotStored), StoredOnceValue(0),
131                   AccessingFunction(0), HasMultipleAccessingFunctions(false),
132                   HasNonInstructionUser(false), HasPHIUser(false) {}
133};
134
135}
136
137/// ConstantIsDead - Return true if the specified constant is (transitively)
138/// dead.  The constant may be used by other constants (e.g. constant arrays and
139/// constant exprs) as long as they are dead, but it cannot be used by anything
140/// else.
141static bool ConstantIsDead(Constant *C) {
142  if (isa<GlobalValue>(C)) return false;
143
144  for (Value::use_iterator UI = C->use_begin(), E = C->use_end(); UI != E; ++UI)
145    if (Constant *CU = dyn_cast<Constant>(*UI)) {
146      if (!ConstantIsDead(CU)) return false;
147    } else
148      return false;
149  return true;
150}
151
152
153/// AnalyzeGlobal - Look at all uses of the global and fill in the GlobalStatus
154/// structure.  If the global has its address taken, return true to indicate we
155/// can't do anything with it.
156///
157static bool AnalyzeGlobal(Value *V, GlobalStatus &GS,
158                          std::set<PHINode*> &PHIUsers) {
159  for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
160    if (ConstantExpr *CE = dyn_cast<ConstantExpr>(*UI)) {
161      GS.HasNonInstructionUser = true;
162
163      if (AnalyzeGlobal(CE, GS, PHIUsers)) return true;
164
165    } else if (Instruction *I = dyn_cast<Instruction>(*UI)) {
166      if (!GS.HasMultipleAccessingFunctions) {
167        Function *F = I->getParent()->getParent();
168        if (GS.AccessingFunction == 0)
169          GS.AccessingFunction = F;
170        else if (GS.AccessingFunction != F)
171          GS.HasMultipleAccessingFunctions = true;
172      }
173      if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
174        GS.isLoaded = true;
175        if (LI->isVolatile()) return true;  // Don't hack on volatile loads.
176      } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
177        // Don't allow a store OF the address, only stores TO the address.
178        if (SI->getOperand(0) == V) return true;
179
180        if (SI->isVolatile()) return true;  // Don't hack on volatile stores.
181
182        // If this is a direct store to the global (i.e., the global is a scalar
183        // value, not an aggregate), keep more specific information about
184        // stores.
185        if (GS.StoredType != GlobalStatus::isStored) {
186          if (GlobalVariable *GV = dyn_cast<GlobalVariable>(SI->getOperand(1))){
187            Value *StoredVal = SI->getOperand(0);
188            if (StoredVal == GV->getInitializer()) {
189              if (GS.StoredType < GlobalStatus::isInitializerStored)
190                GS.StoredType = GlobalStatus::isInitializerStored;
191            } else if (isa<LoadInst>(StoredVal) &&
192                       cast<LoadInst>(StoredVal)->getOperand(0) == GV) {
193              // G = G
194              if (GS.StoredType < GlobalStatus::isInitializerStored)
195                GS.StoredType = GlobalStatus::isInitializerStored;
196            } else if (GS.StoredType < GlobalStatus::isStoredOnce) {
197              GS.StoredType = GlobalStatus::isStoredOnce;
198              GS.StoredOnceValue = StoredVal;
199            } else if (GS.StoredType == GlobalStatus::isStoredOnce &&
200                       GS.StoredOnceValue == StoredVal) {
201              // noop.
202            } else {
203              GS.StoredType = GlobalStatus::isStored;
204            }
205          } else {
206            GS.StoredType = GlobalStatus::isStored;
207          }
208        }
209      } else if (isa<GetElementPtrInst>(I)) {
210        if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
211      } else if (isa<SelectInst>(I)) {
212        if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
213      } else if (PHINode *PN = dyn_cast<PHINode>(I)) {
214        // PHI nodes we can check just like select or GEP instructions, but we
215        // have to be careful about infinite recursion.
216        if (PHIUsers.insert(PN).second)  // Not already visited.
217          if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
218        GS.HasPHIUser = true;
219      } else if (isa<CmpInst>(I)) {
220      } else if (isa<MemCpyInst>(I) || isa<MemMoveInst>(I)) {
221        if (I->getOperand(1) == V)
222          GS.StoredType = GlobalStatus::isStored;
223        if (I->getOperand(2) == V)
224          GS.isLoaded = true;
225      } else if (isa<MemSetInst>(I)) {
226        assert(I->getOperand(1) == V && "Memset only takes one pointer!");
227        GS.StoredType = GlobalStatus::isStored;
228      } else {
229        return true;  // Any other non-load instruction might take address!
230      }
231    } else if (Constant *C = dyn_cast<Constant>(*UI)) {
232      GS.HasNonInstructionUser = true;
233      // We might have a dead and dangling constant hanging off of here.
234      if (!ConstantIsDead(C))
235        return true;
236    } else {
237      GS.HasNonInstructionUser = true;
238      // Otherwise must be some other user.
239      return true;
240    }
241
242  return false;
243}
244
245static Constant *getAggregateConstantElement(Constant *Agg, Constant *Idx) {
246  ConstantInt *CI = dyn_cast<ConstantInt>(Idx);
247  if (!CI) return 0;
248  unsigned IdxV = CI->getZExtValue();
249
250  if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg)) {
251    if (IdxV < CS->getNumOperands()) return CS->getOperand(IdxV);
252  } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg)) {
253    if (IdxV < CA->getNumOperands()) return CA->getOperand(IdxV);
254  } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Agg)) {
255    if (IdxV < CP->getNumOperands()) return CP->getOperand(IdxV);
256  } else if (isa<ConstantAggregateZero>(Agg)) {
257    if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
258      if (IdxV < STy->getNumElements())
259        return Constant::getNullValue(STy->getElementType(IdxV));
260    } else if (const SequentialType *STy =
261               dyn_cast<SequentialType>(Agg->getType())) {
262      return Constant::getNullValue(STy->getElementType());
263    }
264  } else if (isa<UndefValue>(Agg)) {
265    if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
266      if (IdxV < STy->getNumElements())
267        return UndefValue::get(STy->getElementType(IdxV));
268    } else if (const SequentialType *STy =
269               dyn_cast<SequentialType>(Agg->getType())) {
270      return UndefValue::get(STy->getElementType());
271    }
272  }
273  return 0;
274}
275
276
277/// CleanupConstantGlobalUsers - We just marked GV constant.  Loop over all
278/// users of the global, cleaning up the obvious ones.  This is largely just a
279/// quick scan over the use list to clean up the easy and obvious cruft.  This
280/// returns true if it made a change.
281static bool CleanupConstantGlobalUsers(Value *V, Constant *Init) {
282  bool Changed = false;
283  for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;) {
284    User *U = *UI++;
285
286    if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
287      if (Init) {
288        // Replace the load with the initializer.
289        LI->replaceAllUsesWith(Init);
290        LI->eraseFromParent();
291        Changed = true;
292      }
293    } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
294      // Store must be unreachable or storing Init into the global.
295      SI->eraseFromParent();
296      Changed = true;
297    } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
298      if (CE->getOpcode() == Instruction::GetElementPtr) {
299        Constant *SubInit = 0;
300        if (Init)
301          SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
302        Changed |= CleanupConstantGlobalUsers(CE, SubInit);
303      } else if (CE->getOpcode() == Instruction::BitCast &&
304                 isa<PointerType>(CE->getType())) {
305        // Pointer cast, delete any stores and memsets to the global.
306        Changed |= CleanupConstantGlobalUsers(CE, 0);
307      }
308
309      if (CE->use_empty()) {
310        CE->destroyConstant();
311        Changed = true;
312      }
313    } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
314      // Do not transform "gepinst (gep constexpr (GV))" here, because forming
315      // "gepconstexpr (gep constexpr (GV))" will cause the two gep's to fold
316      // and will invalidate our notion of what Init is.
317      Constant *SubInit = 0;
318      if (!isa<ConstantExpr>(GEP->getOperand(0))) {
319        ConstantExpr *CE =
320          dyn_cast_or_null<ConstantExpr>(ConstantFoldInstruction(GEP));
321        if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr)
322          SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
323      }
324      Changed |= CleanupConstantGlobalUsers(GEP, SubInit);
325
326      if (GEP->use_empty()) {
327        GEP->eraseFromParent();
328        Changed = true;
329      }
330    } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv
331      if (MI->getRawDest() == V) {
332        MI->eraseFromParent();
333        Changed = true;
334      }
335
336    } else if (Constant *C = dyn_cast<Constant>(U)) {
337      // If we have a chain of dead constantexprs or other things dangling from
338      // us, and if they are all dead, nuke them without remorse.
339      if (ConstantIsDead(C)) {
340        C->destroyConstant();
341        // This could have invalidated UI, start over from scratch.
342        CleanupConstantGlobalUsers(V, Init);
343        return true;
344      }
345    }
346  }
347  return Changed;
348}
349
350/// isSafeSROAElementUse - Return true if the specified instruction is a safe
351/// user of a derived expression from a global that we want to SROA.
352static bool isSafeSROAElementUse(Value *V) {
353  // We might have a dead and dangling constant hanging off of here.
354  if (Constant *C = dyn_cast<Constant>(V))
355    return ConstantIsDead(C);
356
357  Instruction *I = dyn_cast<Instruction>(V);
358  if (!I) return false;
359
360  // Loads are ok.
361  if (isa<LoadInst>(I)) return true;
362
363  // Stores *to* the pointer are ok.
364  if (StoreInst *SI = dyn_cast<StoreInst>(I))
365    return SI->getOperand(0) != V;
366
367  // Otherwise, it must be a GEP.
368  GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I);
369  if (GEPI == 0) return false;
370
371  if (GEPI->getNumOperands() < 3 || !isa<Constant>(GEPI->getOperand(1)) ||
372      !cast<Constant>(GEPI->getOperand(1))->isNullValue())
373    return false;
374
375  for (Value::use_iterator I = GEPI->use_begin(), E = GEPI->use_end();
376       I != E; ++I)
377    if (!isSafeSROAElementUse(*I))
378      return false;
379  return true;
380}
381
382
383/// IsUserOfGlobalSafeForSRA - U is a direct user of the specified global value.
384/// Look at it and its uses and decide whether it is safe to SROA this global.
385///
386static bool IsUserOfGlobalSafeForSRA(User *U, GlobalValue *GV) {
387  // The user of the global must be a GEP Inst or a ConstantExpr GEP.
388  if (!isa<GetElementPtrInst>(U) &&
389      (!isa<ConstantExpr>(U) ||
390       cast<ConstantExpr>(U)->getOpcode() != Instruction::GetElementPtr))
391    return false;
392
393  // Check to see if this ConstantExpr GEP is SRA'able.  In particular, we
394  // don't like < 3 operand CE's, and we don't like non-constant integer
395  // indices.  This enforces that all uses are 'gep GV, 0, C, ...' for some
396  // value of C.
397  if (U->getNumOperands() < 3 || !isa<Constant>(U->getOperand(1)) ||
398      !cast<Constant>(U->getOperand(1))->isNullValue() ||
399      !isa<ConstantInt>(U->getOperand(2)))
400    return false;
401
402  gep_type_iterator GEPI = gep_type_begin(U), E = gep_type_end(U);
403  ++GEPI;  // Skip over the pointer index.
404
405  // If this is a use of an array allocation, do a bit more checking for sanity.
406  if (const ArrayType *AT = dyn_cast<ArrayType>(*GEPI)) {
407    uint64_t NumElements = AT->getNumElements();
408    ConstantInt *Idx = cast<ConstantInt>(U->getOperand(2));
409
410    // Check to make sure that index falls within the array.  If not,
411    // something funny is going on, so we won't do the optimization.
412    //
413    if (Idx->getZExtValue() >= NumElements)
414      return false;
415
416    // We cannot scalar repl this level of the array unless any array
417    // sub-indices are in-range constants.  In particular, consider:
418    // A[0][i].  We cannot know that the user isn't doing invalid things like
419    // allowing i to index an out-of-range subscript that accesses A[1].
420    //
421    // Scalar replacing *just* the outer index of the array is probably not
422    // going to be a win anyway, so just give up.
423    for (++GEPI; // Skip array index.
424         GEPI != E && (isa<ArrayType>(*GEPI) || isa<VectorType>(*GEPI));
425         ++GEPI) {
426      uint64_t NumElements;
427      if (const ArrayType *SubArrayTy = dyn_cast<ArrayType>(*GEPI))
428        NumElements = SubArrayTy->getNumElements();
429      else
430        NumElements = cast<VectorType>(*GEPI)->getNumElements();
431
432      ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPI.getOperand());
433      if (!IdxVal || IdxVal->getZExtValue() >= NumElements)
434        return false;
435    }
436  }
437
438  for (Value::use_iterator I = U->use_begin(), E = U->use_end(); I != E; ++I)
439    if (!isSafeSROAElementUse(*I))
440      return false;
441  return true;
442}
443
444/// GlobalUsersSafeToSRA - Look at all uses of the global and decide whether it
445/// is safe for us to perform this transformation.
446///
447static bool GlobalUsersSafeToSRA(GlobalValue *GV) {
448  for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end();
449       UI != E; ++UI) {
450    if (!IsUserOfGlobalSafeForSRA(*UI, GV))
451      return false;
452  }
453  return true;
454}
455
456
457/// SRAGlobal - Perform scalar replacement of aggregates on the specified global
458/// variable.  This opens the door for other optimizations by exposing the
459/// behavior of the program in a more fine-grained way.  We have determined that
460/// this transformation is safe already.  We return the first global variable we
461/// insert so that the caller can reprocess it.
462static GlobalVariable *SRAGlobal(GlobalVariable *GV, const TargetData &TD) {
463  // Make sure this global only has simple uses that we can SRA.
464  if (!GlobalUsersSafeToSRA(GV))
465    return 0;
466
467  assert(GV->hasInternalLinkage() && !GV->isConstant());
468  Constant *Init = GV->getInitializer();
469  const Type *Ty = Init->getType();
470
471  std::vector<GlobalVariable*> NewGlobals;
472  Module::GlobalListType &Globals = GV->getParent()->getGlobalList();
473
474  // Get the alignment of the global, either explicit or target-specific.
475  unsigned StartAlignment = GV->getAlignment();
476  if (StartAlignment == 0)
477    StartAlignment = TD.getABITypeAlignment(GV->getType());
478
479  if (const StructType *STy = dyn_cast<StructType>(Ty)) {
480    NewGlobals.reserve(STy->getNumElements());
481    const StructLayout &Layout = *TD.getStructLayout(STy);
482    for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
483      Constant *In = getAggregateConstantElement(Init,
484                                            ConstantInt::get(Type::Int32Ty, i));
485      assert(In && "Couldn't get element of initializer?");
486      GlobalVariable *NGV = new GlobalVariable(STy->getElementType(i), false,
487                                               GlobalVariable::InternalLinkage,
488                                               In, GV->getName()+"."+utostr(i),
489                                               (Module *)NULL,
490                                               GV->isThreadLocal(),
491                                               GV->getType()->getAddressSpace());
492      Globals.insert(GV, NGV);
493      NewGlobals.push_back(NGV);
494
495      // Calculate the known alignment of the field.  If the original aggregate
496      // had 256 byte alignment for example, something might depend on that:
497      // propagate info to each field.
498      uint64_t FieldOffset = Layout.getElementOffset(i);
499      unsigned NewAlign = (unsigned)MinAlign(StartAlignment, FieldOffset);
500      if (NewAlign > TD.getABITypeAlignment(STy->getElementType(i)))
501        NGV->setAlignment(NewAlign);
502    }
503  } else if (const SequentialType *STy = dyn_cast<SequentialType>(Ty)) {
504    unsigned NumElements = 0;
505    if (const ArrayType *ATy = dyn_cast<ArrayType>(STy))
506      NumElements = ATy->getNumElements();
507    else
508      NumElements = cast<VectorType>(STy)->getNumElements();
509
510    if (NumElements > 16 && GV->hasNUsesOrMore(16))
511      return 0; // It's not worth it.
512    NewGlobals.reserve(NumElements);
513
514    uint64_t EltSize = TD.getABITypeSize(STy->getElementType());
515    unsigned EltAlign = TD.getABITypeAlignment(STy->getElementType());
516    for (unsigned i = 0, e = NumElements; i != e; ++i) {
517      Constant *In = getAggregateConstantElement(Init,
518                                            ConstantInt::get(Type::Int32Ty, i));
519      assert(In && "Couldn't get element of initializer?");
520
521      GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false,
522                                               GlobalVariable::InternalLinkage,
523                                               In, GV->getName()+"."+utostr(i),
524                                               (Module *)NULL,
525                                               GV->isThreadLocal(),
526                                               GV->getType()->getAddressSpace());
527      Globals.insert(GV, NGV);
528      NewGlobals.push_back(NGV);
529
530      // Calculate the known alignment of the field.  If the original aggregate
531      // had 256 byte alignment for example, something might depend on that:
532      // propagate info to each field.
533      unsigned NewAlign = (unsigned)MinAlign(StartAlignment, EltSize*i);
534      if (NewAlign > EltAlign)
535        NGV->setAlignment(NewAlign);
536    }
537  }
538
539  if (NewGlobals.empty())
540    return 0;
541
542  DOUT << "PERFORMING GLOBAL SRA ON: " << *GV;
543
544  Constant *NullInt = Constant::getNullValue(Type::Int32Ty);
545
546  // Loop over all of the uses of the global, replacing the constantexpr geps,
547  // with smaller constantexpr geps or direct references.
548  while (!GV->use_empty()) {
549    User *GEP = GV->use_back();
550    assert(((isa<ConstantExpr>(GEP) &&
551             cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)||
552            isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!");
553
554    // Ignore the 1th operand, which has to be zero or else the program is quite
555    // broken (undefined).  Get the 2nd operand, which is the structure or array
556    // index.
557    unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
558    if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access.
559
560    Value *NewPtr = NewGlobals[Val];
561
562    // Form a shorter GEP if needed.
563    if (GEP->getNumOperands() > 3) {
564      if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) {
565        SmallVector<Constant*, 8> Idxs;
566        Idxs.push_back(NullInt);
567        for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i)
568          Idxs.push_back(CE->getOperand(i));
569        NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr),
570                                                &Idxs[0], Idxs.size());
571      } else {
572        GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP);
573        SmallVector<Value*, 8> Idxs;
574        Idxs.push_back(NullInt);
575        for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i)
576          Idxs.push_back(GEPI->getOperand(i));
577        NewPtr = GetElementPtrInst::Create(NewPtr, Idxs.begin(), Idxs.end(),
578                                           GEPI->getName()+"."+utostr(Val), GEPI);
579      }
580    }
581    GEP->replaceAllUsesWith(NewPtr);
582
583    if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP))
584      GEPI->eraseFromParent();
585    else
586      cast<ConstantExpr>(GEP)->destroyConstant();
587  }
588
589  // Delete the old global, now that it is dead.
590  Globals.erase(GV);
591  ++NumSRA;
592
593  // Loop over the new globals array deleting any globals that are obviously
594  // dead.  This can arise due to scalarization of a structure or an array that
595  // has elements that are dead.
596  unsigned FirstGlobal = 0;
597  for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i)
598    if (NewGlobals[i]->use_empty()) {
599      Globals.erase(NewGlobals[i]);
600      if (FirstGlobal == i) ++FirstGlobal;
601    }
602
603  return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : 0;
604}
605
606/// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified
607/// value will trap if the value is dynamically null.  PHIs keeps track of any
608/// phi nodes we've seen to avoid reprocessing them.
609static bool AllUsesOfValueWillTrapIfNull(Value *V,
610                                         SmallPtrSet<PHINode*, 8> &PHIs) {
611  for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
612    if (isa<LoadInst>(*UI)) {
613      // Will trap.
614    } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
615      if (SI->getOperand(0) == V) {
616        //cerr << "NONTRAPPING USE: " << **UI;
617        return false;  // Storing the value.
618      }
619    } else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
620      if (CI->getOperand(0) != V) {
621        //cerr << "NONTRAPPING USE: " << **UI;
622        return false;  // Not calling the ptr
623      }
624    } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
625      if (II->getOperand(0) != V) {
626        //cerr << "NONTRAPPING USE: " << **UI;
627        return false;  // Not calling the ptr
628      }
629    } else if (BitCastInst *CI = dyn_cast<BitCastInst>(*UI)) {
630      if (!AllUsesOfValueWillTrapIfNull(CI, PHIs)) return false;
631    } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) {
632      if (!AllUsesOfValueWillTrapIfNull(GEPI, PHIs)) return false;
633    } else if (PHINode *PN = dyn_cast<PHINode>(*UI)) {
634      // If we've already seen this phi node, ignore it, it has already been
635      // checked.
636      if (PHIs.insert(PN))
637        return AllUsesOfValueWillTrapIfNull(PN, PHIs);
638    } else if (isa<ICmpInst>(*UI) &&
639               isa<ConstantPointerNull>(UI->getOperand(1))) {
640      // Ignore setcc X, null
641    } else {
642      //cerr << "NONTRAPPING USE: " << **UI;
643      return false;
644    }
645  return true;
646}
647
648/// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads
649/// from GV will trap if the loaded value is null.  Note that this also permits
650/// comparisons of the loaded value against null, as a special case.
651static bool AllUsesOfLoadedValueWillTrapIfNull(GlobalVariable *GV) {
652  for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI!=E; ++UI)
653    if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
654      SmallPtrSet<PHINode*, 8> PHIs;
655      if (!AllUsesOfValueWillTrapIfNull(LI, PHIs))
656        return false;
657    } else if (isa<StoreInst>(*UI)) {
658      // Ignore stores to the global.
659    } else {
660      // We don't know or understand this user, bail out.
661      //cerr << "UNKNOWN USER OF GLOBAL!: " << **UI;
662      return false;
663    }
664
665  return true;
666}
667
668static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) {
669  bool Changed = false;
670  for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) {
671    Instruction *I = cast<Instruction>(*UI++);
672    if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
673      LI->setOperand(0, NewV);
674      Changed = true;
675    } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
676      if (SI->getOperand(1) == V) {
677        SI->setOperand(1, NewV);
678        Changed = true;
679      }
680    } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
681      if (I->getOperand(0) == V) {
682        // Calling through the pointer!  Turn into a direct call, but be careful
683        // that the pointer is not also being passed as an argument.
684        I->setOperand(0, NewV);
685        Changed = true;
686        bool PassedAsArg = false;
687        for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i)
688          if (I->getOperand(i) == V) {
689            PassedAsArg = true;
690            I->setOperand(i, NewV);
691          }
692
693        if (PassedAsArg) {
694          // Being passed as an argument also.  Be careful to not invalidate UI!
695          UI = V->use_begin();
696        }
697      }
698    } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
699      Changed |= OptimizeAwayTrappingUsesOfValue(CI,
700                                ConstantExpr::getCast(CI->getOpcode(),
701                                                      NewV, CI->getType()));
702      if (CI->use_empty()) {
703        Changed = true;
704        CI->eraseFromParent();
705      }
706    } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
707      // Should handle GEP here.
708      SmallVector<Constant*, 8> Idxs;
709      Idxs.reserve(GEPI->getNumOperands()-1);
710      for (User::op_iterator i = GEPI->op_begin() + 1, e = GEPI->op_end();
711           i != e; ++i)
712        if (Constant *C = dyn_cast<Constant>(*i))
713          Idxs.push_back(C);
714        else
715          break;
716      if (Idxs.size() == GEPI->getNumOperands()-1)
717        Changed |= OptimizeAwayTrappingUsesOfValue(GEPI,
718                                ConstantExpr::getGetElementPtr(NewV, &Idxs[0],
719                                                               Idxs.size()));
720      if (GEPI->use_empty()) {
721        Changed = true;
722        GEPI->eraseFromParent();
723      }
724    }
725  }
726
727  return Changed;
728}
729
730
731/// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null
732/// value stored into it.  If there are uses of the loaded value that would trap
733/// if the loaded value is dynamically null, then we know that they cannot be
734/// reachable with a null optimize away the load.
735static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV) {
736  std::vector<LoadInst*> Loads;
737  bool Changed = false;
738
739  // Replace all uses of loads with uses of uses of the stored value.
740  for (Value::use_iterator GUI = GV->use_begin(), E = GV->use_end();
741       GUI != E; ++GUI)
742    if (LoadInst *LI = dyn_cast<LoadInst>(*GUI)) {
743      Loads.push_back(LI);
744      Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV);
745    } else {
746      // If we get here we could have stores, selects, or phi nodes whose values
747      // are loaded.
748      assert((isa<StoreInst>(*GUI) || isa<PHINode>(*GUI) ||
749              isa<SelectInst>(*GUI) || isa<ConstantExpr>(*GUI)) &&
750             "Only expect load and stores!");
751    }
752
753  if (Changed) {
754    DOUT << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV;
755    ++NumGlobUses;
756  }
757
758  // Delete all of the loads we can, keeping track of whether we nuked them all!
759  bool AllLoadsGone = true;
760  while (!Loads.empty()) {
761    LoadInst *L = Loads.back();
762    if (L->use_empty()) {
763      L->eraseFromParent();
764      Changed = true;
765    } else {
766      AllLoadsGone = false;
767    }
768    Loads.pop_back();
769  }
770
771  // If we nuked all of the loads, then none of the stores are needed either,
772  // nor is the global.
773  if (AllLoadsGone) {
774    DOUT << "  *** GLOBAL NOW DEAD!\n";
775    CleanupConstantGlobalUsers(GV, 0);
776    if (GV->use_empty()) {
777      GV->eraseFromParent();
778      ++NumDeleted;
779    }
780    Changed = true;
781  }
782  return Changed;
783}
784
785/// ConstantPropUsersOf - Walk the use list of V, constant folding all of the
786/// instructions that are foldable.
787static void ConstantPropUsersOf(Value *V) {
788  for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; )
789    if (Instruction *I = dyn_cast<Instruction>(*UI++))
790      if (Constant *NewC = ConstantFoldInstruction(I)) {
791        I->replaceAllUsesWith(NewC);
792
793        // Advance UI to the next non-I use to avoid invalidating it!
794        // Instructions could multiply use V.
795        while (UI != E && *UI == I)
796          ++UI;
797        I->eraseFromParent();
798      }
799}
800
801/// OptimizeGlobalAddressOfMalloc - This function takes the specified global
802/// variable, and transforms the program as if it always contained the result of
803/// the specified malloc.  Because it is always the result of the specified
804/// malloc, there is no reason to actually DO the malloc.  Instead, turn the
805/// malloc into a global, and any loads of GV as uses of the new global.
806static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV,
807                                                     MallocInst *MI) {
808  DOUT << "PROMOTING MALLOC GLOBAL: " << *GV << "  MALLOC = " << *MI;
809  ConstantInt *NElements = cast<ConstantInt>(MI->getArraySize());
810
811  if (NElements->getZExtValue() != 1) {
812    // If we have an array allocation, transform it to a single element
813    // allocation to make the code below simpler.
814    Type *NewTy = ArrayType::get(MI->getAllocatedType(),
815                                 NElements->getZExtValue());
816    MallocInst *NewMI =
817      new MallocInst(NewTy, Constant::getNullValue(Type::Int32Ty),
818                     MI->getAlignment(), MI->getName(), MI);
819    Value* Indices[2];
820    Indices[0] = Indices[1] = Constant::getNullValue(Type::Int32Ty);
821    Value *NewGEP = GetElementPtrInst::Create(NewMI, Indices, Indices + 2,
822                                              NewMI->getName()+".el0", MI);
823    MI->replaceAllUsesWith(NewGEP);
824    MI->eraseFromParent();
825    MI = NewMI;
826  }
827
828  // Create the new global variable.  The contents of the malloc'd memory is
829  // undefined, so initialize with an undef value.
830  Constant *Init = UndefValue::get(MI->getAllocatedType());
831  GlobalVariable *NewGV = new GlobalVariable(MI->getAllocatedType(), false,
832                                             GlobalValue::InternalLinkage, Init,
833                                             GV->getName()+".body",
834                                             (Module *)NULL,
835                                             GV->isThreadLocal());
836  // FIXME: This new global should have the alignment returned by malloc.  Code
837  // could depend on malloc returning large alignment (on the mac, 16 bytes) but
838  // this would only guarantee some lower alignment.
839  GV->getParent()->getGlobalList().insert(GV, NewGV);
840
841  // Anything that used the malloc now uses the global directly.
842  MI->replaceAllUsesWith(NewGV);
843
844  Constant *RepValue = NewGV;
845  if (NewGV->getType() != GV->getType()->getElementType())
846    RepValue = ConstantExpr::getBitCast(RepValue,
847                                        GV->getType()->getElementType());
848
849  // If there is a comparison against null, we will insert a global bool to
850  // keep track of whether the global was initialized yet or not.
851  GlobalVariable *InitBool =
852    new GlobalVariable(Type::Int1Ty, false, GlobalValue::InternalLinkage,
853                       ConstantInt::getFalse(), GV->getName()+".init",
854                       (Module *)NULL, GV->isThreadLocal());
855  bool InitBoolUsed = false;
856
857  // Loop over all uses of GV, processing them in turn.
858  std::vector<StoreInst*> Stores;
859  while (!GV->use_empty())
860    if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) {
861      while (!LI->use_empty()) {
862        Use &LoadUse = LI->use_begin().getUse();
863        if (!isa<ICmpInst>(LoadUse.getUser()))
864          LoadUse = RepValue;
865        else {
866          ICmpInst *CI = cast<ICmpInst>(LoadUse.getUser());
867          // Replace the cmp X, 0 with a use of the bool value.
868          Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", CI);
869          InitBoolUsed = true;
870          switch (CI->getPredicate()) {
871          default: assert(0 && "Unknown ICmp Predicate!");
872          case ICmpInst::ICMP_ULT:
873          case ICmpInst::ICMP_SLT:
874            LV = ConstantInt::getFalse();   // X < null -> always false
875            break;
876          case ICmpInst::ICMP_ULE:
877          case ICmpInst::ICMP_SLE:
878          case ICmpInst::ICMP_EQ:
879            LV = BinaryOperator::CreateNot(LV, "notinit", CI);
880            break;
881          case ICmpInst::ICMP_NE:
882          case ICmpInst::ICMP_UGE:
883          case ICmpInst::ICMP_SGE:
884          case ICmpInst::ICMP_UGT:
885          case ICmpInst::ICMP_SGT:
886            break;  // no change.
887          }
888          CI->replaceAllUsesWith(LV);
889          CI->eraseFromParent();
890        }
891      }
892      LI->eraseFromParent();
893    } else {
894      StoreInst *SI = cast<StoreInst>(GV->use_back());
895      // The global is initialized when the store to it occurs.
896      new StoreInst(ConstantInt::getTrue(), InitBool, SI);
897      SI->eraseFromParent();
898    }
899
900  // If the initialization boolean was used, insert it, otherwise delete it.
901  if (!InitBoolUsed) {
902    while (!InitBool->use_empty())  // Delete initializations
903      cast<Instruction>(InitBool->use_back())->eraseFromParent();
904    delete InitBool;
905  } else
906    GV->getParent()->getGlobalList().insert(GV, InitBool);
907
908
909  // Now the GV is dead, nuke it and the malloc.
910  GV->eraseFromParent();
911  MI->eraseFromParent();
912
913  // To further other optimizations, loop over all users of NewGV and try to
914  // constant prop them.  This will promote GEP instructions with constant
915  // indices into GEP constant-exprs, which will allow global-opt to hack on it.
916  ConstantPropUsersOf(NewGV);
917  if (RepValue != NewGV)
918    ConstantPropUsersOf(RepValue);
919
920  return NewGV;
921}
922
923/// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking
924/// to make sure that there are no complex uses of V.  We permit simple things
925/// like dereferencing the pointer, but not storing through the address, unless
926/// it is to the specified global.
927static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Instruction *V,
928                                                      GlobalVariable *GV,
929                                              SmallPtrSet<PHINode*, 8> &PHIs) {
930  for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
931    Instruction *Inst = dyn_cast<Instruction>(*UI);
932    if (Inst == 0) return false;
933
934    if (isa<LoadInst>(Inst) || isa<CmpInst>(Inst)) {
935      continue; // Fine, ignore.
936    }
937
938    if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
939      if (SI->getOperand(0) == V && SI->getOperand(1) != GV)
940        return false;  // Storing the pointer itself... bad.
941      continue; // Otherwise, storing through it, or storing into GV... fine.
942    }
943
944    if (isa<GetElementPtrInst>(Inst)) {
945      if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Inst, GV, PHIs))
946        return false;
947      continue;
948    }
949
950    if (PHINode *PN = dyn_cast<PHINode>(Inst)) {
951      // PHIs are ok if all uses are ok.  Don't infinitely recurse through PHI
952      // cycles.
953      if (PHIs.insert(PN))
954        if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(PN, GV, PHIs))
955          return false;
956      continue;
957    }
958
959    if (BitCastInst *BCI = dyn_cast<BitCastInst>(Inst)) {
960      if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(BCI, GV, PHIs))
961        return false;
962      continue;
963    }
964
965    return false;
966  }
967  return true;
968}
969
970/// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV
971/// somewhere.  Transform all uses of the allocation into loads from the
972/// global and uses of the resultant pointer.  Further, delete the store into
973/// GV.  This assumes that these value pass the
974/// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate.
975static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc,
976                                          GlobalVariable *GV) {
977  while (!Alloc->use_empty()) {
978    Instruction *U = cast<Instruction>(*Alloc->use_begin());
979    Instruction *InsertPt = U;
980    if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
981      // If this is the store of the allocation into the global, remove it.
982      if (SI->getOperand(1) == GV) {
983        SI->eraseFromParent();
984        continue;
985      }
986    } else if (PHINode *PN = dyn_cast<PHINode>(U)) {
987      // Insert the load in the corresponding predecessor, not right before the
988      // PHI.
989      unsigned PredNo = Alloc->use_begin().getOperandNo()/2;
990      InsertPt = PN->getIncomingBlock(PredNo)->getTerminator();
991    }
992
993    // Insert a load from the global, and use it instead of the malloc.
994    Value *NL = new LoadInst(GV, GV->getName()+".val", InsertPt);
995    U->replaceUsesOfWith(Alloc, NL);
996  }
997}
998
999/// GlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from
1000/// GV are simple enough to perform HeapSRA, return true.
1001static bool GlobalLoadUsesSimpleEnoughForHeapSRA(GlobalVariable *GV,
1002                                                 MallocInst *MI) {
1003  for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;
1004       ++UI)
1005    if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
1006      // We permit two users of the load: setcc comparing against the null
1007      // pointer, and a getelementptr of a specific form.
1008      for (Value::use_iterator UI = LI->use_begin(), E = LI->use_end();
1009           UI != E; ++UI) {
1010        // Comparison against null is ok.
1011        if (ICmpInst *ICI = dyn_cast<ICmpInst>(*UI)) {
1012          if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
1013            return false;
1014          continue;
1015        }
1016
1017        // getelementptr is also ok, but only a simple form.
1018        if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) {
1019          // Must index into the array and into the struct.
1020          if (GEPI->getNumOperands() < 3)
1021            return false;
1022
1023          // Otherwise the GEP is ok.
1024          continue;
1025        }
1026
1027        if (PHINode *PN = dyn_cast<PHINode>(*UI)) {
1028          // We have a phi of a load from the global.  We can only handle this
1029          // if the other PHI'd values are actually the same.  In this case,
1030          // the rewriter will just drop the phi entirely.
1031          for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1032            Value *IV = PN->getIncomingValue(i);
1033            if (IV == LI) continue;  // Trivial the same.
1034
1035            // If the phi'd value is from the malloc that initializes the value,
1036            // we can xform it.
1037            if (IV == MI) continue;
1038
1039            // Otherwise, we don't know what it is.
1040            return false;
1041          }
1042          return true;
1043        }
1044
1045        // Otherwise we don't know what this is, not ok.
1046        return false;
1047      }
1048    }
1049  return true;
1050}
1051
1052/// GetHeapSROALoad - Return the load for the specified field of the HeapSROA'd
1053/// value, lazily creating it on demand.
1054static Value *GetHeapSROALoad(Instruction *Load, unsigned FieldNo,
1055                              const std::vector<GlobalVariable*> &FieldGlobals,
1056                              std::vector<Value *> &InsertedLoadsForPtr) {
1057  if (InsertedLoadsForPtr.size() <= FieldNo)
1058    InsertedLoadsForPtr.resize(FieldNo+1);
1059  if (InsertedLoadsForPtr[FieldNo] == 0)
1060    InsertedLoadsForPtr[FieldNo] = new LoadInst(FieldGlobals[FieldNo],
1061                                                Load->getName()+".f" +
1062                                                utostr(FieldNo), Load);
1063  return InsertedLoadsForPtr[FieldNo];
1064}
1065
1066/// RewriteHeapSROALoadUser - Given a load instruction and a value derived from
1067/// the load, rewrite the derived value to use the HeapSRoA'd load.
1068static void RewriteHeapSROALoadUser(LoadInst *Load, Instruction *LoadUser,
1069                               const std::vector<GlobalVariable*> &FieldGlobals,
1070                                    std::vector<Value *> &InsertedLoadsForPtr) {
1071  // If this is a comparison against null, handle it.
1072  if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) {
1073    assert(isa<ConstantPointerNull>(SCI->getOperand(1)));
1074    // If we have a setcc of the loaded pointer, we can use a setcc of any
1075    // field.
1076    Value *NPtr;
1077    if (InsertedLoadsForPtr.empty()) {
1078      NPtr = GetHeapSROALoad(Load, 0, FieldGlobals, InsertedLoadsForPtr);
1079    } else {
1080      NPtr = InsertedLoadsForPtr.back();
1081    }
1082
1083    Value *New = new ICmpInst(SCI->getPredicate(), NPtr,
1084                              Constant::getNullValue(NPtr->getType()),
1085                              SCI->getName(), SCI);
1086    SCI->replaceAllUsesWith(New);
1087    SCI->eraseFromParent();
1088    return;
1089  }
1090
1091  // Handle 'getelementptr Ptr, Idx, uint FieldNo ...'
1092  if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) {
1093    assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2))
1094           && "Unexpected GEPI!");
1095
1096    // Load the pointer for this field.
1097    unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
1098    Value *NewPtr = GetHeapSROALoad(Load, FieldNo,
1099                                    FieldGlobals, InsertedLoadsForPtr);
1100
1101    // Create the new GEP idx vector.
1102    SmallVector<Value*, 8> GEPIdx;
1103    GEPIdx.push_back(GEPI->getOperand(1));
1104    GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end());
1105
1106    Value *NGEPI = GetElementPtrInst::Create(NewPtr,
1107                                             GEPIdx.begin(), GEPIdx.end(),
1108                                             GEPI->getName(), GEPI);
1109    GEPI->replaceAllUsesWith(NGEPI);
1110    GEPI->eraseFromParent();
1111    return;
1112  }
1113
1114  // Handle PHI nodes.  PHI nodes must be merging in the same values, plus
1115  // potentially the original malloc.  Insert phi nodes for each field, then
1116  // process uses of the PHI.
1117  PHINode *PN = cast<PHINode>(LoadUser);
1118  std::vector<Value *> PHIsForField;
1119  PHIsForField.resize(FieldGlobals.size());
1120  for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1121    Value *LoadV = GetHeapSROALoad(Load, i, FieldGlobals, InsertedLoadsForPtr);
1122
1123    PHINode *FieldPN = PHINode::Create(LoadV->getType(),
1124                                       PN->getName()+"."+utostr(i), PN);
1125    // Fill in the predecessor values.
1126    for (unsigned pred = 0, e = PN->getNumIncomingValues(); pred != e; ++pred) {
1127      // Each predecessor either uses the load or the original malloc.
1128      Value *InVal = PN->getIncomingValue(pred);
1129      BasicBlock *BB = PN->getIncomingBlock(pred);
1130      Value *NewVal;
1131      if (isa<MallocInst>(InVal)) {
1132        // Insert a reload from the global in the predecessor.
1133        NewVal = GetHeapSROALoad(BB->getTerminator(), i, FieldGlobals,
1134                                 PHIsForField);
1135      } else {
1136        NewVal = InsertedLoadsForPtr[i];
1137      }
1138      FieldPN->addIncoming(NewVal, BB);
1139    }
1140    PHIsForField[i] = FieldPN;
1141  }
1142
1143  // Since PHIsForField specifies a phi for every input value, the lazy inserter
1144  // will never insert a load.
1145  while (!PN->use_empty())
1146    RewriteHeapSROALoadUser(Load, PN->use_back(), FieldGlobals, PHIsForField);
1147  PN->eraseFromParent();
1148}
1149
1150/// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global.  Ptr
1151/// is a value loaded from the global.  Eliminate all uses of Ptr, making them
1152/// use FieldGlobals instead.  All uses of loaded values satisfy
1153/// GlobalLoadUsesSimpleEnoughForHeapSRA.
1154static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load,
1155                             const std::vector<GlobalVariable*> &FieldGlobals) {
1156  std::vector<Value *> InsertedLoadsForPtr;
1157  //InsertedLoadsForPtr.resize(FieldGlobals.size());
1158  while (!Load->use_empty())
1159    RewriteHeapSROALoadUser(Load, Load->use_back(),
1160                            FieldGlobals, InsertedLoadsForPtr);
1161}
1162
1163/// PerformHeapAllocSRoA - MI is an allocation of an array of structures.  Break
1164/// it up into multiple allocations of arrays of the fields.
1165static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, MallocInst *MI){
1166  DOUT << "SROA HEAP ALLOC: " << *GV << "  MALLOC = " << *MI;
1167  const StructType *STy = cast<StructType>(MI->getAllocatedType());
1168
1169  // There is guaranteed to be at least one use of the malloc (storing
1170  // it into GV).  If there are other uses, change them to be uses of
1171  // the global to simplify later code.  This also deletes the store
1172  // into GV.
1173  ReplaceUsesOfMallocWithGlobal(MI, GV);
1174
1175  // Okay, at this point, there are no users of the malloc.  Insert N
1176  // new mallocs at the same place as MI, and N globals.
1177  std::vector<GlobalVariable*> FieldGlobals;
1178  std::vector<MallocInst*> FieldMallocs;
1179
1180  for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){
1181    const Type *FieldTy = STy->getElementType(FieldNo);
1182    const Type *PFieldTy = PointerType::getUnqual(FieldTy);
1183
1184    GlobalVariable *NGV =
1185      new GlobalVariable(PFieldTy, false, GlobalValue::InternalLinkage,
1186                         Constant::getNullValue(PFieldTy),
1187                         GV->getName() + ".f" + utostr(FieldNo), GV,
1188                         GV->isThreadLocal());
1189    FieldGlobals.push_back(NGV);
1190
1191    MallocInst *NMI = new MallocInst(FieldTy, MI->getArraySize(),
1192                                     MI->getName() + ".f" + utostr(FieldNo),MI);
1193    FieldMallocs.push_back(NMI);
1194    new StoreInst(NMI, NGV, MI);
1195  }
1196
1197  // The tricky aspect of this transformation is handling the case when malloc
1198  // fails.  In the original code, malloc failing would set the result pointer
1199  // of malloc to null.  In this case, some mallocs could succeed and others
1200  // could fail.  As such, we emit code that looks like this:
1201  //    F0 = malloc(field0)
1202  //    F1 = malloc(field1)
1203  //    F2 = malloc(field2)
1204  //    if (F0 == 0 || F1 == 0 || F2 == 0) {
1205  //      if (F0) { free(F0); F0 = 0; }
1206  //      if (F1) { free(F1); F1 = 0; }
1207  //      if (F2) { free(F2); F2 = 0; }
1208  //    }
1209  Value *RunningOr = 0;
1210  for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) {
1211    Value *Cond = new ICmpInst(ICmpInst::ICMP_EQ, FieldMallocs[i],
1212                             Constant::getNullValue(FieldMallocs[i]->getType()),
1213                                  "isnull", MI);
1214    if (!RunningOr)
1215      RunningOr = Cond;   // First seteq
1216    else
1217      RunningOr = BinaryOperator::CreateOr(RunningOr, Cond, "tmp", MI);
1218  }
1219
1220  // Split the basic block at the old malloc.
1221  BasicBlock *OrigBB = MI->getParent();
1222  BasicBlock *ContBB = OrigBB->splitBasicBlock(MI, "malloc_cont");
1223
1224  // Create the block to check the first condition.  Put all these blocks at the
1225  // end of the function as they are unlikely to be executed.
1226  BasicBlock *NullPtrBlock = BasicBlock::Create("malloc_ret_null",
1227                                                OrigBB->getParent());
1228
1229  // Remove the uncond branch from OrigBB to ContBB, turning it into a cond
1230  // branch on RunningOr.
1231  OrigBB->getTerminator()->eraseFromParent();
1232  BranchInst::Create(NullPtrBlock, ContBB, RunningOr, OrigBB);
1233
1234  // Within the NullPtrBlock, we need to emit a comparison and branch for each
1235  // pointer, because some may be null while others are not.
1236  for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1237    Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock);
1238    Value *Cmp = new ICmpInst(ICmpInst::ICMP_NE, GVVal,
1239                              Constant::getNullValue(GVVal->getType()),
1240                              "tmp", NullPtrBlock);
1241    BasicBlock *FreeBlock = BasicBlock::Create("free_it", OrigBB->getParent());
1242    BasicBlock *NextBlock = BasicBlock::Create("next", OrigBB->getParent());
1243    BranchInst::Create(FreeBlock, NextBlock, Cmp, NullPtrBlock);
1244
1245    // Fill in FreeBlock.
1246    new FreeInst(GVVal, FreeBlock);
1247    new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i],
1248                  FreeBlock);
1249    BranchInst::Create(NextBlock, FreeBlock);
1250
1251    NullPtrBlock = NextBlock;
1252  }
1253
1254  BranchInst::Create(ContBB, NullPtrBlock);
1255
1256  // MI is no longer needed, remove it.
1257  MI->eraseFromParent();
1258
1259
1260  // Okay, the malloc site is completely handled.  All of the uses of GV are now
1261  // loads, and all uses of those loads are simple.  Rewrite them to use loads
1262  // of the per-field globals instead.
1263  while (!GV->use_empty()) {
1264    if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) {
1265      RewriteUsesOfLoadForHeapSRoA(LI, FieldGlobals);
1266      LI->eraseFromParent();
1267    } else {
1268      // Must be a store of null.
1269      StoreInst *SI = cast<StoreInst>(GV->use_back());
1270      assert(isa<Constant>(SI->getOperand(0)) &&
1271             cast<Constant>(SI->getOperand(0))->isNullValue() &&
1272             "Unexpected heap-sra user!");
1273
1274      // Insert a store of null into each global.
1275      for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1276        Constant *Null =
1277          Constant::getNullValue(FieldGlobals[i]->getType()->getElementType());
1278        new StoreInst(Null, FieldGlobals[i], SI);
1279      }
1280      // Erase the original store.
1281      SI->eraseFromParent();
1282    }
1283  }
1284
1285  // The old global is now dead, remove it.
1286  GV->eraseFromParent();
1287
1288  ++NumHeapSRA;
1289  return FieldGlobals[0];
1290}
1291
1292/// TryToOptimizeStoreOfMallocToGlobal - This function is called when we see a
1293/// pointer global variable with a single value stored it that is a malloc or
1294/// cast of malloc.
1295static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV,
1296                                               MallocInst *MI,
1297                                               Module::global_iterator &GVI,
1298                                               TargetData &TD) {
1299  // If this is a malloc of an abstract type, don't touch it.
1300  if (!MI->getAllocatedType()->isSized())
1301    return false;
1302
1303  // We can't optimize this global unless all uses of it are *known* to be
1304  // of the malloc value, not of the null initializer value (consider a use
1305  // that compares the global's value against zero to see if the malloc has
1306  // been reached).  To do this, we check to see if all uses of the global
1307  // would trap if the global were null: this proves that they must all
1308  // happen after the malloc.
1309  if (!AllUsesOfLoadedValueWillTrapIfNull(GV))
1310    return false;
1311
1312  // We can't optimize this if the malloc itself is used in a complex way,
1313  // for example, being stored into multiple globals.  This allows the
1314  // malloc to be stored into the specified global, loaded setcc'd, and
1315  // GEP'd.  These are all things we could transform to using the global
1316  // for.
1317  {
1318    SmallPtrSet<PHINode*, 8> PHIs;
1319    if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(MI, GV, PHIs))
1320      return false;
1321  }
1322
1323
1324  // If we have a global that is only initialized with a fixed size malloc,
1325  // transform the program to use global memory instead of malloc'd memory.
1326  // This eliminates dynamic allocation, avoids an indirection accessing the
1327  // data, and exposes the resultant global to further GlobalOpt.
1328  if (ConstantInt *NElements = dyn_cast<ConstantInt>(MI->getArraySize())) {
1329    // Restrict this transformation to only working on small allocations
1330    // (2048 bytes currently), as we don't want to introduce a 16M global or
1331    // something.
1332    if (NElements->getZExtValue()*
1333        TD.getABITypeSize(MI->getAllocatedType()) < 2048) {
1334      GVI = OptimizeGlobalAddressOfMalloc(GV, MI);
1335      return true;
1336    }
1337  }
1338
1339  // If the allocation is an array of structures, consider transforming this
1340  // into multiple malloc'd arrays, one for each field.  This is basically
1341  // SRoA for malloc'd memory.
1342  if (const StructType *AllocTy =
1343      dyn_cast<StructType>(MI->getAllocatedType())) {
1344    // This the structure has an unreasonable number of fields, leave it
1345    // alone.
1346    if (AllocTy->getNumElements() <= 16 && AllocTy->getNumElements() > 0 &&
1347        GlobalLoadUsesSimpleEnoughForHeapSRA(GV, MI)) {
1348      GVI = PerformHeapAllocSRoA(GV, MI);
1349      return true;
1350    }
1351  }
1352
1353  return false;
1354}
1355
1356// OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge
1357// that only one value (besides its initializer) is ever stored to the global.
1358static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal,
1359                                     Module::global_iterator &GVI,
1360                                     TargetData &TD) {
1361  if (BitCastInst *CI = dyn_cast<BitCastInst>(StoredOnceVal))
1362    StoredOnceVal = CI->getOperand(0);
1363  else if (GetElementPtrInst *GEPI =dyn_cast<GetElementPtrInst>(StoredOnceVal)){
1364    // "getelementptr Ptr, 0, 0, 0" is really just a cast.
1365    if (GEPI->hasAllZeroIndices())
1366      StoredOnceVal = GEPI->getOperand(0);
1367  }
1368
1369  // If we are dealing with a pointer global that is initialized to null and
1370  // only has one (non-null) value stored into it, then we can optimize any
1371  // users of the loaded value (often calls and loads) that would trap if the
1372  // value was null.
1373  if (isa<PointerType>(GV->getInitializer()->getType()) &&
1374      GV->getInitializer()->isNullValue()) {
1375    if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) {
1376      if (GV->getInitializer()->getType() != SOVC->getType())
1377        SOVC = ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType());
1378
1379      // Optimize away any trapping uses of the loaded value.
1380      if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC))
1381        return true;
1382    } else if (MallocInst *MI = dyn_cast<MallocInst>(StoredOnceVal)) {
1383      if (TryToOptimizeStoreOfMallocToGlobal(GV, MI, GVI, TD))
1384        return true;
1385    }
1386  }
1387
1388  return false;
1389}
1390
1391/// TryToShrinkGlobalToBoolean - At this point, we have learned that the only
1392/// two values ever stored into GV are its initializer and OtherVal.  See if we
1393/// can shrink the global into a boolean and select between the two values
1394/// whenever it is used.  This exposes the values to other scalar optimizations.
1395static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) {
1396  const Type *GVElType = GV->getType()->getElementType();
1397
1398  // If GVElType is already i1, it is already shrunk.  If the type of the GV is
1399  // an FP value or vector, don't do this optimization because a select between
1400  // them is very expensive and unlikely to lead to later simplification.
1401  if (GVElType == Type::Int1Ty || GVElType->isFloatingPoint() ||
1402      isa<VectorType>(GVElType))
1403    return false;
1404
1405  // Walk the use list of the global seeing if all the uses are load or store.
1406  // If there is anything else, bail out.
1407  for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I)
1408    if (!isa<LoadInst>(I) && !isa<StoreInst>(I))
1409      return false;
1410
1411  DOUT << "   *** SHRINKING TO BOOL: " << *GV;
1412
1413  // Create the new global, initializing it to false.
1414  GlobalVariable *NewGV = new GlobalVariable(Type::Int1Ty, false,
1415         GlobalValue::InternalLinkage, ConstantInt::getFalse(),
1416                                             GV->getName()+".b",
1417                                             (Module *)NULL,
1418                                             GV->isThreadLocal());
1419  GV->getParent()->getGlobalList().insert(GV, NewGV);
1420
1421  Constant *InitVal = GV->getInitializer();
1422  assert(InitVal->getType() != Type::Int1Ty && "No reason to shrink to bool!");
1423
1424  // If initialized to zero and storing one into the global, we can use a cast
1425  // instead of a select to synthesize the desired value.
1426  bool IsOneZero = false;
1427  if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal))
1428    IsOneZero = InitVal->isNullValue() && CI->isOne();
1429
1430  while (!GV->use_empty()) {
1431    Instruction *UI = cast<Instruction>(GV->use_back());
1432    if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
1433      // Change the store into a boolean store.
1434      bool StoringOther = SI->getOperand(0) == OtherVal;
1435      // Only do this if we weren't storing a loaded value.
1436      Value *StoreVal;
1437      if (StoringOther || SI->getOperand(0) == InitVal)
1438        StoreVal = ConstantInt::get(Type::Int1Ty, StoringOther);
1439      else {
1440        // Otherwise, we are storing a previously loaded copy.  To do this,
1441        // change the copy from copying the original value to just copying the
1442        // bool.
1443        Instruction *StoredVal = cast<Instruction>(SI->getOperand(0));
1444
1445        // If we're already replaced the input, StoredVal will be a cast or
1446        // select instruction.  If not, it will be a load of the original
1447        // global.
1448        if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
1449          assert(LI->getOperand(0) == GV && "Not a copy!");
1450          // Insert a new load, to preserve the saved value.
1451          StoreVal = new LoadInst(NewGV, LI->getName()+".b", LI);
1452        } else {
1453          assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) &&
1454                 "This is not a form that we understand!");
1455          StoreVal = StoredVal->getOperand(0);
1456          assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!");
1457        }
1458      }
1459      new StoreInst(StoreVal, NewGV, SI);
1460    } else {
1461      // Change the load into a load of bool then a select.
1462      LoadInst *LI = cast<LoadInst>(UI);
1463      LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", LI);
1464      Value *NSI;
1465      if (IsOneZero)
1466        NSI = new ZExtInst(NLI, LI->getType(), "", LI);
1467      else
1468        NSI = SelectInst::Create(NLI, OtherVal, InitVal, "", LI);
1469      NSI->takeName(LI);
1470      LI->replaceAllUsesWith(NSI);
1471    }
1472    UI->eraseFromParent();
1473  }
1474
1475  GV->eraseFromParent();
1476  return true;
1477}
1478
1479
1480/// ProcessInternalGlobal - Analyze the specified global variable and optimize
1481/// it if possible.  If we make a change, return true.
1482bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV,
1483                                      Module::global_iterator &GVI) {
1484  std::set<PHINode*> PHIUsers;
1485  GlobalStatus GS;
1486  GV->removeDeadConstantUsers();
1487
1488  if (GV->use_empty()) {
1489    DOUT << "GLOBAL DEAD: " << *GV;
1490    GV->eraseFromParent();
1491    ++NumDeleted;
1492    return true;
1493  }
1494
1495  if (!AnalyzeGlobal(GV, GS, PHIUsers)) {
1496#if 0
1497    cerr << "Global: " << *GV;
1498    cerr << "  isLoaded = " << GS.isLoaded << "\n";
1499    cerr << "  StoredType = ";
1500    switch (GS.StoredType) {
1501    case GlobalStatus::NotStored: cerr << "NEVER STORED\n"; break;
1502    case GlobalStatus::isInitializerStored: cerr << "INIT STORED\n"; break;
1503    case GlobalStatus::isStoredOnce: cerr << "STORED ONCE\n"; break;
1504    case GlobalStatus::isStored: cerr << "stored\n"; break;
1505    }
1506    if (GS.StoredType == GlobalStatus::isStoredOnce && GS.StoredOnceValue)
1507      cerr << "  StoredOnceValue = " << *GS.StoredOnceValue << "\n";
1508    if (GS.AccessingFunction && !GS.HasMultipleAccessingFunctions)
1509      cerr << "  AccessingFunction = " << GS.AccessingFunction->getName()
1510                << "\n";
1511    cerr << "  HasMultipleAccessingFunctions =  "
1512              << GS.HasMultipleAccessingFunctions << "\n";
1513    cerr << "  HasNonInstructionUser = " << GS.HasNonInstructionUser<<"\n";
1514    cerr << "\n";
1515#endif
1516
1517    // If this is a first class global and has only one accessing function
1518    // and this function is main (which we know is not recursive we can make
1519    // this global a local variable) we replace the global with a local alloca
1520    // in this function.
1521    //
1522    // NOTE: It doesn't make sense to promote non single-value types since we
1523    // are just replacing static memory to stack memory.
1524    if (!GS.HasMultipleAccessingFunctions &&
1525        GS.AccessingFunction && !GS.HasNonInstructionUser &&
1526        GV->getType()->getElementType()->isSingleValueType() &&
1527        GS.AccessingFunction->getName() == "main" &&
1528        GS.AccessingFunction->hasExternalLinkage()) {
1529      DOUT << "LOCALIZING GLOBAL: " << *GV;
1530      Instruction* FirstI = GS.AccessingFunction->getEntryBlock().begin();
1531      const Type* ElemTy = GV->getType()->getElementType();
1532      // FIXME: Pass Global's alignment when globals have alignment
1533      AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), FirstI);
1534      if (!isa<UndefValue>(GV->getInitializer()))
1535        new StoreInst(GV->getInitializer(), Alloca, FirstI);
1536
1537      GV->replaceAllUsesWith(Alloca);
1538      GV->eraseFromParent();
1539      ++NumLocalized;
1540      return true;
1541    }
1542
1543    // If the global is never loaded (but may be stored to), it is dead.
1544    // Delete it now.
1545    if (!GS.isLoaded) {
1546      DOUT << "GLOBAL NEVER LOADED: " << *GV;
1547
1548      // Delete any stores we can find to the global.  We may not be able to
1549      // make it completely dead though.
1550      bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer());
1551
1552      // If the global is dead now, delete it.
1553      if (GV->use_empty()) {
1554        GV->eraseFromParent();
1555        ++NumDeleted;
1556        Changed = true;
1557      }
1558      return Changed;
1559
1560    } else if (GS.StoredType <= GlobalStatus::isInitializerStored) {
1561      DOUT << "MARKING CONSTANT: " << *GV;
1562      GV->setConstant(true);
1563
1564      // Clean up any obviously simplifiable users now.
1565      CleanupConstantGlobalUsers(GV, GV->getInitializer());
1566
1567      // If the global is dead now, just nuke it.
1568      if (GV->use_empty()) {
1569        DOUT << "   *** Marking constant allowed us to simplify "
1570             << "all users and delete global!\n";
1571        GV->eraseFromParent();
1572        ++NumDeleted;
1573      }
1574
1575      ++NumMarked;
1576      return true;
1577    } else if (!GV->getInitializer()->getType()->isSingleValueType()) {
1578      if (GlobalVariable *FirstNewGV = SRAGlobal(GV,
1579                                                 getAnalysis<TargetData>())) {
1580        GVI = FirstNewGV;  // Don't skip the newly produced globals!
1581        return true;
1582      }
1583    } else if (GS.StoredType == GlobalStatus::isStoredOnce) {
1584      // If the initial value for the global was an undef value, and if only
1585      // one other value was stored into it, we can just change the
1586      // initializer to be an undef value, then delete all stores to the
1587      // global.  This allows us to mark it constant.
1588      if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1589        if (isa<UndefValue>(GV->getInitializer())) {
1590          // Change the initial value here.
1591          GV->setInitializer(SOVConstant);
1592
1593          // Clean up any obviously simplifiable users now.
1594          CleanupConstantGlobalUsers(GV, GV->getInitializer());
1595
1596          if (GV->use_empty()) {
1597            DOUT << "   *** Substituting initializer allowed us to "
1598                 << "simplify all users and delete global!\n";
1599            GV->eraseFromParent();
1600            ++NumDeleted;
1601          } else {
1602            GVI = GV;
1603          }
1604          ++NumSubstitute;
1605          return true;
1606        }
1607
1608      // Try to optimize globals based on the knowledge that only one value
1609      // (besides its initializer) is ever stored to the global.
1610      if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI,
1611                                   getAnalysis<TargetData>()))
1612        return true;
1613
1614      // Otherwise, if the global was not a boolean, we can shrink it to be a
1615      // boolean.
1616      if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1617        if (TryToShrinkGlobalToBoolean(GV, SOVConstant)) {
1618          ++NumShrunkToBool;
1619          return true;
1620        }
1621    }
1622  }
1623  return false;
1624}
1625
1626/// OnlyCalledDirectly - Return true if the specified function is only called
1627/// directly.  In other words, its address is never taken.
1628static bool OnlyCalledDirectly(Function *F) {
1629  for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1630    Instruction *User = dyn_cast<Instruction>(*UI);
1631    if (!User) return false;
1632    if (!isa<CallInst>(User) && !isa<InvokeInst>(User)) return false;
1633
1634    // See if the function address is passed as an argument.
1635    for (User::op_iterator i = User->op_begin() + 1, e = User->op_end();
1636         i != e; ++i)
1637      if (*i == F) return false;
1638  }
1639  return true;
1640}
1641
1642/// ChangeCalleesToFastCall - Walk all of the direct calls of the specified
1643/// function, changing them to FastCC.
1644static void ChangeCalleesToFastCall(Function *F) {
1645  for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1646    CallSite User(cast<Instruction>(*UI));
1647    User.setCallingConv(CallingConv::Fast);
1648  }
1649}
1650
1651static AttrListPtr StripNest(const AttrListPtr &Attrs) {
1652  for (unsigned i = 0, e = Attrs.getNumSlots(); i != e; ++i) {
1653    if ((Attrs.getSlot(i).Attrs & Attribute::Nest) == 0)
1654      continue;
1655
1656    // There can be only one.
1657    return Attrs.removeAttr(Attrs.getSlot(i).Index, Attribute::Nest);
1658  }
1659
1660  return Attrs;
1661}
1662
1663static void RemoveNestAttribute(Function *F) {
1664  F->setAttributes(StripNest(F->getAttributes()));
1665  for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1666    CallSite User(cast<Instruction>(*UI));
1667    User.setAttributes(StripNest(User.getAttributes()));
1668  }
1669}
1670
1671bool GlobalOpt::OptimizeFunctions(Module &M) {
1672  bool Changed = false;
1673  // Optimize functions.
1674  for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) {
1675    Function *F = FI++;
1676    F->removeDeadConstantUsers();
1677    if (F->use_empty() && (F->hasInternalLinkage() ||
1678                           F->hasLinkOnceLinkage())) {
1679      M.getFunctionList().erase(F);
1680      Changed = true;
1681      ++NumFnDeleted;
1682    } else if (F->hasInternalLinkage()) {
1683      if (F->getCallingConv() == CallingConv::C && !F->isVarArg() &&
1684          OnlyCalledDirectly(F)) {
1685        // If this function has C calling conventions, is not a varargs
1686        // function, and is only called directly, promote it to use the Fast
1687        // calling convention.
1688        F->setCallingConv(CallingConv::Fast);
1689        ChangeCalleesToFastCall(F);
1690        ++NumFastCallFns;
1691        Changed = true;
1692      }
1693
1694      if (F->getAttributes().hasAttrSomewhere(Attribute::Nest) &&
1695          OnlyCalledDirectly(F)) {
1696        // The function is not used by a trampoline intrinsic, so it is safe
1697        // to remove the 'nest' attribute.
1698        RemoveNestAttribute(F);
1699        ++NumNestRemoved;
1700        Changed = true;
1701      }
1702    }
1703  }
1704  return Changed;
1705}
1706
1707bool GlobalOpt::OptimizeGlobalVars(Module &M) {
1708  bool Changed = false;
1709  for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
1710       GVI != E; ) {
1711    GlobalVariable *GV = GVI++;
1712    if (!GV->isConstant() && GV->hasInternalLinkage() &&
1713        GV->hasInitializer())
1714      Changed |= ProcessInternalGlobal(GV, GVI);
1715  }
1716  return Changed;
1717}
1718
1719/// FindGlobalCtors - Find the llvm.globalctors list, verifying that all
1720/// initializers have an init priority of 65535.
1721GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) {
1722  for (Module::global_iterator I = M.global_begin(), E = M.global_end();
1723       I != E; ++I)
1724    if (I->getName() == "llvm.global_ctors") {
1725      // Found it, verify it's an array of { int, void()* }.
1726      const ArrayType *ATy =dyn_cast<ArrayType>(I->getType()->getElementType());
1727      if (!ATy) return 0;
1728      const StructType *STy = dyn_cast<StructType>(ATy->getElementType());
1729      if (!STy || STy->getNumElements() != 2 ||
1730          STy->getElementType(0) != Type::Int32Ty) return 0;
1731      const PointerType *PFTy = dyn_cast<PointerType>(STy->getElementType(1));
1732      if (!PFTy) return 0;
1733      const FunctionType *FTy = dyn_cast<FunctionType>(PFTy->getElementType());
1734      if (!FTy || FTy->getReturnType() != Type::VoidTy || FTy->isVarArg() ||
1735          FTy->getNumParams() != 0)
1736        return 0;
1737
1738      // Verify that the initializer is simple enough for us to handle.
1739      if (!I->hasInitializer()) return 0;
1740      ConstantArray *CA = dyn_cast<ConstantArray>(I->getInitializer());
1741      if (!CA) return 0;
1742      for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i)
1743        if (ConstantStruct *CS = dyn_cast<ConstantStruct>(*i)) {
1744          if (isa<ConstantPointerNull>(CS->getOperand(1)))
1745            continue;
1746
1747          // Must have a function or null ptr.
1748          if (!isa<Function>(CS->getOperand(1)))
1749            return 0;
1750
1751          // Init priority must be standard.
1752          ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0));
1753          if (!CI || CI->getZExtValue() != 65535)
1754            return 0;
1755        } else {
1756          return 0;
1757        }
1758
1759      return I;
1760    }
1761  return 0;
1762}
1763
1764/// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand,
1765/// return a list of the functions and null terminator as a vector.
1766static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) {
1767  ConstantArray *CA = cast<ConstantArray>(GV->getInitializer());
1768  std::vector<Function*> Result;
1769  Result.reserve(CA->getNumOperands());
1770  for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) {
1771    ConstantStruct *CS = cast<ConstantStruct>(*i);
1772    Result.push_back(dyn_cast<Function>(CS->getOperand(1)));
1773  }
1774  return Result;
1775}
1776
1777/// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the
1778/// specified array, returning the new global to use.
1779static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL,
1780                                          const std::vector<Function*> &Ctors) {
1781  // If we made a change, reassemble the initializer list.
1782  std::vector<Constant*> CSVals;
1783  CSVals.push_back(ConstantInt::get(Type::Int32Ty, 65535));
1784  CSVals.push_back(0);
1785
1786  // Create the new init list.
1787  std::vector<Constant*> CAList;
1788  for (unsigned i = 0, e = Ctors.size(); i != e; ++i) {
1789    if (Ctors[i]) {
1790      CSVals[1] = Ctors[i];
1791    } else {
1792      const Type *FTy = FunctionType::get(Type::VoidTy,
1793                                          std::vector<const Type*>(), false);
1794      const PointerType *PFTy = PointerType::getUnqual(FTy);
1795      CSVals[1] = Constant::getNullValue(PFTy);
1796      CSVals[0] = ConstantInt::get(Type::Int32Ty, 2147483647);
1797    }
1798    CAList.push_back(ConstantStruct::get(CSVals));
1799  }
1800
1801  // Create the array initializer.
1802  const Type *StructTy =
1803    cast<ArrayType>(GCL->getType()->getElementType())->getElementType();
1804  Constant *CA = ConstantArray::get(ArrayType::get(StructTy, CAList.size()),
1805                                    CAList);
1806
1807  // If we didn't change the number of elements, don't create a new GV.
1808  if (CA->getType() == GCL->getInitializer()->getType()) {
1809    GCL->setInitializer(CA);
1810    return GCL;
1811  }
1812
1813  // Create the new global and insert it next to the existing list.
1814  GlobalVariable *NGV = new GlobalVariable(CA->getType(), GCL->isConstant(),
1815                                           GCL->getLinkage(), CA, "",
1816                                           (Module *)NULL,
1817                                           GCL->isThreadLocal());
1818  GCL->getParent()->getGlobalList().insert(GCL, NGV);
1819  NGV->takeName(GCL);
1820
1821  // Nuke the old list, replacing any uses with the new one.
1822  if (!GCL->use_empty()) {
1823    Constant *V = NGV;
1824    if (V->getType() != GCL->getType())
1825      V = ConstantExpr::getBitCast(V, GCL->getType());
1826    GCL->replaceAllUsesWith(V);
1827  }
1828  GCL->eraseFromParent();
1829
1830  if (Ctors.size())
1831    return NGV;
1832  else
1833    return 0;
1834}
1835
1836
1837static Constant *getVal(std::map<Value*, Constant*> &ComputedValues,
1838                        Value *V) {
1839  if (Constant *CV = dyn_cast<Constant>(V)) return CV;
1840  Constant *R = ComputedValues[V];
1841  assert(R && "Reference to an uncomputed value!");
1842  return R;
1843}
1844
1845/// isSimpleEnoughPointerToCommit - Return true if this constant is simple
1846/// enough for us to understand.  In particular, if it is a cast of something,
1847/// we punt.  We basically just support direct accesses to globals and GEP's of
1848/// globals.  This should be kept up to date with CommitValueTo.
1849static bool isSimpleEnoughPointerToCommit(Constant *C) {
1850  if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) {
1851    if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage())
1852      return false;  // do not allow weak/linkonce/dllimport/dllexport linkage.
1853    return !GV->isDeclaration();  // reject external globals.
1854  }
1855  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
1856    // Handle a constantexpr gep.
1857    if (CE->getOpcode() == Instruction::GetElementPtr &&
1858        isa<GlobalVariable>(CE->getOperand(0))) {
1859      GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1860      if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage())
1861        return false;  // do not allow weak/linkonce/dllimport/dllexport linkage.
1862      return GV->hasInitializer() &&
1863             ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
1864    }
1865  return false;
1866}
1867
1868/// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global
1869/// initializer.  This returns 'Init' modified to reflect 'Val' stored into it.
1870/// At this point, the GEP operands of Addr [0, OpNo) have been stepped into.
1871static Constant *EvaluateStoreInto(Constant *Init, Constant *Val,
1872                                   ConstantExpr *Addr, unsigned OpNo) {
1873  // Base case of the recursion.
1874  if (OpNo == Addr->getNumOperands()) {
1875    assert(Val->getType() == Init->getType() && "Type mismatch!");
1876    return Val;
1877  }
1878
1879  if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
1880    std::vector<Constant*> Elts;
1881
1882    // Break up the constant into its elements.
1883    if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
1884      for (User::op_iterator i = CS->op_begin(), e = CS->op_end(); i != e; ++i)
1885        Elts.push_back(cast<Constant>(*i));
1886    } else if (isa<ConstantAggregateZero>(Init)) {
1887      for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1888        Elts.push_back(Constant::getNullValue(STy->getElementType(i)));
1889    } else if (isa<UndefValue>(Init)) {
1890      for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1891        Elts.push_back(UndefValue::get(STy->getElementType(i)));
1892    } else {
1893      assert(0 && "This code is out of sync with "
1894             " ConstantFoldLoadThroughGEPConstantExpr");
1895    }
1896
1897    // Replace the element that we are supposed to.
1898    ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo));
1899    unsigned Idx = CU->getZExtValue();
1900    assert(Idx < STy->getNumElements() && "Struct index out of range!");
1901    Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1);
1902
1903    // Return the modified struct.
1904    return ConstantStruct::get(&Elts[0], Elts.size(), STy->isPacked());
1905  } else {
1906    ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo));
1907    const ArrayType *ATy = cast<ArrayType>(Init->getType());
1908
1909    // Break up the array into elements.
1910    std::vector<Constant*> Elts;
1911    if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
1912      for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i)
1913        Elts.push_back(cast<Constant>(*i));
1914    } else if (isa<ConstantAggregateZero>(Init)) {
1915      Constant *Elt = Constant::getNullValue(ATy->getElementType());
1916      Elts.assign(ATy->getNumElements(), Elt);
1917    } else if (isa<UndefValue>(Init)) {
1918      Constant *Elt = UndefValue::get(ATy->getElementType());
1919      Elts.assign(ATy->getNumElements(), Elt);
1920    } else {
1921      assert(0 && "This code is out of sync with "
1922             " ConstantFoldLoadThroughGEPConstantExpr");
1923    }
1924
1925    assert(CI->getZExtValue() < ATy->getNumElements());
1926    Elts[CI->getZExtValue()] =
1927      EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1);
1928    return ConstantArray::get(ATy, Elts);
1929  }
1930}
1931
1932/// CommitValueTo - We have decided that Addr (which satisfies the predicate
1933/// isSimpleEnoughPointerToCommit) should get Val as its value.  Make it happen.
1934static void CommitValueTo(Constant *Val, Constant *Addr) {
1935  if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
1936    assert(GV->hasInitializer());
1937    GV->setInitializer(Val);
1938    return;
1939  }
1940
1941  ConstantExpr *CE = cast<ConstantExpr>(Addr);
1942  GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1943
1944  Constant *Init = GV->getInitializer();
1945  Init = EvaluateStoreInto(Init, Val, CE, 2);
1946  GV->setInitializer(Init);
1947}
1948
1949/// ComputeLoadResult - Return the value that would be computed by a load from
1950/// P after the stores reflected by 'memory' have been performed.  If we can't
1951/// decide, return null.
1952static Constant *ComputeLoadResult(Constant *P,
1953                                const std::map<Constant*, Constant*> &Memory) {
1954  // If this memory location has been recently stored, use the stored value: it
1955  // is the most up-to-date.
1956  std::map<Constant*, Constant*>::const_iterator I = Memory.find(P);
1957  if (I != Memory.end()) return I->second;
1958
1959  // Access it.
1960  if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
1961    if (GV->hasInitializer())
1962      return GV->getInitializer();
1963    return 0;
1964  }
1965
1966  // Handle a constantexpr getelementptr.
1967  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P))
1968    if (CE->getOpcode() == Instruction::GetElementPtr &&
1969        isa<GlobalVariable>(CE->getOperand(0))) {
1970      GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1971      if (GV->hasInitializer())
1972        return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
1973    }
1974
1975  return 0;  // don't know how to evaluate.
1976}
1977
1978/// EvaluateFunction - Evaluate a call to function F, returning true if
1979/// successful, false if we can't evaluate it.  ActualArgs contains the formal
1980/// arguments for the function.
1981static bool EvaluateFunction(Function *F, Constant *&RetVal,
1982                             const std::vector<Constant*> &ActualArgs,
1983                             std::vector<Function*> &CallStack,
1984                             std::map<Constant*, Constant*> &MutatedMemory,
1985                             std::vector<GlobalVariable*> &AllocaTmps) {
1986  // Check to see if this function is already executing (recursion).  If so,
1987  // bail out.  TODO: we might want to accept limited recursion.
1988  if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end())
1989    return false;
1990
1991  CallStack.push_back(F);
1992
1993  /// Values - As we compute SSA register values, we store their contents here.
1994  std::map<Value*, Constant*> Values;
1995
1996  // Initialize arguments to the incoming values specified.
1997  unsigned ArgNo = 0;
1998  for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E;
1999       ++AI, ++ArgNo)
2000    Values[AI] = ActualArgs[ArgNo];
2001
2002  /// ExecutedBlocks - We only handle non-looping, non-recursive code.  As such,
2003  /// we can only evaluate any one basic block at most once.  This set keeps
2004  /// track of what we have executed so we can detect recursive cases etc.
2005  std::set<BasicBlock*> ExecutedBlocks;
2006
2007  // CurInst - The current instruction we're evaluating.
2008  BasicBlock::iterator CurInst = F->begin()->begin();
2009
2010  // This is the main evaluation loop.
2011  while (1) {
2012    Constant *InstResult = 0;
2013
2014    if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) {
2015      if (SI->isVolatile()) return false;  // no volatile accesses.
2016      Constant *Ptr = getVal(Values, SI->getOperand(1));
2017      if (!isSimpleEnoughPointerToCommit(Ptr))
2018        // If this is too complex for us to commit, reject it.
2019        return false;
2020      Constant *Val = getVal(Values, SI->getOperand(0));
2021      MutatedMemory[Ptr] = Val;
2022    } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) {
2023      InstResult = ConstantExpr::get(BO->getOpcode(),
2024                                     getVal(Values, BO->getOperand(0)),
2025                                     getVal(Values, BO->getOperand(1)));
2026    } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) {
2027      InstResult = ConstantExpr::getCompare(CI->getPredicate(),
2028                                            getVal(Values, CI->getOperand(0)),
2029                                            getVal(Values, CI->getOperand(1)));
2030    } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) {
2031      InstResult = ConstantExpr::getCast(CI->getOpcode(),
2032                                         getVal(Values, CI->getOperand(0)),
2033                                         CI->getType());
2034    } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) {
2035      InstResult = ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)),
2036                                           getVal(Values, SI->getOperand(1)),
2037                                           getVal(Values, SI->getOperand(2)));
2038    } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) {
2039      Constant *P = getVal(Values, GEP->getOperand(0));
2040      SmallVector<Constant*, 8> GEPOps;
2041      for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end();
2042           i != e; ++i)
2043        GEPOps.push_back(getVal(Values, *i));
2044      InstResult = ConstantExpr::getGetElementPtr(P, &GEPOps[0], GEPOps.size());
2045    } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) {
2046      if (LI->isVolatile()) return false;  // no volatile accesses.
2047      InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)),
2048                                     MutatedMemory);
2049      if (InstResult == 0) return false; // Could not evaluate load.
2050    } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) {
2051      if (AI->isArrayAllocation()) return false;  // Cannot handle array allocs.
2052      const Type *Ty = AI->getType()->getElementType();
2053      AllocaTmps.push_back(new GlobalVariable(Ty, false,
2054                                              GlobalValue::InternalLinkage,
2055                                              UndefValue::get(Ty),
2056                                              AI->getName()));
2057      InstResult = AllocaTmps.back();
2058    } else if (CallInst *CI = dyn_cast<CallInst>(CurInst)) {
2059      // Cannot handle inline asm.
2060      if (isa<InlineAsm>(CI->getOperand(0))) return false;
2061
2062      // Resolve function pointers.
2063      Function *Callee = dyn_cast<Function>(getVal(Values, CI->getOperand(0)));
2064      if (!Callee) return false;  // Cannot resolve.
2065
2066      std::vector<Constant*> Formals;
2067      for (User::op_iterator i = CI->op_begin() + 1, e = CI->op_end();
2068           i != e; ++i)
2069        Formals.push_back(getVal(Values, *i));
2070
2071      if (Callee->isDeclaration()) {
2072        // If this is a function we can constant fold, do it.
2073        if (Constant *C = ConstantFoldCall(Callee, &Formals[0],
2074                                           Formals.size())) {
2075          InstResult = C;
2076        } else {
2077          return false;
2078        }
2079      } else {
2080        if (Callee->getFunctionType()->isVarArg())
2081          return false;
2082
2083        Constant *RetVal;
2084
2085        // Execute the call, if successful, use the return value.
2086        if (!EvaluateFunction(Callee, RetVal, Formals, CallStack,
2087                              MutatedMemory, AllocaTmps))
2088          return false;
2089        InstResult = RetVal;
2090      }
2091    } else if (isa<TerminatorInst>(CurInst)) {
2092      BasicBlock *NewBB = 0;
2093      if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) {
2094        if (BI->isUnconditional()) {
2095          NewBB = BI->getSuccessor(0);
2096        } else {
2097          ConstantInt *Cond =
2098            dyn_cast<ConstantInt>(getVal(Values, BI->getCondition()));
2099          if (!Cond) return false;  // Cannot determine.
2100
2101          NewBB = BI->getSuccessor(!Cond->getZExtValue());
2102        }
2103      } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) {
2104        ConstantInt *Val =
2105          dyn_cast<ConstantInt>(getVal(Values, SI->getCondition()));
2106        if (!Val) return false;  // Cannot determine.
2107        NewBB = SI->getSuccessor(SI->findCaseValue(Val));
2108      } else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) {
2109        if (RI->getNumOperands())
2110          RetVal = getVal(Values, RI->getOperand(0));
2111
2112        CallStack.pop_back();  // return from fn.
2113        return true;  // We succeeded at evaluating this ctor!
2114      } else {
2115        // invoke, unwind, unreachable.
2116        return false;  // Cannot handle this terminator.
2117      }
2118
2119      // Okay, we succeeded in evaluating this control flow.  See if we have
2120      // executed the new block before.  If so, we have a looping function,
2121      // which we cannot evaluate in reasonable time.
2122      if (!ExecutedBlocks.insert(NewBB).second)
2123        return false;  // looped!
2124
2125      // Okay, we have never been in this block before.  Check to see if there
2126      // are any PHI nodes.  If so, evaluate them with information about where
2127      // we came from.
2128      BasicBlock *OldBB = CurInst->getParent();
2129      CurInst = NewBB->begin();
2130      PHINode *PN;
2131      for (; (PN = dyn_cast<PHINode>(CurInst)); ++CurInst)
2132        Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB));
2133
2134      // Do NOT increment CurInst.  We know that the terminator had no value.
2135      continue;
2136    } else {
2137      // Did not know how to evaluate this!
2138      return false;
2139    }
2140
2141    if (!CurInst->use_empty())
2142      Values[CurInst] = InstResult;
2143
2144    // Advance program counter.
2145    ++CurInst;
2146  }
2147}
2148
2149/// EvaluateStaticConstructor - Evaluate static constructors in the function, if
2150/// we can.  Return true if we can, false otherwise.
2151static bool EvaluateStaticConstructor(Function *F) {
2152  /// MutatedMemory - For each store we execute, we update this map.  Loads
2153  /// check this to get the most up-to-date value.  If evaluation is successful,
2154  /// this state is committed to the process.
2155  std::map<Constant*, Constant*> MutatedMemory;
2156
2157  /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable
2158  /// to represent its body.  This vector is needed so we can delete the
2159  /// temporary globals when we are done.
2160  std::vector<GlobalVariable*> AllocaTmps;
2161
2162  /// CallStack - This is used to detect recursion.  In pathological situations
2163  /// we could hit exponential behavior, but at least there is nothing
2164  /// unbounded.
2165  std::vector<Function*> CallStack;
2166
2167  // Call the function.
2168  Constant *RetValDummy;
2169  bool EvalSuccess = EvaluateFunction(F, RetValDummy, std::vector<Constant*>(),
2170                                       CallStack, MutatedMemory, AllocaTmps);
2171  if (EvalSuccess) {
2172    // We succeeded at evaluation: commit the result.
2173    DOUT << "FULLY EVALUATED GLOBAL CTOR FUNCTION '"
2174         << F->getName() << "' to " << MutatedMemory.size()
2175         << " stores.\n";
2176    for (std::map<Constant*, Constant*>::iterator I = MutatedMemory.begin(),
2177         E = MutatedMemory.end(); I != E; ++I)
2178      CommitValueTo(I->second, I->first);
2179  }
2180
2181  // At this point, we are done interpreting.  If we created any 'alloca'
2182  // temporaries, release them now.
2183  while (!AllocaTmps.empty()) {
2184    GlobalVariable *Tmp = AllocaTmps.back();
2185    AllocaTmps.pop_back();
2186
2187    // If there are still users of the alloca, the program is doing something
2188    // silly, e.g. storing the address of the alloca somewhere and using it
2189    // later.  Since this is undefined, we'll just make it be null.
2190    if (!Tmp->use_empty())
2191      Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType()));
2192    delete Tmp;
2193  }
2194
2195  return EvalSuccess;
2196}
2197
2198
2199
2200/// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible.
2201/// Return true if anything changed.
2202bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) {
2203  std::vector<Function*> Ctors = ParseGlobalCtors(GCL);
2204  bool MadeChange = false;
2205  if (Ctors.empty()) return false;
2206
2207  // Loop over global ctors, optimizing them when we can.
2208  for (unsigned i = 0; i != Ctors.size(); ++i) {
2209    Function *F = Ctors[i];
2210    // Found a null terminator in the middle of the list, prune off the rest of
2211    // the list.
2212    if (F == 0) {
2213      if (i != Ctors.size()-1) {
2214        Ctors.resize(i+1);
2215        MadeChange = true;
2216      }
2217      break;
2218    }
2219
2220    // We cannot simplify external ctor functions.
2221    if (F->empty()) continue;
2222
2223    // If we can evaluate the ctor at compile time, do.
2224    if (EvaluateStaticConstructor(F)) {
2225      Ctors.erase(Ctors.begin()+i);
2226      MadeChange = true;
2227      --i;
2228      ++NumCtorsEvaluated;
2229      continue;
2230    }
2231  }
2232
2233  if (!MadeChange) return false;
2234
2235  GCL = InstallGlobalCtors(GCL, Ctors);
2236  return true;
2237}
2238
2239bool GlobalOpt::ResolveAliases(Module &M) {
2240  bool Changed = false;
2241
2242  for (Module::alias_iterator I = M.alias_begin(),
2243         E = M.alias_end(); I != E; ++I) {
2244    if (I->use_empty())
2245      continue;
2246
2247    if (const GlobalValue *GV = I->resolveAliasedGlobal())
2248      if (GV != I) {
2249        I->replaceAllUsesWith(const_cast<GlobalValue*>(GV));
2250        Changed = true;
2251      }
2252  }
2253
2254  return Changed;
2255}
2256
2257bool GlobalOpt::runOnModule(Module &M) {
2258  bool Changed = false;
2259
2260  // Try to find the llvm.globalctors list.
2261  GlobalVariable *GlobalCtors = FindGlobalCtors(M);
2262
2263  bool LocalChange = true;
2264  while (LocalChange) {
2265    LocalChange = false;
2266
2267    // Delete functions that are trivially dead, ccc -> fastcc
2268    LocalChange |= OptimizeFunctions(M);
2269
2270    // Optimize global_ctors list.
2271    if (GlobalCtors)
2272      LocalChange |= OptimizeGlobalCtorsList(GlobalCtors);
2273
2274    // Optimize non-address-taken globals.
2275    LocalChange |= OptimizeGlobalVars(M);
2276
2277    // Resolve aliases, when possible.
2278    LocalChange |= ResolveAliases(M);
2279    Changed |= LocalChange;
2280  }
2281
2282  // TODO: Move all global ctors functions to the end of the module for code
2283  // layout.
2284
2285  return Changed;
2286}
2287