GlobalOpt.cpp revision 3ed470da2b05e027282e2f8c67b3739dad11745c
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/Analysis/MemoryBuiltins.h"
27#include "llvm/Target/TargetData.h"
28#include "llvm/Support/CallSite.h"
29#include "llvm/Support/Debug.h"
30#include "llvm/Support/ErrorHandling.h"
31#include "llvm/Support/GetElementPtrTypeIterator.h"
32#include "llvm/Support/MathExtras.h"
33#include "llvm/Support/raw_ostream.h"
34#include "llvm/ADT/DenseMap.h"
35#include "llvm/ADT/SmallPtrSet.h"
36#include "llvm/ADT/SmallVector.h"
37#include "llvm/ADT/Statistic.h"
38#include "llvm/ADT/STLExtras.h"
39#include <algorithm>
40using namespace llvm;
41
42STATISTIC(NumMarked    , "Number of globals marked constant");
43STATISTIC(NumSRA       , "Number of aggregate globals broken into scalars");
44STATISTIC(NumHeapSRA   , "Number of heap objects SRA'd");
45STATISTIC(NumSubstitute,"Number of globals with initializers stored into them");
46STATISTIC(NumDeleted   , "Number of globals deleted");
47STATISTIC(NumFnDeleted , "Number of functions deleted");
48STATISTIC(NumGlobUses  , "Number of global uses devirtualized");
49STATISTIC(NumLocalized , "Number of globals localized");
50STATISTIC(NumShrunkToBool  , "Number of global vars shrunk to booleans");
51STATISTIC(NumFastCallFns   , "Number of functions converted to fastcc");
52STATISTIC(NumCtorsEvaluated, "Number of static ctors evaluated");
53STATISTIC(NumNestRemoved   , "Number of nest attributes removed");
54STATISTIC(NumAliasesResolved, "Number of global aliases resolved");
55STATISTIC(NumAliasesRemoved, "Number of global aliases eliminated");
56
57namespace {
58  struct GlobalOpt : public ModulePass {
59    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
60    }
61    static char ID; // Pass identification, replacement for typeid
62    GlobalOpt() : ModulePass(&ID) {}
63
64    bool runOnModule(Module &M);
65
66  private:
67    GlobalVariable *FindGlobalCtors(Module &M);
68    bool OptimizeFunctions(Module &M);
69    bool OptimizeGlobalVars(Module &M);
70    bool OptimizeGlobalAliases(Module &M);
71    bool OptimizeGlobalCtorsList(GlobalVariable *&GCL);
72    bool ProcessInternalGlobal(GlobalVariable *GV,Module::global_iterator &GVI);
73  };
74}
75
76char GlobalOpt::ID = 0;
77static RegisterPass<GlobalOpt> X("globalopt", "Global Variable Optimizer");
78
79ModulePass *llvm::createGlobalOptimizerPass() { return new GlobalOpt(); }
80
81namespace {
82
83/// GlobalStatus - As we analyze each global, keep track of some information
84/// about it.  If we find out that the address of the global is taken, none of
85/// this info will be accurate.
86struct GlobalStatus {
87  /// isLoaded - True if the global is ever loaded.  If the global isn't ever
88  /// loaded it can be deleted.
89  bool isLoaded;
90
91  /// StoredType - Keep track of what stores to the global look like.
92  ///
93  enum StoredType {
94    /// NotStored - There is no store to this global.  It can thus be marked
95    /// constant.
96    NotStored,
97
98    /// isInitializerStored - This global is stored to, but the only thing
99    /// stored is the constant it was initialized with.  This is only tracked
100    /// for scalar globals.
101    isInitializerStored,
102
103    /// isStoredOnce - This global is stored to, but only its initializer and
104    /// one other value is ever stored to it.  If this global isStoredOnce, we
105    /// track the value stored to it in StoredOnceValue below.  This is only
106    /// tracked for scalar globals.
107    isStoredOnce,
108
109    /// isStored - This global is stored to by multiple values or something else
110    /// that we cannot track.
111    isStored
112  } StoredType;
113
114  /// StoredOnceValue - If only one value (besides the initializer constant) is
115  /// ever stored to this global, keep track of what value it is.
116  Value *StoredOnceValue;
117
118  /// AccessingFunction/HasMultipleAccessingFunctions - These start out
119  /// null/false.  When the first accessing function is noticed, it is recorded.
120  /// When a second different accessing function is noticed,
121  /// HasMultipleAccessingFunctions is set to true.
122  const Function *AccessingFunction;
123  bool HasMultipleAccessingFunctions;
124
125  /// HasNonInstructionUser - Set to true if this global has a user that is not
126  /// an instruction (e.g. a constant expr or GV initializer).
127  bool HasNonInstructionUser;
128
129  /// HasPHIUser - Set to true if this global has a user that is a PHI node.
130  bool HasPHIUser;
131
132  GlobalStatus() : isLoaded(false), StoredType(NotStored), StoredOnceValue(0),
133                   AccessingFunction(0), HasMultipleAccessingFunctions(false),
134                   HasNonInstructionUser(false), HasPHIUser(false) {}
135};
136
137}
138
139// SafeToDestroyConstant - It is safe to destroy a constant iff it is only used
140// by constants itself.  Note that constants cannot be cyclic, so this test is
141// pretty easy to implement recursively.
142//
143static bool SafeToDestroyConstant(const Constant *C) {
144  if (isa<GlobalValue>(C)) return false;
145
146  for (Value::const_use_iterator UI = C->use_begin(), E = C->use_end(); UI != E; ++UI)
147    if (const Constant *CU = dyn_cast<Constant>(*UI)) {
148      if (!SafeToDestroyConstant(CU)) return false;
149    } else
150      return false;
151  return true;
152}
153
154
155/// AnalyzeGlobal - Look at all uses of the global and fill in the GlobalStatus
156/// structure.  If the global has its address taken, return true to indicate we
157/// can't do anything with it.
158///
159static bool AnalyzeGlobal(const Value *V, GlobalStatus &GS,
160                          SmallPtrSet<const PHINode*, 16> &PHIUsers) {
161  for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
162    if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(*UI)) {
163      GS.HasNonInstructionUser = true;
164
165      if (AnalyzeGlobal(CE, GS, PHIUsers)) return true;
166
167    } else if (const Instruction *I = dyn_cast<Instruction>(*UI)) {
168      if (!GS.HasMultipleAccessingFunctions) {
169        const Function *F = I->getParent()->getParent();
170        if (GS.AccessingFunction == 0)
171          GS.AccessingFunction = F;
172        else if (GS.AccessingFunction != F)
173          GS.HasMultipleAccessingFunctions = true;
174      }
175      if (const LoadInst *LI = dyn_cast<LoadInst>(I)) {
176        GS.isLoaded = true;
177        if (LI->isVolatile()) return true;  // Don't hack on volatile loads.
178      } else if (const StoreInst *SI = dyn_cast<StoreInst>(I)) {
179        // Don't allow a store OF the address, only stores TO the address.
180        if (SI->getOperand(0) == V) return true;
181
182        if (SI->isVolatile()) return true;  // Don't hack on volatile stores.
183
184        // If this is a direct store to the global (i.e., the global is a scalar
185        // value, not an aggregate), keep more specific information about
186        // stores.
187        if (GS.StoredType != GlobalStatus::isStored) {
188          if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(SI->getOperand(1))){
189            Value *StoredVal = SI->getOperand(0);
190            if (StoredVal == GV->getInitializer()) {
191              if (GS.StoredType < GlobalStatus::isInitializerStored)
192                GS.StoredType = GlobalStatus::isInitializerStored;
193            } else if (isa<LoadInst>(StoredVal) &&
194                       cast<LoadInst>(StoredVal)->getOperand(0) == GV) {
195              if (GS.StoredType < GlobalStatus::isInitializerStored)
196                GS.StoredType = GlobalStatus::isInitializerStored;
197            } else if (GS.StoredType < GlobalStatus::isStoredOnce) {
198              GS.StoredType = GlobalStatus::isStoredOnce;
199              GS.StoredOnceValue = StoredVal;
200            } else if (GS.StoredType == GlobalStatus::isStoredOnce &&
201                       GS.StoredOnceValue == StoredVal) {
202              // noop.
203            } else {
204              GS.StoredType = GlobalStatus::isStored;
205            }
206          } else {
207            GS.StoredType = GlobalStatus::isStored;
208          }
209        }
210      } else if (isa<GetElementPtrInst>(I)) {
211        if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
212      } else if (isa<SelectInst>(I)) {
213        if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
214      } else if (const PHINode *PN = dyn_cast<PHINode>(I)) {
215        // PHI nodes we can check just like select or GEP instructions, but we
216        // have to be careful about infinite recursion.
217        if (PHIUsers.insert(PN))  // Not already visited.
218          if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
219        GS.HasPHIUser = true;
220      } else if (isa<CmpInst>(I)) {
221      } else if (isa<MemTransferInst>(I)) {
222        if (I->getOperand(1) == V)
223          GS.StoredType = GlobalStatus::isStored;
224        if (I->getOperand(2) == V)
225          GS.isLoaded = true;
226      } else if (isa<MemSetInst>(I)) {
227        assert(I->getOperand(1) == V && "Memset only takes one pointer!");
228        GS.StoredType = GlobalStatus::isStored;
229      } else {
230        return true;  // Any other non-load instruction might take address!
231      }
232    } else if (const Constant *C = dyn_cast<Constant>(*UI)) {
233      GS.HasNonInstructionUser = true;
234      // We might have a dead and dangling constant hanging off of here.
235      if (!SafeToDestroyConstant(C))
236        return true;
237    } else {
238      GS.HasNonInstructionUser = true;
239      // Otherwise must be some other user.
240      return true;
241    }
242
243  return false;
244}
245
246static Constant *getAggregateConstantElement(Constant *Agg, Constant *Idx) {
247  ConstantInt *CI = dyn_cast<ConstantInt>(Idx);
248  if (!CI) return 0;
249  unsigned IdxV = CI->getZExtValue();
250
251  if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg)) {
252    if (IdxV < CS->getNumOperands()) return CS->getOperand(IdxV);
253  } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg)) {
254    if (IdxV < CA->getNumOperands()) return CA->getOperand(IdxV);
255  } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Agg)) {
256    if (IdxV < CP->getNumOperands()) return CP->getOperand(IdxV);
257  } else if (isa<ConstantAggregateZero>(Agg)) {
258    if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
259      if (IdxV < STy->getNumElements())
260        return Constant::getNullValue(STy->getElementType(IdxV));
261    } else if (const SequentialType *STy =
262               dyn_cast<SequentialType>(Agg->getType())) {
263      return Constant::getNullValue(STy->getElementType());
264    }
265  } else if (isa<UndefValue>(Agg)) {
266    if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
267      if (IdxV < STy->getNumElements())
268        return UndefValue::get(STy->getElementType(IdxV));
269    } else if (const SequentialType *STy =
270               dyn_cast<SequentialType>(Agg->getType())) {
271      return UndefValue::get(STy->getElementType());
272    }
273  }
274  return 0;
275}
276
277
278/// CleanupConstantGlobalUsers - We just marked GV constant.  Loop over all
279/// users of the global, cleaning up the obvious ones.  This is largely just a
280/// quick scan over the use list to clean up the easy and obvious cruft.  This
281/// returns true if it made a change.
282static bool CleanupConstantGlobalUsers(Value *V, Constant *Init) {
283  bool Changed = false;
284  for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;) {
285    User *U = *UI++;
286
287    if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
288      if (Init) {
289        // Replace the load with the initializer.
290        LI->replaceAllUsesWith(Init);
291        LI->eraseFromParent();
292        Changed = true;
293      }
294    } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
295      // Store must be unreachable or storing Init into the global.
296      SI->eraseFromParent();
297      Changed = true;
298    } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
299      if (CE->getOpcode() == Instruction::GetElementPtr) {
300        Constant *SubInit = 0;
301        if (Init)
302          SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
303        Changed |= CleanupConstantGlobalUsers(CE, SubInit);
304      } else if (CE->getOpcode() == Instruction::BitCast &&
305                 CE->getType()->isPointerTy()) {
306        // Pointer cast, delete any stores and memsets to the global.
307        Changed |= CleanupConstantGlobalUsers(CE, 0);
308      }
309
310      if (CE->use_empty()) {
311        CE->destroyConstant();
312        Changed = true;
313      }
314    } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
315      // Do not transform "gepinst (gep constexpr (GV))" here, because forming
316      // "gepconstexpr (gep constexpr (GV))" will cause the two gep's to fold
317      // and will invalidate our notion of what Init is.
318      Constant *SubInit = 0;
319      if (!isa<ConstantExpr>(GEP->getOperand(0))) {
320        ConstantExpr *CE =
321          dyn_cast_or_null<ConstantExpr>(ConstantFoldInstruction(GEP));
322        if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr)
323          SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
324      }
325      Changed |= CleanupConstantGlobalUsers(GEP, SubInit);
326
327      if (GEP->use_empty()) {
328        GEP->eraseFromParent();
329        Changed = true;
330      }
331    } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv
332      if (MI->getRawDest() == V) {
333        MI->eraseFromParent();
334        Changed = true;
335      }
336
337    } else if (Constant *C = dyn_cast<Constant>(U)) {
338      // If we have a chain of dead constantexprs or other things dangling from
339      // us, and if they are all dead, nuke them without remorse.
340      if (SafeToDestroyConstant(C)) {
341        C->destroyConstant();
342        // This could have invalidated UI, start over from scratch.
343        CleanupConstantGlobalUsers(V, Init);
344        return true;
345      }
346    }
347  }
348  return Changed;
349}
350
351/// isSafeSROAElementUse - Return true if the specified instruction is a safe
352/// user of a derived expression from a global that we want to SROA.
353static bool isSafeSROAElementUse(Value *V) {
354  // We might have a dead and dangling constant hanging off of here.
355  if (Constant *C = dyn_cast<Constant>(V))
356    return SafeToDestroyConstant(C);
357
358  Instruction *I = dyn_cast<Instruction>(V);
359  if (!I) return false;
360
361  // Loads are ok.
362  if (isa<LoadInst>(I)) return true;
363
364  // Stores *to* the pointer are ok.
365  if (StoreInst *SI = dyn_cast<StoreInst>(I))
366    return SI->getOperand(0) != V;
367
368  // Otherwise, it must be a GEP.
369  GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I);
370  if (GEPI == 0) return false;
371
372  if (GEPI->getNumOperands() < 3 || !isa<Constant>(GEPI->getOperand(1)) ||
373      !cast<Constant>(GEPI->getOperand(1))->isNullValue())
374    return false;
375
376  for (Value::use_iterator I = GEPI->use_begin(), E = GEPI->use_end();
377       I != E; ++I)
378    if (!isSafeSROAElementUse(*I))
379      return false;
380  return true;
381}
382
383
384/// IsUserOfGlobalSafeForSRA - U is a direct user of the specified global value.
385/// Look at it and its uses and decide whether it is safe to SROA this global.
386///
387static bool IsUserOfGlobalSafeForSRA(User *U, GlobalValue *GV) {
388  // The user of the global must be a GEP Inst or a ConstantExpr GEP.
389  if (!isa<GetElementPtrInst>(U) &&
390      (!isa<ConstantExpr>(U) ||
391       cast<ConstantExpr>(U)->getOpcode() != Instruction::GetElementPtr))
392    return false;
393
394  // Check to see if this ConstantExpr GEP is SRA'able.  In particular, we
395  // don't like < 3 operand CE's, and we don't like non-constant integer
396  // indices.  This enforces that all uses are 'gep GV, 0, C, ...' for some
397  // value of C.
398  if (U->getNumOperands() < 3 || !isa<Constant>(U->getOperand(1)) ||
399      !cast<Constant>(U->getOperand(1))->isNullValue() ||
400      !isa<ConstantInt>(U->getOperand(2)))
401    return false;
402
403  gep_type_iterator GEPI = gep_type_begin(U), E = gep_type_end(U);
404  ++GEPI;  // Skip over the pointer index.
405
406  // If this is a use of an array allocation, do a bit more checking for sanity.
407  if (const ArrayType *AT = dyn_cast<ArrayType>(*GEPI)) {
408    uint64_t NumElements = AT->getNumElements();
409    ConstantInt *Idx = cast<ConstantInt>(U->getOperand(2));
410
411    // Check to make sure that index falls within the array.  If not,
412    // something funny is going on, so we won't do the optimization.
413    //
414    if (Idx->getZExtValue() >= NumElements)
415      return false;
416
417    // We cannot scalar repl this level of the array unless any array
418    // sub-indices are in-range constants.  In particular, consider:
419    // A[0][i].  We cannot know that the user isn't doing invalid things like
420    // allowing i to index an out-of-range subscript that accesses A[1].
421    //
422    // Scalar replacing *just* the outer index of the array is probably not
423    // going to be a win anyway, so just give up.
424    for (++GEPI; // Skip array index.
425         GEPI != E;
426         ++GEPI) {
427      uint64_t NumElements;
428      if (const ArrayType *SubArrayTy = dyn_cast<ArrayType>(*GEPI))
429        NumElements = SubArrayTy->getNumElements();
430      else if (const VectorType *SubVectorTy = dyn_cast<VectorType>(*GEPI))
431        NumElements = SubVectorTy->getNumElements();
432      else {
433        assert((*GEPI)->isStructTy() &&
434               "Indexed GEP type is not array, vector, or struct!");
435        continue;
436      }
437
438      ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPI.getOperand());
439      if (!IdxVal || IdxVal->getZExtValue() >= NumElements)
440        return false;
441    }
442  }
443
444  for (Value::use_iterator I = U->use_begin(), E = U->use_end(); I != E; ++I)
445    if (!isSafeSROAElementUse(*I))
446      return false;
447  return true;
448}
449
450/// GlobalUsersSafeToSRA - Look at all uses of the global and decide whether it
451/// is safe for us to perform this transformation.
452///
453static bool GlobalUsersSafeToSRA(GlobalValue *GV) {
454  for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end();
455       UI != E; ++UI) {
456    if (!IsUserOfGlobalSafeForSRA(*UI, GV))
457      return false;
458  }
459  return true;
460}
461
462
463/// SRAGlobal - Perform scalar replacement of aggregates on the specified global
464/// variable.  This opens the door for other optimizations by exposing the
465/// behavior of the program in a more fine-grained way.  We have determined that
466/// this transformation is safe already.  We return the first global variable we
467/// insert so that the caller can reprocess it.
468static GlobalVariable *SRAGlobal(GlobalVariable *GV, const TargetData &TD) {
469  // Make sure this global only has simple uses that we can SRA.
470  if (!GlobalUsersSafeToSRA(GV))
471    return 0;
472
473  assert(GV->hasLocalLinkage() && !GV->isConstant());
474  Constant *Init = GV->getInitializer();
475  const Type *Ty = Init->getType();
476
477  std::vector<GlobalVariable*> NewGlobals;
478  Module::GlobalListType &Globals = GV->getParent()->getGlobalList();
479
480  // Get the alignment of the global, either explicit or target-specific.
481  unsigned StartAlignment = GV->getAlignment();
482  if (StartAlignment == 0)
483    StartAlignment = TD.getABITypeAlignment(GV->getType());
484
485  if (const StructType *STy = dyn_cast<StructType>(Ty)) {
486    NewGlobals.reserve(STy->getNumElements());
487    const StructLayout &Layout = *TD.getStructLayout(STy);
488    for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
489      Constant *In = getAggregateConstantElement(Init,
490                    ConstantInt::get(Type::getInt32Ty(STy->getContext()), i));
491      assert(In && "Couldn't get element of initializer?");
492      GlobalVariable *NGV = new GlobalVariable(STy->getElementType(i), false,
493                                               GlobalVariable::InternalLinkage,
494                                               In, GV->getName()+"."+Twine(i),
495                                               GV->isThreadLocal(),
496                                              GV->getType()->getAddressSpace());
497      Globals.insert(GV, NGV);
498      NewGlobals.push_back(NGV);
499
500      // Calculate the known alignment of the field.  If the original aggregate
501      // had 256 byte alignment for example, something might depend on that:
502      // propagate info to each field.
503      uint64_t FieldOffset = Layout.getElementOffset(i);
504      unsigned NewAlign = (unsigned)MinAlign(StartAlignment, FieldOffset);
505      if (NewAlign > TD.getABITypeAlignment(STy->getElementType(i)))
506        NGV->setAlignment(NewAlign);
507    }
508  } else if (const SequentialType *STy = dyn_cast<SequentialType>(Ty)) {
509    unsigned NumElements = 0;
510    if (const ArrayType *ATy = dyn_cast<ArrayType>(STy))
511      NumElements = ATy->getNumElements();
512    else
513      NumElements = cast<VectorType>(STy)->getNumElements();
514
515    if (NumElements > 16 && GV->hasNUsesOrMore(16))
516      return 0; // It's not worth it.
517    NewGlobals.reserve(NumElements);
518
519    uint64_t EltSize = TD.getTypeAllocSize(STy->getElementType());
520    unsigned EltAlign = TD.getABITypeAlignment(STy->getElementType());
521    for (unsigned i = 0, e = NumElements; i != e; ++i) {
522      Constant *In = getAggregateConstantElement(Init,
523                    ConstantInt::get(Type::getInt32Ty(Init->getContext()), i));
524      assert(In && "Couldn't get element of initializer?");
525
526      GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false,
527                                               GlobalVariable::InternalLinkage,
528                                               In, GV->getName()+"."+Twine(i),
529                                               GV->isThreadLocal(),
530                                              GV->getType()->getAddressSpace());
531      Globals.insert(GV, NGV);
532      NewGlobals.push_back(NGV);
533
534      // Calculate the known alignment of the field.  If the original aggregate
535      // had 256 byte alignment for example, something might depend on that:
536      // propagate info to each field.
537      unsigned NewAlign = (unsigned)MinAlign(StartAlignment, EltSize*i);
538      if (NewAlign > EltAlign)
539        NGV->setAlignment(NewAlign);
540    }
541  }
542
543  if (NewGlobals.empty())
544    return 0;
545
546  DEBUG(dbgs() << "PERFORMING GLOBAL SRA ON: " << *GV);
547
548  Constant *NullInt =Constant::getNullValue(Type::getInt32Ty(GV->getContext()));
549
550  // Loop over all of the uses of the global, replacing the constantexpr geps,
551  // with smaller constantexpr geps or direct references.
552  while (!GV->use_empty()) {
553    User *GEP = GV->use_back();
554    assert(((isa<ConstantExpr>(GEP) &&
555             cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)||
556            isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!");
557
558    // Ignore the 1th operand, which has to be zero or else the program is quite
559    // broken (undefined).  Get the 2nd operand, which is the structure or array
560    // index.
561    unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
562    if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access.
563
564    Value *NewPtr = NewGlobals[Val];
565
566    // Form a shorter GEP if needed.
567    if (GEP->getNumOperands() > 3) {
568      if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) {
569        SmallVector<Constant*, 8> Idxs;
570        Idxs.push_back(NullInt);
571        for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i)
572          Idxs.push_back(CE->getOperand(i));
573        NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr),
574                                                &Idxs[0], Idxs.size());
575      } else {
576        GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP);
577        SmallVector<Value*, 8> Idxs;
578        Idxs.push_back(NullInt);
579        for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i)
580          Idxs.push_back(GEPI->getOperand(i));
581        NewPtr = GetElementPtrInst::Create(NewPtr, Idxs.begin(), Idxs.end(),
582                                           GEPI->getName()+"."+Twine(Val),GEPI);
583      }
584    }
585    GEP->replaceAllUsesWith(NewPtr);
586
587    if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP))
588      GEPI->eraseFromParent();
589    else
590      cast<ConstantExpr>(GEP)->destroyConstant();
591  }
592
593  // Delete the old global, now that it is dead.
594  Globals.erase(GV);
595  ++NumSRA;
596
597  // Loop over the new globals array deleting any globals that are obviously
598  // dead.  This can arise due to scalarization of a structure or an array that
599  // has elements that are dead.
600  unsigned FirstGlobal = 0;
601  for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i)
602    if (NewGlobals[i]->use_empty()) {
603      Globals.erase(NewGlobals[i]);
604      if (FirstGlobal == i) ++FirstGlobal;
605    }
606
607  return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : 0;
608}
609
610/// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified
611/// value will trap if the value is dynamically null.  PHIs keeps track of any
612/// phi nodes we've seen to avoid reprocessing them.
613static bool AllUsesOfValueWillTrapIfNull(Value *V,
614                                         SmallPtrSet<PHINode*, 8> &PHIs) {
615  for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
616    if (isa<LoadInst>(*UI)) {
617      // Will trap.
618    } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
619      if (SI->getOperand(0) == V) {
620        //cerr << "NONTRAPPING USE: " << **UI;
621        return false;  // Storing the value.
622      }
623    } else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
624      if (CI->getCalledValue() != V) {
625        //cerr << "NONTRAPPING USE: " << **UI;
626        return false;  // Not calling the ptr
627      }
628    } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
629      if (II->getCalledValue() != V) {
630        //cerr << "NONTRAPPING USE: " << **UI;
631        return false;  // Not calling the ptr
632      }
633    } else if (BitCastInst *CI = dyn_cast<BitCastInst>(*UI)) {
634      if (!AllUsesOfValueWillTrapIfNull(CI, PHIs)) return false;
635    } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) {
636      if (!AllUsesOfValueWillTrapIfNull(GEPI, PHIs)) return false;
637    } else if (PHINode *PN = dyn_cast<PHINode>(*UI)) {
638      // If we've already seen this phi node, ignore it, it has already been
639      // checked.
640      if (PHIs.insert(PN) && !AllUsesOfValueWillTrapIfNull(PN, PHIs))
641        return false;
642    } else if (isa<ICmpInst>(*UI) &&
643               isa<ConstantPointerNull>(UI->getOperand(1))) {
644      // Ignore icmp X, null
645    } else {
646      //cerr << "NONTRAPPING USE: " << **UI;
647      return false;
648    }
649  return true;
650}
651
652/// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads
653/// from GV will trap if the loaded value is null.  Note that this also permits
654/// comparisons of the loaded value against null, as a special case.
655static bool AllUsesOfLoadedValueWillTrapIfNull(GlobalVariable *GV) {
656  for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI!=E; ++UI)
657    if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
658      SmallPtrSet<PHINode*, 8> PHIs;
659      if (!AllUsesOfValueWillTrapIfNull(LI, PHIs))
660        return false;
661    } else if (isa<StoreInst>(*UI)) {
662      // Ignore stores to the global.
663    } else {
664      // We don't know or understand this user, bail out.
665      //cerr << "UNKNOWN USER OF GLOBAL!: " << **UI;
666      return false;
667    }
668
669  return true;
670}
671
672static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) {
673  bool Changed = false;
674  for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) {
675    Instruction *I = cast<Instruction>(*UI++);
676    if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
677      LI->setOperand(0, NewV);
678      Changed = true;
679    } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
680      if (SI->getOperand(1) == V) {
681        SI->setOperand(1, NewV);
682        Changed = true;
683      }
684    } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
685      if (I->getOperand(0) == V) {
686        // Calling through the pointer!  Turn into a direct call, but be careful
687        // that the pointer is not also being passed as an argument.
688        I->setOperand(0, NewV);
689        Changed = true;
690        bool PassedAsArg = false;
691        for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i)
692          if (I->getOperand(i) == V) {
693            PassedAsArg = true;
694            I->setOperand(i, NewV);
695          }
696
697        if (PassedAsArg) {
698          // Being passed as an argument also.  Be careful to not invalidate UI!
699          UI = V->use_begin();
700        }
701      }
702    } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
703      Changed |= OptimizeAwayTrappingUsesOfValue(CI,
704                                ConstantExpr::getCast(CI->getOpcode(),
705                                                      NewV, CI->getType()));
706      if (CI->use_empty()) {
707        Changed = true;
708        CI->eraseFromParent();
709      }
710    } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
711      // Should handle GEP here.
712      SmallVector<Constant*, 8> Idxs;
713      Idxs.reserve(GEPI->getNumOperands()-1);
714      for (User::op_iterator i = GEPI->op_begin() + 1, e = GEPI->op_end();
715           i != e; ++i)
716        if (Constant *C = dyn_cast<Constant>(*i))
717          Idxs.push_back(C);
718        else
719          break;
720      if (Idxs.size() == GEPI->getNumOperands()-1)
721        Changed |= OptimizeAwayTrappingUsesOfValue(GEPI,
722                          ConstantExpr::getGetElementPtr(NewV, &Idxs[0],
723                                                        Idxs.size()));
724      if (GEPI->use_empty()) {
725        Changed = true;
726        GEPI->eraseFromParent();
727      }
728    }
729  }
730
731  return Changed;
732}
733
734
735/// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null
736/// value stored into it.  If there are uses of the loaded value that would trap
737/// if the loaded value is dynamically null, then we know that they cannot be
738/// reachable with a null optimize away the load.
739static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV) {
740  bool Changed = false;
741
742  // Keep track of whether we are able to remove all the uses of the global
743  // other than the store that defines it.
744  bool AllNonStoreUsesGone = true;
745
746  // Replace all uses of loads with uses of uses of the stored value.
747  for (Value::use_iterator GUI = GV->use_begin(), E = GV->use_end(); GUI != E;){
748    User *GlobalUser = *GUI++;
749    if (LoadInst *LI = dyn_cast<LoadInst>(GlobalUser)) {
750      Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV);
751      // If we were able to delete all uses of the loads
752      if (LI->use_empty()) {
753        LI->eraseFromParent();
754        Changed = true;
755      } else {
756        AllNonStoreUsesGone = false;
757      }
758    } else if (isa<StoreInst>(GlobalUser)) {
759      // Ignore the store that stores "LV" to the global.
760      assert(GlobalUser->getOperand(1) == GV &&
761             "Must be storing *to* the global");
762    } else {
763      AllNonStoreUsesGone = false;
764
765      // If we get here we could have other crazy uses that are transitively
766      // loaded.
767      assert((isa<PHINode>(GlobalUser) || isa<SelectInst>(GlobalUser) ||
768              isa<ConstantExpr>(GlobalUser)) && "Only expect load and stores!");
769    }
770  }
771
772  if (Changed) {
773    DEBUG(dbgs() << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV);
774    ++NumGlobUses;
775  }
776
777  // If we nuked all of the loads, then none of the stores are needed either,
778  // nor is the global.
779  if (AllNonStoreUsesGone) {
780    DEBUG(dbgs() << "  *** GLOBAL NOW DEAD!\n");
781    CleanupConstantGlobalUsers(GV, 0);
782    if (GV->use_empty()) {
783      GV->eraseFromParent();
784      ++NumDeleted;
785    }
786    Changed = true;
787  }
788  return Changed;
789}
790
791/// ConstantPropUsersOf - Walk the use list of V, constant folding all of the
792/// instructions that are foldable.
793static void ConstantPropUsersOf(Value *V) {
794  for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; )
795    if (Instruction *I = dyn_cast<Instruction>(*UI++))
796      if (Constant *NewC = ConstantFoldInstruction(I)) {
797        I->replaceAllUsesWith(NewC);
798
799        // Advance UI to the next non-I use to avoid invalidating it!
800        // Instructions could multiply use V.
801        while (UI != E && *UI == I)
802          ++UI;
803        I->eraseFromParent();
804      }
805}
806
807/// OptimizeGlobalAddressOfMalloc - This function takes the specified global
808/// variable, and transforms the program as if it always contained the result of
809/// the specified malloc.  Because it is always the result of the specified
810/// malloc, there is no reason to actually DO the malloc.  Instead, turn the
811/// malloc into a global, and any loads of GV as uses of the new global.
812static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV,
813                                                     CallInst *CI,
814                                                     const Type *AllocTy,
815                                                     ConstantInt *NElements,
816                                                     TargetData* TD) {
817  DEBUG(errs() << "PROMOTING GLOBAL: " << *GV << "  CALL = " << *CI << '\n');
818
819  const Type *GlobalType;
820  if (NElements->getZExtValue() == 1)
821    GlobalType = AllocTy;
822  else
823    // If we have an array allocation, the global variable is of an array.
824    GlobalType = ArrayType::get(AllocTy, NElements->getZExtValue());
825
826  // Create the new global variable.  The contents of the malloc'd memory is
827  // undefined, so initialize with an undef value.
828  GlobalVariable *NewGV = new GlobalVariable(*GV->getParent(),
829                                             GlobalType, false,
830                                             GlobalValue::InternalLinkage,
831                                             UndefValue::get(GlobalType),
832                                             GV->getName()+".body",
833                                             GV,
834                                             GV->isThreadLocal());
835
836  // If there are bitcast users of the malloc (which is typical, usually we have
837  // a malloc + bitcast) then replace them with uses of the new global.  Update
838  // other users to use the global as well.
839  BitCastInst *TheBC = 0;
840  while (!CI->use_empty()) {
841    Instruction *User = cast<Instruction>(CI->use_back());
842    if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
843      if (BCI->getType() == NewGV->getType()) {
844        BCI->replaceAllUsesWith(NewGV);
845        BCI->eraseFromParent();
846      } else {
847        BCI->setOperand(0, NewGV);
848      }
849    } else {
850      if (TheBC == 0)
851        TheBC = new BitCastInst(NewGV, CI->getType(), "newgv", CI);
852      User->replaceUsesOfWith(CI, TheBC);
853    }
854  }
855
856  Constant *RepValue = NewGV;
857  if (NewGV->getType() != GV->getType()->getElementType())
858    RepValue = ConstantExpr::getBitCast(RepValue,
859                                        GV->getType()->getElementType());
860
861  // If there is a comparison against null, we will insert a global bool to
862  // keep track of whether the global was initialized yet or not.
863  GlobalVariable *InitBool =
864    new GlobalVariable(Type::getInt1Ty(GV->getContext()), false,
865                       GlobalValue::InternalLinkage,
866                       ConstantInt::getFalse(GV->getContext()),
867                       GV->getName()+".init", GV->isThreadLocal());
868  bool InitBoolUsed = false;
869
870  // Loop over all uses of GV, processing them in turn.
871  while (!GV->use_empty()) {
872    if (StoreInst *SI = dyn_cast<StoreInst>(GV->use_back())) {
873      // The global is initialized when the store to it occurs.
874      new StoreInst(ConstantInt::getTrue(GV->getContext()), InitBool, SI);
875      SI->eraseFromParent();
876      continue;
877    }
878
879    LoadInst *LI = cast<LoadInst>(GV->use_back());
880    while (!LI->use_empty()) {
881      Use &LoadUse = LI->use_begin().getUse();
882      if (!isa<ICmpInst>(LoadUse.getUser())) {
883        LoadUse = RepValue;
884        continue;
885      }
886
887      ICmpInst *ICI = cast<ICmpInst>(LoadUse.getUser());
888      // Replace the cmp X, 0 with a use of the bool value.
889      Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", ICI);
890      InitBoolUsed = true;
891      switch (ICI->getPredicate()) {
892      default: llvm_unreachable("Unknown ICmp Predicate!");
893      case ICmpInst::ICMP_ULT:
894      case ICmpInst::ICMP_SLT:   // X < null -> always false
895        LV = ConstantInt::getFalse(GV->getContext());
896        break;
897      case ICmpInst::ICMP_ULE:
898      case ICmpInst::ICMP_SLE:
899      case ICmpInst::ICMP_EQ:
900        LV = BinaryOperator::CreateNot(LV, "notinit", ICI);
901        break;
902      case ICmpInst::ICMP_NE:
903      case ICmpInst::ICMP_UGE:
904      case ICmpInst::ICMP_SGE:
905      case ICmpInst::ICMP_UGT:
906      case ICmpInst::ICMP_SGT:
907        break;  // no change.
908      }
909      ICI->replaceAllUsesWith(LV);
910      ICI->eraseFromParent();
911    }
912    LI->eraseFromParent();
913  }
914
915  // If the initialization boolean was used, insert it, otherwise delete it.
916  if (!InitBoolUsed) {
917    while (!InitBool->use_empty())  // Delete initializations
918      cast<StoreInst>(InitBool->use_back())->eraseFromParent();
919    delete InitBool;
920  } else
921    GV->getParent()->getGlobalList().insert(GV, InitBool);
922
923  // Now the GV is dead, nuke it and the malloc..
924  GV->eraseFromParent();
925  CI->eraseFromParent();
926
927  // To further other optimizations, loop over all users of NewGV and try to
928  // constant prop them.  This will promote GEP instructions with constant
929  // indices into GEP constant-exprs, which will allow global-opt to hack on it.
930  ConstantPropUsersOf(NewGV);
931  if (RepValue != NewGV)
932    ConstantPropUsersOf(RepValue);
933
934  return NewGV;
935}
936
937/// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking
938/// to make sure that there are no complex uses of V.  We permit simple things
939/// like dereferencing the pointer, but not storing through the address, unless
940/// it is to the specified global.
941static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Instruction *V,
942                                                      GlobalVariable *GV,
943                                              SmallPtrSet<PHINode*, 8> &PHIs) {
944  for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
945    Instruction *Inst = cast<Instruction>(*UI);
946
947    if (isa<LoadInst>(Inst) || isa<CmpInst>(Inst)) {
948      continue; // Fine, ignore.
949    }
950
951    if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
952      if (SI->getOperand(0) == V && SI->getOperand(1) != GV)
953        return false;  // Storing the pointer itself... bad.
954      continue; // Otherwise, storing through it, or storing into GV... fine.
955    }
956
957    if (isa<GetElementPtrInst>(Inst)) {
958      if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Inst, GV, PHIs))
959        return false;
960      continue;
961    }
962
963    if (PHINode *PN = dyn_cast<PHINode>(Inst)) {
964      // PHIs are ok if all uses are ok.  Don't infinitely recurse through PHI
965      // cycles.
966      if (PHIs.insert(PN))
967        if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(PN, GV, PHIs))
968          return false;
969      continue;
970    }
971
972    if (BitCastInst *BCI = dyn_cast<BitCastInst>(Inst)) {
973      if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(BCI, GV, PHIs))
974        return false;
975      continue;
976    }
977
978    return false;
979  }
980  return true;
981}
982
983/// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV
984/// somewhere.  Transform all uses of the allocation into loads from the
985/// global and uses of the resultant pointer.  Further, delete the store into
986/// GV.  This assumes that these value pass the
987/// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate.
988static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc,
989                                          GlobalVariable *GV) {
990  while (!Alloc->use_empty()) {
991    Instruction *U = cast<Instruction>(*Alloc->use_begin());
992    Instruction *InsertPt = U;
993    if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
994      // If this is the store of the allocation into the global, remove it.
995      if (SI->getOperand(1) == GV) {
996        SI->eraseFromParent();
997        continue;
998      }
999    } else if (PHINode *PN = dyn_cast<PHINode>(U)) {
1000      // Insert the load in the corresponding predecessor, not right before the
1001      // PHI.
1002      InsertPt = PN->getIncomingBlock(Alloc->use_begin())->getTerminator();
1003    } else if (isa<BitCastInst>(U)) {
1004      // Must be bitcast between the malloc and store to initialize the global.
1005      ReplaceUsesOfMallocWithGlobal(U, GV);
1006      U->eraseFromParent();
1007      continue;
1008    } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
1009      // If this is a "GEP bitcast" and the user is a store to the global, then
1010      // just process it as a bitcast.
1011      if (GEPI->hasAllZeroIndices() && GEPI->hasOneUse())
1012        if (StoreInst *SI = dyn_cast<StoreInst>(GEPI->use_back()))
1013          if (SI->getOperand(1) == GV) {
1014            // Must be bitcast GEP between the malloc and store to initialize
1015            // the global.
1016            ReplaceUsesOfMallocWithGlobal(GEPI, GV);
1017            GEPI->eraseFromParent();
1018            continue;
1019          }
1020    }
1021
1022    // Insert a load from the global, and use it instead of the malloc.
1023    Value *NL = new LoadInst(GV, GV->getName()+".val", InsertPt);
1024    U->replaceUsesOfWith(Alloc, NL);
1025  }
1026}
1027
1028/// LoadUsesSimpleEnoughForHeapSRA - Verify that all uses of V (a load, or a phi
1029/// of a load) are simple enough to perform heap SRA on.  This permits GEP's
1030/// that index through the array and struct field, icmps of null, and PHIs.
1031static bool LoadUsesSimpleEnoughForHeapSRA(const Value *V,
1032                              SmallPtrSet<const PHINode*, 32> &LoadUsingPHIs,
1033                              SmallPtrSet<const PHINode*, 32> &LoadUsingPHIsPerLoad) {
1034  // We permit two users of the load: setcc comparing against the null
1035  // pointer, and a getelementptr of a specific form.
1036  for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
1037    const Instruction *User = cast<Instruction>(*UI);
1038
1039    // Comparison against null is ok.
1040    if (const ICmpInst *ICI = dyn_cast<ICmpInst>(User)) {
1041      if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
1042        return false;
1043      continue;
1044    }
1045
1046    // getelementptr is also ok, but only a simple form.
1047    if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1048      // Must index into the array and into the struct.
1049      if (GEPI->getNumOperands() < 3)
1050        return false;
1051
1052      // Otherwise the GEP is ok.
1053      continue;
1054    }
1055
1056    if (const PHINode *PN = dyn_cast<PHINode>(User)) {
1057      if (!LoadUsingPHIsPerLoad.insert(PN))
1058        // This means some phi nodes are dependent on each other.
1059        // Avoid infinite looping!
1060        return false;
1061      if (!LoadUsingPHIs.insert(PN))
1062        // If we have already analyzed this PHI, then it is safe.
1063        continue;
1064
1065      // Make sure all uses of the PHI are simple enough to transform.
1066      if (!LoadUsesSimpleEnoughForHeapSRA(PN,
1067                                          LoadUsingPHIs, LoadUsingPHIsPerLoad))
1068        return false;
1069
1070      continue;
1071    }
1072
1073    // Otherwise we don't know what this is, not ok.
1074    return false;
1075  }
1076
1077  return true;
1078}
1079
1080
1081/// AllGlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from
1082/// GV are simple enough to perform HeapSRA, return true.
1083static bool AllGlobalLoadUsesSimpleEnoughForHeapSRA(const GlobalVariable *GV,
1084                                                    Instruction *StoredVal) {
1085  SmallPtrSet<const PHINode*, 32> LoadUsingPHIs;
1086  SmallPtrSet<const PHINode*, 32> LoadUsingPHIsPerLoad;
1087  for (Value::const_use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;
1088       ++UI)
1089    if (const LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
1090      if (!LoadUsesSimpleEnoughForHeapSRA(LI, LoadUsingPHIs,
1091                                          LoadUsingPHIsPerLoad))
1092        return false;
1093      LoadUsingPHIsPerLoad.clear();
1094    }
1095
1096  // If we reach here, we know that all uses of the loads and transitive uses
1097  // (through PHI nodes) are simple enough to transform.  However, we don't know
1098  // that all inputs the to the PHI nodes are in the same equivalence sets.
1099  // Check to verify that all operands of the PHIs are either PHIS that can be
1100  // transformed, loads from GV, or MI itself.
1101  for (SmallPtrSet<const PHINode*, 32>::const_iterator I = LoadUsingPHIs.begin(),
1102       E = LoadUsingPHIs.end(); I != E; ++I) {
1103    const PHINode *PN = *I;
1104    for (unsigned op = 0, e = PN->getNumIncomingValues(); op != e; ++op) {
1105      Value *InVal = PN->getIncomingValue(op);
1106
1107      // PHI of the stored value itself is ok.
1108      if (InVal == StoredVal) continue;
1109
1110      if (const PHINode *InPN = dyn_cast<PHINode>(InVal)) {
1111        // One of the PHIs in our set is (optimistically) ok.
1112        if (LoadUsingPHIs.count(InPN))
1113          continue;
1114        return false;
1115      }
1116
1117      // Load from GV is ok.
1118      if (const LoadInst *LI = dyn_cast<LoadInst>(InVal))
1119        if (LI->getOperand(0) == GV)
1120          continue;
1121
1122      // UNDEF? NULL?
1123
1124      // Anything else is rejected.
1125      return false;
1126    }
1127  }
1128
1129  return true;
1130}
1131
1132static Value *GetHeapSROAValue(Value *V, unsigned FieldNo,
1133               DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1134                   std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1135  std::vector<Value*> &FieldVals = InsertedScalarizedValues[V];
1136
1137  if (FieldNo >= FieldVals.size())
1138    FieldVals.resize(FieldNo+1);
1139
1140  // If we already have this value, just reuse the previously scalarized
1141  // version.
1142  if (Value *FieldVal = FieldVals[FieldNo])
1143    return FieldVal;
1144
1145  // Depending on what instruction this is, we have several cases.
1146  Value *Result;
1147  if (LoadInst *LI = dyn_cast<LoadInst>(V)) {
1148    // This is a scalarized version of the load from the global.  Just create
1149    // a new Load of the scalarized global.
1150    Result = new LoadInst(GetHeapSROAValue(LI->getOperand(0), FieldNo,
1151                                           InsertedScalarizedValues,
1152                                           PHIsToRewrite),
1153                          LI->getName()+".f"+Twine(FieldNo), LI);
1154  } else if (PHINode *PN = dyn_cast<PHINode>(V)) {
1155    // PN's type is pointer to struct.  Make a new PHI of pointer to struct
1156    // field.
1157    const StructType *ST =
1158      cast<StructType>(cast<PointerType>(PN->getType())->getElementType());
1159
1160    Result =
1161     PHINode::Create(PointerType::getUnqual(ST->getElementType(FieldNo)),
1162                     PN->getName()+".f"+Twine(FieldNo), PN);
1163    PHIsToRewrite.push_back(std::make_pair(PN, FieldNo));
1164  } else {
1165    llvm_unreachable("Unknown usable value");
1166    Result = 0;
1167  }
1168
1169  return FieldVals[FieldNo] = Result;
1170}
1171
1172/// RewriteHeapSROALoadUser - Given a load instruction and a value derived from
1173/// the load, rewrite the derived value to use the HeapSRoA'd load.
1174static void RewriteHeapSROALoadUser(Instruction *LoadUser,
1175             DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1176                   std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1177  // If this is a comparison against null, handle it.
1178  if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) {
1179    assert(isa<ConstantPointerNull>(SCI->getOperand(1)));
1180    // If we have a setcc of the loaded pointer, we can use a setcc of any
1181    // field.
1182    Value *NPtr = GetHeapSROAValue(SCI->getOperand(0), 0,
1183                                   InsertedScalarizedValues, PHIsToRewrite);
1184
1185    Value *New = new ICmpInst(SCI, SCI->getPredicate(), NPtr,
1186                              Constant::getNullValue(NPtr->getType()),
1187                              SCI->getName());
1188    SCI->replaceAllUsesWith(New);
1189    SCI->eraseFromParent();
1190    return;
1191  }
1192
1193  // Handle 'getelementptr Ptr, Idx, i32 FieldNo ...'
1194  if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) {
1195    assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2))
1196           && "Unexpected GEPI!");
1197
1198    // Load the pointer for this field.
1199    unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
1200    Value *NewPtr = GetHeapSROAValue(GEPI->getOperand(0), FieldNo,
1201                                     InsertedScalarizedValues, PHIsToRewrite);
1202
1203    // Create the new GEP idx vector.
1204    SmallVector<Value*, 8> GEPIdx;
1205    GEPIdx.push_back(GEPI->getOperand(1));
1206    GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end());
1207
1208    Value *NGEPI = GetElementPtrInst::Create(NewPtr,
1209                                             GEPIdx.begin(), GEPIdx.end(),
1210                                             GEPI->getName(), GEPI);
1211    GEPI->replaceAllUsesWith(NGEPI);
1212    GEPI->eraseFromParent();
1213    return;
1214  }
1215
1216  // Recursively transform the users of PHI nodes.  This will lazily create the
1217  // PHIs that are needed for individual elements.  Keep track of what PHIs we
1218  // see in InsertedScalarizedValues so that we don't get infinite loops (very
1219  // antisocial).  If the PHI is already in InsertedScalarizedValues, it has
1220  // already been seen first by another load, so its uses have already been
1221  // processed.
1222  PHINode *PN = cast<PHINode>(LoadUser);
1223  bool Inserted;
1224  DenseMap<Value*, std::vector<Value*> >::iterator InsertPos;
1225  tie(InsertPos, Inserted) =
1226    InsertedScalarizedValues.insert(std::make_pair(PN, std::vector<Value*>()));
1227  if (!Inserted) return;
1228
1229  // If this is the first time we've seen this PHI, recursively process all
1230  // users.
1231  for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end(); UI != E; ) {
1232    Instruction *User = cast<Instruction>(*UI++);
1233    RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite);
1234  }
1235}
1236
1237/// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global.  Ptr
1238/// is a value loaded from the global.  Eliminate all uses of Ptr, making them
1239/// use FieldGlobals instead.  All uses of loaded values satisfy
1240/// AllGlobalLoadUsesSimpleEnoughForHeapSRA.
1241static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load,
1242               DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1243                   std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1244  for (Value::use_iterator UI = Load->use_begin(), E = Load->use_end();
1245       UI != E; ) {
1246    Instruction *User = cast<Instruction>(*UI++);
1247    RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite);
1248  }
1249
1250  if (Load->use_empty()) {
1251    Load->eraseFromParent();
1252    InsertedScalarizedValues.erase(Load);
1253  }
1254}
1255
1256/// PerformHeapAllocSRoA - CI is an allocation of an array of structures.  Break
1257/// it up into multiple allocations of arrays of the fields.
1258static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, CallInst *CI,
1259                                            Value* NElems, TargetData *TD) {
1260  DEBUG(dbgs() << "SROA HEAP ALLOC: " << *GV << "  MALLOC = " << *CI << '\n');
1261  const Type* MAT = getMallocAllocatedType(CI);
1262  const StructType *STy = cast<StructType>(MAT);
1263
1264  // There is guaranteed to be at least one use of the malloc (storing
1265  // it into GV).  If there are other uses, change them to be uses of
1266  // the global to simplify later code.  This also deletes the store
1267  // into GV.
1268  ReplaceUsesOfMallocWithGlobal(CI, GV);
1269
1270  // Okay, at this point, there are no users of the malloc.  Insert N
1271  // new mallocs at the same place as CI, and N globals.
1272  std::vector<Value*> FieldGlobals;
1273  std::vector<Value*> FieldMallocs;
1274
1275  for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){
1276    const Type *FieldTy = STy->getElementType(FieldNo);
1277    const PointerType *PFieldTy = PointerType::getUnqual(FieldTy);
1278
1279    GlobalVariable *NGV =
1280      new GlobalVariable(*GV->getParent(),
1281                         PFieldTy, false, GlobalValue::InternalLinkage,
1282                         Constant::getNullValue(PFieldTy),
1283                         GV->getName() + ".f" + Twine(FieldNo), GV,
1284                         GV->isThreadLocal());
1285    FieldGlobals.push_back(NGV);
1286
1287    unsigned TypeSize = TD->getTypeAllocSize(FieldTy);
1288    if (const StructType *ST = dyn_cast<StructType>(FieldTy))
1289      TypeSize = TD->getStructLayout(ST)->getSizeInBytes();
1290    const Type *IntPtrTy = TD->getIntPtrType(CI->getContext());
1291    Value *NMI = CallInst::CreateMalloc(CI, IntPtrTy, FieldTy,
1292                                        ConstantInt::get(IntPtrTy, TypeSize),
1293                                        NElems,
1294                                        CI->getName() + ".f" + Twine(FieldNo));
1295    FieldMallocs.push_back(NMI);
1296    new StoreInst(NMI, NGV, CI);
1297  }
1298
1299  // The tricky aspect of this transformation is handling the case when malloc
1300  // fails.  In the original code, malloc failing would set the result pointer
1301  // of malloc to null.  In this case, some mallocs could succeed and others
1302  // could fail.  As such, we emit code that looks like this:
1303  //    F0 = malloc(field0)
1304  //    F1 = malloc(field1)
1305  //    F2 = malloc(field2)
1306  //    if (F0 == 0 || F1 == 0 || F2 == 0) {
1307  //      if (F0) { free(F0); F0 = 0; }
1308  //      if (F1) { free(F1); F1 = 0; }
1309  //      if (F2) { free(F2); F2 = 0; }
1310  //    }
1311  // The malloc can also fail if its argument is too large.
1312  Constant *ConstantZero = ConstantInt::get(CI->getOperand(1)->getType(), 0);
1313  Value *RunningOr = new ICmpInst(CI, ICmpInst::ICMP_SLT, CI->getOperand(1),
1314                                  ConstantZero, "isneg");
1315  for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) {
1316    Value *Cond = new ICmpInst(CI, ICmpInst::ICMP_EQ, FieldMallocs[i],
1317                             Constant::getNullValue(FieldMallocs[i]->getType()),
1318                               "isnull");
1319    RunningOr = BinaryOperator::CreateOr(RunningOr, Cond, "tmp", CI);
1320  }
1321
1322  // Split the basic block at the old malloc.
1323  BasicBlock *OrigBB = CI->getParent();
1324  BasicBlock *ContBB = OrigBB->splitBasicBlock(CI, "malloc_cont");
1325
1326  // Create the block to check the first condition.  Put all these blocks at the
1327  // end of the function as they are unlikely to be executed.
1328  BasicBlock *NullPtrBlock = BasicBlock::Create(OrigBB->getContext(),
1329                                                "malloc_ret_null",
1330                                                OrigBB->getParent());
1331
1332  // Remove the uncond branch from OrigBB to ContBB, turning it into a cond
1333  // branch on RunningOr.
1334  OrigBB->getTerminator()->eraseFromParent();
1335  BranchInst::Create(NullPtrBlock, ContBB, RunningOr, OrigBB);
1336
1337  // Within the NullPtrBlock, we need to emit a comparison and branch for each
1338  // pointer, because some may be null while others are not.
1339  for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1340    Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock);
1341    Value *Cmp = new ICmpInst(*NullPtrBlock, ICmpInst::ICMP_NE, GVVal,
1342                              Constant::getNullValue(GVVal->getType()),
1343                              "tmp");
1344    BasicBlock *FreeBlock = BasicBlock::Create(Cmp->getContext(), "free_it",
1345                                               OrigBB->getParent());
1346    BasicBlock *NextBlock = BasicBlock::Create(Cmp->getContext(), "next",
1347                                               OrigBB->getParent());
1348    Instruction *BI = BranchInst::Create(FreeBlock, NextBlock,
1349                                         Cmp, NullPtrBlock);
1350
1351    // Fill in FreeBlock.
1352    CallInst::CreateFree(GVVal, BI);
1353    new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i],
1354                  FreeBlock);
1355    BranchInst::Create(NextBlock, FreeBlock);
1356
1357    NullPtrBlock = NextBlock;
1358  }
1359
1360  BranchInst::Create(ContBB, NullPtrBlock);
1361
1362  // CI is no longer needed, remove it.
1363  CI->eraseFromParent();
1364
1365  /// InsertedScalarizedLoads - As we process loads, if we can't immediately
1366  /// update all uses of the load, keep track of what scalarized loads are
1367  /// inserted for a given load.
1368  DenseMap<Value*, std::vector<Value*> > InsertedScalarizedValues;
1369  InsertedScalarizedValues[GV] = FieldGlobals;
1370
1371  std::vector<std::pair<PHINode*, unsigned> > PHIsToRewrite;
1372
1373  // Okay, the malloc site is completely handled.  All of the uses of GV are now
1374  // loads, and all uses of those loads are simple.  Rewrite them to use loads
1375  // of the per-field globals instead.
1376  for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;) {
1377    Instruction *User = cast<Instruction>(*UI++);
1378
1379    if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1380      RewriteUsesOfLoadForHeapSRoA(LI, InsertedScalarizedValues, PHIsToRewrite);
1381      continue;
1382    }
1383
1384    // Must be a store of null.
1385    StoreInst *SI = cast<StoreInst>(User);
1386    assert(isa<ConstantPointerNull>(SI->getOperand(0)) &&
1387           "Unexpected heap-sra user!");
1388
1389    // Insert a store of null into each global.
1390    for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1391      const PointerType *PT = cast<PointerType>(FieldGlobals[i]->getType());
1392      Constant *Null = Constant::getNullValue(PT->getElementType());
1393      new StoreInst(Null, FieldGlobals[i], SI);
1394    }
1395    // Erase the original store.
1396    SI->eraseFromParent();
1397  }
1398
1399  // While we have PHIs that are interesting to rewrite, do it.
1400  while (!PHIsToRewrite.empty()) {
1401    PHINode *PN = PHIsToRewrite.back().first;
1402    unsigned FieldNo = PHIsToRewrite.back().second;
1403    PHIsToRewrite.pop_back();
1404    PHINode *FieldPN = cast<PHINode>(InsertedScalarizedValues[PN][FieldNo]);
1405    assert(FieldPN->getNumIncomingValues() == 0 &&"Already processed this phi");
1406
1407    // Add all the incoming values.  This can materialize more phis.
1408    for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1409      Value *InVal = PN->getIncomingValue(i);
1410      InVal = GetHeapSROAValue(InVal, FieldNo, InsertedScalarizedValues,
1411                               PHIsToRewrite);
1412      FieldPN->addIncoming(InVal, PN->getIncomingBlock(i));
1413    }
1414  }
1415
1416  // Drop all inter-phi links and any loads that made it this far.
1417  for (DenseMap<Value*, std::vector<Value*> >::iterator
1418       I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1419       I != E; ++I) {
1420    if (PHINode *PN = dyn_cast<PHINode>(I->first))
1421      PN->dropAllReferences();
1422    else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1423      LI->dropAllReferences();
1424  }
1425
1426  // Delete all the phis and loads now that inter-references are dead.
1427  for (DenseMap<Value*, std::vector<Value*> >::iterator
1428       I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1429       I != E; ++I) {
1430    if (PHINode *PN = dyn_cast<PHINode>(I->first))
1431      PN->eraseFromParent();
1432    else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1433      LI->eraseFromParent();
1434  }
1435
1436  // The old global is now dead, remove it.
1437  GV->eraseFromParent();
1438
1439  ++NumHeapSRA;
1440  return cast<GlobalVariable>(FieldGlobals[0]);
1441}
1442
1443/// TryToOptimizeStoreOfMallocToGlobal - This function is called when we see a
1444/// pointer global variable with a single value stored it that is a malloc or
1445/// cast of malloc.
1446static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV,
1447                                               CallInst *CI,
1448                                               const Type *AllocTy,
1449                                               Module::global_iterator &GVI,
1450                                               TargetData *TD) {
1451  // If this is a malloc of an abstract type, don't touch it.
1452  if (!AllocTy->isSized())
1453    return false;
1454
1455  // We can't optimize this global unless all uses of it are *known* to be
1456  // of the malloc value, not of the null initializer value (consider a use
1457  // that compares the global's value against zero to see if the malloc has
1458  // been reached).  To do this, we check to see if all uses of the global
1459  // would trap if the global were null: this proves that they must all
1460  // happen after the malloc.
1461  if (!AllUsesOfLoadedValueWillTrapIfNull(GV))
1462    return false;
1463
1464  // We can't optimize this if the malloc itself is used in a complex way,
1465  // for example, being stored into multiple globals.  This allows the
1466  // malloc to be stored into the specified global, loaded setcc'd, and
1467  // GEP'd.  These are all things we could transform to using the global
1468  // for.
1469  {
1470    SmallPtrSet<PHINode*, 8> PHIs;
1471    if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(CI, GV, PHIs))
1472      return false;
1473  }
1474
1475  // If we have a global that is only initialized with a fixed size malloc,
1476  // transform the program to use global memory instead of malloc'd memory.
1477  // This eliminates dynamic allocation, avoids an indirection accessing the
1478  // data, and exposes the resultant global to further GlobalOpt.
1479  // We cannot optimize the malloc if we cannot determine malloc array size.
1480  if (Value *NElems = getMallocArraySize(CI, TD, true)) {
1481    if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems))
1482      // Restrict this transformation to only working on small allocations
1483      // (2048 bytes currently), as we don't want to introduce a 16M global or
1484      // something.
1485      if (TD &&
1486          NElements->getZExtValue() * TD->getTypeAllocSize(AllocTy) < 2048) {
1487        GVI = OptimizeGlobalAddressOfMalloc(GV, CI, AllocTy, NElements, TD);
1488        return true;
1489      }
1490
1491    // If the allocation is an array of structures, consider transforming this
1492    // into multiple malloc'd arrays, one for each field.  This is basically
1493    // SRoA for malloc'd memory.
1494
1495    // If this is an allocation of a fixed size array of structs, analyze as a
1496    // variable size array.  malloc [100 x struct],1 -> malloc struct, 100
1497    if (NElems == ConstantInt::get(CI->getOperand(1)->getType(), 1))
1498      if (const ArrayType *AT = dyn_cast<ArrayType>(AllocTy))
1499        AllocTy = AT->getElementType();
1500
1501    if (const StructType *AllocSTy = dyn_cast<StructType>(AllocTy)) {
1502      // This the structure has an unreasonable number of fields, leave it
1503      // alone.
1504      if (AllocSTy->getNumElements() <= 16 && AllocSTy->getNumElements() != 0 &&
1505          AllGlobalLoadUsesSimpleEnoughForHeapSRA(GV, CI)) {
1506
1507        // If this is a fixed size array, transform the Malloc to be an alloc of
1508        // structs.  malloc [100 x struct],1 -> malloc struct, 100
1509        if (const ArrayType *AT =
1510                              dyn_cast<ArrayType>(getMallocAllocatedType(CI))) {
1511          const Type *IntPtrTy = TD->getIntPtrType(CI->getContext());
1512          unsigned TypeSize = TD->getStructLayout(AllocSTy)->getSizeInBytes();
1513          Value *AllocSize = ConstantInt::get(IntPtrTy, TypeSize);
1514          Value *NumElements = ConstantInt::get(IntPtrTy, AT->getNumElements());
1515          Instruction *Malloc = CallInst::CreateMalloc(CI, IntPtrTy, AllocSTy,
1516                                                       AllocSize, NumElements,
1517                                                       CI->getName());
1518          Instruction *Cast = new BitCastInst(Malloc, CI->getType(), "tmp", CI);
1519          CI->replaceAllUsesWith(Cast);
1520          CI->eraseFromParent();
1521          CI = dyn_cast<BitCastInst>(Malloc) ?
1522               extractMallocCallFromBitCast(Malloc) : cast<CallInst>(Malloc);
1523        }
1524
1525        GVI = PerformHeapAllocSRoA(GV, CI, getMallocArraySize(CI, TD, true),TD);
1526        return true;
1527      }
1528    }
1529  }
1530
1531  return false;
1532}
1533
1534// OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge
1535// that only one value (besides its initializer) is ever stored to the global.
1536static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal,
1537                                     Module::global_iterator &GVI,
1538                                     TargetData *TD) {
1539  // Ignore no-op GEPs and bitcasts.
1540  StoredOnceVal = StoredOnceVal->stripPointerCasts();
1541
1542  // If we are dealing with a pointer global that is initialized to null and
1543  // only has one (non-null) value stored into it, then we can optimize any
1544  // users of the loaded value (often calls and loads) that would trap if the
1545  // value was null.
1546  if (GV->getInitializer()->getType()->isPointerTy() &&
1547      GV->getInitializer()->isNullValue()) {
1548    if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) {
1549      if (GV->getInitializer()->getType() != SOVC->getType())
1550        SOVC =
1551         ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType());
1552
1553      // Optimize away any trapping uses of the loaded value.
1554      if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC))
1555        return true;
1556    } else if (CallInst *CI = extractMallocCall(StoredOnceVal)) {
1557      const Type* MallocType = getMallocAllocatedType(CI);
1558      if (MallocType && TryToOptimizeStoreOfMallocToGlobal(GV, CI, MallocType,
1559                                                           GVI, TD))
1560        return true;
1561    }
1562  }
1563
1564  return false;
1565}
1566
1567/// TryToShrinkGlobalToBoolean - At this point, we have learned that the only
1568/// two values ever stored into GV are its initializer and OtherVal.  See if we
1569/// can shrink the global into a boolean and select between the two values
1570/// whenever it is used.  This exposes the values to other scalar optimizations.
1571static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) {
1572  const Type *GVElType = GV->getType()->getElementType();
1573
1574  // If GVElType is already i1, it is already shrunk.  If the type of the GV is
1575  // an FP value, pointer or vector, don't do this optimization because a select
1576  // between them is very expensive and unlikely to lead to later
1577  // simplification.  In these cases, we typically end up with "cond ? v1 : v2"
1578  // where v1 and v2 both require constant pool loads, a big loss.
1579  if (GVElType == Type::getInt1Ty(GV->getContext()) ||
1580      GVElType->isFloatingPointTy() ||
1581      GVElType->isPointerTy() || GVElType->isVectorTy())
1582    return false;
1583
1584  // Walk the use list of the global seeing if all the uses are load or store.
1585  // If there is anything else, bail out.
1586  for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I)
1587    if (!isa<LoadInst>(I) && !isa<StoreInst>(I))
1588      return false;
1589
1590  DEBUG(dbgs() << "   *** SHRINKING TO BOOL: " << *GV);
1591
1592  // Create the new global, initializing it to false.
1593  GlobalVariable *NewGV = new GlobalVariable(Type::getInt1Ty(GV->getContext()),
1594                                             false,
1595                                             GlobalValue::InternalLinkage,
1596                                        ConstantInt::getFalse(GV->getContext()),
1597                                             GV->getName()+".b",
1598                                             GV->isThreadLocal());
1599  GV->getParent()->getGlobalList().insert(GV, NewGV);
1600
1601  Constant *InitVal = GV->getInitializer();
1602  assert(InitVal->getType() != Type::getInt1Ty(GV->getContext()) &&
1603         "No reason to shrink to bool!");
1604
1605  // If initialized to zero and storing one into the global, we can use a cast
1606  // instead of a select to synthesize the desired value.
1607  bool IsOneZero = false;
1608  if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal))
1609    IsOneZero = InitVal->isNullValue() && CI->isOne();
1610
1611  while (!GV->use_empty()) {
1612    Instruction *UI = cast<Instruction>(GV->use_back());
1613    if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
1614      // Change the store into a boolean store.
1615      bool StoringOther = SI->getOperand(0) == OtherVal;
1616      // Only do this if we weren't storing a loaded value.
1617      Value *StoreVal;
1618      if (StoringOther || SI->getOperand(0) == InitVal)
1619        StoreVal = ConstantInt::get(Type::getInt1Ty(GV->getContext()),
1620                                    StoringOther);
1621      else {
1622        // Otherwise, we are storing a previously loaded copy.  To do this,
1623        // change the copy from copying the original value to just copying the
1624        // bool.
1625        Instruction *StoredVal = cast<Instruction>(SI->getOperand(0));
1626
1627        // If we're already replaced the input, StoredVal will be a cast or
1628        // select instruction.  If not, it will be a load of the original
1629        // global.
1630        if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
1631          assert(LI->getOperand(0) == GV && "Not a copy!");
1632          // Insert a new load, to preserve the saved value.
1633          StoreVal = new LoadInst(NewGV, LI->getName()+".b", LI);
1634        } else {
1635          assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) &&
1636                 "This is not a form that we understand!");
1637          StoreVal = StoredVal->getOperand(0);
1638          assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!");
1639        }
1640      }
1641      new StoreInst(StoreVal, NewGV, SI);
1642    } else {
1643      // Change the load into a load of bool then a select.
1644      LoadInst *LI = cast<LoadInst>(UI);
1645      LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", LI);
1646      Value *NSI;
1647      if (IsOneZero)
1648        NSI = new ZExtInst(NLI, LI->getType(), "", LI);
1649      else
1650        NSI = SelectInst::Create(NLI, OtherVal, InitVal, "", LI);
1651      NSI->takeName(LI);
1652      LI->replaceAllUsesWith(NSI);
1653    }
1654    UI->eraseFromParent();
1655  }
1656
1657  GV->eraseFromParent();
1658  return true;
1659}
1660
1661
1662/// ProcessInternalGlobal - Analyze the specified global variable and optimize
1663/// it if possible.  If we make a change, return true.
1664bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV,
1665                                      Module::global_iterator &GVI) {
1666  SmallPtrSet<const PHINode*, 16> PHIUsers;
1667  GlobalStatus GS;
1668  GV->removeDeadConstantUsers();
1669
1670  if (GV->use_empty()) {
1671    DEBUG(dbgs() << "GLOBAL DEAD: " << *GV);
1672    GV->eraseFromParent();
1673    ++NumDeleted;
1674    return true;
1675  }
1676
1677  if (!AnalyzeGlobal(GV, GS, PHIUsers)) {
1678#if 0
1679    DEBUG(dbgs() << "Global: " << *GV);
1680    DEBUG(dbgs() << "  isLoaded = " << GS.isLoaded << "\n");
1681    DEBUG(dbgs() << "  StoredType = ");
1682    switch (GS.StoredType) {
1683    case GlobalStatus::NotStored: DEBUG(dbgs() << "NEVER STORED\n"); break;
1684    case GlobalStatus::isInitializerStored: DEBUG(dbgs() << "INIT STORED\n");
1685                                            break;
1686    case GlobalStatus::isStoredOnce: DEBUG(dbgs() << "STORED ONCE\n"); break;
1687    case GlobalStatus::isStored: DEBUG(dbgs() << "stored\n"); break;
1688    }
1689    if (GS.StoredType == GlobalStatus::isStoredOnce && GS.StoredOnceValue)
1690      DEBUG(dbgs() << "  StoredOnceValue = " << *GS.StoredOnceValue << "\n");
1691    if (GS.AccessingFunction && !GS.HasMultipleAccessingFunctions)
1692      DEBUG(dbgs() << "  AccessingFunction = " << GS.AccessingFunction->getName()
1693                  << "\n");
1694    DEBUG(dbgs() << "  HasMultipleAccessingFunctions =  "
1695                 << GS.HasMultipleAccessingFunctions << "\n");
1696    DEBUG(dbgs() << "  HasNonInstructionUser = "
1697                 << GS.HasNonInstructionUser<<"\n");
1698    DEBUG(dbgs() << "\n");
1699#endif
1700
1701    // If this is a first class global and has only one accessing function
1702    // and this function is main (which we know is not recursive we can make
1703    // this global a local variable) we replace the global with a local alloca
1704    // in this function.
1705    //
1706    // NOTE: It doesn't make sense to promote non single-value types since we
1707    // are just replacing static memory to stack memory.
1708    //
1709    // If the global is in different address space, don't bring it to stack.
1710    if (!GS.HasMultipleAccessingFunctions &&
1711        GS.AccessingFunction && !GS.HasNonInstructionUser &&
1712        GV->getType()->getElementType()->isSingleValueType() &&
1713        GS.AccessingFunction->getName() == "main" &&
1714        GS.AccessingFunction->hasExternalLinkage() &&
1715        GV->getType()->getAddressSpace() == 0) {
1716      DEBUG(dbgs() << "LOCALIZING GLOBAL: " << *GV);
1717      Instruction& FirstI = const_cast<Instruction&>(*GS.AccessingFunction
1718                                                     ->getEntryBlock().begin());
1719      const Type* ElemTy = GV->getType()->getElementType();
1720      // FIXME: Pass Global's alignment when globals have alignment
1721      AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), &FirstI);
1722      if (!isa<UndefValue>(GV->getInitializer()))
1723        new StoreInst(GV->getInitializer(), Alloca, &FirstI);
1724
1725      GV->replaceAllUsesWith(Alloca);
1726      GV->eraseFromParent();
1727      ++NumLocalized;
1728      return true;
1729    }
1730
1731    // If the global is never loaded (but may be stored to), it is dead.
1732    // Delete it now.
1733    if (!GS.isLoaded) {
1734      DEBUG(dbgs() << "GLOBAL NEVER LOADED: " << *GV);
1735
1736      // Delete any stores we can find to the global.  We may not be able to
1737      // make it completely dead though.
1738      bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer());
1739
1740      // If the global is dead now, delete it.
1741      if (GV->use_empty()) {
1742        GV->eraseFromParent();
1743        ++NumDeleted;
1744        Changed = true;
1745      }
1746      return Changed;
1747
1748    } else if (GS.StoredType <= GlobalStatus::isInitializerStored) {
1749      DEBUG(dbgs() << "MARKING CONSTANT: " << *GV);
1750      GV->setConstant(true);
1751
1752      // Clean up any obviously simplifiable users now.
1753      CleanupConstantGlobalUsers(GV, GV->getInitializer());
1754
1755      // If the global is dead now, just nuke it.
1756      if (GV->use_empty()) {
1757        DEBUG(dbgs() << "   *** Marking constant allowed us to simplify "
1758                     << "all users and delete global!\n");
1759        GV->eraseFromParent();
1760        ++NumDeleted;
1761      }
1762
1763      ++NumMarked;
1764      return true;
1765    } else if (!GV->getInitializer()->getType()->isSingleValueType()) {
1766      if (TargetData *TD = getAnalysisIfAvailable<TargetData>())
1767        if (GlobalVariable *FirstNewGV = SRAGlobal(GV, *TD)) {
1768          GVI = FirstNewGV;  // Don't skip the newly produced globals!
1769          return true;
1770        }
1771    } else if (GS.StoredType == GlobalStatus::isStoredOnce) {
1772      // If the initial value for the global was an undef value, and if only
1773      // one other value was stored into it, we can just change the
1774      // initializer to be the stored value, then delete all stores to the
1775      // global.  This allows us to mark it constant.
1776      if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1777        if (isa<UndefValue>(GV->getInitializer())) {
1778          // Change the initial value here.
1779          GV->setInitializer(SOVConstant);
1780
1781          // Clean up any obviously simplifiable users now.
1782          CleanupConstantGlobalUsers(GV, GV->getInitializer());
1783
1784          if (GV->use_empty()) {
1785            DEBUG(dbgs() << "   *** Substituting initializer allowed us to "
1786                         << "simplify all users and delete global!\n");
1787            GV->eraseFromParent();
1788            ++NumDeleted;
1789          } else {
1790            GVI = GV;
1791          }
1792          ++NumSubstitute;
1793          return true;
1794        }
1795
1796      // Try to optimize globals based on the knowledge that only one value
1797      // (besides its initializer) is ever stored to the global.
1798      if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI,
1799                                   getAnalysisIfAvailable<TargetData>()))
1800        return true;
1801
1802      // Otherwise, if the global was not a boolean, we can shrink it to be a
1803      // boolean.
1804      if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1805        if (TryToShrinkGlobalToBoolean(GV, SOVConstant)) {
1806          ++NumShrunkToBool;
1807          return true;
1808        }
1809    }
1810  }
1811  return false;
1812}
1813
1814/// ChangeCalleesToFastCall - Walk all of the direct calls of the specified
1815/// function, changing them to FastCC.
1816static void ChangeCalleesToFastCall(Function *F) {
1817  for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1818    CallSite User(cast<Instruction>(*UI));
1819    User.setCallingConv(CallingConv::Fast);
1820  }
1821}
1822
1823static AttrListPtr StripNest(const AttrListPtr &Attrs) {
1824  for (unsigned i = 0, e = Attrs.getNumSlots(); i != e; ++i) {
1825    if ((Attrs.getSlot(i).Attrs & Attribute::Nest) == 0)
1826      continue;
1827
1828    // There can be only one.
1829    return Attrs.removeAttr(Attrs.getSlot(i).Index, Attribute::Nest);
1830  }
1831
1832  return Attrs;
1833}
1834
1835static void RemoveNestAttribute(Function *F) {
1836  F->setAttributes(StripNest(F->getAttributes()));
1837  for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1838    CallSite User(cast<Instruction>(*UI));
1839    User.setAttributes(StripNest(User.getAttributes()));
1840  }
1841}
1842
1843bool GlobalOpt::OptimizeFunctions(Module &M) {
1844  bool Changed = false;
1845  // Optimize functions.
1846  for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) {
1847    Function *F = FI++;
1848    // Functions without names cannot be referenced outside this module.
1849    if (!F->hasName() && !F->isDeclaration())
1850      F->setLinkage(GlobalValue::InternalLinkage);
1851    F->removeDeadConstantUsers();
1852    if (F->use_empty() && (F->hasLocalLinkage() || F->hasLinkOnceLinkage())) {
1853      F->eraseFromParent();
1854      Changed = true;
1855      ++NumFnDeleted;
1856    } else if (F->hasLocalLinkage()) {
1857      if (F->getCallingConv() == CallingConv::C && !F->isVarArg() &&
1858          !F->hasAddressTaken()) {
1859        // If this function has C calling conventions, is not a varargs
1860        // function, and is only called directly, promote it to use the Fast
1861        // calling convention.
1862        F->setCallingConv(CallingConv::Fast);
1863        ChangeCalleesToFastCall(F);
1864        ++NumFastCallFns;
1865        Changed = true;
1866      }
1867
1868      if (F->getAttributes().hasAttrSomewhere(Attribute::Nest) &&
1869          !F->hasAddressTaken()) {
1870        // The function is not used by a trampoline intrinsic, so it is safe
1871        // to remove the 'nest' attribute.
1872        RemoveNestAttribute(F);
1873        ++NumNestRemoved;
1874        Changed = true;
1875      }
1876    }
1877  }
1878  return Changed;
1879}
1880
1881bool GlobalOpt::OptimizeGlobalVars(Module &M) {
1882  bool Changed = false;
1883  for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
1884       GVI != E; ) {
1885    GlobalVariable *GV = GVI++;
1886    // Global variables without names cannot be referenced outside this module.
1887    if (!GV->hasName() && !GV->isDeclaration())
1888      GV->setLinkage(GlobalValue::InternalLinkage);
1889    // Simplify the initializer.
1890    if (GV->hasInitializer())
1891      if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GV->getInitializer())) {
1892        TargetData *TD = getAnalysisIfAvailable<TargetData>();
1893        Constant *New = ConstantFoldConstantExpression(CE, TD);
1894        if (New && New != CE)
1895          GV->setInitializer(New);
1896      }
1897    // Do more involved optimizations if the global is internal.
1898    if (!GV->isConstant() && GV->hasLocalLinkage() &&
1899        GV->hasInitializer())
1900      Changed |= ProcessInternalGlobal(GV, GVI);
1901  }
1902  return Changed;
1903}
1904
1905/// FindGlobalCtors - Find the llvm.globalctors list, verifying that all
1906/// initializers have an init priority of 65535.
1907GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) {
1908  for (Module::global_iterator I = M.global_begin(), E = M.global_end();
1909       I != E; ++I)
1910    if (I->getName() == "llvm.global_ctors") {
1911      // Found it, verify it's an array of { int, void()* }.
1912      const ArrayType *ATy =dyn_cast<ArrayType>(I->getType()->getElementType());
1913      if (!ATy) return 0;
1914      const StructType *STy = dyn_cast<StructType>(ATy->getElementType());
1915      if (!STy || STy->getNumElements() != 2 ||
1916          !STy->getElementType(0)->isIntegerTy(32)) return 0;
1917      const PointerType *PFTy = dyn_cast<PointerType>(STy->getElementType(1));
1918      if (!PFTy) return 0;
1919      const FunctionType *FTy = dyn_cast<FunctionType>(PFTy->getElementType());
1920      if (!FTy || !FTy->getReturnType()->isVoidTy() ||
1921          FTy->isVarArg() || FTy->getNumParams() != 0)
1922        return 0;
1923
1924      // Verify that the initializer is simple enough for us to handle.
1925      if (!I->hasDefinitiveInitializer()) return 0;
1926      ConstantArray *CA = dyn_cast<ConstantArray>(I->getInitializer());
1927      if (!CA) return 0;
1928      for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i)
1929        if (ConstantStruct *CS = dyn_cast<ConstantStruct>(*i)) {
1930          if (isa<ConstantPointerNull>(CS->getOperand(1)))
1931            continue;
1932
1933          // Must have a function or null ptr.
1934          if (!isa<Function>(CS->getOperand(1)))
1935            return 0;
1936
1937          // Init priority must be standard.
1938          ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0));
1939          if (!CI || CI->getZExtValue() != 65535)
1940            return 0;
1941        } else {
1942          return 0;
1943        }
1944
1945      return I;
1946    }
1947  return 0;
1948}
1949
1950/// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand,
1951/// return a list of the functions and null terminator as a vector.
1952static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) {
1953  ConstantArray *CA = cast<ConstantArray>(GV->getInitializer());
1954  std::vector<Function*> Result;
1955  Result.reserve(CA->getNumOperands());
1956  for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) {
1957    ConstantStruct *CS = cast<ConstantStruct>(*i);
1958    Result.push_back(dyn_cast<Function>(CS->getOperand(1)));
1959  }
1960  return Result;
1961}
1962
1963/// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the
1964/// specified array, returning the new global to use.
1965static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL,
1966                                          const std::vector<Function*> &Ctors) {
1967  // If we made a change, reassemble the initializer list.
1968  std::vector<Constant*> CSVals;
1969  CSVals.push_back(ConstantInt::get(Type::getInt32Ty(GCL->getContext()),65535));
1970  CSVals.push_back(0);
1971
1972  // Create the new init list.
1973  std::vector<Constant*> CAList;
1974  for (unsigned i = 0, e = Ctors.size(); i != e; ++i) {
1975    if (Ctors[i]) {
1976      CSVals[1] = Ctors[i];
1977    } else {
1978      const Type *FTy = FunctionType::get(Type::getVoidTy(GCL->getContext()),
1979                                          false);
1980      const PointerType *PFTy = PointerType::getUnqual(FTy);
1981      CSVals[1] = Constant::getNullValue(PFTy);
1982      CSVals[0] = ConstantInt::get(Type::getInt32Ty(GCL->getContext()),
1983                                   2147483647);
1984    }
1985    CAList.push_back(ConstantStruct::get(GCL->getContext(), CSVals, false));
1986  }
1987
1988  // Create the array initializer.
1989  const Type *StructTy =
1990      cast<ArrayType>(GCL->getType()->getElementType())->getElementType();
1991  Constant *CA = ConstantArray::get(ArrayType::get(StructTy,
1992                                                   CAList.size()), CAList);
1993
1994  // If we didn't change the number of elements, don't create a new GV.
1995  if (CA->getType() == GCL->getInitializer()->getType()) {
1996    GCL->setInitializer(CA);
1997    return GCL;
1998  }
1999
2000  // Create the new global and insert it next to the existing list.
2001  GlobalVariable *NGV = new GlobalVariable(CA->getType(), GCL->isConstant(),
2002                                           GCL->getLinkage(), CA, "",
2003                                           GCL->isThreadLocal());
2004  GCL->getParent()->getGlobalList().insert(GCL, NGV);
2005  NGV->takeName(GCL);
2006
2007  // Nuke the old list, replacing any uses with the new one.
2008  if (!GCL->use_empty()) {
2009    Constant *V = NGV;
2010    if (V->getType() != GCL->getType())
2011      V = ConstantExpr::getBitCast(V, GCL->getType());
2012    GCL->replaceAllUsesWith(V);
2013  }
2014  GCL->eraseFromParent();
2015
2016  if (Ctors.size())
2017    return NGV;
2018  else
2019    return 0;
2020}
2021
2022
2023static Constant *getVal(DenseMap<Value*, Constant*> &ComputedValues,
2024                        Value *V) {
2025  if (Constant *CV = dyn_cast<Constant>(V)) return CV;
2026  Constant *R = ComputedValues[V];
2027  assert(R && "Reference to an uncomputed value!");
2028  return R;
2029}
2030
2031/// isSimpleEnoughPointerToCommit - Return true if this constant is simple
2032/// enough for us to understand.  In particular, if it is a cast of something,
2033/// we punt.  We basically just support direct accesses to globals and GEP's of
2034/// globals.  This should be kept up to date with CommitValueTo.
2035static bool isSimpleEnoughPointerToCommit(Constant *C) {
2036  // Conservatively, avoid aggregate types. This is because we don't
2037  // want to worry about them partially overlapping other stores.
2038  if (!cast<PointerType>(C->getType())->getElementType()->isSingleValueType())
2039    return false;
2040
2041  if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
2042    // Do not allow weak/linkonce/dllimport/dllexport linkage or
2043    // external globals.
2044    return GV->hasDefinitiveInitializer();
2045
2046  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2047    // Handle a constantexpr gep.
2048    if (CE->getOpcode() == Instruction::GetElementPtr &&
2049        isa<GlobalVariable>(CE->getOperand(0)) &&
2050        cast<GEPOperator>(CE)->isInBounds()) {
2051      GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2052      // Do not allow weak/linkonce/dllimport/dllexport linkage or
2053      // external globals.
2054      if (!GV->hasDefinitiveInitializer())
2055        return false;
2056
2057      // The first index must be zero.
2058      ConstantInt *CI = dyn_cast<ConstantInt>(*next(CE->op_begin()));
2059      if (!CI || !CI->isZero()) return false;
2060
2061      // The remaining indices must be compile-time known integers within the
2062      // notional bounds of the corresponding static array types.
2063      if (!CE->isGEPWithNoNotionalOverIndexing())
2064        return false;
2065
2066      return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
2067    }
2068  return false;
2069}
2070
2071/// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global
2072/// initializer.  This returns 'Init' modified to reflect 'Val' stored into it.
2073/// At this point, the GEP operands of Addr [0, OpNo) have been stepped into.
2074static Constant *EvaluateStoreInto(Constant *Init, Constant *Val,
2075                                   ConstantExpr *Addr, unsigned OpNo) {
2076  // Base case of the recursion.
2077  if (OpNo == Addr->getNumOperands()) {
2078    assert(Val->getType() == Init->getType() && "Type mismatch!");
2079    return Val;
2080  }
2081
2082  std::vector<Constant*> Elts;
2083  if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
2084
2085    // Break up the constant into its elements.
2086    if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
2087      for (User::op_iterator i = CS->op_begin(), e = CS->op_end(); i != e; ++i)
2088        Elts.push_back(cast<Constant>(*i));
2089    } else if (isa<ConstantAggregateZero>(Init)) {
2090      for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
2091        Elts.push_back(Constant::getNullValue(STy->getElementType(i)));
2092    } else if (isa<UndefValue>(Init)) {
2093      for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
2094        Elts.push_back(UndefValue::get(STy->getElementType(i)));
2095    } else {
2096      llvm_unreachable("This code is out of sync with "
2097             " ConstantFoldLoadThroughGEPConstantExpr");
2098    }
2099
2100    // Replace the element that we are supposed to.
2101    ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo));
2102    unsigned Idx = CU->getZExtValue();
2103    assert(Idx < STy->getNumElements() && "Struct index out of range!");
2104    Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1);
2105
2106    // Return the modified struct.
2107    return ConstantStruct::get(Init->getContext(), &Elts[0], Elts.size(),
2108                               STy->isPacked());
2109  } else {
2110    ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo));
2111    const SequentialType *InitTy = cast<SequentialType>(Init->getType());
2112
2113    uint64_t NumElts;
2114    if (const ArrayType *ATy = dyn_cast<ArrayType>(InitTy))
2115      NumElts = ATy->getNumElements();
2116    else
2117      NumElts = cast<VectorType>(InitTy)->getNumElements();
2118
2119
2120    // Break up the array into elements.
2121    if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
2122      for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i)
2123        Elts.push_back(cast<Constant>(*i));
2124    } else if (ConstantVector *CV = dyn_cast<ConstantVector>(Init)) {
2125      for (User::op_iterator i = CV->op_begin(), e = CV->op_end(); i != e; ++i)
2126        Elts.push_back(cast<Constant>(*i));
2127    } else if (isa<ConstantAggregateZero>(Init)) {
2128      Elts.assign(NumElts, Constant::getNullValue(InitTy->getElementType()));
2129    } else {
2130      assert(isa<UndefValue>(Init) && "This code is out of sync with "
2131             " ConstantFoldLoadThroughGEPConstantExpr");
2132      Elts.assign(NumElts, UndefValue::get(InitTy->getElementType()));
2133    }
2134
2135    assert(CI->getZExtValue() < NumElts);
2136    Elts[CI->getZExtValue()] =
2137      EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1);
2138
2139    if (Init->getType()->isArrayTy())
2140      return ConstantArray::get(cast<ArrayType>(InitTy), Elts);
2141    else
2142      return ConstantVector::get(&Elts[0], Elts.size());
2143  }
2144}
2145
2146/// CommitValueTo - We have decided that Addr (which satisfies the predicate
2147/// isSimpleEnoughPointerToCommit) should get Val as its value.  Make it happen.
2148static void CommitValueTo(Constant *Val, Constant *Addr) {
2149  if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
2150    assert(GV->hasInitializer());
2151    GV->setInitializer(Val);
2152    return;
2153  }
2154
2155  ConstantExpr *CE = cast<ConstantExpr>(Addr);
2156  GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2157  GV->setInitializer(EvaluateStoreInto(GV->getInitializer(), Val, CE, 2));
2158}
2159
2160/// ComputeLoadResult - Return the value that would be computed by a load from
2161/// P after the stores reflected by 'memory' have been performed.  If we can't
2162/// decide, return null.
2163static Constant *ComputeLoadResult(Constant *P,
2164                                const DenseMap<Constant*, Constant*> &Memory) {
2165  // If this memory location has been recently stored, use the stored value: it
2166  // is the most up-to-date.
2167  DenseMap<Constant*, Constant*>::const_iterator I = Memory.find(P);
2168  if (I != Memory.end()) return I->second;
2169
2170  // Access it.
2171  if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
2172    if (GV->hasDefinitiveInitializer())
2173      return GV->getInitializer();
2174    return 0;
2175  }
2176
2177  // Handle a constantexpr getelementptr.
2178  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P))
2179    if (CE->getOpcode() == Instruction::GetElementPtr &&
2180        isa<GlobalVariable>(CE->getOperand(0))) {
2181      GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2182      if (GV->hasDefinitiveInitializer())
2183        return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
2184    }
2185
2186  return 0;  // don't know how to evaluate.
2187}
2188
2189/// EvaluateFunction - Evaluate a call to function F, returning true if
2190/// successful, false if we can't evaluate it.  ActualArgs contains the formal
2191/// arguments for the function.
2192static bool EvaluateFunction(Function *F, Constant *&RetVal,
2193                             const SmallVectorImpl<Constant*> &ActualArgs,
2194                             std::vector<Function*> &CallStack,
2195                             DenseMap<Constant*, Constant*> &MutatedMemory,
2196                             std::vector<GlobalVariable*> &AllocaTmps) {
2197  // Check to see if this function is already executing (recursion).  If so,
2198  // bail out.  TODO: we might want to accept limited recursion.
2199  if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end())
2200    return false;
2201
2202  CallStack.push_back(F);
2203
2204  /// Values - As we compute SSA register values, we store their contents here.
2205  DenseMap<Value*, Constant*> Values;
2206
2207  // Initialize arguments to the incoming values specified.
2208  unsigned ArgNo = 0;
2209  for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E;
2210       ++AI, ++ArgNo)
2211    Values[AI] = ActualArgs[ArgNo];
2212
2213  /// ExecutedBlocks - We only handle non-looping, non-recursive code.  As such,
2214  /// we can only evaluate any one basic block at most once.  This set keeps
2215  /// track of what we have executed so we can detect recursive cases etc.
2216  SmallPtrSet<BasicBlock*, 32> ExecutedBlocks;
2217
2218  // CurInst - The current instruction we're evaluating.
2219  BasicBlock::iterator CurInst = F->begin()->begin();
2220
2221  // This is the main evaluation loop.
2222  while (1) {
2223    Constant *InstResult = 0;
2224
2225    if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) {
2226      if (SI->isVolatile()) return false;  // no volatile accesses.
2227      Constant *Ptr = getVal(Values, SI->getOperand(1));
2228      if (!isSimpleEnoughPointerToCommit(Ptr))
2229        // If this is too complex for us to commit, reject it.
2230        return false;
2231      Constant *Val = getVal(Values, SI->getOperand(0));
2232      MutatedMemory[Ptr] = Val;
2233    } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) {
2234      InstResult = ConstantExpr::get(BO->getOpcode(),
2235                                     getVal(Values, BO->getOperand(0)),
2236                                     getVal(Values, BO->getOperand(1)));
2237    } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) {
2238      InstResult = ConstantExpr::getCompare(CI->getPredicate(),
2239                                            getVal(Values, CI->getOperand(0)),
2240                                            getVal(Values, CI->getOperand(1)));
2241    } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) {
2242      InstResult = ConstantExpr::getCast(CI->getOpcode(),
2243                                         getVal(Values, CI->getOperand(0)),
2244                                         CI->getType());
2245    } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) {
2246      InstResult =
2247            ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)),
2248                                           getVal(Values, SI->getOperand(1)),
2249                                           getVal(Values, SI->getOperand(2)));
2250    } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) {
2251      Constant *P = getVal(Values, GEP->getOperand(0));
2252      SmallVector<Constant*, 8> GEPOps;
2253      for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end();
2254           i != e; ++i)
2255        GEPOps.push_back(getVal(Values, *i));
2256      InstResult = cast<GEPOperator>(GEP)->isInBounds() ?
2257          ConstantExpr::getInBoundsGetElementPtr(P, &GEPOps[0], GEPOps.size()) :
2258          ConstantExpr::getGetElementPtr(P, &GEPOps[0], GEPOps.size());
2259    } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) {
2260      if (LI->isVolatile()) return false;  // no volatile accesses.
2261      InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)),
2262                                     MutatedMemory);
2263      if (InstResult == 0) return false; // Could not evaluate load.
2264    } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) {
2265      if (AI->isArrayAllocation()) return false;  // Cannot handle array allocs.
2266      const Type *Ty = AI->getType()->getElementType();
2267      AllocaTmps.push_back(new GlobalVariable(Ty, false,
2268                                              GlobalValue::InternalLinkage,
2269                                              UndefValue::get(Ty),
2270                                              AI->getName()));
2271      InstResult = AllocaTmps.back();
2272    } else if (CallInst *CI = dyn_cast<CallInst>(CurInst)) {
2273
2274      // Debug info can safely be ignored here.
2275      if (isa<DbgInfoIntrinsic>(CI)) {
2276        ++CurInst;
2277        continue;
2278      }
2279
2280      // Cannot handle inline asm.
2281      if (isa<InlineAsm>(CI->getOperand(0))) return false;
2282
2283      // Resolve function pointers.
2284      Function *Callee = dyn_cast<Function>(getVal(Values, CI->getOperand(0)));
2285      if (!Callee) return false;  // Cannot resolve.
2286
2287      SmallVector<Constant*, 8> Formals;
2288      for (User::op_iterator i = CI->op_begin() + 1, e = CI->op_end();
2289           i != e; ++i)
2290        Formals.push_back(getVal(Values, *i));
2291
2292      if (Callee->isDeclaration()) {
2293        // If this is a function we can constant fold, do it.
2294        if (Constant *C = ConstantFoldCall(Callee, Formals.data(),
2295                                           Formals.size())) {
2296          InstResult = C;
2297        } else {
2298          return false;
2299        }
2300      } else {
2301        if (Callee->getFunctionType()->isVarArg())
2302          return false;
2303
2304        Constant *RetVal;
2305        // Execute the call, if successful, use the return value.
2306        if (!EvaluateFunction(Callee, RetVal, Formals, CallStack,
2307                              MutatedMemory, AllocaTmps))
2308          return false;
2309        InstResult = RetVal;
2310      }
2311    } else if (isa<TerminatorInst>(CurInst)) {
2312      BasicBlock *NewBB = 0;
2313      if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) {
2314        if (BI->isUnconditional()) {
2315          NewBB = BI->getSuccessor(0);
2316        } else {
2317          ConstantInt *Cond =
2318            dyn_cast<ConstantInt>(getVal(Values, BI->getCondition()));
2319          if (!Cond) return false;  // Cannot determine.
2320
2321          NewBB = BI->getSuccessor(!Cond->getZExtValue());
2322        }
2323      } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) {
2324        ConstantInt *Val =
2325          dyn_cast<ConstantInt>(getVal(Values, SI->getCondition()));
2326        if (!Val) return false;  // Cannot determine.
2327        NewBB = SI->getSuccessor(SI->findCaseValue(Val));
2328      } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(CurInst)) {
2329        Value *Val = getVal(Values, IBI->getAddress())->stripPointerCasts();
2330        if (BlockAddress *BA = dyn_cast<BlockAddress>(Val))
2331          NewBB = BA->getBasicBlock();
2332        else
2333          return false;  // Cannot determine.
2334      } else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) {
2335        if (RI->getNumOperands())
2336          RetVal = getVal(Values, RI->getOperand(0));
2337
2338        CallStack.pop_back();  // return from fn.
2339        return true;  // We succeeded at evaluating this ctor!
2340      } else {
2341        // invoke, unwind, unreachable.
2342        return false;  // Cannot handle this terminator.
2343      }
2344
2345      // Okay, we succeeded in evaluating this control flow.  See if we have
2346      // executed the new block before.  If so, we have a looping function,
2347      // which we cannot evaluate in reasonable time.
2348      if (!ExecutedBlocks.insert(NewBB))
2349        return false;  // looped!
2350
2351      // Okay, we have never been in this block before.  Check to see if there
2352      // are any PHI nodes.  If so, evaluate them with information about where
2353      // we came from.
2354      BasicBlock *OldBB = CurInst->getParent();
2355      CurInst = NewBB->begin();
2356      PHINode *PN;
2357      for (; (PN = dyn_cast<PHINode>(CurInst)); ++CurInst)
2358        Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB));
2359
2360      // Do NOT increment CurInst.  We know that the terminator had no value.
2361      continue;
2362    } else {
2363      // Did not know how to evaluate this!
2364      return false;
2365    }
2366
2367    if (!CurInst->use_empty())
2368      Values[CurInst] = InstResult;
2369
2370    // Advance program counter.
2371    ++CurInst;
2372  }
2373}
2374
2375/// EvaluateStaticConstructor - Evaluate static constructors in the function, if
2376/// we can.  Return true if we can, false otherwise.
2377static bool EvaluateStaticConstructor(Function *F) {
2378  /// MutatedMemory - For each store we execute, we update this map.  Loads
2379  /// check this to get the most up-to-date value.  If evaluation is successful,
2380  /// this state is committed to the process.
2381  DenseMap<Constant*, Constant*> MutatedMemory;
2382
2383  /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable
2384  /// to represent its body.  This vector is needed so we can delete the
2385  /// temporary globals when we are done.
2386  std::vector<GlobalVariable*> AllocaTmps;
2387
2388  /// CallStack - This is used to detect recursion.  In pathological situations
2389  /// we could hit exponential behavior, but at least there is nothing
2390  /// unbounded.
2391  std::vector<Function*> CallStack;
2392
2393  // Call the function.
2394  Constant *RetValDummy;
2395  bool EvalSuccess = EvaluateFunction(F, RetValDummy,
2396                                      SmallVector<Constant*, 0>(), CallStack,
2397                                      MutatedMemory, AllocaTmps);
2398  if (EvalSuccess) {
2399    // We succeeded at evaluation: commit the result.
2400    DEBUG(dbgs() << "FULLY EVALUATED GLOBAL CTOR FUNCTION '"
2401          << F->getName() << "' to " << MutatedMemory.size()
2402          << " stores.\n");
2403    for (DenseMap<Constant*, Constant*>::iterator I = MutatedMemory.begin(),
2404         E = MutatedMemory.end(); I != E; ++I)
2405      CommitValueTo(I->second, I->first);
2406  }
2407
2408  // At this point, we are done interpreting.  If we created any 'alloca'
2409  // temporaries, release them now.
2410  while (!AllocaTmps.empty()) {
2411    GlobalVariable *Tmp = AllocaTmps.back();
2412    AllocaTmps.pop_back();
2413
2414    // If there are still users of the alloca, the program is doing something
2415    // silly, e.g. storing the address of the alloca somewhere and using it
2416    // later.  Since this is undefined, we'll just make it be null.
2417    if (!Tmp->use_empty())
2418      Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType()));
2419    delete Tmp;
2420  }
2421
2422  return EvalSuccess;
2423}
2424
2425
2426
2427/// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible.
2428/// Return true if anything changed.
2429bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) {
2430  std::vector<Function*> Ctors = ParseGlobalCtors(GCL);
2431  bool MadeChange = false;
2432  if (Ctors.empty()) return false;
2433
2434  // Loop over global ctors, optimizing them when we can.
2435  for (unsigned i = 0; i != Ctors.size(); ++i) {
2436    Function *F = Ctors[i];
2437    // Found a null terminator in the middle of the list, prune off the rest of
2438    // the list.
2439    if (F == 0) {
2440      if (i != Ctors.size()-1) {
2441        Ctors.resize(i+1);
2442        MadeChange = true;
2443      }
2444      break;
2445    }
2446
2447    // We cannot simplify external ctor functions.
2448    if (F->empty()) continue;
2449
2450    // If we can evaluate the ctor at compile time, do.
2451    if (EvaluateStaticConstructor(F)) {
2452      Ctors.erase(Ctors.begin()+i);
2453      MadeChange = true;
2454      --i;
2455      ++NumCtorsEvaluated;
2456      continue;
2457    }
2458  }
2459
2460  if (!MadeChange) return false;
2461
2462  GCL = InstallGlobalCtors(GCL, Ctors);
2463  return true;
2464}
2465
2466bool GlobalOpt::OptimizeGlobalAliases(Module &M) {
2467  bool Changed = false;
2468
2469  for (Module::alias_iterator I = M.alias_begin(), E = M.alias_end();
2470       I != E;) {
2471    Module::alias_iterator J = I++;
2472    // Aliases without names cannot be referenced outside this module.
2473    if (!J->hasName() && !J->isDeclaration())
2474      J->setLinkage(GlobalValue::InternalLinkage);
2475    // If the aliasee may change at link time, nothing can be done - bail out.
2476    if (J->mayBeOverridden())
2477      continue;
2478
2479    Constant *Aliasee = J->getAliasee();
2480    GlobalValue *Target = cast<GlobalValue>(Aliasee->stripPointerCasts());
2481    Target->removeDeadConstantUsers();
2482    bool hasOneUse = Target->hasOneUse() && Aliasee->hasOneUse();
2483
2484    // Make all users of the alias use the aliasee instead.
2485    if (!J->use_empty()) {
2486      J->replaceAllUsesWith(Aliasee);
2487      ++NumAliasesResolved;
2488      Changed = true;
2489    }
2490
2491    // If the alias is externally visible, we may still be able to simplify it.
2492    if (!J->hasLocalLinkage()) {
2493      // If the aliasee has internal linkage, give it the name and linkage
2494      // of the alias, and delete the alias.  This turns:
2495      //   define internal ... @f(...)
2496      //   @a = alias ... @f
2497      // into:
2498      //   define ... @a(...)
2499      if (!Target->hasLocalLinkage())
2500        continue;
2501
2502      // Do not perform the transform if multiple aliases potentially target the
2503      // aliasee.  This check also ensures that it is safe to replace the section
2504      // and other attributes of the aliasee with those of the alias.
2505      if (!hasOneUse)
2506        continue;
2507
2508      // Give the aliasee the name, linkage and other attributes of the alias.
2509      Target->takeName(J);
2510      Target->setLinkage(J->getLinkage());
2511      Target->GlobalValue::copyAttributesFrom(J);
2512    }
2513
2514    // Delete the alias.
2515    M.getAliasList().erase(J);
2516    ++NumAliasesRemoved;
2517    Changed = true;
2518  }
2519
2520  return Changed;
2521}
2522
2523bool GlobalOpt::runOnModule(Module &M) {
2524  bool Changed = false;
2525
2526  // Try to find the llvm.globalctors list.
2527  GlobalVariable *GlobalCtors = FindGlobalCtors(M);
2528
2529  bool LocalChange = true;
2530  while (LocalChange) {
2531    LocalChange = false;
2532
2533    // Delete functions that are trivially dead, ccc -> fastcc
2534    LocalChange |= OptimizeFunctions(M);
2535
2536    // Optimize global_ctors list.
2537    if (GlobalCtors)
2538      LocalChange |= OptimizeGlobalCtorsList(GlobalCtors);
2539
2540    // Optimize non-address-taken globals.
2541    LocalChange |= OptimizeGlobalVars(M);
2542
2543    // Resolve aliases, when possible.
2544    LocalChange |= OptimizeGlobalAliases(M);
2545    Changed |= LocalChange;
2546  }
2547
2548  // TODO: Move all global ctors functions to the end of the module for code
2549  // layout.
2550
2551  return Changed;
2552}
2553