GlobalOpt.cpp revision 3be3ab75d6545698dc54d1197b8bd93ed4325add
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      CallSite CS(I);
686      if (CS.getCalledValue() == V) {
687        // Calling through the pointer!  Turn into a direct call, but be careful
688        // that the pointer is not also being passed as an argument.
689        CS.setCalledFunction(NewV);
690        Changed = true;
691        bool PassedAsArg = false;
692        for (unsigned i = 0, e = CS.arg_size(); i != e; ++i)
693          if (CS.getArgument(i) == V) {
694            PassedAsArg = true;
695            CS.setArgument(i, NewV);
696          }
697
698        if (PassedAsArg) {
699          // Being passed as an argument also.  Be careful to not invalidate UI!
700          UI = V->use_begin();
701        }
702      }
703    } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
704      Changed |= OptimizeAwayTrappingUsesOfValue(CI,
705                                ConstantExpr::getCast(CI->getOpcode(),
706                                                      NewV, CI->getType()));
707      if (CI->use_empty()) {
708        Changed = true;
709        CI->eraseFromParent();
710      }
711    } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
712      // Should handle GEP here.
713      SmallVector<Constant*, 8> Idxs;
714      Idxs.reserve(GEPI->getNumOperands()-1);
715      for (User::op_iterator i = GEPI->op_begin() + 1, e = GEPI->op_end();
716           i != e; ++i)
717        if (Constant *C = dyn_cast<Constant>(*i))
718          Idxs.push_back(C);
719        else
720          break;
721      if (Idxs.size() == GEPI->getNumOperands()-1)
722        Changed |= OptimizeAwayTrappingUsesOfValue(GEPI,
723                          ConstantExpr::getGetElementPtr(NewV, &Idxs[0],
724                                                        Idxs.size()));
725      if (GEPI->use_empty()) {
726        Changed = true;
727        GEPI->eraseFromParent();
728      }
729    }
730  }
731
732  return Changed;
733}
734
735
736/// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null
737/// value stored into it.  If there are uses of the loaded value that would trap
738/// if the loaded value is dynamically null, then we know that they cannot be
739/// reachable with a null optimize away the load.
740static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV) {
741  bool Changed = false;
742
743  // Keep track of whether we are able to remove all the uses of the global
744  // other than the store that defines it.
745  bool AllNonStoreUsesGone = true;
746
747  // Replace all uses of loads with uses of uses of the stored value.
748  for (Value::use_iterator GUI = GV->use_begin(), E = GV->use_end(); GUI != E;){
749    User *GlobalUser = *GUI++;
750    if (LoadInst *LI = dyn_cast<LoadInst>(GlobalUser)) {
751      Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV);
752      // If we were able to delete all uses of the loads
753      if (LI->use_empty()) {
754        LI->eraseFromParent();
755        Changed = true;
756      } else {
757        AllNonStoreUsesGone = false;
758      }
759    } else if (isa<StoreInst>(GlobalUser)) {
760      // Ignore the store that stores "LV" to the global.
761      assert(GlobalUser->getOperand(1) == GV &&
762             "Must be storing *to* the global");
763    } else {
764      AllNonStoreUsesGone = false;
765
766      // If we get here we could have other crazy uses that are transitively
767      // loaded.
768      assert((isa<PHINode>(GlobalUser) || isa<SelectInst>(GlobalUser) ||
769              isa<ConstantExpr>(GlobalUser)) && "Only expect load and stores!");
770    }
771  }
772
773  if (Changed) {
774    DEBUG(dbgs() << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV);
775    ++NumGlobUses;
776  }
777
778  // If we nuked all of the loads, then none of the stores are needed either,
779  // nor is the global.
780  if (AllNonStoreUsesGone) {
781    DEBUG(dbgs() << "  *** GLOBAL NOW DEAD!\n");
782    CleanupConstantGlobalUsers(GV, 0);
783    if (GV->use_empty()) {
784      GV->eraseFromParent();
785      ++NumDeleted;
786    }
787    Changed = true;
788  }
789  return Changed;
790}
791
792/// ConstantPropUsersOf - Walk the use list of V, constant folding all of the
793/// instructions that are foldable.
794static void ConstantPropUsersOf(Value *V) {
795  for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; )
796    if (Instruction *I = dyn_cast<Instruction>(*UI++))
797      if (Constant *NewC = ConstantFoldInstruction(I)) {
798        I->replaceAllUsesWith(NewC);
799
800        // Advance UI to the next non-I use to avoid invalidating it!
801        // Instructions could multiply use V.
802        while (UI != E && *UI == I)
803          ++UI;
804        I->eraseFromParent();
805      }
806}
807
808/// OptimizeGlobalAddressOfMalloc - This function takes the specified global
809/// variable, and transforms the program as if it always contained the result of
810/// the specified malloc.  Because it is always the result of the specified
811/// malloc, there is no reason to actually DO the malloc.  Instead, turn the
812/// malloc into a global, and any loads of GV as uses of the new global.
813static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV,
814                                                     CallInst *CI,
815                                                     const Type *AllocTy,
816                                                     ConstantInt *NElements,
817                                                     TargetData* TD) {
818  DEBUG(errs() << "PROMOTING GLOBAL: " << *GV << "  CALL = " << *CI << '\n');
819
820  const Type *GlobalType;
821  if (NElements->getZExtValue() == 1)
822    GlobalType = AllocTy;
823  else
824    // If we have an array allocation, the global variable is of an array.
825    GlobalType = ArrayType::get(AllocTy, NElements->getZExtValue());
826
827  // Create the new global variable.  The contents of the malloc'd memory is
828  // undefined, so initialize with an undef value.
829  GlobalVariable *NewGV = new GlobalVariable(*GV->getParent(),
830                                             GlobalType, false,
831                                             GlobalValue::InternalLinkage,
832                                             UndefValue::get(GlobalType),
833                                             GV->getName()+".body",
834                                             GV,
835                                             GV->isThreadLocal());
836
837  // If there are bitcast users of the malloc (which is typical, usually we have
838  // a malloc + bitcast) then replace them with uses of the new global.  Update
839  // other users to use the global as well.
840  BitCastInst *TheBC = 0;
841  while (!CI->use_empty()) {
842    Instruction *User = cast<Instruction>(CI->use_back());
843    if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
844      if (BCI->getType() == NewGV->getType()) {
845        BCI->replaceAllUsesWith(NewGV);
846        BCI->eraseFromParent();
847      } else {
848        BCI->setOperand(0, NewGV);
849      }
850    } else {
851      if (TheBC == 0)
852        TheBC = new BitCastInst(NewGV, CI->getType(), "newgv", CI);
853      User->replaceUsesOfWith(CI, TheBC);
854    }
855  }
856
857  Constant *RepValue = NewGV;
858  if (NewGV->getType() != GV->getType()->getElementType())
859    RepValue = ConstantExpr::getBitCast(RepValue,
860                                        GV->getType()->getElementType());
861
862  // If there is a comparison against null, we will insert a global bool to
863  // keep track of whether the global was initialized yet or not.
864  GlobalVariable *InitBool =
865    new GlobalVariable(Type::getInt1Ty(GV->getContext()), false,
866                       GlobalValue::InternalLinkage,
867                       ConstantInt::getFalse(GV->getContext()),
868                       GV->getName()+".init", GV->isThreadLocal());
869  bool InitBoolUsed = false;
870
871  // Loop over all uses of GV, processing them in turn.
872  while (!GV->use_empty()) {
873    if (StoreInst *SI = dyn_cast<StoreInst>(GV->use_back())) {
874      // The global is initialized when the store to it occurs.
875      new StoreInst(ConstantInt::getTrue(GV->getContext()), InitBool, SI);
876      SI->eraseFromParent();
877      continue;
878    }
879
880    LoadInst *LI = cast<LoadInst>(GV->use_back());
881    while (!LI->use_empty()) {
882      Use &LoadUse = LI->use_begin().getUse();
883      if (!isa<ICmpInst>(LoadUse.getUser())) {
884        LoadUse = RepValue;
885        continue;
886      }
887
888      ICmpInst *ICI = cast<ICmpInst>(LoadUse.getUser());
889      // Replace the cmp X, 0 with a use of the bool value.
890      Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", ICI);
891      InitBoolUsed = true;
892      switch (ICI->getPredicate()) {
893      default: llvm_unreachable("Unknown ICmp Predicate!");
894      case ICmpInst::ICMP_ULT:
895      case ICmpInst::ICMP_SLT:   // X < null -> always false
896        LV = ConstantInt::getFalse(GV->getContext());
897        break;
898      case ICmpInst::ICMP_ULE:
899      case ICmpInst::ICMP_SLE:
900      case ICmpInst::ICMP_EQ:
901        LV = BinaryOperator::CreateNot(LV, "notinit", ICI);
902        break;
903      case ICmpInst::ICMP_NE:
904      case ICmpInst::ICMP_UGE:
905      case ICmpInst::ICMP_SGE:
906      case ICmpInst::ICMP_UGT:
907      case ICmpInst::ICMP_SGT:
908        break;  // no change.
909      }
910      ICI->replaceAllUsesWith(LV);
911      ICI->eraseFromParent();
912    }
913    LI->eraseFromParent();
914  }
915
916  // If the initialization boolean was used, insert it, otherwise delete it.
917  if (!InitBoolUsed) {
918    while (!InitBool->use_empty())  // Delete initializations
919      cast<StoreInst>(InitBool->use_back())->eraseFromParent();
920    delete InitBool;
921  } else
922    GV->getParent()->getGlobalList().insert(GV, InitBool);
923
924  // Now the GV is dead, nuke it and the malloc..
925  GV->eraseFromParent();
926  CI->eraseFromParent();
927
928  // To further other optimizations, loop over all users of NewGV and try to
929  // constant prop them.  This will promote GEP instructions with constant
930  // indices into GEP constant-exprs, which will allow global-opt to hack on it.
931  ConstantPropUsersOf(NewGV);
932  if (RepValue != NewGV)
933    ConstantPropUsersOf(RepValue);
934
935  return NewGV;
936}
937
938/// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking
939/// to make sure that there are no complex uses of V.  We permit simple things
940/// like dereferencing the pointer, but not storing through the address, unless
941/// it is to the specified global.
942static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Instruction *V,
943                                                      GlobalVariable *GV,
944                                              SmallPtrSet<PHINode*, 8> &PHIs) {
945  for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
946    Instruction *Inst = cast<Instruction>(*UI);
947
948    if (isa<LoadInst>(Inst) || isa<CmpInst>(Inst)) {
949      continue; // Fine, ignore.
950    }
951
952    if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
953      if (SI->getOperand(0) == V && SI->getOperand(1) != GV)
954        return false;  // Storing the pointer itself... bad.
955      continue; // Otherwise, storing through it, or storing into GV... fine.
956    }
957
958    if (isa<GetElementPtrInst>(Inst)) {
959      if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Inst, GV, PHIs))
960        return false;
961      continue;
962    }
963
964    if (PHINode *PN = dyn_cast<PHINode>(Inst)) {
965      // PHIs are ok if all uses are ok.  Don't infinitely recurse through PHI
966      // cycles.
967      if (PHIs.insert(PN))
968        if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(PN, GV, PHIs))
969          return false;
970      continue;
971    }
972
973    if (BitCastInst *BCI = dyn_cast<BitCastInst>(Inst)) {
974      if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(BCI, GV, PHIs))
975        return false;
976      continue;
977    }
978
979    return false;
980  }
981  return true;
982}
983
984/// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV
985/// somewhere.  Transform all uses of the allocation into loads from the
986/// global and uses of the resultant pointer.  Further, delete the store into
987/// GV.  This assumes that these value pass the
988/// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate.
989static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc,
990                                          GlobalVariable *GV) {
991  while (!Alloc->use_empty()) {
992    Instruction *U = cast<Instruction>(*Alloc->use_begin());
993    Instruction *InsertPt = U;
994    if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
995      // If this is the store of the allocation into the global, remove it.
996      if (SI->getOperand(1) == GV) {
997        SI->eraseFromParent();
998        continue;
999      }
1000    } else if (PHINode *PN = dyn_cast<PHINode>(U)) {
1001      // Insert the load in the corresponding predecessor, not right before the
1002      // PHI.
1003      InsertPt = PN->getIncomingBlock(Alloc->use_begin())->getTerminator();
1004    } else if (isa<BitCastInst>(U)) {
1005      // Must be bitcast between the malloc and store to initialize the global.
1006      ReplaceUsesOfMallocWithGlobal(U, GV);
1007      U->eraseFromParent();
1008      continue;
1009    } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
1010      // If this is a "GEP bitcast" and the user is a store to the global, then
1011      // just process it as a bitcast.
1012      if (GEPI->hasAllZeroIndices() && GEPI->hasOneUse())
1013        if (StoreInst *SI = dyn_cast<StoreInst>(GEPI->use_back()))
1014          if (SI->getOperand(1) == GV) {
1015            // Must be bitcast GEP between the malloc and store to initialize
1016            // the global.
1017            ReplaceUsesOfMallocWithGlobal(GEPI, GV);
1018            GEPI->eraseFromParent();
1019            continue;
1020          }
1021    }
1022
1023    // Insert a load from the global, and use it instead of the malloc.
1024    Value *NL = new LoadInst(GV, GV->getName()+".val", InsertPt);
1025    U->replaceUsesOfWith(Alloc, NL);
1026  }
1027}
1028
1029/// LoadUsesSimpleEnoughForHeapSRA - Verify that all uses of V (a load, or a phi
1030/// of a load) are simple enough to perform heap SRA on.  This permits GEP's
1031/// that index through the array and struct field, icmps of null, and PHIs.
1032static bool LoadUsesSimpleEnoughForHeapSRA(const Value *V,
1033                              SmallPtrSet<const PHINode*, 32> &LoadUsingPHIs,
1034                              SmallPtrSet<const PHINode*, 32> &LoadUsingPHIsPerLoad) {
1035  // We permit two users of the load: setcc comparing against the null
1036  // pointer, and a getelementptr of a specific form.
1037  for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
1038    const Instruction *User = cast<Instruction>(*UI);
1039
1040    // Comparison against null is ok.
1041    if (const ICmpInst *ICI = dyn_cast<ICmpInst>(User)) {
1042      if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
1043        return false;
1044      continue;
1045    }
1046
1047    // getelementptr is also ok, but only a simple form.
1048    if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1049      // Must index into the array and into the struct.
1050      if (GEPI->getNumOperands() < 3)
1051        return false;
1052
1053      // Otherwise the GEP is ok.
1054      continue;
1055    }
1056
1057    if (const PHINode *PN = dyn_cast<PHINode>(User)) {
1058      if (!LoadUsingPHIsPerLoad.insert(PN))
1059        // This means some phi nodes are dependent on each other.
1060        // Avoid infinite looping!
1061        return false;
1062      if (!LoadUsingPHIs.insert(PN))
1063        // If we have already analyzed this PHI, then it is safe.
1064        continue;
1065
1066      // Make sure all uses of the PHI are simple enough to transform.
1067      if (!LoadUsesSimpleEnoughForHeapSRA(PN,
1068                                          LoadUsingPHIs, LoadUsingPHIsPerLoad))
1069        return false;
1070
1071      continue;
1072    }
1073
1074    // Otherwise we don't know what this is, not ok.
1075    return false;
1076  }
1077
1078  return true;
1079}
1080
1081
1082/// AllGlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from
1083/// GV are simple enough to perform HeapSRA, return true.
1084static bool AllGlobalLoadUsesSimpleEnoughForHeapSRA(const GlobalVariable *GV,
1085                                                    Instruction *StoredVal) {
1086  SmallPtrSet<const PHINode*, 32> LoadUsingPHIs;
1087  SmallPtrSet<const PHINode*, 32> LoadUsingPHIsPerLoad;
1088  for (Value::const_use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;
1089       ++UI)
1090    if (const LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
1091      if (!LoadUsesSimpleEnoughForHeapSRA(LI, LoadUsingPHIs,
1092                                          LoadUsingPHIsPerLoad))
1093        return false;
1094      LoadUsingPHIsPerLoad.clear();
1095    }
1096
1097  // If we reach here, we know that all uses of the loads and transitive uses
1098  // (through PHI nodes) are simple enough to transform.  However, we don't know
1099  // that all inputs the to the PHI nodes are in the same equivalence sets.
1100  // Check to verify that all operands of the PHIs are either PHIS that can be
1101  // transformed, loads from GV, or MI itself.
1102  for (SmallPtrSet<const PHINode*, 32>::const_iterator I = LoadUsingPHIs.begin(),
1103       E = LoadUsingPHIs.end(); I != E; ++I) {
1104    const PHINode *PN = *I;
1105    for (unsigned op = 0, e = PN->getNumIncomingValues(); op != e; ++op) {
1106      Value *InVal = PN->getIncomingValue(op);
1107
1108      // PHI of the stored value itself is ok.
1109      if (InVal == StoredVal) continue;
1110
1111      if (const PHINode *InPN = dyn_cast<PHINode>(InVal)) {
1112        // One of the PHIs in our set is (optimistically) ok.
1113        if (LoadUsingPHIs.count(InPN))
1114          continue;
1115        return false;
1116      }
1117
1118      // Load from GV is ok.
1119      if (const LoadInst *LI = dyn_cast<LoadInst>(InVal))
1120        if (LI->getOperand(0) == GV)
1121          continue;
1122
1123      // UNDEF? NULL?
1124
1125      // Anything else is rejected.
1126      return false;
1127    }
1128  }
1129
1130  return true;
1131}
1132
1133static Value *GetHeapSROAValue(Value *V, unsigned FieldNo,
1134               DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1135                   std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1136  std::vector<Value*> &FieldVals = InsertedScalarizedValues[V];
1137
1138  if (FieldNo >= FieldVals.size())
1139    FieldVals.resize(FieldNo+1);
1140
1141  // If we already have this value, just reuse the previously scalarized
1142  // version.
1143  if (Value *FieldVal = FieldVals[FieldNo])
1144    return FieldVal;
1145
1146  // Depending on what instruction this is, we have several cases.
1147  Value *Result;
1148  if (LoadInst *LI = dyn_cast<LoadInst>(V)) {
1149    // This is a scalarized version of the load from the global.  Just create
1150    // a new Load of the scalarized global.
1151    Result = new LoadInst(GetHeapSROAValue(LI->getOperand(0), FieldNo,
1152                                           InsertedScalarizedValues,
1153                                           PHIsToRewrite),
1154                          LI->getName()+".f"+Twine(FieldNo), LI);
1155  } else if (PHINode *PN = dyn_cast<PHINode>(V)) {
1156    // PN's type is pointer to struct.  Make a new PHI of pointer to struct
1157    // field.
1158    const StructType *ST =
1159      cast<StructType>(cast<PointerType>(PN->getType())->getElementType());
1160
1161    Result =
1162     PHINode::Create(PointerType::getUnqual(ST->getElementType(FieldNo)),
1163                     PN->getName()+".f"+Twine(FieldNo), PN);
1164    PHIsToRewrite.push_back(std::make_pair(PN, FieldNo));
1165  } else {
1166    llvm_unreachable("Unknown usable value");
1167    Result = 0;
1168  }
1169
1170  return FieldVals[FieldNo] = Result;
1171}
1172
1173/// RewriteHeapSROALoadUser - Given a load instruction and a value derived from
1174/// the load, rewrite the derived value to use the HeapSRoA'd load.
1175static void RewriteHeapSROALoadUser(Instruction *LoadUser,
1176             DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1177                   std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1178  // If this is a comparison against null, handle it.
1179  if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) {
1180    assert(isa<ConstantPointerNull>(SCI->getOperand(1)));
1181    // If we have a setcc of the loaded pointer, we can use a setcc of any
1182    // field.
1183    Value *NPtr = GetHeapSROAValue(SCI->getOperand(0), 0,
1184                                   InsertedScalarizedValues, PHIsToRewrite);
1185
1186    Value *New = new ICmpInst(SCI, SCI->getPredicate(), NPtr,
1187                              Constant::getNullValue(NPtr->getType()),
1188                              SCI->getName());
1189    SCI->replaceAllUsesWith(New);
1190    SCI->eraseFromParent();
1191    return;
1192  }
1193
1194  // Handle 'getelementptr Ptr, Idx, i32 FieldNo ...'
1195  if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) {
1196    assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2))
1197           && "Unexpected GEPI!");
1198
1199    // Load the pointer for this field.
1200    unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
1201    Value *NewPtr = GetHeapSROAValue(GEPI->getOperand(0), FieldNo,
1202                                     InsertedScalarizedValues, PHIsToRewrite);
1203
1204    // Create the new GEP idx vector.
1205    SmallVector<Value*, 8> GEPIdx;
1206    GEPIdx.push_back(GEPI->getOperand(1));
1207    GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end());
1208
1209    Value *NGEPI = GetElementPtrInst::Create(NewPtr,
1210                                             GEPIdx.begin(), GEPIdx.end(),
1211                                             GEPI->getName(), GEPI);
1212    GEPI->replaceAllUsesWith(NGEPI);
1213    GEPI->eraseFromParent();
1214    return;
1215  }
1216
1217  // Recursively transform the users of PHI nodes.  This will lazily create the
1218  // PHIs that are needed for individual elements.  Keep track of what PHIs we
1219  // see in InsertedScalarizedValues so that we don't get infinite loops (very
1220  // antisocial).  If the PHI is already in InsertedScalarizedValues, it has
1221  // already been seen first by another load, so its uses have already been
1222  // processed.
1223  PHINode *PN = cast<PHINode>(LoadUser);
1224  bool Inserted;
1225  DenseMap<Value*, std::vector<Value*> >::iterator InsertPos;
1226  tie(InsertPos, Inserted) =
1227    InsertedScalarizedValues.insert(std::make_pair(PN, std::vector<Value*>()));
1228  if (!Inserted) return;
1229
1230  // If this is the first time we've seen this PHI, recursively process all
1231  // users.
1232  for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end(); UI != E; ) {
1233    Instruction *User = cast<Instruction>(*UI++);
1234    RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite);
1235  }
1236}
1237
1238/// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global.  Ptr
1239/// is a value loaded from the global.  Eliminate all uses of Ptr, making them
1240/// use FieldGlobals instead.  All uses of loaded values satisfy
1241/// AllGlobalLoadUsesSimpleEnoughForHeapSRA.
1242static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load,
1243               DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1244                   std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1245  for (Value::use_iterator UI = Load->use_begin(), E = Load->use_end();
1246       UI != E; ) {
1247    Instruction *User = cast<Instruction>(*UI++);
1248    RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite);
1249  }
1250
1251  if (Load->use_empty()) {
1252    Load->eraseFromParent();
1253    InsertedScalarizedValues.erase(Load);
1254  }
1255}
1256
1257/// PerformHeapAllocSRoA - CI is an allocation of an array of structures.  Break
1258/// it up into multiple allocations of arrays of the fields.
1259static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, CallInst *CI,
1260                                            Value* NElems, TargetData *TD) {
1261  DEBUG(dbgs() << "SROA HEAP ALLOC: " << *GV << "  MALLOC = " << *CI << '\n');
1262  const Type* MAT = getMallocAllocatedType(CI);
1263  const StructType *STy = cast<StructType>(MAT);
1264
1265  // There is guaranteed to be at least one use of the malloc (storing
1266  // it into GV).  If there are other uses, change them to be uses of
1267  // the global to simplify later code.  This also deletes the store
1268  // into GV.
1269  ReplaceUsesOfMallocWithGlobal(CI, GV);
1270
1271  // Okay, at this point, there are no users of the malloc.  Insert N
1272  // new mallocs at the same place as CI, and N globals.
1273  std::vector<Value*> FieldGlobals;
1274  std::vector<Value*> FieldMallocs;
1275
1276  for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){
1277    const Type *FieldTy = STy->getElementType(FieldNo);
1278    const PointerType *PFieldTy = PointerType::getUnqual(FieldTy);
1279
1280    GlobalVariable *NGV =
1281      new GlobalVariable(*GV->getParent(),
1282                         PFieldTy, false, GlobalValue::InternalLinkage,
1283                         Constant::getNullValue(PFieldTy),
1284                         GV->getName() + ".f" + Twine(FieldNo), GV,
1285                         GV->isThreadLocal());
1286    FieldGlobals.push_back(NGV);
1287
1288    unsigned TypeSize = TD->getTypeAllocSize(FieldTy);
1289    if (const StructType *ST = dyn_cast<StructType>(FieldTy))
1290      TypeSize = TD->getStructLayout(ST)->getSizeInBytes();
1291    const Type *IntPtrTy = TD->getIntPtrType(CI->getContext());
1292    Value *NMI = CallInst::CreateMalloc(CI, IntPtrTy, FieldTy,
1293                                        ConstantInt::get(IntPtrTy, TypeSize),
1294                                        NElems,
1295                                        CI->getName() + ".f" + Twine(FieldNo));
1296    FieldMallocs.push_back(NMI);
1297    new StoreInst(NMI, NGV, CI);
1298  }
1299
1300  // The tricky aspect of this transformation is handling the case when malloc
1301  // fails.  In the original code, malloc failing would set the result pointer
1302  // of malloc to null.  In this case, some mallocs could succeed and others
1303  // could fail.  As such, we emit code that looks like this:
1304  //    F0 = malloc(field0)
1305  //    F1 = malloc(field1)
1306  //    F2 = malloc(field2)
1307  //    if (F0 == 0 || F1 == 0 || F2 == 0) {
1308  //      if (F0) { free(F0); F0 = 0; }
1309  //      if (F1) { free(F1); F1 = 0; }
1310  //      if (F2) { free(F2); F2 = 0; }
1311  //    }
1312  // The malloc can also fail if its argument is too large.
1313  Constant *ConstantZero = ConstantInt::get(CI->getOperand(1)->getType(), 0);
1314  Value *RunningOr = new ICmpInst(CI, ICmpInst::ICMP_SLT, CI->getOperand(1),
1315                                  ConstantZero, "isneg");
1316  for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) {
1317    Value *Cond = new ICmpInst(CI, ICmpInst::ICMP_EQ, FieldMallocs[i],
1318                             Constant::getNullValue(FieldMallocs[i]->getType()),
1319                               "isnull");
1320    RunningOr = BinaryOperator::CreateOr(RunningOr, Cond, "tmp", CI);
1321  }
1322
1323  // Split the basic block at the old malloc.
1324  BasicBlock *OrigBB = CI->getParent();
1325  BasicBlock *ContBB = OrigBB->splitBasicBlock(CI, "malloc_cont");
1326
1327  // Create the block to check the first condition.  Put all these blocks at the
1328  // end of the function as they are unlikely to be executed.
1329  BasicBlock *NullPtrBlock = BasicBlock::Create(OrigBB->getContext(),
1330                                                "malloc_ret_null",
1331                                                OrigBB->getParent());
1332
1333  // Remove the uncond branch from OrigBB to ContBB, turning it into a cond
1334  // branch on RunningOr.
1335  OrigBB->getTerminator()->eraseFromParent();
1336  BranchInst::Create(NullPtrBlock, ContBB, RunningOr, OrigBB);
1337
1338  // Within the NullPtrBlock, we need to emit a comparison and branch for each
1339  // pointer, because some may be null while others are not.
1340  for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1341    Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock);
1342    Value *Cmp = new ICmpInst(*NullPtrBlock, ICmpInst::ICMP_NE, GVVal,
1343                              Constant::getNullValue(GVVal->getType()),
1344                              "tmp");
1345    BasicBlock *FreeBlock = BasicBlock::Create(Cmp->getContext(), "free_it",
1346                                               OrigBB->getParent());
1347    BasicBlock *NextBlock = BasicBlock::Create(Cmp->getContext(), "next",
1348                                               OrigBB->getParent());
1349    Instruction *BI = BranchInst::Create(FreeBlock, NextBlock,
1350                                         Cmp, NullPtrBlock);
1351
1352    // Fill in FreeBlock.
1353    CallInst::CreateFree(GVVal, BI);
1354    new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i],
1355                  FreeBlock);
1356    BranchInst::Create(NextBlock, FreeBlock);
1357
1358    NullPtrBlock = NextBlock;
1359  }
1360
1361  BranchInst::Create(ContBB, NullPtrBlock);
1362
1363  // CI is no longer needed, remove it.
1364  CI->eraseFromParent();
1365
1366  /// InsertedScalarizedLoads - As we process loads, if we can't immediately
1367  /// update all uses of the load, keep track of what scalarized loads are
1368  /// inserted for a given load.
1369  DenseMap<Value*, std::vector<Value*> > InsertedScalarizedValues;
1370  InsertedScalarizedValues[GV] = FieldGlobals;
1371
1372  std::vector<std::pair<PHINode*, unsigned> > PHIsToRewrite;
1373
1374  // Okay, the malloc site is completely handled.  All of the uses of GV are now
1375  // loads, and all uses of those loads are simple.  Rewrite them to use loads
1376  // of the per-field globals instead.
1377  for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;) {
1378    Instruction *User = cast<Instruction>(*UI++);
1379
1380    if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1381      RewriteUsesOfLoadForHeapSRoA(LI, InsertedScalarizedValues, PHIsToRewrite);
1382      continue;
1383    }
1384
1385    // Must be a store of null.
1386    StoreInst *SI = cast<StoreInst>(User);
1387    assert(isa<ConstantPointerNull>(SI->getOperand(0)) &&
1388           "Unexpected heap-sra user!");
1389
1390    // Insert a store of null into each global.
1391    for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1392      const PointerType *PT = cast<PointerType>(FieldGlobals[i]->getType());
1393      Constant *Null = Constant::getNullValue(PT->getElementType());
1394      new StoreInst(Null, FieldGlobals[i], SI);
1395    }
1396    // Erase the original store.
1397    SI->eraseFromParent();
1398  }
1399
1400  // While we have PHIs that are interesting to rewrite, do it.
1401  while (!PHIsToRewrite.empty()) {
1402    PHINode *PN = PHIsToRewrite.back().first;
1403    unsigned FieldNo = PHIsToRewrite.back().second;
1404    PHIsToRewrite.pop_back();
1405    PHINode *FieldPN = cast<PHINode>(InsertedScalarizedValues[PN][FieldNo]);
1406    assert(FieldPN->getNumIncomingValues() == 0 &&"Already processed this phi");
1407
1408    // Add all the incoming values.  This can materialize more phis.
1409    for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1410      Value *InVal = PN->getIncomingValue(i);
1411      InVal = GetHeapSROAValue(InVal, FieldNo, InsertedScalarizedValues,
1412                               PHIsToRewrite);
1413      FieldPN->addIncoming(InVal, PN->getIncomingBlock(i));
1414    }
1415  }
1416
1417  // Drop all inter-phi links and any loads that made it this far.
1418  for (DenseMap<Value*, std::vector<Value*> >::iterator
1419       I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1420       I != E; ++I) {
1421    if (PHINode *PN = dyn_cast<PHINode>(I->first))
1422      PN->dropAllReferences();
1423    else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1424      LI->dropAllReferences();
1425  }
1426
1427  // Delete all the phis and loads now that inter-references are dead.
1428  for (DenseMap<Value*, std::vector<Value*> >::iterator
1429       I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1430       I != E; ++I) {
1431    if (PHINode *PN = dyn_cast<PHINode>(I->first))
1432      PN->eraseFromParent();
1433    else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1434      LI->eraseFromParent();
1435  }
1436
1437  // The old global is now dead, remove it.
1438  GV->eraseFromParent();
1439
1440  ++NumHeapSRA;
1441  return cast<GlobalVariable>(FieldGlobals[0]);
1442}
1443
1444/// TryToOptimizeStoreOfMallocToGlobal - This function is called when we see a
1445/// pointer global variable with a single value stored it that is a malloc or
1446/// cast of malloc.
1447static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV,
1448                                               CallInst *CI,
1449                                               const Type *AllocTy,
1450                                               Module::global_iterator &GVI,
1451                                               TargetData *TD) {
1452  // If this is a malloc of an abstract type, don't touch it.
1453  if (!AllocTy->isSized())
1454    return false;
1455
1456  // We can't optimize this global unless all uses of it are *known* to be
1457  // of the malloc value, not of the null initializer value (consider a use
1458  // that compares the global's value against zero to see if the malloc has
1459  // been reached).  To do this, we check to see if all uses of the global
1460  // would trap if the global were null: this proves that they must all
1461  // happen after the malloc.
1462  if (!AllUsesOfLoadedValueWillTrapIfNull(GV))
1463    return false;
1464
1465  // We can't optimize this if the malloc itself is used in a complex way,
1466  // for example, being stored into multiple globals.  This allows the
1467  // malloc to be stored into the specified global, loaded setcc'd, and
1468  // GEP'd.  These are all things we could transform to using the global
1469  // for.
1470  {
1471    SmallPtrSet<PHINode*, 8> PHIs;
1472    if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(CI, GV, PHIs))
1473      return false;
1474  }
1475
1476  // If we have a global that is only initialized with a fixed size malloc,
1477  // transform the program to use global memory instead of malloc'd memory.
1478  // This eliminates dynamic allocation, avoids an indirection accessing the
1479  // data, and exposes the resultant global to further GlobalOpt.
1480  // We cannot optimize the malloc if we cannot determine malloc array size.
1481  if (Value *NElems = getMallocArraySize(CI, TD, true)) {
1482    if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems))
1483      // Restrict this transformation to only working on small allocations
1484      // (2048 bytes currently), as we don't want to introduce a 16M global or
1485      // something.
1486      if (TD &&
1487          NElements->getZExtValue() * TD->getTypeAllocSize(AllocTy) < 2048) {
1488        GVI = OptimizeGlobalAddressOfMalloc(GV, CI, AllocTy, NElements, TD);
1489        return true;
1490      }
1491
1492    // If the allocation is an array of structures, consider transforming this
1493    // into multiple malloc'd arrays, one for each field.  This is basically
1494    // SRoA for malloc'd memory.
1495
1496    // If this is an allocation of a fixed size array of structs, analyze as a
1497    // variable size array.  malloc [100 x struct],1 -> malloc struct, 100
1498    if (NElems == ConstantInt::get(CI->getOperand(1)->getType(), 1))
1499      if (const ArrayType *AT = dyn_cast<ArrayType>(AllocTy))
1500        AllocTy = AT->getElementType();
1501
1502    if (const StructType *AllocSTy = dyn_cast<StructType>(AllocTy)) {
1503      // This the structure has an unreasonable number of fields, leave it
1504      // alone.
1505      if (AllocSTy->getNumElements() <= 16 && AllocSTy->getNumElements() != 0 &&
1506          AllGlobalLoadUsesSimpleEnoughForHeapSRA(GV, CI)) {
1507
1508        // If this is a fixed size array, transform the Malloc to be an alloc of
1509        // structs.  malloc [100 x struct],1 -> malloc struct, 100
1510        if (const ArrayType *AT =
1511                              dyn_cast<ArrayType>(getMallocAllocatedType(CI))) {
1512          const Type *IntPtrTy = TD->getIntPtrType(CI->getContext());
1513          unsigned TypeSize = TD->getStructLayout(AllocSTy)->getSizeInBytes();
1514          Value *AllocSize = ConstantInt::get(IntPtrTy, TypeSize);
1515          Value *NumElements = ConstantInt::get(IntPtrTy, AT->getNumElements());
1516          Instruction *Malloc = CallInst::CreateMalloc(CI, IntPtrTy, AllocSTy,
1517                                                       AllocSize, NumElements,
1518                                                       CI->getName());
1519          Instruction *Cast = new BitCastInst(Malloc, CI->getType(), "tmp", CI);
1520          CI->replaceAllUsesWith(Cast);
1521          CI->eraseFromParent();
1522          CI = dyn_cast<BitCastInst>(Malloc) ?
1523               extractMallocCallFromBitCast(Malloc) : cast<CallInst>(Malloc);
1524        }
1525
1526        GVI = PerformHeapAllocSRoA(GV, CI, getMallocArraySize(CI, TD, true),TD);
1527        return true;
1528      }
1529    }
1530  }
1531
1532  return false;
1533}
1534
1535// OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge
1536// that only one value (besides its initializer) is ever stored to the global.
1537static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal,
1538                                     Module::global_iterator &GVI,
1539                                     TargetData *TD) {
1540  // Ignore no-op GEPs and bitcasts.
1541  StoredOnceVal = StoredOnceVal->stripPointerCasts();
1542
1543  // If we are dealing with a pointer global that is initialized to null and
1544  // only has one (non-null) value stored into it, then we can optimize any
1545  // users of the loaded value (often calls and loads) that would trap if the
1546  // value was null.
1547  if (GV->getInitializer()->getType()->isPointerTy() &&
1548      GV->getInitializer()->isNullValue()) {
1549    if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) {
1550      if (GV->getInitializer()->getType() != SOVC->getType())
1551        SOVC =
1552         ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType());
1553
1554      // Optimize away any trapping uses of the loaded value.
1555      if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC))
1556        return true;
1557    } else if (CallInst *CI = extractMallocCall(StoredOnceVal)) {
1558      const Type* MallocType = getMallocAllocatedType(CI);
1559      if (MallocType && TryToOptimizeStoreOfMallocToGlobal(GV, CI, MallocType,
1560                                                           GVI, TD))
1561        return true;
1562    }
1563  }
1564
1565  return false;
1566}
1567
1568/// TryToShrinkGlobalToBoolean - At this point, we have learned that the only
1569/// two values ever stored into GV are its initializer and OtherVal.  See if we
1570/// can shrink the global into a boolean and select between the two values
1571/// whenever it is used.  This exposes the values to other scalar optimizations.
1572static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) {
1573  const Type *GVElType = GV->getType()->getElementType();
1574
1575  // If GVElType is already i1, it is already shrunk.  If the type of the GV is
1576  // an FP value, pointer or vector, don't do this optimization because a select
1577  // between them is very expensive and unlikely to lead to later
1578  // simplification.  In these cases, we typically end up with "cond ? v1 : v2"
1579  // where v1 and v2 both require constant pool loads, a big loss.
1580  if (GVElType == Type::getInt1Ty(GV->getContext()) ||
1581      GVElType->isFloatingPointTy() ||
1582      GVElType->isPointerTy() || GVElType->isVectorTy())
1583    return false;
1584
1585  // Walk the use list of the global seeing if all the uses are load or store.
1586  // If there is anything else, bail out.
1587  for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I)
1588    if (!isa<LoadInst>(I) && !isa<StoreInst>(I))
1589      return false;
1590
1591  DEBUG(dbgs() << "   *** SHRINKING TO BOOL: " << *GV);
1592
1593  // Create the new global, initializing it to false.
1594  GlobalVariable *NewGV = new GlobalVariable(Type::getInt1Ty(GV->getContext()),
1595                                             false,
1596                                             GlobalValue::InternalLinkage,
1597                                        ConstantInt::getFalse(GV->getContext()),
1598                                             GV->getName()+".b",
1599                                             GV->isThreadLocal());
1600  GV->getParent()->getGlobalList().insert(GV, NewGV);
1601
1602  Constant *InitVal = GV->getInitializer();
1603  assert(InitVal->getType() != Type::getInt1Ty(GV->getContext()) &&
1604         "No reason to shrink to bool!");
1605
1606  // If initialized to zero and storing one into the global, we can use a cast
1607  // instead of a select to synthesize the desired value.
1608  bool IsOneZero = false;
1609  if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal))
1610    IsOneZero = InitVal->isNullValue() && CI->isOne();
1611
1612  while (!GV->use_empty()) {
1613    Instruction *UI = cast<Instruction>(GV->use_back());
1614    if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
1615      // Change the store into a boolean store.
1616      bool StoringOther = SI->getOperand(0) == OtherVal;
1617      // Only do this if we weren't storing a loaded value.
1618      Value *StoreVal;
1619      if (StoringOther || SI->getOperand(0) == InitVal)
1620        StoreVal = ConstantInt::get(Type::getInt1Ty(GV->getContext()),
1621                                    StoringOther);
1622      else {
1623        // Otherwise, we are storing a previously loaded copy.  To do this,
1624        // change the copy from copying the original value to just copying the
1625        // bool.
1626        Instruction *StoredVal = cast<Instruction>(SI->getOperand(0));
1627
1628        // If we're already replaced the input, StoredVal will be a cast or
1629        // select instruction.  If not, it will be a load of the original
1630        // global.
1631        if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
1632          assert(LI->getOperand(0) == GV && "Not a copy!");
1633          // Insert a new load, to preserve the saved value.
1634          StoreVal = new LoadInst(NewGV, LI->getName()+".b", LI);
1635        } else {
1636          assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) &&
1637                 "This is not a form that we understand!");
1638          StoreVal = StoredVal->getOperand(0);
1639          assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!");
1640        }
1641      }
1642      new StoreInst(StoreVal, NewGV, SI);
1643    } else {
1644      // Change the load into a load of bool then a select.
1645      LoadInst *LI = cast<LoadInst>(UI);
1646      LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", LI);
1647      Value *NSI;
1648      if (IsOneZero)
1649        NSI = new ZExtInst(NLI, LI->getType(), "", LI);
1650      else
1651        NSI = SelectInst::Create(NLI, OtherVal, InitVal, "", LI);
1652      NSI->takeName(LI);
1653      LI->replaceAllUsesWith(NSI);
1654    }
1655    UI->eraseFromParent();
1656  }
1657
1658  GV->eraseFromParent();
1659  return true;
1660}
1661
1662
1663/// ProcessInternalGlobal - Analyze the specified global variable and optimize
1664/// it if possible.  If we make a change, return true.
1665bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV,
1666                                      Module::global_iterator &GVI) {
1667  SmallPtrSet<const PHINode*, 16> PHIUsers;
1668  GlobalStatus GS;
1669  GV->removeDeadConstantUsers();
1670
1671  if (GV->use_empty()) {
1672    DEBUG(dbgs() << "GLOBAL DEAD: " << *GV);
1673    GV->eraseFromParent();
1674    ++NumDeleted;
1675    return true;
1676  }
1677
1678  if (!AnalyzeGlobal(GV, GS, PHIUsers)) {
1679#if 0
1680    DEBUG(dbgs() << "Global: " << *GV);
1681    DEBUG(dbgs() << "  isLoaded = " << GS.isLoaded << "\n");
1682    DEBUG(dbgs() << "  StoredType = ");
1683    switch (GS.StoredType) {
1684    case GlobalStatus::NotStored: DEBUG(dbgs() << "NEVER STORED\n"); break;
1685    case GlobalStatus::isInitializerStored: DEBUG(dbgs() << "INIT STORED\n");
1686                                            break;
1687    case GlobalStatus::isStoredOnce: DEBUG(dbgs() << "STORED ONCE\n"); break;
1688    case GlobalStatus::isStored: DEBUG(dbgs() << "stored\n"); break;
1689    }
1690    if (GS.StoredType == GlobalStatus::isStoredOnce && GS.StoredOnceValue)
1691      DEBUG(dbgs() << "  StoredOnceValue = " << *GS.StoredOnceValue << "\n");
1692    if (GS.AccessingFunction && !GS.HasMultipleAccessingFunctions)
1693      DEBUG(dbgs() << "  AccessingFunction = " << GS.AccessingFunction->getName()
1694                  << "\n");
1695    DEBUG(dbgs() << "  HasMultipleAccessingFunctions =  "
1696                 << GS.HasMultipleAccessingFunctions << "\n");
1697    DEBUG(dbgs() << "  HasNonInstructionUser = "
1698                 << GS.HasNonInstructionUser<<"\n");
1699    DEBUG(dbgs() << "\n");
1700#endif
1701
1702    // If this is a first class global and has only one accessing function
1703    // and this function is main (which we know is not recursive we can make
1704    // this global a local variable) we replace the global with a local alloca
1705    // in this function.
1706    //
1707    // NOTE: It doesn't make sense to promote non single-value types since we
1708    // are just replacing static memory to stack memory.
1709    //
1710    // If the global is in different address space, don't bring it to stack.
1711    if (!GS.HasMultipleAccessingFunctions &&
1712        GS.AccessingFunction && !GS.HasNonInstructionUser &&
1713        GV->getType()->getElementType()->isSingleValueType() &&
1714        GS.AccessingFunction->getName() == "main" &&
1715        GS.AccessingFunction->hasExternalLinkage() &&
1716        GV->getType()->getAddressSpace() == 0) {
1717      DEBUG(dbgs() << "LOCALIZING GLOBAL: " << *GV);
1718      Instruction& FirstI = const_cast<Instruction&>(*GS.AccessingFunction
1719                                                     ->getEntryBlock().begin());
1720      const Type* ElemTy = GV->getType()->getElementType();
1721      // FIXME: Pass Global's alignment when globals have alignment
1722      AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), &FirstI);
1723      if (!isa<UndefValue>(GV->getInitializer()))
1724        new StoreInst(GV->getInitializer(), Alloca, &FirstI);
1725
1726      GV->replaceAllUsesWith(Alloca);
1727      GV->eraseFromParent();
1728      ++NumLocalized;
1729      return true;
1730    }
1731
1732    // If the global is never loaded (but may be stored to), it is dead.
1733    // Delete it now.
1734    if (!GS.isLoaded) {
1735      DEBUG(dbgs() << "GLOBAL NEVER LOADED: " << *GV);
1736
1737      // Delete any stores we can find to the global.  We may not be able to
1738      // make it completely dead though.
1739      bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer());
1740
1741      // If the global is dead now, delete it.
1742      if (GV->use_empty()) {
1743        GV->eraseFromParent();
1744        ++NumDeleted;
1745        Changed = true;
1746      }
1747      return Changed;
1748
1749    } else if (GS.StoredType <= GlobalStatus::isInitializerStored) {
1750      DEBUG(dbgs() << "MARKING CONSTANT: " << *GV);
1751      GV->setConstant(true);
1752
1753      // Clean up any obviously simplifiable users now.
1754      CleanupConstantGlobalUsers(GV, GV->getInitializer());
1755
1756      // If the global is dead now, just nuke it.
1757      if (GV->use_empty()) {
1758        DEBUG(dbgs() << "   *** Marking constant allowed us to simplify "
1759                     << "all users and delete global!\n");
1760        GV->eraseFromParent();
1761        ++NumDeleted;
1762      }
1763
1764      ++NumMarked;
1765      return true;
1766    } else if (!GV->getInitializer()->getType()->isSingleValueType()) {
1767      if (TargetData *TD = getAnalysisIfAvailable<TargetData>())
1768        if (GlobalVariable *FirstNewGV = SRAGlobal(GV, *TD)) {
1769          GVI = FirstNewGV;  // Don't skip the newly produced globals!
1770          return true;
1771        }
1772    } else if (GS.StoredType == GlobalStatus::isStoredOnce) {
1773      // If the initial value for the global was an undef value, and if only
1774      // one other value was stored into it, we can just change the
1775      // initializer to be the stored value, then delete all stores to the
1776      // global.  This allows us to mark it constant.
1777      if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1778        if (isa<UndefValue>(GV->getInitializer())) {
1779          // Change the initial value here.
1780          GV->setInitializer(SOVConstant);
1781
1782          // Clean up any obviously simplifiable users now.
1783          CleanupConstantGlobalUsers(GV, GV->getInitializer());
1784
1785          if (GV->use_empty()) {
1786            DEBUG(dbgs() << "   *** Substituting initializer allowed us to "
1787                         << "simplify all users and delete global!\n");
1788            GV->eraseFromParent();
1789            ++NumDeleted;
1790          } else {
1791            GVI = GV;
1792          }
1793          ++NumSubstitute;
1794          return true;
1795        }
1796
1797      // Try to optimize globals based on the knowledge that only one value
1798      // (besides its initializer) is ever stored to the global.
1799      if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI,
1800                                   getAnalysisIfAvailable<TargetData>()))
1801        return true;
1802
1803      // Otherwise, if the global was not a boolean, we can shrink it to be a
1804      // boolean.
1805      if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1806        if (TryToShrinkGlobalToBoolean(GV, SOVConstant)) {
1807          ++NumShrunkToBool;
1808          return true;
1809        }
1810    }
1811  }
1812  return false;
1813}
1814
1815/// ChangeCalleesToFastCall - Walk all of the direct calls of the specified
1816/// function, changing them to FastCC.
1817static void ChangeCalleesToFastCall(Function *F) {
1818  for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1819    CallSite User(cast<Instruction>(*UI));
1820    User.setCallingConv(CallingConv::Fast);
1821  }
1822}
1823
1824static AttrListPtr StripNest(const AttrListPtr &Attrs) {
1825  for (unsigned i = 0, e = Attrs.getNumSlots(); i != e; ++i) {
1826    if ((Attrs.getSlot(i).Attrs & Attribute::Nest) == 0)
1827      continue;
1828
1829    // There can be only one.
1830    return Attrs.removeAttr(Attrs.getSlot(i).Index, Attribute::Nest);
1831  }
1832
1833  return Attrs;
1834}
1835
1836static void RemoveNestAttribute(Function *F) {
1837  F->setAttributes(StripNest(F->getAttributes()));
1838  for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1839    CallSite User(cast<Instruction>(*UI));
1840    User.setAttributes(StripNest(User.getAttributes()));
1841  }
1842}
1843
1844bool GlobalOpt::OptimizeFunctions(Module &M) {
1845  bool Changed = false;
1846  // Optimize functions.
1847  for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) {
1848    Function *F = FI++;
1849    // Functions without names cannot be referenced outside this module.
1850    if (!F->hasName() && !F->isDeclaration())
1851      F->setLinkage(GlobalValue::InternalLinkage);
1852    F->removeDeadConstantUsers();
1853    if (F->use_empty() && (F->hasLocalLinkage() || F->hasLinkOnceLinkage())) {
1854      F->eraseFromParent();
1855      Changed = true;
1856      ++NumFnDeleted;
1857    } else if (F->hasLocalLinkage()) {
1858      if (F->getCallingConv() == CallingConv::C && !F->isVarArg() &&
1859          !F->hasAddressTaken()) {
1860        // If this function has C calling conventions, is not a varargs
1861        // function, and is only called directly, promote it to use the Fast
1862        // calling convention.
1863        F->setCallingConv(CallingConv::Fast);
1864        ChangeCalleesToFastCall(F);
1865        ++NumFastCallFns;
1866        Changed = true;
1867      }
1868
1869      if (F->getAttributes().hasAttrSomewhere(Attribute::Nest) &&
1870          !F->hasAddressTaken()) {
1871        // The function is not used by a trampoline intrinsic, so it is safe
1872        // to remove the 'nest' attribute.
1873        RemoveNestAttribute(F);
1874        ++NumNestRemoved;
1875        Changed = true;
1876      }
1877    }
1878  }
1879  return Changed;
1880}
1881
1882bool GlobalOpt::OptimizeGlobalVars(Module &M) {
1883  bool Changed = false;
1884  for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
1885       GVI != E; ) {
1886    GlobalVariable *GV = GVI++;
1887    // Global variables without names cannot be referenced outside this module.
1888    if (!GV->hasName() && !GV->isDeclaration())
1889      GV->setLinkage(GlobalValue::InternalLinkage);
1890    // Simplify the initializer.
1891    if (GV->hasInitializer())
1892      if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GV->getInitializer())) {
1893        TargetData *TD = getAnalysisIfAvailable<TargetData>();
1894        Constant *New = ConstantFoldConstantExpression(CE, TD);
1895        if (New && New != CE)
1896          GV->setInitializer(New);
1897      }
1898    // Do more involved optimizations if the global is internal.
1899    if (!GV->isConstant() && GV->hasLocalLinkage() &&
1900        GV->hasInitializer())
1901      Changed |= ProcessInternalGlobal(GV, GVI);
1902  }
1903  return Changed;
1904}
1905
1906/// FindGlobalCtors - Find the llvm.globalctors list, verifying that all
1907/// initializers have an init priority of 65535.
1908GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) {
1909  for (Module::global_iterator I = M.global_begin(), E = M.global_end();
1910       I != E; ++I)
1911    if (I->getName() == "llvm.global_ctors") {
1912      // Found it, verify it's an array of { int, void()* }.
1913      const ArrayType *ATy =dyn_cast<ArrayType>(I->getType()->getElementType());
1914      if (!ATy) return 0;
1915      const StructType *STy = dyn_cast<StructType>(ATy->getElementType());
1916      if (!STy || STy->getNumElements() != 2 ||
1917          !STy->getElementType(0)->isIntegerTy(32)) return 0;
1918      const PointerType *PFTy = dyn_cast<PointerType>(STy->getElementType(1));
1919      if (!PFTy) return 0;
1920      const FunctionType *FTy = dyn_cast<FunctionType>(PFTy->getElementType());
1921      if (!FTy || !FTy->getReturnType()->isVoidTy() ||
1922          FTy->isVarArg() || FTy->getNumParams() != 0)
1923        return 0;
1924
1925      // Verify that the initializer is simple enough for us to handle.
1926      if (!I->hasDefinitiveInitializer()) return 0;
1927      ConstantArray *CA = dyn_cast<ConstantArray>(I->getInitializer());
1928      if (!CA) return 0;
1929      for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i)
1930        if (ConstantStruct *CS = dyn_cast<ConstantStruct>(*i)) {
1931          if (isa<ConstantPointerNull>(CS->getOperand(1)))
1932            continue;
1933
1934          // Must have a function or null ptr.
1935          if (!isa<Function>(CS->getOperand(1)))
1936            return 0;
1937
1938          // Init priority must be standard.
1939          ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0));
1940          if (!CI || CI->getZExtValue() != 65535)
1941            return 0;
1942        } else {
1943          return 0;
1944        }
1945
1946      return I;
1947    }
1948  return 0;
1949}
1950
1951/// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand,
1952/// return a list of the functions and null terminator as a vector.
1953static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) {
1954  ConstantArray *CA = cast<ConstantArray>(GV->getInitializer());
1955  std::vector<Function*> Result;
1956  Result.reserve(CA->getNumOperands());
1957  for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) {
1958    ConstantStruct *CS = cast<ConstantStruct>(*i);
1959    Result.push_back(dyn_cast<Function>(CS->getOperand(1)));
1960  }
1961  return Result;
1962}
1963
1964/// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the
1965/// specified array, returning the new global to use.
1966static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL,
1967                                          const std::vector<Function*> &Ctors) {
1968  // If we made a change, reassemble the initializer list.
1969  std::vector<Constant*> CSVals;
1970  CSVals.push_back(ConstantInt::get(Type::getInt32Ty(GCL->getContext()),65535));
1971  CSVals.push_back(0);
1972
1973  // Create the new init list.
1974  std::vector<Constant*> CAList;
1975  for (unsigned i = 0, e = Ctors.size(); i != e; ++i) {
1976    if (Ctors[i]) {
1977      CSVals[1] = Ctors[i];
1978    } else {
1979      const Type *FTy = FunctionType::get(Type::getVoidTy(GCL->getContext()),
1980                                          false);
1981      const PointerType *PFTy = PointerType::getUnqual(FTy);
1982      CSVals[1] = Constant::getNullValue(PFTy);
1983      CSVals[0] = ConstantInt::get(Type::getInt32Ty(GCL->getContext()),
1984                                   2147483647);
1985    }
1986    CAList.push_back(ConstantStruct::get(GCL->getContext(), CSVals, false));
1987  }
1988
1989  // Create the array initializer.
1990  const Type *StructTy =
1991      cast<ArrayType>(GCL->getType()->getElementType())->getElementType();
1992  Constant *CA = ConstantArray::get(ArrayType::get(StructTy,
1993                                                   CAList.size()), CAList);
1994
1995  // If we didn't change the number of elements, don't create a new GV.
1996  if (CA->getType() == GCL->getInitializer()->getType()) {
1997    GCL->setInitializer(CA);
1998    return GCL;
1999  }
2000
2001  // Create the new global and insert it next to the existing list.
2002  GlobalVariable *NGV = new GlobalVariable(CA->getType(), GCL->isConstant(),
2003                                           GCL->getLinkage(), CA, "",
2004                                           GCL->isThreadLocal());
2005  GCL->getParent()->getGlobalList().insert(GCL, NGV);
2006  NGV->takeName(GCL);
2007
2008  // Nuke the old list, replacing any uses with the new one.
2009  if (!GCL->use_empty()) {
2010    Constant *V = NGV;
2011    if (V->getType() != GCL->getType())
2012      V = ConstantExpr::getBitCast(V, GCL->getType());
2013    GCL->replaceAllUsesWith(V);
2014  }
2015  GCL->eraseFromParent();
2016
2017  if (Ctors.size())
2018    return NGV;
2019  else
2020    return 0;
2021}
2022
2023
2024static Constant *getVal(DenseMap<Value*, Constant*> &ComputedValues,
2025                        Value *V) {
2026  if (Constant *CV = dyn_cast<Constant>(V)) return CV;
2027  Constant *R = ComputedValues[V];
2028  assert(R && "Reference to an uncomputed value!");
2029  return R;
2030}
2031
2032/// isSimpleEnoughPointerToCommit - Return true if this constant is simple
2033/// enough for us to understand.  In particular, if it is a cast of something,
2034/// we punt.  We basically just support direct accesses to globals and GEP's of
2035/// globals.  This should be kept up to date with CommitValueTo.
2036static bool isSimpleEnoughPointerToCommit(Constant *C) {
2037  // Conservatively, avoid aggregate types. This is because we don't
2038  // want to worry about them partially overlapping other stores.
2039  if (!cast<PointerType>(C->getType())->getElementType()->isSingleValueType())
2040    return false;
2041
2042  if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
2043    // Do not allow weak/linkonce/dllimport/dllexport linkage or
2044    // external globals.
2045    return GV->hasDefinitiveInitializer();
2046
2047  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2048    // Handle a constantexpr gep.
2049    if (CE->getOpcode() == Instruction::GetElementPtr &&
2050        isa<GlobalVariable>(CE->getOperand(0)) &&
2051        cast<GEPOperator>(CE)->isInBounds()) {
2052      GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2053      // Do not allow weak/linkonce/dllimport/dllexport linkage or
2054      // external globals.
2055      if (!GV->hasDefinitiveInitializer())
2056        return false;
2057
2058      // The first index must be zero.
2059      ConstantInt *CI = dyn_cast<ConstantInt>(*next(CE->op_begin()));
2060      if (!CI || !CI->isZero()) return false;
2061
2062      // The remaining indices must be compile-time known integers within the
2063      // notional bounds of the corresponding static array types.
2064      if (!CE->isGEPWithNoNotionalOverIndexing())
2065        return false;
2066
2067      return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
2068    }
2069  return false;
2070}
2071
2072/// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global
2073/// initializer.  This returns 'Init' modified to reflect 'Val' stored into it.
2074/// At this point, the GEP operands of Addr [0, OpNo) have been stepped into.
2075static Constant *EvaluateStoreInto(Constant *Init, Constant *Val,
2076                                   ConstantExpr *Addr, unsigned OpNo) {
2077  // Base case of the recursion.
2078  if (OpNo == Addr->getNumOperands()) {
2079    assert(Val->getType() == Init->getType() && "Type mismatch!");
2080    return Val;
2081  }
2082
2083  std::vector<Constant*> Elts;
2084  if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
2085
2086    // Break up the constant into its elements.
2087    if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
2088      for (User::op_iterator i = CS->op_begin(), e = CS->op_end(); i != e; ++i)
2089        Elts.push_back(cast<Constant>(*i));
2090    } else if (isa<ConstantAggregateZero>(Init)) {
2091      for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
2092        Elts.push_back(Constant::getNullValue(STy->getElementType(i)));
2093    } else if (isa<UndefValue>(Init)) {
2094      for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
2095        Elts.push_back(UndefValue::get(STy->getElementType(i)));
2096    } else {
2097      llvm_unreachable("This code is out of sync with "
2098             " ConstantFoldLoadThroughGEPConstantExpr");
2099    }
2100
2101    // Replace the element that we are supposed to.
2102    ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo));
2103    unsigned Idx = CU->getZExtValue();
2104    assert(Idx < STy->getNumElements() && "Struct index out of range!");
2105    Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1);
2106
2107    // Return the modified struct.
2108    return ConstantStruct::get(Init->getContext(), &Elts[0], Elts.size(),
2109                               STy->isPacked());
2110  } else {
2111    ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo));
2112    const SequentialType *InitTy = cast<SequentialType>(Init->getType());
2113
2114    uint64_t NumElts;
2115    if (const ArrayType *ATy = dyn_cast<ArrayType>(InitTy))
2116      NumElts = ATy->getNumElements();
2117    else
2118      NumElts = cast<VectorType>(InitTy)->getNumElements();
2119
2120
2121    // Break up the array into elements.
2122    if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
2123      for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i)
2124        Elts.push_back(cast<Constant>(*i));
2125    } else if (ConstantVector *CV = dyn_cast<ConstantVector>(Init)) {
2126      for (User::op_iterator i = CV->op_begin(), e = CV->op_end(); i != e; ++i)
2127        Elts.push_back(cast<Constant>(*i));
2128    } else if (isa<ConstantAggregateZero>(Init)) {
2129      Elts.assign(NumElts, Constant::getNullValue(InitTy->getElementType()));
2130    } else {
2131      assert(isa<UndefValue>(Init) && "This code is out of sync with "
2132             " ConstantFoldLoadThroughGEPConstantExpr");
2133      Elts.assign(NumElts, UndefValue::get(InitTy->getElementType()));
2134    }
2135
2136    assert(CI->getZExtValue() < NumElts);
2137    Elts[CI->getZExtValue()] =
2138      EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1);
2139
2140    if (Init->getType()->isArrayTy())
2141      return ConstantArray::get(cast<ArrayType>(InitTy), Elts);
2142    else
2143      return ConstantVector::get(&Elts[0], Elts.size());
2144  }
2145}
2146
2147/// CommitValueTo - We have decided that Addr (which satisfies the predicate
2148/// isSimpleEnoughPointerToCommit) should get Val as its value.  Make it happen.
2149static void CommitValueTo(Constant *Val, Constant *Addr) {
2150  if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
2151    assert(GV->hasInitializer());
2152    GV->setInitializer(Val);
2153    return;
2154  }
2155
2156  ConstantExpr *CE = cast<ConstantExpr>(Addr);
2157  GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2158  GV->setInitializer(EvaluateStoreInto(GV->getInitializer(), Val, CE, 2));
2159}
2160
2161/// ComputeLoadResult - Return the value that would be computed by a load from
2162/// P after the stores reflected by 'memory' have been performed.  If we can't
2163/// decide, return null.
2164static Constant *ComputeLoadResult(Constant *P,
2165                                const DenseMap<Constant*, Constant*> &Memory) {
2166  // If this memory location has been recently stored, use the stored value: it
2167  // is the most up-to-date.
2168  DenseMap<Constant*, Constant*>::const_iterator I = Memory.find(P);
2169  if (I != Memory.end()) return I->second;
2170
2171  // Access it.
2172  if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
2173    if (GV->hasDefinitiveInitializer())
2174      return GV->getInitializer();
2175    return 0;
2176  }
2177
2178  // Handle a constantexpr getelementptr.
2179  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P))
2180    if (CE->getOpcode() == Instruction::GetElementPtr &&
2181        isa<GlobalVariable>(CE->getOperand(0))) {
2182      GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2183      if (GV->hasDefinitiveInitializer())
2184        return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
2185    }
2186
2187  return 0;  // don't know how to evaluate.
2188}
2189
2190/// EvaluateFunction - Evaluate a call to function F, returning true if
2191/// successful, false if we can't evaluate it.  ActualArgs contains the formal
2192/// arguments for the function.
2193static bool EvaluateFunction(Function *F, Constant *&RetVal,
2194                             const SmallVectorImpl<Constant*> &ActualArgs,
2195                             std::vector<Function*> &CallStack,
2196                             DenseMap<Constant*, Constant*> &MutatedMemory,
2197                             std::vector<GlobalVariable*> &AllocaTmps) {
2198  // Check to see if this function is already executing (recursion).  If so,
2199  // bail out.  TODO: we might want to accept limited recursion.
2200  if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end())
2201    return false;
2202
2203  CallStack.push_back(F);
2204
2205  /// Values - As we compute SSA register values, we store their contents here.
2206  DenseMap<Value*, Constant*> Values;
2207
2208  // Initialize arguments to the incoming values specified.
2209  unsigned ArgNo = 0;
2210  for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E;
2211       ++AI, ++ArgNo)
2212    Values[AI] = ActualArgs[ArgNo];
2213
2214  /// ExecutedBlocks - We only handle non-looping, non-recursive code.  As such,
2215  /// we can only evaluate any one basic block at most once.  This set keeps
2216  /// track of what we have executed so we can detect recursive cases etc.
2217  SmallPtrSet<BasicBlock*, 32> ExecutedBlocks;
2218
2219  // CurInst - The current instruction we're evaluating.
2220  BasicBlock::iterator CurInst = F->begin()->begin();
2221
2222  // This is the main evaluation loop.
2223  while (1) {
2224    Constant *InstResult = 0;
2225
2226    if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) {
2227      if (SI->isVolatile()) return false;  // no volatile accesses.
2228      Constant *Ptr = getVal(Values, SI->getOperand(1));
2229      if (!isSimpleEnoughPointerToCommit(Ptr))
2230        // If this is too complex for us to commit, reject it.
2231        return false;
2232      Constant *Val = getVal(Values, SI->getOperand(0));
2233      MutatedMemory[Ptr] = Val;
2234    } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) {
2235      InstResult = ConstantExpr::get(BO->getOpcode(),
2236                                     getVal(Values, BO->getOperand(0)),
2237                                     getVal(Values, BO->getOperand(1)));
2238    } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) {
2239      InstResult = ConstantExpr::getCompare(CI->getPredicate(),
2240                                            getVal(Values, CI->getOperand(0)),
2241                                            getVal(Values, CI->getOperand(1)));
2242    } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) {
2243      InstResult = ConstantExpr::getCast(CI->getOpcode(),
2244                                         getVal(Values, CI->getOperand(0)),
2245                                         CI->getType());
2246    } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) {
2247      InstResult =
2248            ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)),
2249                                           getVal(Values, SI->getOperand(1)),
2250                                           getVal(Values, SI->getOperand(2)));
2251    } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) {
2252      Constant *P = getVal(Values, GEP->getOperand(0));
2253      SmallVector<Constant*, 8> GEPOps;
2254      for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end();
2255           i != e; ++i)
2256        GEPOps.push_back(getVal(Values, *i));
2257      InstResult = cast<GEPOperator>(GEP)->isInBounds() ?
2258          ConstantExpr::getInBoundsGetElementPtr(P, &GEPOps[0], GEPOps.size()) :
2259          ConstantExpr::getGetElementPtr(P, &GEPOps[0], GEPOps.size());
2260    } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) {
2261      if (LI->isVolatile()) return false;  // no volatile accesses.
2262      InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)),
2263                                     MutatedMemory);
2264      if (InstResult == 0) return false; // Could not evaluate load.
2265    } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) {
2266      if (AI->isArrayAllocation()) return false;  // Cannot handle array allocs.
2267      const Type *Ty = AI->getType()->getElementType();
2268      AllocaTmps.push_back(new GlobalVariable(Ty, false,
2269                                              GlobalValue::InternalLinkage,
2270                                              UndefValue::get(Ty),
2271                                              AI->getName()));
2272      InstResult = AllocaTmps.back();
2273    } else if (CallInst *CI = dyn_cast<CallInst>(CurInst)) {
2274
2275      // Debug info can safely be ignored here.
2276      if (isa<DbgInfoIntrinsic>(CI)) {
2277        ++CurInst;
2278        continue;
2279      }
2280
2281      // Cannot handle inline asm.
2282      if (isa<InlineAsm>(CI->getOperand(0))) return false;
2283
2284      // Resolve function pointers.
2285      Function *Callee = dyn_cast<Function>(getVal(Values, CI->getOperand(0)));
2286      if (!Callee) return false;  // Cannot resolve.
2287
2288      SmallVector<Constant*, 8> Formals;
2289      for (User::op_iterator i = CI->op_begin() + 1, e = CI->op_end();
2290           i != e; ++i)
2291        Formals.push_back(getVal(Values, *i));
2292
2293      if (Callee->isDeclaration()) {
2294        // If this is a function we can constant fold, do it.
2295        if (Constant *C = ConstantFoldCall(Callee, Formals.data(),
2296                                           Formals.size())) {
2297          InstResult = C;
2298        } else {
2299          return false;
2300        }
2301      } else {
2302        if (Callee->getFunctionType()->isVarArg())
2303          return false;
2304
2305        Constant *RetVal;
2306        // Execute the call, if successful, use the return value.
2307        if (!EvaluateFunction(Callee, RetVal, Formals, CallStack,
2308                              MutatedMemory, AllocaTmps))
2309          return false;
2310        InstResult = RetVal;
2311      }
2312    } else if (isa<TerminatorInst>(CurInst)) {
2313      BasicBlock *NewBB = 0;
2314      if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) {
2315        if (BI->isUnconditional()) {
2316          NewBB = BI->getSuccessor(0);
2317        } else {
2318          ConstantInt *Cond =
2319            dyn_cast<ConstantInt>(getVal(Values, BI->getCondition()));
2320          if (!Cond) return false;  // Cannot determine.
2321
2322          NewBB = BI->getSuccessor(!Cond->getZExtValue());
2323        }
2324      } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) {
2325        ConstantInt *Val =
2326          dyn_cast<ConstantInt>(getVal(Values, SI->getCondition()));
2327        if (!Val) return false;  // Cannot determine.
2328        NewBB = SI->getSuccessor(SI->findCaseValue(Val));
2329      } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(CurInst)) {
2330        Value *Val = getVal(Values, IBI->getAddress())->stripPointerCasts();
2331        if (BlockAddress *BA = dyn_cast<BlockAddress>(Val))
2332          NewBB = BA->getBasicBlock();
2333        else
2334          return false;  // Cannot determine.
2335      } else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) {
2336        if (RI->getNumOperands())
2337          RetVal = getVal(Values, RI->getOperand(0));
2338
2339        CallStack.pop_back();  // return from fn.
2340        return true;  // We succeeded at evaluating this ctor!
2341      } else {
2342        // invoke, unwind, unreachable.
2343        return false;  // Cannot handle this terminator.
2344      }
2345
2346      // Okay, we succeeded in evaluating this control flow.  See if we have
2347      // executed the new block before.  If so, we have a looping function,
2348      // which we cannot evaluate in reasonable time.
2349      if (!ExecutedBlocks.insert(NewBB))
2350        return false;  // looped!
2351
2352      // Okay, we have never been in this block before.  Check to see if there
2353      // are any PHI nodes.  If so, evaluate them with information about where
2354      // we came from.
2355      BasicBlock *OldBB = CurInst->getParent();
2356      CurInst = NewBB->begin();
2357      PHINode *PN;
2358      for (; (PN = dyn_cast<PHINode>(CurInst)); ++CurInst)
2359        Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB));
2360
2361      // Do NOT increment CurInst.  We know that the terminator had no value.
2362      continue;
2363    } else {
2364      // Did not know how to evaluate this!
2365      return false;
2366    }
2367
2368    if (!CurInst->use_empty())
2369      Values[CurInst] = InstResult;
2370
2371    // Advance program counter.
2372    ++CurInst;
2373  }
2374}
2375
2376/// EvaluateStaticConstructor - Evaluate static constructors in the function, if
2377/// we can.  Return true if we can, false otherwise.
2378static bool EvaluateStaticConstructor(Function *F) {
2379  /// MutatedMemory - For each store we execute, we update this map.  Loads
2380  /// check this to get the most up-to-date value.  If evaluation is successful,
2381  /// this state is committed to the process.
2382  DenseMap<Constant*, Constant*> MutatedMemory;
2383
2384  /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable
2385  /// to represent its body.  This vector is needed so we can delete the
2386  /// temporary globals when we are done.
2387  std::vector<GlobalVariable*> AllocaTmps;
2388
2389  /// CallStack - This is used to detect recursion.  In pathological situations
2390  /// we could hit exponential behavior, but at least there is nothing
2391  /// unbounded.
2392  std::vector<Function*> CallStack;
2393
2394  // Call the function.
2395  Constant *RetValDummy;
2396  bool EvalSuccess = EvaluateFunction(F, RetValDummy,
2397                                      SmallVector<Constant*, 0>(), CallStack,
2398                                      MutatedMemory, AllocaTmps);
2399  if (EvalSuccess) {
2400    // We succeeded at evaluation: commit the result.
2401    DEBUG(dbgs() << "FULLY EVALUATED GLOBAL CTOR FUNCTION '"
2402          << F->getName() << "' to " << MutatedMemory.size()
2403          << " stores.\n");
2404    for (DenseMap<Constant*, Constant*>::iterator I = MutatedMemory.begin(),
2405         E = MutatedMemory.end(); I != E; ++I)
2406      CommitValueTo(I->second, I->first);
2407  }
2408
2409  // At this point, we are done interpreting.  If we created any 'alloca'
2410  // temporaries, release them now.
2411  while (!AllocaTmps.empty()) {
2412    GlobalVariable *Tmp = AllocaTmps.back();
2413    AllocaTmps.pop_back();
2414
2415    // If there are still users of the alloca, the program is doing something
2416    // silly, e.g. storing the address of the alloca somewhere and using it
2417    // later.  Since this is undefined, we'll just make it be null.
2418    if (!Tmp->use_empty())
2419      Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType()));
2420    delete Tmp;
2421  }
2422
2423  return EvalSuccess;
2424}
2425
2426
2427
2428/// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible.
2429/// Return true if anything changed.
2430bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) {
2431  std::vector<Function*> Ctors = ParseGlobalCtors(GCL);
2432  bool MadeChange = false;
2433  if (Ctors.empty()) return false;
2434
2435  // Loop over global ctors, optimizing them when we can.
2436  for (unsigned i = 0; i != Ctors.size(); ++i) {
2437    Function *F = Ctors[i];
2438    // Found a null terminator in the middle of the list, prune off the rest of
2439    // the list.
2440    if (F == 0) {
2441      if (i != Ctors.size()-1) {
2442        Ctors.resize(i+1);
2443        MadeChange = true;
2444      }
2445      break;
2446    }
2447
2448    // We cannot simplify external ctor functions.
2449    if (F->empty()) continue;
2450
2451    // If we can evaluate the ctor at compile time, do.
2452    if (EvaluateStaticConstructor(F)) {
2453      Ctors.erase(Ctors.begin()+i);
2454      MadeChange = true;
2455      --i;
2456      ++NumCtorsEvaluated;
2457      continue;
2458    }
2459  }
2460
2461  if (!MadeChange) return false;
2462
2463  GCL = InstallGlobalCtors(GCL, Ctors);
2464  return true;
2465}
2466
2467bool GlobalOpt::OptimizeGlobalAliases(Module &M) {
2468  bool Changed = false;
2469
2470  for (Module::alias_iterator I = M.alias_begin(), E = M.alias_end();
2471       I != E;) {
2472    Module::alias_iterator J = I++;
2473    // Aliases without names cannot be referenced outside this module.
2474    if (!J->hasName() && !J->isDeclaration())
2475      J->setLinkage(GlobalValue::InternalLinkage);
2476    // If the aliasee may change at link time, nothing can be done - bail out.
2477    if (J->mayBeOverridden())
2478      continue;
2479
2480    Constant *Aliasee = J->getAliasee();
2481    GlobalValue *Target = cast<GlobalValue>(Aliasee->stripPointerCasts());
2482    Target->removeDeadConstantUsers();
2483    bool hasOneUse = Target->hasOneUse() && Aliasee->hasOneUse();
2484
2485    // Make all users of the alias use the aliasee instead.
2486    if (!J->use_empty()) {
2487      J->replaceAllUsesWith(Aliasee);
2488      ++NumAliasesResolved;
2489      Changed = true;
2490    }
2491
2492    // If the alias is externally visible, we may still be able to simplify it.
2493    if (!J->hasLocalLinkage()) {
2494      // If the aliasee has internal linkage, give it the name and linkage
2495      // of the alias, and delete the alias.  This turns:
2496      //   define internal ... @f(...)
2497      //   @a = alias ... @f
2498      // into:
2499      //   define ... @a(...)
2500      if (!Target->hasLocalLinkage())
2501        continue;
2502
2503      // Do not perform the transform if multiple aliases potentially target the
2504      // aliasee.  This check also ensures that it is safe to replace the section
2505      // and other attributes of the aliasee with those of the alias.
2506      if (!hasOneUse)
2507        continue;
2508
2509      // Give the aliasee the name, linkage and other attributes of the alias.
2510      Target->takeName(J);
2511      Target->setLinkage(J->getLinkage());
2512      Target->GlobalValue::copyAttributesFrom(J);
2513    }
2514
2515    // Delete the alias.
2516    M.getAliasList().erase(J);
2517    ++NumAliasesRemoved;
2518    Changed = true;
2519  }
2520
2521  return Changed;
2522}
2523
2524bool GlobalOpt::runOnModule(Module &M) {
2525  bool Changed = false;
2526
2527  // Try to find the llvm.globalctors list.
2528  GlobalVariable *GlobalCtors = FindGlobalCtors(M);
2529
2530  bool LocalChange = true;
2531  while (LocalChange) {
2532    LocalChange = false;
2533
2534    // Delete functions that are trivially dead, ccc -> fastcc
2535    LocalChange |= OptimizeFunctions(M);
2536
2537    // Optimize global_ctors list.
2538    if (GlobalCtors)
2539      LocalChange |= OptimizeGlobalCtorsList(GlobalCtors);
2540
2541    // Optimize non-address-taken globals.
2542    LocalChange |= OptimizeGlobalVars(M);
2543
2544    // Resolve aliases, when possible.
2545    LocalChange |= OptimizeGlobalAliases(M);
2546    Changed |= LocalChange;
2547  }
2548
2549  // TODO: Move all global ctors functions to the end of the module for code
2550  // layout.
2551
2552  return Changed;
2553}
2554