GlobalOpt.cpp revision e7f431f9d4c41cb3cfc4cee1d2d5a3cda7f58a0c
1//===- GlobalOpt.cpp - Optimize Global Variables --------------------------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This pass transforms simple global variables that never have their address
11// taken.  If obviously true, it marks read/write globals as constant, deletes
12// variables only stored to, etc.
13//
14//===----------------------------------------------------------------------===//
15
16#define DEBUG_TYPE "globalopt"
17#include "llvm/Transforms/IPO.h"
18#include "llvm/CallingConv.h"
19#include "llvm/Constants.h"
20#include "llvm/DerivedTypes.h"
21#include "llvm/Instructions.h"
22#include "llvm/IntrinsicInst.h"
23#include "llvm/Module.h"
24#include "llvm/Pass.h"
25#include "llvm/Analysis/ConstantFolding.h"
26#include "llvm/Target/TargetData.h"
27#include "llvm/Support/CallSite.h"
28#include "llvm/Support/Compiler.h"
29#include "llvm/Support/Debug.h"
30#include "llvm/Support/GetElementPtrTypeIterator.h"
31#include "llvm/Support/MathExtras.h"
32#include "llvm/ADT/DenseMap.h"
33#include "llvm/ADT/SmallPtrSet.h"
34#include "llvm/ADT/SmallVector.h"
35#include "llvm/ADT/Statistic.h"
36#include "llvm/ADT/StringExtras.h"
37#include "llvm/ADT/STLExtras.h"
38#include <algorithm>
39using namespace llvm;
40
41STATISTIC(NumMarked    , "Number of globals marked constant");
42STATISTIC(NumSRA       , "Number of aggregate globals broken into scalars");
43STATISTIC(NumHeapSRA   , "Number of heap objects SRA'd");
44STATISTIC(NumSubstitute,"Number of globals with initializers stored into them");
45STATISTIC(NumDeleted   , "Number of globals deleted");
46STATISTIC(NumFnDeleted , "Number of functions deleted");
47STATISTIC(NumGlobUses  , "Number of global uses devirtualized");
48STATISTIC(NumLocalized , "Number of globals localized");
49STATISTIC(NumShrunkToBool  , "Number of global vars shrunk to booleans");
50STATISTIC(NumFastCallFns   , "Number of functions converted to fastcc");
51STATISTIC(NumCtorsEvaluated, "Number of static ctors evaluated");
52STATISTIC(NumNestRemoved   , "Number of nest attributes removed");
53STATISTIC(NumAliasesResolved, "Number of global aliases resolved");
54STATISTIC(NumAliasesRemoved, "Number of global aliases eliminated");
55
56namespace {
57  struct VISIBILITY_HIDDEN GlobalOpt : public ModulePass {
58    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
59      AU.addRequired<TargetData>();
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 ResolveAliases(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 VISIBILITY_HIDDEN 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  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/// ConstantIsDead - Return true if the specified constant is (transitively)
140/// dead.  The constant may be used by other constants (e.g. constant arrays and
141/// constant exprs) as long as they are dead, but it cannot be used by anything
142/// else.
143static bool ConstantIsDead(Constant *C) {
144  if (isa<GlobalValue>(C)) return false;
145
146  for (Value::use_iterator UI = C->use_begin(), E = C->use_end(); UI != E; ++UI)
147    if (Constant *CU = dyn_cast<Constant>(*UI)) {
148      if (!ConstantIsDead(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(Value *V, GlobalStatus &GS,
160                          SmallPtrSet<PHINode*, 16> &PHIUsers) {
161  for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
162    if (ConstantExpr *CE = dyn_cast<ConstantExpr>(*UI)) {
163      GS.HasNonInstructionUser = true;
164
165      if (AnalyzeGlobal(CE, GS, PHIUsers)) return true;
166
167    } else if (Instruction *I = dyn_cast<Instruction>(*UI)) {
168      if (!GS.HasMultipleAccessingFunctions) {
169        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 (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 (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 (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              // G = G
196              if (GS.StoredType < GlobalStatus::isInitializerStored)
197                GS.StoredType = GlobalStatus::isInitializerStored;
198            } else if (GS.StoredType < GlobalStatus::isStoredOnce) {
199              GS.StoredType = GlobalStatus::isStoredOnce;
200              GS.StoredOnceValue = StoredVal;
201            } else if (GS.StoredType == GlobalStatus::isStoredOnce &&
202                       GS.StoredOnceValue == StoredVal) {
203              // noop.
204            } else {
205              GS.StoredType = GlobalStatus::isStored;
206            }
207          } else {
208            GS.StoredType = GlobalStatus::isStored;
209          }
210        }
211      } else if (isa<GetElementPtrInst>(I)) {
212        if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
213      } else if (isa<SelectInst>(I)) {
214        if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
215      } else if (PHINode *PN = dyn_cast<PHINode>(I)) {
216        // PHI nodes we can check just like select or GEP instructions, but we
217        // have to be careful about infinite recursion.
218        if (PHIUsers.insert(PN))  // Not already visited.
219          if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
220        GS.HasPHIUser = true;
221      } else if (isa<CmpInst>(I)) {
222      } else if (isa<MemCpyInst>(I) || isa<MemMoveInst>(I)) {
223        if (I->getOperand(1) == V)
224          GS.StoredType = GlobalStatus::isStored;
225        if (I->getOperand(2) == V)
226          GS.isLoaded = true;
227      } else if (isa<MemSetInst>(I)) {
228        assert(I->getOperand(1) == V && "Memset only takes one pointer!");
229        GS.StoredType = GlobalStatus::isStored;
230      } else {
231        return true;  // Any other non-load instruction might take address!
232      }
233    } else if (Constant *C = dyn_cast<Constant>(*UI)) {
234      GS.HasNonInstructionUser = true;
235      // We might have a dead and dangling constant hanging off of here.
236      if (!ConstantIsDead(C))
237        return true;
238    } else {
239      GS.HasNonInstructionUser = true;
240      // Otherwise must be some other user.
241      return true;
242    }
243
244  return false;
245}
246
247static Constant *getAggregateConstantElement(Constant *Agg, Constant *Idx) {
248  ConstantInt *CI = dyn_cast<ConstantInt>(Idx);
249  if (!CI) return 0;
250  unsigned IdxV = CI->getZExtValue();
251
252  if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg)) {
253    if (IdxV < CS->getNumOperands()) return CS->getOperand(IdxV);
254  } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg)) {
255    if (IdxV < CA->getNumOperands()) return CA->getOperand(IdxV);
256  } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Agg)) {
257    if (IdxV < CP->getNumOperands()) return CP->getOperand(IdxV);
258  } else if (isa<ConstantAggregateZero>(Agg)) {
259    if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
260      if (IdxV < STy->getNumElements())
261        return Constant::getNullValue(STy->getElementType(IdxV));
262    } else if (const SequentialType *STy =
263               dyn_cast<SequentialType>(Agg->getType())) {
264      return Constant::getNullValue(STy->getElementType());
265    }
266  } else if (isa<UndefValue>(Agg)) {
267    if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
268      if (IdxV < STy->getNumElements())
269        return UndefValue::get(STy->getElementType(IdxV));
270    } else if (const SequentialType *STy =
271               dyn_cast<SequentialType>(Agg->getType())) {
272      return UndefValue::get(STy->getElementType());
273    }
274  }
275  return 0;
276}
277
278
279/// CleanupConstantGlobalUsers - We just marked GV constant.  Loop over all
280/// users of the global, cleaning up the obvious ones.  This is largely just a
281/// quick scan over the use list to clean up the easy and obvious cruft.  This
282/// returns true if it made a change.
283static bool CleanupConstantGlobalUsers(Value *V, Constant *Init) {
284  bool Changed = false;
285  for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;) {
286    User *U = *UI++;
287
288    if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
289      if (Init) {
290        // Replace the load with the initializer.
291        LI->replaceAllUsesWith(Init);
292        LI->eraseFromParent();
293        Changed = true;
294      }
295    } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
296      // Store must be unreachable or storing Init into the global.
297      SI->eraseFromParent();
298      Changed = true;
299    } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
300      if (CE->getOpcode() == Instruction::GetElementPtr) {
301        Constant *SubInit = 0;
302        if (Init)
303          SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
304        Changed |= CleanupConstantGlobalUsers(CE, SubInit);
305      } else if (CE->getOpcode() == Instruction::BitCast &&
306                 isa<PointerType>(CE->getType())) {
307        // Pointer cast, delete any stores and memsets to the global.
308        Changed |= CleanupConstantGlobalUsers(CE, 0);
309      }
310
311      if (CE->use_empty()) {
312        CE->destroyConstant();
313        Changed = true;
314      }
315    } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
316      // Do not transform "gepinst (gep constexpr (GV))" here, because forming
317      // "gepconstexpr (gep constexpr (GV))" will cause the two gep's to fold
318      // and will invalidate our notion of what Init is.
319      Constant *SubInit = 0;
320      if (!isa<ConstantExpr>(GEP->getOperand(0))) {
321        ConstantExpr *CE =
322          dyn_cast_or_null<ConstantExpr>(ConstantFoldInstruction(GEP));
323        if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr)
324          SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
325      }
326      Changed |= CleanupConstantGlobalUsers(GEP, SubInit);
327
328      if (GEP->use_empty()) {
329        GEP->eraseFromParent();
330        Changed = true;
331      }
332    } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv
333      if (MI->getRawDest() == V) {
334        MI->eraseFromParent();
335        Changed = true;
336      }
337
338    } else if (Constant *C = dyn_cast<Constant>(U)) {
339      // If we have a chain of dead constantexprs or other things dangling from
340      // us, and if they are all dead, nuke them without remorse.
341      if (ConstantIsDead(C)) {
342        C->destroyConstant();
343        // This could have invalidated UI, start over from scratch.
344        CleanupConstantGlobalUsers(V, Init);
345        return true;
346      }
347    }
348  }
349  return Changed;
350}
351
352/// isSafeSROAElementUse - Return true if the specified instruction is a safe
353/// user of a derived expression from a global that we want to SROA.
354static bool isSafeSROAElementUse(Value *V) {
355  // We might have a dead and dangling constant hanging off of here.
356  if (Constant *C = dyn_cast<Constant>(V))
357    return ConstantIsDead(C);
358
359  Instruction *I = dyn_cast<Instruction>(V);
360  if (!I) return false;
361
362  // Loads are ok.
363  if (isa<LoadInst>(I)) return true;
364
365  // Stores *to* the pointer are ok.
366  if (StoreInst *SI = dyn_cast<StoreInst>(I))
367    return SI->getOperand(0) != V;
368
369  // Otherwise, it must be a GEP.
370  GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I);
371  if (GEPI == 0) return false;
372
373  if (GEPI->getNumOperands() < 3 || !isa<Constant>(GEPI->getOperand(1)) ||
374      !cast<Constant>(GEPI->getOperand(1))->isNullValue())
375    return false;
376
377  for (Value::use_iterator I = GEPI->use_begin(), E = GEPI->use_end();
378       I != E; ++I)
379    if (!isSafeSROAElementUse(*I))
380      return false;
381  return true;
382}
383
384
385/// IsUserOfGlobalSafeForSRA - U is a direct user of the specified global value.
386/// Look at it and its uses and decide whether it is safe to SROA this global.
387///
388static bool IsUserOfGlobalSafeForSRA(User *U, GlobalValue *GV) {
389  // The user of the global must be a GEP Inst or a ConstantExpr GEP.
390  if (!isa<GetElementPtrInst>(U) &&
391      (!isa<ConstantExpr>(U) ||
392       cast<ConstantExpr>(U)->getOpcode() != Instruction::GetElementPtr))
393    return false;
394
395  // Check to see if this ConstantExpr GEP is SRA'able.  In particular, we
396  // don't like < 3 operand CE's, and we don't like non-constant integer
397  // indices.  This enforces that all uses are 'gep GV, 0, C, ...' for some
398  // value of C.
399  if (U->getNumOperands() < 3 || !isa<Constant>(U->getOperand(1)) ||
400      !cast<Constant>(U->getOperand(1))->isNullValue() ||
401      !isa<ConstantInt>(U->getOperand(2)))
402    return false;
403
404  gep_type_iterator GEPI = gep_type_begin(U), E = gep_type_end(U);
405  ++GEPI;  // Skip over the pointer index.
406
407  // If this is a use of an array allocation, do a bit more checking for sanity.
408  if (const ArrayType *AT = dyn_cast<ArrayType>(*GEPI)) {
409    uint64_t NumElements = AT->getNumElements();
410    ConstantInt *Idx = cast<ConstantInt>(U->getOperand(2));
411
412    // Check to make sure that index falls within the array.  If not,
413    // something funny is going on, so we won't do the optimization.
414    //
415    if (Idx->getZExtValue() >= NumElements)
416      return false;
417
418    // We cannot scalar repl this level of the array unless any array
419    // sub-indices are in-range constants.  In particular, consider:
420    // A[0][i].  We cannot know that the user isn't doing invalid things like
421    // allowing i to index an out-of-range subscript that accesses A[1].
422    //
423    // Scalar replacing *just* the outer index of the array is probably not
424    // going to be a win anyway, so just give up.
425    for (++GEPI; // Skip array index.
426         GEPI != E && (isa<ArrayType>(*GEPI) || isa<VectorType>(*GEPI));
427         ++GEPI) {
428      uint64_t NumElements;
429      if (const ArrayType *SubArrayTy = dyn_cast<ArrayType>(*GEPI))
430        NumElements = SubArrayTy->getNumElements();
431      else
432        NumElements = cast<VectorType>(*GEPI)->getNumElements();
433
434      ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPI.getOperand());
435      if (!IdxVal || IdxVal->getZExtValue() >= NumElements)
436        return false;
437    }
438  }
439
440  for (Value::use_iterator I = U->use_begin(), E = U->use_end(); I != E; ++I)
441    if (!isSafeSROAElementUse(*I))
442      return false;
443  return true;
444}
445
446/// GlobalUsersSafeToSRA - Look at all uses of the global and decide whether it
447/// is safe for us to perform this transformation.
448///
449static bool GlobalUsersSafeToSRA(GlobalValue *GV) {
450  for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end();
451       UI != E; ++UI) {
452    if (!IsUserOfGlobalSafeForSRA(*UI, GV))
453      return false;
454  }
455  return true;
456}
457
458
459/// SRAGlobal - Perform scalar replacement of aggregates on the specified global
460/// variable.  This opens the door for other optimizations by exposing the
461/// behavior of the program in a more fine-grained way.  We have determined that
462/// this transformation is safe already.  We return the first global variable we
463/// insert so that the caller can reprocess it.
464static GlobalVariable *SRAGlobal(GlobalVariable *GV, const TargetData &TD) {
465  // Make sure this global only has simple uses that we can SRA.
466  if (!GlobalUsersSafeToSRA(GV))
467    return 0;
468
469  assert(GV->hasLocalLinkage() && !GV->isConstant());
470  Constant *Init = GV->getInitializer();
471  const Type *Ty = Init->getType();
472
473  std::vector<GlobalVariable*> NewGlobals;
474  Module::GlobalListType &Globals = GV->getParent()->getGlobalList();
475
476  // Get the alignment of the global, either explicit or target-specific.
477  unsigned StartAlignment = GV->getAlignment();
478  if (StartAlignment == 0)
479    StartAlignment = TD.getABITypeAlignment(GV->getType());
480
481  if (const StructType *STy = dyn_cast<StructType>(Ty)) {
482    NewGlobals.reserve(STy->getNumElements());
483    const StructLayout &Layout = *TD.getStructLayout(STy);
484    for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
485      Constant *In = getAggregateConstantElement(Init,
486                                            ConstantInt::get(Type::Int32Ty, i));
487      assert(In && "Couldn't get element of initializer?");
488      GlobalVariable *NGV = new GlobalVariable(STy->getElementType(i), false,
489                                               GlobalVariable::InternalLinkage,
490                                               In, GV->getName()+"."+utostr(i),
491                                               (Module *)NULL,
492                                               GV->isThreadLocal(),
493                                               GV->getType()->getAddressSpace());
494      Globals.insert(GV, NGV);
495      NewGlobals.push_back(NGV);
496
497      // Calculate the known alignment of the field.  If the original aggregate
498      // had 256 byte alignment for example, something might depend on that:
499      // propagate info to each field.
500      uint64_t FieldOffset = Layout.getElementOffset(i);
501      unsigned NewAlign = (unsigned)MinAlign(StartAlignment, FieldOffset);
502      if (NewAlign > TD.getABITypeAlignment(STy->getElementType(i)))
503        NGV->setAlignment(NewAlign);
504    }
505  } else if (const SequentialType *STy = dyn_cast<SequentialType>(Ty)) {
506    unsigned NumElements = 0;
507    if (const ArrayType *ATy = dyn_cast<ArrayType>(STy))
508      NumElements = ATy->getNumElements();
509    else
510      NumElements = cast<VectorType>(STy)->getNumElements();
511
512    if (NumElements > 16 && GV->hasNUsesOrMore(16))
513      return 0; // It's not worth it.
514    NewGlobals.reserve(NumElements);
515
516    uint64_t EltSize = TD.getTypePaddedSize(STy->getElementType());
517    unsigned EltAlign = TD.getABITypeAlignment(STy->getElementType());
518    for (unsigned i = 0, e = NumElements; i != e; ++i) {
519      Constant *In = getAggregateConstantElement(Init,
520                                            ConstantInt::get(Type::Int32Ty, i));
521      assert(In && "Couldn't get element of initializer?");
522
523      GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false,
524                                               GlobalVariable::InternalLinkage,
525                                               In, GV->getName()+"."+utostr(i),
526                                               (Module *)NULL,
527                                               GV->isThreadLocal(),
528                                               GV->getType()->getAddressSpace());
529      Globals.insert(GV, NGV);
530      NewGlobals.push_back(NGV);
531
532      // Calculate the known alignment of the field.  If the original aggregate
533      // had 256 byte alignment for example, something might depend on that:
534      // propagate info to each field.
535      unsigned NewAlign = (unsigned)MinAlign(StartAlignment, EltSize*i);
536      if (NewAlign > EltAlign)
537        NGV->setAlignment(NewAlign);
538    }
539  }
540
541  if (NewGlobals.empty())
542    return 0;
543
544  DOUT << "PERFORMING GLOBAL SRA ON: " << *GV;
545
546  Constant *NullInt = Constant::getNullValue(Type::Int32Ty);
547
548  // Loop over all of the uses of the global, replacing the constantexpr geps,
549  // with smaller constantexpr geps or direct references.
550  while (!GV->use_empty()) {
551    User *GEP = GV->use_back();
552    assert(((isa<ConstantExpr>(GEP) &&
553             cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)||
554            isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!");
555
556    // Ignore the 1th operand, which has to be zero or else the program is quite
557    // broken (undefined).  Get the 2nd operand, which is the structure or array
558    // index.
559    unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
560    if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access.
561
562    Value *NewPtr = NewGlobals[Val];
563
564    // Form a shorter GEP if needed.
565    if (GEP->getNumOperands() > 3) {
566      if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) {
567        SmallVector<Constant*, 8> Idxs;
568        Idxs.push_back(NullInt);
569        for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i)
570          Idxs.push_back(CE->getOperand(i));
571        NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr),
572                                                &Idxs[0], Idxs.size());
573      } else {
574        GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP);
575        SmallVector<Value*, 8> Idxs;
576        Idxs.push_back(NullInt);
577        for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i)
578          Idxs.push_back(GEPI->getOperand(i));
579        NewPtr = GetElementPtrInst::Create(NewPtr, Idxs.begin(), Idxs.end(),
580                                           GEPI->getName()+"."+utostr(Val), GEPI);
581      }
582    }
583    GEP->replaceAllUsesWith(NewPtr);
584
585    if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP))
586      GEPI->eraseFromParent();
587    else
588      cast<ConstantExpr>(GEP)->destroyConstant();
589  }
590
591  // Delete the old global, now that it is dead.
592  Globals.erase(GV);
593  ++NumSRA;
594
595  // Loop over the new globals array deleting any globals that are obviously
596  // dead.  This can arise due to scalarization of a structure or an array that
597  // has elements that are dead.
598  unsigned FirstGlobal = 0;
599  for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i)
600    if (NewGlobals[i]->use_empty()) {
601      Globals.erase(NewGlobals[i]);
602      if (FirstGlobal == i) ++FirstGlobal;
603    }
604
605  return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : 0;
606}
607
608/// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified
609/// value will trap if the value is dynamically null.  PHIs keeps track of any
610/// phi nodes we've seen to avoid reprocessing them.
611static bool AllUsesOfValueWillTrapIfNull(Value *V,
612                                         SmallPtrSet<PHINode*, 8> &PHIs) {
613  for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
614    if (isa<LoadInst>(*UI)) {
615      // Will trap.
616    } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
617      if (SI->getOperand(0) == V) {
618        //cerr << "NONTRAPPING USE: " << **UI;
619        return false;  // Storing the value.
620      }
621    } else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
622      if (CI->getOperand(0) != V) {
623        //cerr << "NONTRAPPING USE: " << **UI;
624        return false;  // Not calling the ptr
625      }
626    } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
627      if (II->getOperand(0) != V) {
628        //cerr << "NONTRAPPING USE: " << **UI;
629        return false;  // Not calling the ptr
630      }
631    } else if (BitCastInst *CI = dyn_cast<BitCastInst>(*UI)) {
632      if (!AllUsesOfValueWillTrapIfNull(CI, PHIs)) return false;
633    } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) {
634      if (!AllUsesOfValueWillTrapIfNull(GEPI, PHIs)) return false;
635    } else if (PHINode *PN = dyn_cast<PHINode>(*UI)) {
636      // If we've already seen this phi node, ignore it, it has already been
637      // checked.
638      if (PHIs.insert(PN))
639        return AllUsesOfValueWillTrapIfNull(PN, PHIs);
640    } else if (isa<ICmpInst>(*UI) &&
641               isa<ConstantPointerNull>(UI->getOperand(1))) {
642      // Ignore setcc X, null
643    } else {
644      //cerr << "NONTRAPPING USE: " << **UI;
645      return false;
646    }
647  return true;
648}
649
650/// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads
651/// from GV will trap if the loaded value is null.  Note that this also permits
652/// comparisons of the loaded value against null, as a special case.
653static bool AllUsesOfLoadedValueWillTrapIfNull(GlobalVariable *GV) {
654  for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI!=E; ++UI)
655    if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
656      SmallPtrSet<PHINode*, 8> PHIs;
657      if (!AllUsesOfValueWillTrapIfNull(LI, PHIs))
658        return false;
659    } else if (isa<StoreInst>(*UI)) {
660      // Ignore stores to the global.
661    } else {
662      // We don't know or understand this user, bail out.
663      //cerr << "UNKNOWN USER OF GLOBAL!: " << **UI;
664      return false;
665    }
666
667  return true;
668}
669
670static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) {
671  bool Changed = false;
672  for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) {
673    Instruction *I = cast<Instruction>(*UI++);
674    if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
675      LI->setOperand(0, NewV);
676      Changed = true;
677    } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
678      if (SI->getOperand(1) == V) {
679        SI->setOperand(1, NewV);
680        Changed = true;
681      }
682    } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
683      if (I->getOperand(0) == V) {
684        // Calling through the pointer!  Turn into a direct call, but be careful
685        // that the pointer is not also being passed as an argument.
686        I->setOperand(0, NewV);
687        Changed = true;
688        bool PassedAsArg = false;
689        for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i)
690          if (I->getOperand(i) == V) {
691            PassedAsArg = true;
692            I->setOperand(i, NewV);
693          }
694
695        if (PassedAsArg) {
696          // Being passed as an argument also.  Be careful to not invalidate UI!
697          UI = V->use_begin();
698        }
699      }
700    } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
701      Changed |= OptimizeAwayTrappingUsesOfValue(CI,
702                                ConstantExpr::getCast(CI->getOpcode(),
703                                                      NewV, CI->getType()));
704      if (CI->use_empty()) {
705        Changed = true;
706        CI->eraseFromParent();
707      }
708    } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
709      // Should handle GEP here.
710      SmallVector<Constant*, 8> Idxs;
711      Idxs.reserve(GEPI->getNumOperands()-1);
712      for (User::op_iterator i = GEPI->op_begin() + 1, e = GEPI->op_end();
713           i != e; ++i)
714        if (Constant *C = dyn_cast<Constant>(*i))
715          Idxs.push_back(C);
716        else
717          break;
718      if (Idxs.size() == GEPI->getNumOperands()-1)
719        Changed |= OptimizeAwayTrappingUsesOfValue(GEPI,
720                                ConstantExpr::getGetElementPtr(NewV, &Idxs[0],
721                                                               Idxs.size()));
722      if (GEPI->use_empty()) {
723        Changed = true;
724        GEPI->eraseFromParent();
725      }
726    }
727  }
728
729  return Changed;
730}
731
732
733/// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null
734/// value stored into it.  If there are uses of the loaded value that would trap
735/// if the loaded value is dynamically null, then we know that they cannot be
736/// reachable with a null optimize away the load.
737static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV) {
738  bool Changed = false;
739
740  // Keep track of whether we are able to remove all the uses of the global
741  // other than the store that defines it.
742  bool AllNonStoreUsesGone = true;
743
744  // Replace all uses of loads with uses of uses of the stored value.
745  for (Value::use_iterator GUI = GV->use_begin(), E = GV->use_end(); GUI != E;){
746    User *GlobalUser = *GUI++;
747    if (LoadInst *LI = dyn_cast<LoadInst>(GlobalUser)) {
748      Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV);
749      // If we were able to delete all uses of the loads
750      if (LI->use_empty()) {
751        LI->eraseFromParent();
752        Changed = true;
753      } else {
754        AllNonStoreUsesGone = false;
755      }
756    } else if (isa<StoreInst>(GlobalUser)) {
757      // Ignore the store that stores "LV" to the global.
758      assert(GlobalUser->getOperand(1) == GV &&
759             "Must be storing *to* the global");
760    } else {
761      AllNonStoreUsesGone = false;
762
763      // If we get here we could have other crazy uses that are transitively
764      // loaded.
765      assert((isa<PHINode>(GlobalUser) || isa<SelectInst>(GlobalUser) ||
766              isa<ConstantExpr>(GlobalUser)) && "Only expect load and stores!");
767    }
768  }
769
770  if (Changed) {
771    DOUT << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV;
772    ++NumGlobUses;
773  }
774
775  // If we nuked all of the loads, then none of the stores are needed either,
776  // nor is the global.
777  if (AllNonStoreUsesGone) {
778    DOUT << "  *** GLOBAL NOW DEAD!\n";
779    CleanupConstantGlobalUsers(GV, 0);
780    if (GV->use_empty()) {
781      GV->eraseFromParent();
782      ++NumDeleted;
783    }
784    Changed = true;
785  }
786  return Changed;
787}
788
789/// ConstantPropUsersOf - Walk the use list of V, constant folding all of the
790/// instructions that are foldable.
791static void ConstantPropUsersOf(Value *V) {
792  for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; )
793    if (Instruction *I = dyn_cast<Instruction>(*UI++))
794      if (Constant *NewC = ConstantFoldInstruction(I)) {
795        I->replaceAllUsesWith(NewC);
796
797        // Advance UI to the next non-I use to avoid invalidating it!
798        // Instructions could multiply use V.
799        while (UI != E && *UI == I)
800          ++UI;
801        I->eraseFromParent();
802      }
803}
804
805/// OptimizeGlobalAddressOfMalloc - This function takes the specified global
806/// variable, and transforms the program as if it always contained the result of
807/// the specified malloc.  Because it is always the result of the specified
808/// malloc, there is no reason to actually DO the malloc.  Instead, turn the
809/// malloc into a global, and any loads of GV as uses of the new global.
810static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV,
811                                                     MallocInst *MI) {
812  DOUT << "PROMOTING MALLOC GLOBAL: " << *GV << "  MALLOC = " << *MI;
813  ConstantInt *NElements = cast<ConstantInt>(MI->getArraySize());
814
815  if (NElements->getZExtValue() != 1) {
816    // If we have an array allocation, transform it to a single element
817    // allocation to make the code below simpler.
818    Type *NewTy = ArrayType::get(MI->getAllocatedType(),
819                                 NElements->getZExtValue());
820    MallocInst *NewMI =
821      new MallocInst(NewTy, Constant::getNullValue(Type::Int32Ty),
822                     MI->getAlignment(), MI->getName(), MI);
823    Value* Indices[2];
824    Indices[0] = Indices[1] = Constant::getNullValue(Type::Int32Ty);
825    Value *NewGEP = GetElementPtrInst::Create(NewMI, Indices, Indices + 2,
826                                              NewMI->getName()+".el0", MI);
827    MI->replaceAllUsesWith(NewGEP);
828    MI->eraseFromParent();
829    MI = NewMI;
830  }
831
832  // Create the new global variable.  The contents of the malloc'd memory is
833  // undefined, so initialize with an undef value.
834  Constant *Init = UndefValue::get(MI->getAllocatedType());
835  GlobalVariable *NewGV = new GlobalVariable(MI->getAllocatedType(), false,
836                                             GlobalValue::InternalLinkage, Init,
837                                             GV->getName()+".body",
838                                             (Module *)NULL,
839                                             GV->isThreadLocal());
840  // FIXME: This new global should have the alignment returned by malloc.  Code
841  // could depend on malloc returning large alignment (on the mac, 16 bytes) but
842  // this would only guarantee some lower alignment.
843  GV->getParent()->getGlobalList().insert(GV, NewGV);
844
845  // Anything that used the malloc now uses the global directly.
846  MI->replaceAllUsesWith(NewGV);
847
848  Constant *RepValue = NewGV;
849  if (NewGV->getType() != GV->getType()->getElementType())
850    RepValue = ConstantExpr::getBitCast(RepValue,
851                                        GV->getType()->getElementType());
852
853  // If there is a comparison against null, we will insert a global bool to
854  // keep track of whether the global was initialized yet or not.
855  GlobalVariable *InitBool =
856    new GlobalVariable(Type::Int1Ty, false, GlobalValue::InternalLinkage,
857                       ConstantInt::getFalse(), GV->getName()+".init",
858                       (Module *)NULL, GV->isThreadLocal());
859  bool InitBoolUsed = false;
860
861  // Loop over all uses of GV, processing them in turn.
862  std::vector<StoreInst*> Stores;
863  while (!GV->use_empty())
864    if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) {
865      while (!LI->use_empty()) {
866        Use &LoadUse = LI->use_begin().getUse();
867        if (!isa<ICmpInst>(LoadUse.getUser()))
868          LoadUse = RepValue;
869        else {
870          ICmpInst *CI = cast<ICmpInst>(LoadUse.getUser());
871          // Replace the cmp X, 0 with a use of the bool value.
872          Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", CI);
873          InitBoolUsed = true;
874          switch (CI->getPredicate()) {
875          default: assert(0 && "Unknown ICmp Predicate!");
876          case ICmpInst::ICMP_ULT:
877          case ICmpInst::ICMP_SLT:
878            LV = ConstantInt::getFalse();   // X < null -> always false
879            break;
880          case ICmpInst::ICMP_ULE:
881          case ICmpInst::ICMP_SLE:
882          case ICmpInst::ICMP_EQ:
883            LV = BinaryOperator::CreateNot(LV, "notinit", CI);
884            break;
885          case ICmpInst::ICMP_NE:
886          case ICmpInst::ICMP_UGE:
887          case ICmpInst::ICMP_SGE:
888          case ICmpInst::ICMP_UGT:
889          case ICmpInst::ICMP_SGT:
890            break;  // no change.
891          }
892          CI->replaceAllUsesWith(LV);
893          CI->eraseFromParent();
894        }
895      }
896      LI->eraseFromParent();
897    } else {
898      StoreInst *SI = cast<StoreInst>(GV->use_back());
899      // The global is initialized when the store to it occurs.
900      new StoreInst(ConstantInt::getTrue(), InitBool, SI);
901      SI->eraseFromParent();
902    }
903
904  // If the initialization boolean was used, insert it, otherwise delete it.
905  if (!InitBoolUsed) {
906    while (!InitBool->use_empty())  // Delete initializations
907      cast<Instruction>(InitBool->use_back())->eraseFromParent();
908    delete InitBool;
909  } else
910    GV->getParent()->getGlobalList().insert(GV, InitBool);
911
912
913  // Now the GV is dead, nuke it and the malloc.
914  GV->eraseFromParent();
915  MI->eraseFromParent();
916
917  // To further other optimizations, loop over all users of NewGV and try to
918  // constant prop them.  This will promote GEP instructions with constant
919  // indices into GEP constant-exprs, which will allow global-opt to hack on it.
920  ConstantPropUsersOf(NewGV);
921  if (RepValue != NewGV)
922    ConstantPropUsersOf(RepValue);
923
924  return NewGV;
925}
926
927/// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking
928/// to make sure that there are no complex uses of V.  We permit simple things
929/// like dereferencing the pointer, but not storing through the address, unless
930/// it is to the specified global.
931static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Instruction *V,
932                                                      GlobalVariable *GV,
933                                              SmallPtrSet<PHINode*, 8> &PHIs) {
934  for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
935    Instruction *Inst = dyn_cast<Instruction>(*UI);
936    if (Inst == 0) return false;
937
938    if (isa<LoadInst>(Inst) || isa<CmpInst>(Inst)) {
939      continue; // Fine, ignore.
940    }
941
942    if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
943      if (SI->getOperand(0) == V && SI->getOperand(1) != GV)
944        return false;  // Storing the pointer itself... bad.
945      continue; // Otherwise, storing through it, or storing into GV... fine.
946    }
947
948    if (isa<GetElementPtrInst>(Inst)) {
949      if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Inst, GV, PHIs))
950        return false;
951      continue;
952    }
953
954    if (PHINode *PN = dyn_cast<PHINode>(Inst)) {
955      // PHIs are ok if all uses are ok.  Don't infinitely recurse through PHI
956      // cycles.
957      if (PHIs.insert(PN))
958        if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(PN, GV, PHIs))
959          return false;
960      continue;
961    }
962
963    if (BitCastInst *BCI = dyn_cast<BitCastInst>(Inst)) {
964      if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(BCI, GV, PHIs))
965        return false;
966      continue;
967    }
968
969    return false;
970  }
971  return true;
972}
973
974/// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV
975/// somewhere.  Transform all uses of the allocation into loads from the
976/// global and uses of the resultant pointer.  Further, delete the store into
977/// GV.  This assumes that these value pass the
978/// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate.
979static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc,
980                                          GlobalVariable *GV) {
981  while (!Alloc->use_empty()) {
982    Instruction *U = cast<Instruction>(*Alloc->use_begin());
983    Instruction *InsertPt = U;
984    if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
985      // If this is the store of the allocation into the global, remove it.
986      if (SI->getOperand(1) == GV) {
987        SI->eraseFromParent();
988        continue;
989      }
990    } else if (PHINode *PN = dyn_cast<PHINode>(U)) {
991      // Insert the load in the corresponding predecessor, not right before the
992      // PHI.
993      InsertPt = PN->getIncomingBlock(Alloc->use_begin())->getTerminator();
994    } else if (isa<BitCastInst>(U)) {
995      // Must be bitcast between the malloc and store to initialize the global.
996      ReplaceUsesOfMallocWithGlobal(U, GV);
997      U->eraseFromParent();
998      continue;
999    } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
1000      // If this is a "GEP bitcast" and the user is a store to the global, then
1001      // just process it as a bitcast.
1002      if (GEPI->hasAllZeroIndices() && GEPI->hasOneUse())
1003        if (StoreInst *SI = dyn_cast<StoreInst>(GEPI->use_back()))
1004          if (SI->getOperand(1) == GV) {
1005            // Must be bitcast GEP between the malloc and store to initialize
1006            // the global.
1007            ReplaceUsesOfMallocWithGlobal(GEPI, GV);
1008            GEPI->eraseFromParent();
1009            continue;
1010          }
1011    }
1012
1013    // Insert a load from the global, and use it instead of the malloc.
1014    Value *NL = new LoadInst(GV, GV->getName()+".val", InsertPt);
1015    U->replaceUsesOfWith(Alloc, NL);
1016  }
1017}
1018
1019/// LoadUsesSimpleEnoughForHeapSRA - Verify that all uses of V (a load, or a phi
1020/// of a load) are simple enough to perform heap SRA on.  This permits GEP's
1021/// that index through the array and struct field, icmps of null, and PHIs.
1022static bool LoadUsesSimpleEnoughForHeapSRA(Value *V,
1023                                     SmallPtrSet<PHINode*, 32> &LoadUsingPHIs) {
1024  // We permit two users of the load: setcc comparing against the null
1025  // pointer, and a getelementptr of a specific form.
1026  for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
1027    Instruction *User = cast<Instruction>(*UI);
1028
1029    // Comparison against null is ok.
1030    if (ICmpInst *ICI = dyn_cast<ICmpInst>(User)) {
1031      if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
1032        return false;
1033      continue;
1034    }
1035
1036    // getelementptr is also ok, but only a simple form.
1037    if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1038      // Must index into the array and into the struct.
1039      if (GEPI->getNumOperands() < 3)
1040        return false;
1041
1042      // Otherwise the GEP is ok.
1043      continue;
1044    }
1045
1046    if (PHINode *PN = dyn_cast<PHINode>(User)) {
1047      // If we have already recursively analyzed this PHI, then it is safe.
1048      if (LoadUsingPHIs.insert(PN))
1049        continue;
1050
1051      // Make sure all uses of the PHI are simple enough to transform.
1052      if (!LoadUsesSimpleEnoughForHeapSRA(PN, LoadUsingPHIs))
1053        return false;
1054
1055      continue;
1056    }
1057
1058    // Otherwise we don't know what this is, not ok.
1059    return false;
1060  }
1061
1062  return true;
1063}
1064
1065
1066/// AllGlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from
1067/// GV are simple enough to perform HeapSRA, return true.
1068static bool AllGlobalLoadUsesSimpleEnoughForHeapSRA(GlobalVariable *GV,
1069                                                    MallocInst *MI) {
1070  SmallPtrSet<PHINode*, 32> LoadUsingPHIs;
1071  for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;
1072       ++UI)
1073    if (LoadInst *LI = dyn_cast<LoadInst>(*UI))
1074      if (!LoadUsesSimpleEnoughForHeapSRA(LI, LoadUsingPHIs))
1075        return false;
1076
1077  // If we reach here, we know that all uses of the loads and transitive uses
1078  // (through PHI nodes) are simple enough to transform.  However, we don't know
1079  // that all inputs the to the PHI nodes are in the same equivalence sets.
1080  // Check to verify that all operands of the PHIs are either PHIS that can be
1081  // transformed, loads from GV, or MI itself.
1082  for (SmallPtrSet<PHINode*, 32>::iterator I = LoadUsingPHIs.begin(),
1083       E = LoadUsingPHIs.end(); I != E; ++I) {
1084    PHINode *PN = *I;
1085    for (unsigned op = 0, e = PN->getNumIncomingValues(); op != e; ++op) {
1086      Value *InVal = PN->getIncomingValue(op);
1087
1088      // PHI of the stored value itself is ok.
1089      if (InVal == MI) continue;
1090
1091      if (PHINode *InPN = dyn_cast<PHINode>(InVal)) {
1092        // One of the PHIs in our set is (optimistically) ok.
1093        if (LoadUsingPHIs.count(InPN))
1094          continue;
1095        return false;
1096      }
1097
1098      // Load from GV is ok.
1099      if (LoadInst *LI = dyn_cast<LoadInst>(InVal))
1100        if (LI->getOperand(0) == GV)
1101          continue;
1102
1103      // UNDEF? NULL?
1104
1105      // Anything else is rejected.
1106      return false;
1107    }
1108  }
1109
1110  return true;
1111}
1112
1113static Value *GetHeapSROAValue(Value *V, unsigned FieldNo,
1114               DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1115                   std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1116  std::vector<Value*> &FieldVals = InsertedScalarizedValues[V];
1117
1118  if (FieldNo >= FieldVals.size())
1119    FieldVals.resize(FieldNo+1);
1120
1121  // If we already have this value, just reuse the previously scalarized
1122  // version.
1123  if (Value *FieldVal = FieldVals[FieldNo])
1124    return FieldVal;
1125
1126  // Depending on what instruction this is, we have several cases.
1127  Value *Result;
1128  if (LoadInst *LI = dyn_cast<LoadInst>(V)) {
1129    // This is a scalarized version of the load from the global.  Just create
1130    // a new Load of the scalarized global.
1131    Result = new LoadInst(GetHeapSROAValue(LI->getOperand(0), FieldNo,
1132                                           InsertedScalarizedValues,
1133                                           PHIsToRewrite),
1134                          LI->getName()+".f" + utostr(FieldNo), LI);
1135  } else if (PHINode *PN = dyn_cast<PHINode>(V)) {
1136    // PN's type is pointer to struct.  Make a new PHI of pointer to struct
1137    // field.
1138    const StructType *ST =
1139      cast<StructType>(cast<PointerType>(PN->getType())->getElementType());
1140
1141    Result =PHINode::Create(PointerType::getUnqual(ST->getElementType(FieldNo)),
1142                            PN->getName()+".f"+utostr(FieldNo), PN);
1143    PHIsToRewrite.push_back(std::make_pair(PN, FieldNo));
1144  } else {
1145    assert(0 && "Unknown usable value");
1146    Result = 0;
1147  }
1148
1149  return FieldVals[FieldNo] = Result;
1150}
1151
1152/// RewriteHeapSROALoadUser - Given a load instruction and a value derived from
1153/// the load, rewrite the derived value to use the HeapSRoA'd load.
1154static void RewriteHeapSROALoadUser(Instruction *LoadUser,
1155             DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1156                   std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1157  // If this is a comparison against null, handle it.
1158  if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) {
1159    assert(isa<ConstantPointerNull>(SCI->getOperand(1)));
1160    // If we have a setcc of the loaded pointer, we can use a setcc of any
1161    // field.
1162    Value *NPtr = GetHeapSROAValue(SCI->getOperand(0), 0,
1163                                   InsertedScalarizedValues, PHIsToRewrite);
1164
1165    Value *New = new ICmpInst(SCI->getPredicate(), NPtr,
1166                              Constant::getNullValue(NPtr->getType()),
1167                              SCI->getName(), SCI);
1168    SCI->replaceAllUsesWith(New);
1169    SCI->eraseFromParent();
1170    return;
1171  }
1172
1173  // Handle 'getelementptr Ptr, Idx, i32 FieldNo ...'
1174  if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) {
1175    assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2))
1176           && "Unexpected GEPI!");
1177
1178    // Load the pointer for this field.
1179    unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
1180    Value *NewPtr = GetHeapSROAValue(GEPI->getOperand(0), FieldNo,
1181                                     InsertedScalarizedValues, PHIsToRewrite);
1182
1183    // Create the new GEP idx vector.
1184    SmallVector<Value*, 8> GEPIdx;
1185    GEPIdx.push_back(GEPI->getOperand(1));
1186    GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end());
1187
1188    Value *NGEPI = GetElementPtrInst::Create(NewPtr,
1189                                             GEPIdx.begin(), GEPIdx.end(),
1190                                             GEPI->getName(), GEPI);
1191    GEPI->replaceAllUsesWith(NGEPI);
1192    GEPI->eraseFromParent();
1193    return;
1194  }
1195
1196  // Recursively transform the users of PHI nodes.  This will lazily create the
1197  // PHIs that are needed for individual elements.  Keep track of what PHIs we
1198  // see in InsertedScalarizedValues so that we don't get infinite loops (very
1199  // antisocial).  If the PHI is already in InsertedScalarizedValues, it has
1200  // already been seen first by another load, so its uses have already been
1201  // processed.
1202  PHINode *PN = cast<PHINode>(LoadUser);
1203  bool Inserted;
1204  DenseMap<Value*, std::vector<Value*> >::iterator InsertPos;
1205  tie(InsertPos, Inserted) =
1206    InsertedScalarizedValues.insert(std::make_pair(PN, std::vector<Value*>()));
1207  if (!Inserted) return;
1208
1209  // If this is the first time we've seen this PHI, recursively process all
1210  // users.
1211  for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end(); UI != E; ) {
1212    Instruction *User = cast<Instruction>(*UI++);
1213    RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite);
1214  }
1215}
1216
1217/// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global.  Ptr
1218/// is a value loaded from the global.  Eliminate all uses of Ptr, making them
1219/// use FieldGlobals instead.  All uses of loaded values satisfy
1220/// AllGlobalLoadUsesSimpleEnoughForHeapSRA.
1221static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load,
1222               DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1223                   std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1224  for (Value::use_iterator UI = Load->use_begin(), E = Load->use_end();
1225       UI != E; ) {
1226    Instruction *User = cast<Instruction>(*UI++);
1227    RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite);
1228  }
1229
1230  if (Load->use_empty()) {
1231    Load->eraseFromParent();
1232    InsertedScalarizedValues.erase(Load);
1233  }
1234}
1235
1236/// PerformHeapAllocSRoA - MI is an allocation of an array of structures.  Break
1237/// it up into multiple allocations of arrays of the fields.
1238static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, MallocInst *MI){
1239  DOUT << "SROA HEAP ALLOC: " << *GV << "  MALLOC = " << *MI;
1240  const StructType *STy = cast<StructType>(MI->getAllocatedType());
1241
1242  // There is guaranteed to be at least one use of the malloc (storing
1243  // it into GV).  If there are other uses, change them to be uses of
1244  // the global to simplify later code.  This also deletes the store
1245  // into GV.
1246  ReplaceUsesOfMallocWithGlobal(MI, GV);
1247
1248  // Okay, at this point, there are no users of the malloc.  Insert N
1249  // new mallocs at the same place as MI, and N globals.
1250  std::vector<Value*> FieldGlobals;
1251  std::vector<MallocInst*> FieldMallocs;
1252
1253  for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){
1254    const Type *FieldTy = STy->getElementType(FieldNo);
1255    const Type *PFieldTy = PointerType::getUnqual(FieldTy);
1256
1257    GlobalVariable *NGV =
1258      new GlobalVariable(PFieldTy, false, GlobalValue::InternalLinkage,
1259                         Constant::getNullValue(PFieldTy),
1260                         GV->getName() + ".f" + utostr(FieldNo), GV,
1261                         GV->isThreadLocal());
1262    FieldGlobals.push_back(NGV);
1263
1264    MallocInst *NMI = new MallocInst(FieldTy, MI->getArraySize(),
1265                                     MI->getName() + ".f" + utostr(FieldNo),MI);
1266    FieldMallocs.push_back(NMI);
1267    new StoreInst(NMI, NGV, MI);
1268  }
1269
1270  // The tricky aspect of this transformation is handling the case when malloc
1271  // fails.  In the original code, malloc failing would set the result pointer
1272  // of malloc to null.  In this case, some mallocs could succeed and others
1273  // could fail.  As such, we emit code that looks like this:
1274  //    F0 = malloc(field0)
1275  //    F1 = malloc(field1)
1276  //    F2 = malloc(field2)
1277  //    if (F0 == 0 || F1 == 0 || F2 == 0) {
1278  //      if (F0) { free(F0); F0 = 0; }
1279  //      if (F1) { free(F1); F1 = 0; }
1280  //      if (F2) { free(F2); F2 = 0; }
1281  //    }
1282  Value *RunningOr = 0;
1283  for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) {
1284    Value *Cond = new ICmpInst(ICmpInst::ICMP_EQ, FieldMallocs[i],
1285                             Constant::getNullValue(FieldMallocs[i]->getType()),
1286                                  "isnull", MI);
1287    if (!RunningOr)
1288      RunningOr = Cond;   // First seteq
1289    else
1290      RunningOr = BinaryOperator::CreateOr(RunningOr, Cond, "tmp", MI);
1291  }
1292
1293  // Split the basic block at the old malloc.
1294  BasicBlock *OrigBB = MI->getParent();
1295  BasicBlock *ContBB = OrigBB->splitBasicBlock(MI, "malloc_cont");
1296
1297  // Create the block to check the first condition.  Put all these blocks at the
1298  // end of the function as they are unlikely to be executed.
1299  BasicBlock *NullPtrBlock = BasicBlock::Create("malloc_ret_null",
1300                                                OrigBB->getParent());
1301
1302  // Remove the uncond branch from OrigBB to ContBB, turning it into a cond
1303  // branch on RunningOr.
1304  OrigBB->getTerminator()->eraseFromParent();
1305  BranchInst::Create(NullPtrBlock, ContBB, RunningOr, OrigBB);
1306
1307  // Within the NullPtrBlock, we need to emit a comparison and branch for each
1308  // pointer, because some may be null while others are not.
1309  for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1310    Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock);
1311    Value *Cmp = new ICmpInst(ICmpInst::ICMP_NE, GVVal,
1312                              Constant::getNullValue(GVVal->getType()),
1313                              "tmp", NullPtrBlock);
1314    BasicBlock *FreeBlock = BasicBlock::Create("free_it", OrigBB->getParent());
1315    BasicBlock *NextBlock = BasicBlock::Create("next", OrigBB->getParent());
1316    BranchInst::Create(FreeBlock, NextBlock, Cmp, NullPtrBlock);
1317
1318    // Fill in FreeBlock.
1319    new FreeInst(GVVal, FreeBlock);
1320    new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i],
1321                  FreeBlock);
1322    BranchInst::Create(NextBlock, FreeBlock);
1323
1324    NullPtrBlock = NextBlock;
1325  }
1326
1327  BranchInst::Create(ContBB, NullPtrBlock);
1328
1329  // MI is no longer needed, remove it.
1330  MI->eraseFromParent();
1331
1332  /// InsertedScalarizedLoads - As we process loads, if we can't immediately
1333  /// update all uses of the load, keep track of what scalarized loads are
1334  /// inserted for a given load.
1335  DenseMap<Value*, std::vector<Value*> > InsertedScalarizedValues;
1336  InsertedScalarizedValues[GV] = FieldGlobals;
1337
1338  std::vector<std::pair<PHINode*, unsigned> > PHIsToRewrite;
1339
1340  // Okay, the malloc site is completely handled.  All of the uses of GV are now
1341  // loads, and all uses of those loads are simple.  Rewrite them to use loads
1342  // of the per-field globals instead.
1343  for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;) {
1344    Instruction *User = cast<Instruction>(*UI++);
1345
1346    if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1347      RewriteUsesOfLoadForHeapSRoA(LI, InsertedScalarizedValues, PHIsToRewrite);
1348      continue;
1349    }
1350
1351    // Must be a store of null.
1352    StoreInst *SI = cast<StoreInst>(User);
1353    assert(isa<ConstantPointerNull>(SI->getOperand(0)) &&
1354           "Unexpected heap-sra user!");
1355
1356    // Insert a store of null into each global.
1357    for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1358      const PointerType *PT = cast<PointerType>(FieldGlobals[i]->getType());
1359      Constant *Null = Constant::getNullValue(PT->getElementType());
1360      new StoreInst(Null, FieldGlobals[i], SI);
1361    }
1362    // Erase the original store.
1363    SI->eraseFromParent();
1364  }
1365
1366  // While we have PHIs that are interesting to rewrite, do it.
1367  while (!PHIsToRewrite.empty()) {
1368    PHINode *PN = PHIsToRewrite.back().first;
1369    unsigned FieldNo = PHIsToRewrite.back().second;
1370    PHIsToRewrite.pop_back();
1371    PHINode *FieldPN = cast<PHINode>(InsertedScalarizedValues[PN][FieldNo]);
1372    assert(FieldPN->getNumIncomingValues() == 0 &&"Already processed this phi");
1373
1374    // Add all the incoming values.  This can materialize more phis.
1375    for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1376      Value *InVal = PN->getIncomingValue(i);
1377      InVal = GetHeapSROAValue(InVal, FieldNo, InsertedScalarizedValues,
1378                               PHIsToRewrite);
1379      FieldPN->addIncoming(InVal, PN->getIncomingBlock(i));
1380    }
1381  }
1382
1383  // Drop all inter-phi links and any loads that made it this far.
1384  for (DenseMap<Value*, std::vector<Value*> >::iterator
1385       I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1386       I != E; ++I) {
1387    if (PHINode *PN = dyn_cast<PHINode>(I->first))
1388      PN->dropAllReferences();
1389    else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1390      LI->dropAllReferences();
1391  }
1392
1393  // Delete all the phis and loads now that inter-references are dead.
1394  for (DenseMap<Value*, std::vector<Value*> >::iterator
1395       I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1396       I != E; ++I) {
1397    if (PHINode *PN = dyn_cast<PHINode>(I->first))
1398      PN->eraseFromParent();
1399    else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1400      LI->eraseFromParent();
1401  }
1402
1403  // The old global is now dead, remove it.
1404  GV->eraseFromParent();
1405
1406  ++NumHeapSRA;
1407  return cast<GlobalVariable>(FieldGlobals[0]);
1408}
1409
1410/// TryToOptimizeStoreOfMallocToGlobal - This function is called when we see a
1411/// pointer global variable with a single value stored it that is a malloc or
1412/// cast of malloc.
1413static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV,
1414                                               MallocInst *MI,
1415                                               Module::global_iterator &GVI,
1416                                               TargetData &TD) {
1417  // If this is a malloc of an abstract type, don't touch it.
1418  if (!MI->getAllocatedType()->isSized())
1419    return false;
1420
1421  // We can't optimize this global unless all uses of it are *known* to be
1422  // of the malloc value, not of the null initializer value (consider a use
1423  // that compares the global's value against zero to see if the malloc has
1424  // been reached).  To do this, we check to see if all uses of the global
1425  // would trap if the global were null: this proves that they must all
1426  // happen after the malloc.
1427  if (!AllUsesOfLoadedValueWillTrapIfNull(GV))
1428    return false;
1429
1430  // We can't optimize this if the malloc itself is used in a complex way,
1431  // for example, being stored into multiple globals.  This allows the
1432  // malloc to be stored into the specified global, loaded setcc'd, and
1433  // GEP'd.  These are all things we could transform to using the global
1434  // for.
1435  {
1436    SmallPtrSet<PHINode*, 8> PHIs;
1437    if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(MI, GV, PHIs))
1438      return false;
1439  }
1440
1441
1442  // If we have a global that is only initialized with a fixed size malloc,
1443  // transform the program to use global memory instead of malloc'd memory.
1444  // This eliminates dynamic allocation, avoids an indirection accessing the
1445  // data, and exposes the resultant global to further GlobalOpt.
1446  if (ConstantInt *NElements = dyn_cast<ConstantInt>(MI->getArraySize())) {
1447    // Restrict this transformation to only working on small allocations
1448    // (2048 bytes currently), as we don't want to introduce a 16M global or
1449    // something.
1450    if (NElements->getZExtValue()*
1451        TD.getTypePaddedSize(MI->getAllocatedType()) < 2048) {
1452      GVI = OptimizeGlobalAddressOfMalloc(GV, MI);
1453      return true;
1454    }
1455  }
1456
1457  // If the allocation is an array of structures, consider transforming this
1458  // into multiple malloc'd arrays, one for each field.  This is basically
1459  // SRoA for malloc'd memory.
1460  const Type *AllocTy = MI->getAllocatedType();
1461
1462  // If this is an allocation of a fixed size array of structs, analyze as a
1463  // variable size array.  malloc [100 x struct],1 -> malloc struct, 100
1464  if (!MI->isArrayAllocation())
1465    if (const ArrayType *AT = dyn_cast<ArrayType>(AllocTy))
1466      AllocTy = AT->getElementType();
1467
1468  if (const StructType *AllocSTy = dyn_cast<StructType>(AllocTy)) {
1469    // This the structure has an unreasonable number of fields, leave it
1470    // alone.
1471    if (AllocSTy->getNumElements() <= 16 && AllocSTy->getNumElements() != 0 &&
1472        AllGlobalLoadUsesSimpleEnoughForHeapSRA(GV, MI)) {
1473
1474      // If this is a fixed size array, transform the Malloc to be an alloc of
1475      // structs.  malloc [100 x struct],1 -> malloc struct, 100
1476      if (const ArrayType *AT = dyn_cast<ArrayType>(MI->getAllocatedType())) {
1477        MallocInst *NewMI =
1478          new MallocInst(AllocSTy,
1479                         ConstantInt::get(Type::Int32Ty, AT->getNumElements()),
1480                         "", MI);
1481        NewMI->takeName(MI);
1482        Value *Cast = new BitCastInst(NewMI, MI->getType(), "tmp", MI);
1483        MI->replaceAllUsesWith(Cast);
1484        MI->eraseFromParent();
1485        MI = NewMI;
1486      }
1487
1488      GVI = PerformHeapAllocSRoA(GV, MI);
1489      return true;
1490    }
1491  }
1492
1493  return false;
1494}
1495
1496// OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge
1497// that only one value (besides its initializer) is ever stored to the global.
1498static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal,
1499                                     Module::global_iterator &GVI,
1500                                     TargetData &TD) {
1501  // Ignore no-op GEPs and bitcasts.
1502  StoredOnceVal = StoredOnceVal->stripPointerCasts();
1503
1504  // If we are dealing with a pointer global that is initialized to null and
1505  // only has one (non-null) value stored into it, then we can optimize any
1506  // users of the loaded value (often calls and loads) that would trap if the
1507  // value was null.
1508  if (isa<PointerType>(GV->getInitializer()->getType()) &&
1509      GV->getInitializer()->isNullValue()) {
1510    if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) {
1511      if (GV->getInitializer()->getType() != SOVC->getType())
1512        SOVC = ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType());
1513
1514      // Optimize away any trapping uses of the loaded value.
1515      if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC))
1516        return true;
1517    } else if (MallocInst *MI = dyn_cast<MallocInst>(StoredOnceVal)) {
1518      if (TryToOptimizeStoreOfMallocToGlobal(GV, MI, GVI, TD))
1519        return true;
1520    }
1521  }
1522
1523  return false;
1524}
1525
1526/// TryToShrinkGlobalToBoolean - At this point, we have learned that the only
1527/// two values ever stored into GV are its initializer and OtherVal.  See if we
1528/// can shrink the global into a boolean and select between the two values
1529/// whenever it is used.  This exposes the values to other scalar optimizations.
1530static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) {
1531  const Type *GVElType = GV->getType()->getElementType();
1532
1533  // If GVElType is already i1, it is already shrunk.  If the type of the GV is
1534  // an FP value or vector, don't do this optimization because a select between
1535  // them is very expensive and unlikely to lead to later simplification.
1536  if (GVElType == Type::Int1Ty || GVElType->isFloatingPoint() ||
1537      isa<VectorType>(GVElType))
1538    return false;
1539
1540  // Walk the use list of the global seeing if all the uses are load or store.
1541  // If there is anything else, bail out.
1542  for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I)
1543    if (!isa<LoadInst>(I) && !isa<StoreInst>(I))
1544      return false;
1545
1546  DOUT << "   *** SHRINKING TO BOOL: " << *GV;
1547
1548  // Create the new global, initializing it to false.
1549  GlobalVariable *NewGV = new GlobalVariable(Type::Int1Ty, false,
1550         GlobalValue::InternalLinkage, ConstantInt::getFalse(),
1551                                             GV->getName()+".b",
1552                                             (Module *)NULL,
1553                                             GV->isThreadLocal());
1554  GV->getParent()->getGlobalList().insert(GV, NewGV);
1555
1556  Constant *InitVal = GV->getInitializer();
1557  assert(InitVal->getType() != Type::Int1Ty && "No reason to shrink to bool!");
1558
1559  // If initialized to zero and storing one into the global, we can use a cast
1560  // instead of a select to synthesize the desired value.
1561  bool IsOneZero = false;
1562  if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal))
1563    IsOneZero = InitVal->isNullValue() && CI->isOne();
1564
1565  while (!GV->use_empty()) {
1566    Instruction *UI = cast<Instruction>(GV->use_back());
1567    if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
1568      // Change the store into a boolean store.
1569      bool StoringOther = SI->getOperand(0) == OtherVal;
1570      // Only do this if we weren't storing a loaded value.
1571      Value *StoreVal;
1572      if (StoringOther || SI->getOperand(0) == InitVal)
1573        StoreVal = ConstantInt::get(Type::Int1Ty, StoringOther);
1574      else {
1575        // Otherwise, we are storing a previously loaded copy.  To do this,
1576        // change the copy from copying the original value to just copying the
1577        // bool.
1578        Instruction *StoredVal = cast<Instruction>(SI->getOperand(0));
1579
1580        // If we're already replaced the input, StoredVal will be a cast or
1581        // select instruction.  If not, it will be a load of the original
1582        // global.
1583        if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
1584          assert(LI->getOperand(0) == GV && "Not a copy!");
1585          // Insert a new load, to preserve the saved value.
1586          StoreVal = new LoadInst(NewGV, LI->getName()+".b", LI);
1587        } else {
1588          assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) &&
1589                 "This is not a form that we understand!");
1590          StoreVal = StoredVal->getOperand(0);
1591          assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!");
1592        }
1593      }
1594      new StoreInst(StoreVal, NewGV, SI);
1595    } else {
1596      // Change the load into a load of bool then a select.
1597      LoadInst *LI = cast<LoadInst>(UI);
1598      LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", LI);
1599      Value *NSI;
1600      if (IsOneZero)
1601        NSI = new ZExtInst(NLI, LI->getType(), "", LI);
1602      else
1603        NSI = SelectInst::Create(NLI, OtherVal, InitVal, "", LI);
1604      NSI->takeName(LI);
1605      LI->replaceAllUsesWith(NSI);
1606    }
1607    UI->eraseFromParent();
1608  }
1609
1610  GV->eraseFromParent();
1611  return true;
1612}
1613
1614
1615/// ProcessInternalGlobal - Analyze the specified global variable and optimize
1616/// it if possible.  If we make a change, return true.
1617bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV,
1618                                      Module::global_iterator &GVI) {
1619  SmallPtrSet<PHINode*, 16> PHIUsers;
1620  GlobalStatus GS;
1621  GV->removeDeadConstantUsers();
1622
1623  if (GV->use_empty()) {
1624    DOUT << "GLOBAL DEAD: " << *GV;
1625    GV->eraseFromParent();
1626    ++NumDeleted;
1627    return true;
1628  }
1629
1630  if (!AnalyzeGlobal(GV, GS, PHIUsers)) {
1631#if 0
1632    cerr << "Global: " << *GV;
1633    cerr << "  isLoaded = " << GS.isLoaded << "\n";
1634    cerr << "  StoredType = ";
1635    switch (GS.StoredType) {
1636    case GlobalStatus::NotStored: cerr << "NEVER STORED\n"; break;
1637    case GlobalStatus::isInitializerStored: cerr << "INIT STORED\n"; break;
1638    case GlobalStatus::isStoredOnce: cerr << "STORED ONCE\n"; break;
1639    case GlobalStatus::isStored: cerr << "stored\n"; break;
1640    }
1641    if (GS.StoredType == GlobalStatus::isStoredOnce && GS.StoredOnceValue)
1642      cerr << "  StoredOnceValue = " << *GS.StoredOnceValue << "\n";
1643    if (GS.AccessingFunction && !GS.HasMultipleAccessingFunctions)
1644      cerr << "  AccessingFunction = " << GS.AccessingFunction->getName()
1645                << "\n";
1646    cerr << "  HasMultipleAccessingFunctions =  "
1647              << GS.HasMultipleAccessingFunctions << "\n";
1648    cerr << "  HasNonInstructionUser = " << GS.HasNonInstructionUser<<"\n";
1649    cerr << "\n";
1650#endif
1651
1652    // If this is a first class global and has only one accessing function
1653    // and this function is main (which we know is not recursive we can make
1654    // this global a local variable) we replace the global with a local alloca
1655    // in this function.
1656    //
1657    // NOTE: It doesn't make sense to promote non single-value types since we
1658    // are just replacing static memory to stack memory.
1659    if (!GS.HasMultipleAccessingFunctions &&
1660        GS.AccessingFunction && !GS.HasNonInstructionUser &&
1661        GV->getType()->getElementType()->isSingleValueType() &&
1662        GS.AccessingFunction->getName() == "main" &&
1663        GS.AccessingFunction->hasExternalLinkage()) {
1664      DOUT << "LOCALIZING GLOBAL: " << *GV;
1665      Instruction* FirstI = GS.AccessingFunction->getEntryBlock().begin();
1666      const Type* ElemTy = GV->getType()->getElementType();
1667      // FIXME: Pass Global's alignment when globals have alignment
1668      AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), FirstI);
1669      if (!isa<UndefValue>(GV->getInitializer()))
1670        new StoreInst(GV->getInitializer(), Alloca, FirstI);
1671
1672      GV->replaceAllUsesWith(Alloca);
1673      GV->eraseFromParent();
1674      ++NumLocalized;
1675      return true;
1676    }
1677
1678    // If the global is never loaded (but may be stored to), it is dead.
1679    // Delete it now.
1680    if (!GS.isLoaded) {
1681      DOUT << "GLOBAL NEVER LOADED: " << *GV;
1682
1683      // Delete any stores we can find to the global.  We may not be able to
1684      // make it completely dead though.
1685      bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer());
1686
1687      // If the global is dead now, delete it.
1688      if (GV->use_empty()) {
1689        GV->eraseFromParent();
1690        ++NumDeleted;
1691        Changed = true;
1692      }
1693      return Changed;
1694
1695    } else if (GS.StoredType <= GlobalStatus::isInitializerStored) {
1696      DOUT << "MARKING CONSTANT: " << *GV;
1697      GV->setConstant(true);
1698
1699      // Clean up any obviously simplifiable users now.
1700      CleanupConstantGlobalUsers(GV, GV->getInitializer());
1701
1702      // If the global is dead now, just nuke it.
1703      if (GV->use_empty()) {
1704        DOUT << "   *** Marking constant allowed us to simplify "
1705             << "all users and delete global!\n";
1706        GV->eraseFromParent();
1707        ++NumDeleted;
1708      }
1709
1710      ++NumMarked;
1711      return true;
1712    } else if (!GV->getInitializer()->getType()->isSingleValueType()) {
1713      if (GlobalVariable *FirstNewGV = SRAGlobal(GV,
1714                                                 getAnalysis<TargetData>())) {
1715        GVI = FirstNewGV;  // Don't skip the newly produced globals!
1716        return true;
1717      }
1718    } else if (GS.StoredType == GlobalStatus::isStoredOnce) {
1719      // If the initial value for the global was an undef value, and if only
1720      // one other value was stored into it, we can just change the
1721      // initializer to be the stored value, then delete all stores to the
1722      // global.  This allows us to mark it constant.
1723      if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1724        if (isa<UndefValue>(GV->getInitializer())) {
1725          // Change the initial value here.
1726          GV->setInitializer(SOVConstant);
1727
1728          // Clean up any obviously simplifiable users now.
1729          CleanupConstantGlobalUsers(GV, GV->getInitializer());
1730
1731          if (GV->use_empty()) {
1732            DOUT << "   *** Substituting initializer allowed us to "
1733                 << "simplify all users and delete global!\n";
1734            GV->eraseFromParent();
1735            ++NumDeleted;
1736          } else {
1737            GVI = GV;
1738          }
1739          ++NumSubstitute;
1740          return true;
1741        }
1742
1743      // Try to optimize globals based on the knowledge that only one value
1744      // (besides its initializer) is ever stored to the global.
1745      if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI,
1746                                   getAnalysis<TargetData>()))
1747        return true;
1748
1749      // Otherwise, if the global was not a boolean, we can shrink it to be a
1750      // boolean.
1751      if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1752        if (TryToShrinkGlobalToBoolean(GV, SOVConstant)) {
1753          ++NumShrunkToBool;
1754          return true;
1755        }
1756    }
1757  }
1758  return false;
1759}
1760
1761/// OnlyCalledDirectly - Return true if the specified function is only called
1762/// directly.  In other words, its address is never taken.
1763static bool OnlyCalledDirectly(Function *F) {
1764  for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1765    Instruction *User = dyn_cast<Instruction>(*UI);
1766    if (!User) return false;
1767    if (!isa<CallInst>(User) && !isa<InvokeInst>(User)) return false;
1768
1769    // See if the function address is passed as an argument.
1770    for (User::op_iterator i = User->op_begin() + 1, e = User->op_end();
1771         i != e; ++i)
1772      if (*i == F) return false;
1773  }
1774  return true;
1775}
1776
1777/// ChangeCalleesToFastCall - Walk all of the direct calls of the specified
1778/// function, changing them to FastCC.
1779static void ChangeCalleesToFastCall(Function *F) {
1780  for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1781    CallSite User(cast<Instruction>(*UI));
1782    User.setCallingConv(CallingConv::Fast);
1783  }
1784}
1785
1786static AttrListPtr StripNest(const AttrListPtr &Attrs) {
1787  for (unsigned i = 0, e = Attrs.getNumSlots(); i != e; ++i) {
1788    if ((Attrs.getSlot(i).Attrs & Attribute::Nest) == 0)
1789      continue;
1790
1791    // There can be only one.
1792    return Attrs.removeAttr(Attrs.getSlot(i).Index, Attribute::Nest);
1793  }
1794
1795  return Attrs;
1796}
1797
1798static void RemoveNestAttribute(Function *F) {
1799  F->setAttributes(StripNest(F->getAttributes()));
1800  for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1801    CallSite User(cast<Instruction>(*UI));
1802    User.setAttributes(StripNest(User.getAttributes()));
1803  }
1804}
1805
1806bool GlobalOpt::OptimizeFunctions(Module &M) {
1807  bool Changed = false;
1808  // Optimize functions.
1809  for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) {
1810    Function *F = FI++;
1811    F->removeDeadConstantUsers();
1812    if (F->use_empty() && (F->hasLocalLinkage() ||
1813                           F->hasLinkOnceLinkage())) {
1814      M.getFunctionList().erase(F);
1815      Changed = true;
1816      ++NumFnDeleted;
1817    } else if (F->hasLocalLinkage()) {
1818      if (F->getCallingConv() == CallingConv::C && !F->isVarArg() &&
1819          OnlyCalledDirectly(F)) {
1820        // If this function has C calling conventions, is not a varargs
1821        // function, and is only called directly, promote it to use the Fast
1822        // calling convention.
1823        F->setCallingConv(CallingConv::Fast);
1824        ChangeCalleesToFastCall(F);
1825        ++NumFastCallFns;
1826        Changed = true;
1827      }
1828
1829      if (F->getAttributes().hasAttrSomewhere(Attribute::Nest) &&
1830          OnlyCalledDirectly(F)) {
1831        // The function is not used by a trampoline intrinsic, so it is safe
1832        // to remove the 'nest' attribute.
1833        RemoveNestAttribute(F);
1834        ++NumNestRemoved;
1835        Changed = true;
1836      }
1837    }
1838  }
1839  return Changed;
1840}
1841
1842bool GlobalOpt::OptimizeGlobalVars(Module &M) {
1843  bool Changed = false;
1844  for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
1845       GVI != E; ) {
1846    GlobalVariable *GV = GVI++;
1847    if (!GV->isConstant() && GV->hasLocalLinkage() &&
1848        GV->hasInitializer())
1849      Changed |= ProcessInternalGlobal(GV, GVI);
1850  }
1851  return Changed;
1852}
1853
1854/// FindGlobalCtors - Find the llvm.globalctors list, verifying that all
1855/// initializers have an init priority of 65535.
1856GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) {
1857  for (Module::global_iterator I = M.global_begin(), E = M.global_end();
1858       I != E; ++I)
1859    if (I->getName() == "llvm.global_ctors") {
1860      // Found it, verify it's an array of { int, void()* }.
1861      const ArrayType *ATy =dyn_cast<ArrayType>(I->getType()->getElementType());
1862      if (!ATy) return 0;
1863      const StructType *STy = dyn_cast<StructType>(ATy->getElementType());
1864      if (!STy || STy->getNumElements() != 2 ||
1865          STy->getElementType(0) != Type::Int32Ty) return 0;
1866      const PointerType *PFTy = dyn_cast<PointerType>(STy->getElementType(1));
1867      if (!PFTy) return 0;
1868      const FunctionType *FTy = dyn_cast<FunctionType>(PFTy->getElementType());
1869      if (!FTy || FTy->getReturnType() != Type::VoidTy || FTy->isVarArg() ||
1870          FTy->getNumParams() != 0)
1871        return 0;
1872
1873      // Verify that the initializer is simple enough for us to handle.
1874      if (!I->hasInitializer()) return 0;
1875      ConstantArray *CA = dyn_cast<ConstantArray>(I->getInitializer());
1876      if (!CA) return 0;
1877      for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i)
1878        if (ConstantStruct *CS = dyn_cast<ConstantStruct>(*i)) {
1879          if (isa<ConstantPointerNull>(CS->getOperand(1)))
1880            continue;
1881
1882          // Must have a function or null ptr.
1883          if (!isa<Function>(CS->getOperand(1)))
1884            return 0;
1885
1886          // Init priority must be standard.
1887          ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0));
1888          if (!CI || CI->getZExtValue() != 65535)
1889            return 0;
1890        } else {
1891          return 0;
1892        }
1893
1894      return I;
1895    }
1896  return 0;
1897}
1898
1899/// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand,
1900/// return a list of the functions and null terminator as a vector.
1901static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) {
1902  ConstantArray *CA = cast<ConstantArray>(GV->getInitializer());
1903  std::vector<Function*> Result;
1904  Result.reserve(CA->getNumOperands());
1905  for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) {
1906    ConstantStruct *CS = cast<ConstantStruct>(*i);
1907    Result.push_back(dyn_cast<Function>(CS->getOperand(1)));
1908  }
1909  return Result;
1910}
1911
1912/// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the
1913/// specified array, returning the new global to use.
1914static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL,
1915                                          const std::vector<Function*> &Ctors) {
1916  // If we made a change, reassemble the initializer list.
1917  std::vector<Constant*> CSVals;
1918  CSVals.push_back(ConstantInt::get(Type::Int32Ty, 65535));
1919  CSVals.push_back(0);
1920
1921  // Create the new init list.
1922  std::vector<Constant*> CAList;
1923  for (unsigned i = 0, e = Ctors.size(); i != e; ++i) {
1924    if (Ctors[i]) {
1925      CSVals[1] = Ctors[i];
1926    } else {
1927      const Type *FTy = FunctionType::get(Type::VoidTy,
1928                                          std::vector<const Type*>(), false);
1929      const PointerType *PFTy = PointerType::getUnqual(FTy);
1930      CSVals[1] = Constant::getNullValue(PFTy);
1931      CSVals[0] = ConstantInt::get(Type::Int32Ty, 2147483647);
1932    }
1933    CAList.push_back(ConstantStruct::get(CSVals));
1934  }
1935
1936  // Create the array initializer.
1937  const Type *StructTy =
1938    cast<ArrayType>(GCL->getType()->getElementType())->getElementType();
1939  Constant *CA = ConstantArray::get(ArrayType::get(StructTy, CAList.size()),
1940                                    CAList);
1941
1942  // If we didn't change the number of elements, don't create a new GV.
1943  if (CA->getType() == GCL->getInitializer()->getType()) {
1944    GCL->setInitializer(CA);
1945    return GCL;
1946  }
1947
1948  // Create the new global and insert it next to the existing list.
1949  GlobalVariable *NGV = new GlobalVariable(CA->getType(), GCL->isConstant(),
1950                                           GCL->getLinkage(), CA, "",
1951                                           (Module *)NULL,
1952                                           GCL->isThreadLocal());
1953  GCL->getParent()->getGlobalList().insert(GCL, NGV);
1954  NGV->takeName(GCL);
1955
1956  // Nuke the old list, replacing any uses with the new one.
1957  if (!GCL->use_empty()) {
1958    Constant *V = NGV;
1959    if (V->getType() != GCL->getType())
1960      V = ConstantExpr::getBitCast(V, GCL->getType());
1961    GCL->replaceAllUsesWith(V);
1962  }
1963  GCL->eraseFromParent();
1964
1965  if (Ctors.size())
1966    return NGV;
1967  else
1968    return 0;
1969}
1970
1971
1972static Constant *getVal(DenseMap<Value*, Constant*> &ComputedValues,
1973                        Value *V) {
1974  if (Constant *CV = dyn_cast<Constant>(V)) return CV;
1975  Constant *R = ComputedValues[V];
1976  assert(R && "Reference to an uncomputed value!");
1977  return R;
1978}
1979
1980/// isSimpleEnoughPointerToCommit - Return true if this constant is simple
1981/// enough for us to understand.  In particular, if it is a cast of something,
1982/// we punt.  We basically just support direct accesses to globals and GEP's of
1983/// globals.  This should be kept up to date with CommitValueTo.
1984static bool isSimpleEnoughPointerToCommit(Constant *C) {
1985  if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) {
1986    if (!GV->hasExternalLinkage() && !GV->hasLocalLinkage())
1987      return false;  // do not allow weak/linkonce/dllimport/dllexport linkage.
1988    return !GV->isDeclaration();  // reject external globals.
1989  }
1990  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
1991    // Handle a constantexpr gep.
1992    if (CE->getOpcode() == Instruction::GetElementPtr &&
1993        isa<GlobalVariable>(CE->getOperand(0))) {
1994      GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1995      if (!GV->hasExternalLinkage() && !GV->hasLocalLinkage())
1996        return false;  // do not allow weak/linkonce/dllimport/dllexport linkage.
1997      return GV->hasInitializer() &&
1998             ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
1999    }
2000  return false;
2001}
2002
2003/// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global
2004/// initializer.  This returns 'Init' modified to reflect 'Val' stored into it.
2005/// At this point, the GEP operands of Addr [0, OpNo) have been stepped into.
2006static Constant *EvaluateStoreInto(Constant *Init, Constant *Val,
2007                                   ConstantExpr *Addr, unsigned OpNo) {
2008  // Base case of the recursion.
2009  if (OpNo == Addr->getNumOperands()) {
2010    assert(Val->getType() == Init->getType() && "Type mismatch!");
2011    return Val;
2012  }
2013
2014  if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
2015    std::vector<Constant*> Elts;
2016
2017    // Break up the constant into its elements.
2018    if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
2019      for (User::op_iterator i = CS->op_begin(), e = CS->op_end(); i != e; ++i)
2020        Elts.push_back(cast<Constant>(*i));
2021    } else if (isa<ConstantAggregateZero>(Init)) {
2022      for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
2023        Elts.push_back(Constant::getNullValue(STy->getElementType(i)));
2024    } else if (isa<UndefValue>(Init)) {
2025      for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
2026        Elts.push_back(UndefValue::get(STy->getElementType(i)));
2027    } else {
2028      assert(0 && "This code is out of sync with "
2029             " ConstantFoldLoadThroughGEPConstantExpr");
2030    }
2031
2032    // Replace the element that we are supposed to.
2033    ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo));
2034    unsigned Idx = CU->getZExtValue();
2035    assert(Idx < STy->getNumElements() && "Struct index out of range!");
2036    Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1);
2037
2038    // Return the modified struct.
2039    return ConstantStruct::get(&Elts[0], Elts.size(), STy->isPacked());
2040  } else {
2041    ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo));
2042    const ArrayType *ATy = cast<ArrayType>(Init->getType());
2043
2044    // Break up the array into elements.
2045    std::vector<Constant*> Elts;
2046    if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
2047      for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i)
2048        Elts.push_back(cast<Constant>(*i));
2049    } else if (isa<ConstantAggregateZero>(Init)) {
2050      Constant *Elt = Constant::getNullValue(ATy->getElementType());
2051      Elts.assign(ATy->getNumElements(), Elt);
2052    } else if (isa<UndefValue>(Init)) {
2053      Constant *Elt = UndefValue::get(ATy->getElementType());
2054      Elts.assign(ATy->getNumElements(), Elt);
2055    } else {
2056      assert(0 && "This code is out of sync with "
2057             " ConstantFoldLoadThroughGEPConstantExpr");
2058    }
2059
2060    assert(CI->getZExtValue() < ATy->getNumElements());
2061    Elts[CI->getZExtValue()] =
2062      EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1);
2063    return ConstantArray::get(ATy, Elts);
2064  }
2065}
2066
2067/// CommitValueTo - We have decided that Addr (which satisfies the predicate
2068/// isSimpleEnoughPointerToCommit) should get Val as its value.  Make it happen.
2069static void CommitValueTo(Constant *Val, Constant *Addr) {
2070  if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
2071    assert(GV->hasInitializer());
2072    GV->setInitializer(Val);
2073    return;
2074  }
2075
2076  ConstantExpr *CE = cast<ConstantExpr>(Addr);
2077  GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2078
2079  Constant *Init = GV->getInitializer();
2080  Init = EvaluateStoreInto(Init, Val, CE, 2);
2081  GV->setInitializer(Init);
2082}
2083
2084/// ComputeLoadResult - Return the value that would be computed by a load from
2085/// P after the stores reflected by 'memory' have been performed.  If we can't
2086/// decide, return null.
2087static Constant *ComputeLoadResult(Constant *P,
2088                                const DenseMap<Constant*, Constant*> &Memory) {
2089  // If this memory location has been recently stored, use the stored value: it
2090  // is the most up-to-date.
2091  DenseMap<Constant*, Constant*>::const_iterator I = Memory.find(P);
2092  if (I != Memory.end()) return I->second;
2093
2094  // Access it.
2095  if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
2096    if (GV->hasInitializer())
2097      return GV->getInitializer();
2098    return 0;
2099  }
2100
2101  // Handle a constantexpr getelementptr.
2102  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P))
2103    if (CE->getOpcode() == Instruction::GetElementPtr &&
2104        isa<GlobalVariable>(CE->getOperand(0))) {
2105      GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2106      if (GV->hasInitializer())
2107        return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
2108    }
2109
2110  return 0;  // don't know how to evaluate.
2111}
2112
2113/// EvaluateFunction - Evaluate a call to function F, returning true if
2114/// successful, false if we can't evaluate it.  ActualArgs contains the formal
2115/// arguments for the function.
2116static bool EvaluateFunction(Function *F, Constant *&RetVal,
2117                             const std::vector<Constant*> &ActualArgs,
2118                             std::vector<Function*> &CallStack,
2119                             DenseMap<Constant*, Constant*> &MutatedMemory,
2120                             std::vector<GlobalVariable*> &AllocaTmps) {
2121  // Check to see if this function is already executing (recursion).  If so,
2122  // bail out.  TODO: we might want to accept limited recursion.
2123  if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end())
2124    return false;
2125
2126  CallStack.push_back(F);
2127
2128  /// Values - As we compute SSA register values, we store their contents here.
2129  DenseMap<Value*, Constant*> Values;
2130
2131  // Initialize arguments to the incoming values specified.
2132  unsigned ArgNo = 0;
2133  for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E;
2134       ++AI, ++ArgNo)
2135    Values[AI] = ActualArgs[ArgNo];
2136
2137  /// ExecutedBlocks - We only handle non-looping, non-recursive code.  As such,
2138  /// we can only evaluate any one basic block at most once.  This set keeps
2139  /// track of what we have executed so we can detect recursive cases etc.
2140  SmallPtrSet<BasicBlock*, 32> ExecutedBlocks;
2141
2142  // CurInst - The current instruction we're evaluating.
2143  BasicBlock::iterator CurInst = F->begin()->begin();
2144
2145  // This is the main evaluation loop.
2146  while (1) {
2147    Constant *InstResult = 0;
2148
2149    if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) {
2150      if (SI->isVolatile()) return false;  // no volatile accesses.
2151      Constant *Ptr = getVal(Values, SI->getOperand(1));
2152      if (!isSimpleEnoughPointerToCommit(Ptr))
2153        // If this is too complex for us to commit, reject it.
2154        return false;
2155      Constant *Val = getVal(Values, SI->getOperand(0));
2156      MutatedMemory[Ptr] = Val;
2157    } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) {
2158      InstResult = ConstantExpr::get(BO->getOpcode(),
2159                                     getVal(Values, BO->getOperand(0)),
2160                                     getVal(Values, BO->getOperand(1)));
2161    } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) {
2162      InstResult = ConstantExpr::getCompare(CI->getPredicate(),
2163                                            getVal(Values, CI->getOperand(0)),
2164                                            getVal(Values, CI->getOperand(1)));
2165    } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) {
2166      InstResult = ConstantExpr::getCast(CI->getOpcode(),
2167                                         getVal(Values, CI->getOperand(0)),
2168                                         CI->getType());
2169    } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) {
2170      InstResult = ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)),
2171                                           getVal(Values, SI->getOperand(1)),
2172                                           getVal(Values, SI->getOperand(2)));
2173    } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) {
2174      Constant *P = getVal(Values, GEP->getOperand(0));
2175      SmallVector<Constant*, 8> GEPOps;
2176      for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end();
2177           i != e; ++i)
2178        GEPOps.push_back(getVal(Values, *i));
2179      InstResult = ConstantExpr::getGetElementPtr(P, &GEPOps[0], GEPOps.size());
2180    } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) {
2181      if (LI->isVolatile()) return false;  // no volatile accesses.
2182      InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)),
2183                                     MutatedMemory);
2184      if (InstResult == 0) return false; // Could not evaluate load.
2185    } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) {
2186      if (AI->isArrayAllocation()) return false;  // Cannot handle array allocs.
2187      const Type *Ty = AI->getType()->getElementType();
2188      AllocaTmps.push_back(new GlobalVariable(Ty, false,
2189                                              GlobalValue::InternalLinkage,
2190                                              UndefValue::get(Ty),
2191                                              AI->getName()));
2192      InstResult = AllocaTmps.back();
2193    } else if (CallInst *CI = dyn_cast<CallInst>(CurInst)) {
2194      // Cannot handle inline asm.
2195      if (isa<InlineAsm>(CI->getOperand(0))) return false;
2196
2197      // Resolve function pointers.
2198      Function *Callee = dyn_cast<Function>(getVal(Values, CI->getOperand(0)));
2199      if (!Callee) return false;  // Cannot resolve.
2200
2201      std::vector<Constant*> Formals;
2202      for (User::op_iterator i = CI->op_begin() + 1, e = CI->op_end();
2203           i != e; ++i)
2204        Formals.push_back(getVal(Values, *i));
2205
2206      if (Callee->isDeclaration()) {
2207        // If this is a function we can constant fold, do it.
2208        if (Constant *C = ConstantFoldCall(Callee, &Formals[0],
2209                                           Formals.size())) {
2210          InstResult = C;
2211        } else {
2212          return false;
2213        }
2214      } else {
2215        if (Callee->getFunctionType()->isVarArg())
2216          return false;
2217
2218        Constant *RetVal;
2219
2220        // Execute the call, if successful, use the return value.
2221        if (!EvaluateFunction(Callee, RetVal, Formals, CallStack,
2222                              MutatedMemory, AllocaTmps))
2223          return false;
2224        InstResult = RetVal;
2225      }
2226    } else if (isa<TerminatorInst>(CurInst)) {
2227      BasicBlock *NewBB = 0;
2228      if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) {
2229        if (BI->isUnconditional()) {
2230          NewBB = BI->getSuccessor(0);
2231        } else {
2232          ConstantInt *Cond =
2233            dyn_cast<ConstantInt>(getVal(Values, BI->getCondition()));
2234          if (!Cond) return false;  // Cannot determine.
2235
2236          NewBB = BI->getSuccessor(!Cond->getZExtValue());
2237        }
2238      } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) {
2239        ConstantInt *Val =
2240          dyn_cast<ConstantInt>(getVal(Values, SI->getCondition()));
2241        if (!Val) return false;  // Cannot determine.
2242        NewBB = SI->getSuccessor(SI->findCaseValue(Val));
2243      } else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) {
2244        if (RI->getNumOperands())
2245          RetVal = getVal(Values, RI->getOperand(0));
2246
2247        CallStack.pop_back();  // return from fn.
2248        return true;  // We succeeded at evaluating this ctor!
2249      } else {
2250        // invoke, unwind, unreachable.
2251        return false;  // Cannot handle this terminator.
2252      }
2253
2254      // Okay, we succeeded in evaluating this control flow.  See if we have
2255      // executed the new block before.  If so, we have a looping function,
2256      // which we cannot evaluate in reasonable time.
2257      if (!ExecutedBlocks.insert(NewBB))
2258        return false;  // looped!
2259
2260      // Okay, we have never been in this block before.  Check to see if there
2261      // are any PHI nodes.  If so, evaluate them with information about where
2262      // we came from.
2263      BasicBlock *OldBB = CurInst->getParent();
2264      CurInst = NewBB->begin();
2265      PHINode *PN;
2266      for (; (PN = dyn_cast<PHINode>(CurInst)); ++CurInst)
2267        Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB));
2268
2269      // Do NOT increment CurInst.  We know that the terminator had no value.
2270      continue;
2271    } else {
2272      // Did not know how to evaluate this!
2273      return false;
2274    }
2275
2276    if (!CurInst->use_empty())
2277      Values[CurInst] = InstResult;
2278
2279    // Advance program counter.
2280    ++CurInst;
2281  }
2282}
2283
2284/// EvaluateStaticConstructor - Evaluate static constructors in the function, if
2285/// we can.  Return true if we can, false otherwise.
2286static bool EvaluateStaticConstructor(Function *F) {
2287  /// MutatedMemory - For each store we execute, we update this map.  Loads
2288  /// check this to get the most up-to-date value.  If evaluation is successful,
2289  /// this state is committed to the process.
2290  DenseMap<Constant*, Constant*> MutatedMemory;
2291
2292  /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable
2293  /// to represent its body.  This vector is needed so we can delete the
2294  /// temporary globals when we are done.
2295  std::vector<GlobalVariable*> AllocaTmps;
2296
2297  /// CallStack - This is used to detect recursion.  In pathological situations
2298  /// we could hit exponential behavior, but at least there is nothing
2299  /// unbounded.
2300  std::vector<Function*> CallStack;
2301
2302  // Call the function.
2303  Constant *RetValDummy;
2304  bool EvalSuccess = EvaluateFunction(F, RetValDummy, std::vector<Constant*>(),
2305                                       CallStack, MutatedMemory, AllocaTmps);
2306  if (EvalSuccess) {
2307    // We succeeded at evaluation: commit the result.
2308    DOUT << "FULLY EVALUATED GLOBAL CTOR FUNCTION '"
2309         << F->getName() << "' to " << MutatedMemory.size()
2310         << " stores.\n";
2311    for (DenseMap<Constant*, Constant*>::iterator I = MutatedMemory.begin(),
2312         E = MutatedMemory.end(); I != E; ++I)
2313      CommitValueTo(I->second, I->first);
2314  }
2315
2316  // At this point, we are done interpreting.  If we created any 'alloca'
2317  // temporaries, release them now.
2318  while (!AllocaTmps.empty()) {
2319    GlobalVariable *Tmp = AllocaTmps.back();
2320    AllocaTmps.pop_back();
2321
2322    // If there are still users of the alloca, the program is doing something
2323    // silly, e.g. storing the address of the alloca somewhere and using it
2324    // later.  Since this is undefined, we'll just make it be null.
2325    if (!Tmp->use_empty())
2326      Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType()));
2327    delete Tmp;
2328  }
2329
2330  return EvalSuccess;
2331}
2332
2333
2334
2335/// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible.
2336/// Return true if anything changed.
2337bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) {
2338  std::vector<Function*> Ctors = ParseGlobalCtors(GCL);
2339  bool MadeChange = false;
2340  if (Ctors.empty()) return false;
2341
2342  // Loop over global ctors, optimizing them when we can.
2343  for (unsigned i = 0; i != Ctors.size(); ++i) {
2344    Function *F = Ctors[i];
2345    // Found a null terminator in the middle of the list, prune off the rest of
2346    // the list.
2347    if (F == 0) {
2348      if (i != Ctors.size()-1) {
2349        Ctors.resize(i+1);
2350        MadeChange = true;
2351      }
2352      break;
2353    }
2354
2355    // We cannot simplify external ctor functions.
2356    if (F->empty()) continue;
2357
2358    // If we can evaluate the ctor at compile time, do.
2359    if (EvaluateStaticConstructor(F)) {
2360      Ctors.erase(Ctors.begin()+i);
2361      MadeChange = true;
2362      --i;
2363      ++NumCtorsEvaluated;
2364      continue;
2365    }
2366  }
2367
2368  if (!MadeChange) return false;
2369
2370  GCL = InstallGlobalCtors(GCL, Ctors);
2371  return true;
2372}
2373
2374bool GlobalOpt::ResolveAliases(Module &M) {
2375  bool Changed = false;
2376
2377  for (Module::alias_iterator I = M.alias_begin(), E = M.alias_end();
2378       I != E;) {
2379    Module::alias_iterator J = I++;
2380    // If the aliasee may change at link time, nothing can be done - bail out.
2381    if (J->mayBeOverridden())
2382      continue;
2383
2384    Constant *Aliasee = J->getAliasee();
2385    GlobalValue *Target = cast<GlobalValue>(Aliasee->stripPointerCasts());
2386    bool hasOneUse = Target->hasOneUse() && Aliasee->hasOneUse();
2387
2388    // Make all users of the alias use the aliasee instead.
2389    if (!J->use_empty()) {
2390      J->replaceAllUsesWith(Aliasee);
2391      ++NumAliasesResolved;
2392      Changed = true;
2393    }
2394
2395    // If the aliasee has internal linkage, give it the name and linkage
2396    // of the alias, and delete the alias.  This turns:
2397    //   define internal ... @f(...)
2398    //   @a = alias ... @f
2399    // into:
2400    //   define ... @a(...)
2401    if (!Target->hasInternalLinkage())
2402      continue;
2403
2404    // The transform is only useful if the alias does not have internal linkage.
2405    if (J->hasInternalLinkage())
2406      continue;
2407
2408    // Be conservative and do not perform the transform if multiple aliases
2409    // potentially target the aliasee.  TODO: Make this more aggressive.
2410    if (!hasOneUse)
2411      continue;
2412
2413    // Do not perform the transform if it would change the visibility.
2414    if (J->getVisibility() != Target->getVisibility())
2415      continue;
2416
2417    // Do not perform the transform if it would change the section.
2418    if (J->hasSection() != Target->hasSection() ||
2419        (J->hasSection() && J->getSection() != Target->getSection()))
2420      continue;
2421
2422    // Give the aliasee the name and linkage of the alias.
2423    Target->takeName(J);
2424    Target->setLinkage(J->getLinkage());
2425
2426    // The alignment is the only remaining attribute that may not match.
2427    Target->setAlignment(std::max(J->getAlignment(), Target->getAlignment()));
2428
2429    // Delete the alias.
2430    M.getAliasList().erase(J);
2431    ++NumAliasesRemoved;
2432    Changed = true;
2433  }
2434
2435  return Changed;
2436}
2437
2438bool GlobalOpt::runOnModule(Module &M) {
2439  bool Changed = false;
2440
2441  // Try to find the llvm.globalctors list.
2442  GlobalVariable *GlobalCtors = FindGlobalCtors(M);
2443
2444  bool LocalChange = true;
2445  while (LocalChange) {
2446    LocalChange = false;
2447
2448    // Delete functions that are trivially dead, ccc -> fastcc
2449    LocalChange |= OptimizeFunctions(M);
2450
2451    // Optimize global_ctors list.
2452    if (GlobalCtors)
2453      LocalChange |= OptimizeGlobalCtorsList(GlobalCtors);
2454
2455    // Optimize non-address-taken globals.
2456    LocalChange |= OptimizeGlobalVars(M);
2457
2458    // Resolve aliases, when possible.
2459    LocalChange |= ResolveAliases(M);
2460    Changed |= LocalChange;
2461  }
2462
2463  // TODO: Move all global ctors functions to the end of the module for code
2464  // layout.
2465
2466  return Changed;
2467}
2468