1 //===- GlobalOpt.cpp - Optimize Global Variables --------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
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.
14 //===----------------------------------------------------------------------===//
16 #define DEBUG_TYPE "globalopt"
17 #include "llvm/Transforms/IPO.h"
18 #include "llvm/CallingConv.h"
19 #include "llvm/Constants.h"
20 #include "llvm/DerivedTypes.h"
21 #include "llvm/Instructions.h"
22 #include "llvm/IntrinsicInst.h"
23 #include "llvm/Module.h"
24 #include "llvm/Pass.h"
25 #include "llvm/Analysis/ConstantFolding.h"
26 #include "llvm/Analysis/MemoryBuiltins.h"
27 #include "llvm/Target/TargetData.h"
28 #include "llvm/Support/CallSite.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/Support/ErrorHandling.h"
31 #include "llvm/Support/GetElementPtrTypeIterator.h"
32 #include "llvm/Support/MathExtras.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/ADT/DenseMap.h"
35 #include "llvm/ADT/SmallPtrSet.h"
36 #include "llvm/ADT/SmallVector.h"
37 #include "llvm/ADT/Statistic.h"
38 #include "llvm/ADT/STLExtras.h"
42 STATISTIC(NumMarked , "Number of globals marked constant");
43 STATISTIC(NumUnnamed , "Number of globals marked unnamed_addr");
44 STATISTIC(NumSRA , "Number of aggregate globals broken into scalars");
45 STATISTIC(NumHeapSRA , "Number of heap objects SRA'd");
46 STATISTIC(NumSubstitute,"Number of globals with initializers stored into them");
47 STATISTIC(NumDeleted , "Number of globals deleted");
48 STATISTIC(NumFnDeleted , "Number of functions deleted");
49 STATISTIC(NumGlobUses , "Number of global uses devirtualized");
50 STATISTIC(NumLocalized , "Number of globals localized");
51 STATISTIC(NumShrunkToBool , "Number of global vars shrunk to booleans");
52 STATISTIC(NumFastCallFns , "Number of functions converted to fastcc");
53 STATISTIC(NumCtorsEvaluated, "Number of static ctors evaluated");
54 STATISTIC(NumNestRemoved , "Number of nest attributes removed");
55 STATISTIC(NumAliasesResolved, "Number of global aliases resolved");
56 STATISTIC(NumAliasesRemoved, "Number of global aliases eliminated");
57 STATISTIC(NumCXXDtorsRemoved, "Number of global C++ destructors removed");
61 struct GlobalOpt : public ModulePass {
62 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
64 static char ID; // Pass identification, replacement for typeid
65 GlobalOpt() : ModulePass(ID) {
66 initializeGlobalOptPass(*PassRegistry::getPassRegistry());
69 bool runOnModule(Module &M);
72 GlobalVariable *FindGlobalCtors(Module &M);
73 bool OptimizeFunctions(Module &M);
74 bool OptimizeGlobalVars(Module &M);
75 bool OptimizeGlobalAliases(Module &M);
76 bool OptimizeGlobalCtorsList(GlobalVariable *&GCL);
77 bool ProcessGlobal(GlobalVariable *GV,Module::global_iterator &GVI);
78 bool ProcessInternalGlobal(GlobalVariable *GV,Module::global_iterator &GVI,
79 const SmallPtrSet<const PHINode*, 16> &PHIUsers,
80 const GlobalStatus &GS);
81 bool OptimizeEmptyGlobalCXXDtors(Function *CXAAtExitFn);
85 char GlobalOpt::ID = 0;
86 INITIALIZE_PASS(GlobalOpt, "globalopt",
87 "Global Variable Optimizer", false, false)
89 ModulePass *llvm::createGlobalOptimizerPass() { return new GlobalOpt(); }
93 /// GlobalStatus - As we analyze each global, keep track of some information
94 /// about it. If we find out that the address of the global is taken, none of
95 /// this info will be accurate.
97 /// isCompared - True if the global's address is used in a comparison.
100 /// isLoaded - True if the global is ever loaded. If the global isn't ever
101 /// loaded it can be deleted.
104 /// StoredType - Keep track of what stores to the global look like.
107 /// NotStored - There is no store to this global. It can thus be marked
111 /// isInitializerStored - This global is stored to, but the only thing
112 /// stored is the constant it was initialized with. This is only tracked
113 /// for scalar globals.
116 /// isStoredOnce - This global is stored to, but only its initializer and
117 /// one other value is ever stored to it. If this global isStoredOnce, we
118 /// track the value stored to it in StoredOnceValue below. This is only
119 /// tracked for scalar globals.
122 /// isStored - This global is stored to by multiple values or something else
123 /// that we cannot track.
127 /// StoredOnceValue - If only one value (besides the initializer constant) is
128 /// ever stored to this global, keep track of what value it is.
129 Value *StoredOnceValue;
131 /// AccessingFunction/HasMultipleAccessingFunctions - These start out
132 /// null/false. When the first accessing function is noticed, it is recorded.
133 /// When a second different accessing function is noticed,
134 /// HasMultipleAccessingFunctions is set to true.
135 const Function *AccessingFunction;
136 bool HasMultipleAccessingFunctions;
138 /// HasNonInstructionUser - Set to true if this global has a user that is not
139 /// an instruction (e.g. a constant expr or GV initializer).
140 bool HasNonInstructionUser;
142 /// HasPHIUser - Set to true if this global has a user that is a PHI node.
145 GlobalStatus() : isCompared(false), isLoaded(false), StoredType(NotStored),
146 StoredOnceValue(0), AccessingFunction(0),
147 HasMultipleAccessingFunctions(false), HasNonInstructionUser(false),
153 // SafeToDestroyConstant - It is safe to destroy a constant iff it is only used
154 // by constants itself. Note that constants cannot be cyclic, so this test is
155 // pretty easy to implement recursively.
157 static bool SafeToDestroyConstant(const Constant *C) {
158 if (isa<GlobalValue>(C)) return false;
160 for (Value::const_use_iterator UI = C->use_begin(), E = C->use_end(); UI != E;
162 if (const Constant *CU = dyn_cast<Constant>(*UI)) {
163 if (!SafeToDestroyConstant(CU)) return false;
170 /// AnalyzeGlobal - Look at all uses of the global and fill in the GlobalStatus
171 /// structure. If the global has its address taken, return true to indicate we
172 /// can't do anything with it.
174 static bool AnalyzeGlobal(const Value *V, GlobalStatus &GS,
175 SmallPtrSet<const PHINode*, 16> &PHIUsers) {
176 for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;
179 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
180 GS.HasNonInstructionUser = true;
182 // If the result of the constantexpr isn't pointer type, then we won't
183 // know to expect it in various places. Just reject early.
184 if (!isa<PointerType>(CE->getType())) return true;
186 if (AnalyzeGlobal(CE, GS, PHIUsers)) return true;
187 } else if (const Instruction *I = dyn_cast<Instruction>(U)) {
188 if (!GS.HasMultipleAccessingFunctions) {
189 const Function *F = I->getParent()->getParent();
190 if (GS.AccessingFunction == 0)
191 GS.AccessingFunction = F;
192 else if (GS.AccessingFunction != F)
193 GS.HasMultipleAccessingFunctions = true;
195 if (const LoadInst *LI = dyn_cast<LoadInst>(I)) {
197 if (LI->isVolatile()) return true; // Don't hack on volatile loads.
198 } else if (const StoreInst *SI = dyn_cast<StoreInst>(I)) {
199 // Don't allow a store OF the address, only stores TO the address.
200 if (SI->getOperand(0) == V) return true;
202 if (SI->isVolatile()) return true; // Don't hack on volatile stores.
204 // If this is a direct store to the global (i.e., the global is a scalar
205 // value, not an aggregate), keep more specific information about
207 if (GS.StoredType != GlobalStatus::isStored) {
208 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(
209 SI->getOperand(1))) {
210 Value *StoredVal = SI->getOperand(0);
211 if (StoredVal == GV->getInitializer()) {
212 if (GS.StoredType < GlobalStatus::isInitializerStored)
213 GS.StoredType = GlobalStatus::isInitializerStored;
214 } else if (isa<LoadInst>(StoredVal) &&
215 cast<LoadInst>(StoredVal)->getOperand(0) == GV) {
216 if (GS.StoredType < GlobalStatus::isInitializerStored)
217 GS.StoredType = GlobalStatus::isInitializerStored;
218 } else if (GS.StoredType < GlobalStatus::isStoredOnce) {
219 GS.StoredType = GlobalStatus::isStoredOnce;
220 GS.StoredOnceValue = StoredVal;
221 } else if (GS.StoredType == GlobalStatus::isStoredOnce &&
222 GS.StoredOnceValue == StoredVal) {
225 GS.StoredType = GlobalStatus::isStored;
228 GS.StoredType = GlobalStatus::isStored;
231 } else if (isa<GetElementPtrInst>(I)) {
232 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
233 } else if (isa<SelectInst>(I)) {
234 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
235 } else if (const PHINode *PN = dyn_cast<PHINode>(I)) {
236 // PHI nodes we can check just like select or GEP instructions, but we
237 // have to be careful about infinite recursion.
238 if (PHIUsers.insert(PN)) // Not already visited.
239 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
240 GS.HasPHIUser = true;
241 } else if (isa<CmpInst>(I)) {
242 GS.isCompared = true;
243 } else if (isa<MemTransferInst>(I)) {
244 const MemTransferInst *MTI = cast<MemTransferInst>(I);
245 if (MTI->getArgOperand(0) == V)
246 GS.StoredType = GlobalStatus::isStored;
247 if (MTI->getArgOperand(1) == V)
249 } else if (isa<MemSetInst>(I)) {
250 assert(cast<MemSetInst>(I)->getArgOperand(0) == V &&
251 "Memset only takes one pointer!");
252 GS.StoredType = GlobalStatus::isStored;
254 return true; // Any other non-load instruction might take address!
256 } else if (const Constant *C = dyn_cast<Constant>(U)) {
257 GS.HasNonInstructionUser = true;
258 // We might have a dead and dangling constant hanging off of here.
259 if (!SafeToDestroyConstant(C))
262 GS.HasNonInstructionUser = true;
263 // Otherwise must be some other user.
271 static Constant *getAggregateConstantElement(Constant *Agg, Constant *Idx) {
272 ConstantInt *CI = dyn_cast<ConstantInt>(Idx);
274 unsigned IdxV = CI->getZExtValue();
276 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg)) {
277 if (IdxV < CS->getNumOperands()) return CS->getOperand(IdxV);
278 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg)) {
279 if (IdxV < CA->getNumOperands()) return CA->getOperand(IdxV);
280 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Agg)) {
281 if (IdxV < CP->getNumOperands()) return CP->getOperand(IdxV);
282 } else if (isa<ConstantAggregateZero>(Agg)) {
283 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
284 if (IdxV < STy->getNumElements())
285 return Constant::getNullValue(STy->getElementType(IdxV));
286 } else if (const SequentialType *STy =
287 dyn_cast<SequentialType>(Agg->getType())) {
288 return Constant::getNullValue(STy->getElementType());
290 } else if (isa<UndefValue>(Agg)) {
291 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
292 if (IdxV < STy->getNumElements())
293 return UndefValue::get(STy->getElementType(IdxV));
294 } else if (const SequentialType *STy =
295 dyn_cast<SequentialType>(Agg->getType())) {
296 return UndefValue::get(STy->getElementType());
303 /// CleanupConstantGlobalUsers - We just marked GV constant. Loop over all
304 /// users of the global, cleaning up the obvious ones. This is largely just a
305 /// quick scan over the use list to clean up the easy and obvious cruft. This
306 /// returns true if it made a change.
307 static bool CleanupConstantGlobalUsers(Value *V, Constant *Init) {
308 bool Changed = false;
309 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;) {
312 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
314 // Replace the load with the initializer.
315 LI->replaceAllUsesWith(Init);
316 LI->eraseFromParent();
319 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
320 // Store must be unreachable or storing Init into the global.
321 SI->eraseFromParent();
323 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
324 if (CE->getOpcode() == Instruction::GetElementPtr) {
325 Constant *SubInit = 0;
327 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
328 Changed |= CleanupConstantGlobalUsers(CE, SubInit);
329 } else if (CE->getOpcode() == Instruction::BitCast &&
330 CE->getType()->isPointerTy()) {
331 // Pointer cast, delete any stores and memsets to the global.
332 Changed |= CleanupConstantGlobalUsers(CE, 0);
335 if (CE->use_empty()) {
336 CE->destroyConstant();
339 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
340 // Do not transform "gepinst (gep constexpr (GV))" here, because forming
341 // "gepconstexpr (gep constexpr (GV))" will cause the two gep's to fold
342 // and will invalidate our notion of what Init is.
343 Constant *SubInit = 0;
344 if (!isa<ConstantExpr>(GEP->getOperand(0))) {
346 dyn_cast_or_null<ConstantExpr>(ConstantFoldInstruction(GEP));
347 if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr)
348 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
350 Changed |= CleanupConstantGlobalUsers(GEP, SubInit);
352 if (GEP->use_empty()) {
353 GEP->eraseFromParent();
356 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv
357 if (MI->getRawDest() == V) {
358 MI->eraseFromParent();
362 } else if (Constant *C = dyn_cast<Constant>(U)) {
363 // If we have a chain of dead constantexprs or other things dangling from
364 // us, and if they are all dead, nuke them without remorse.
365 if (SafeToDestroyConstant(C)) {
366 C->destroyConstant();
367 // This could have invalidated UI, start over from scratch.
368 CleanupConstantGlobalUsers(V, Init);
376 /// isSafeSROAElementUse - Return true if the specified instruction is a safe
377 /// user of a derived expression from a global that we want to SROA.
378 static bool isSafeSROAElementUse(Value *V) {
379 // We might have a dead and dangling constant hanging off of here.
380 if (Constant *C = dyn_cast<Constant>(V))
381 return SafeToDestroyConstant(C);
383 Instruction *I = dyn_cast<Instruction>(V);
384 if (!I) return false;
387 if (isa<LoadInst>(I)) return true;
389 // Stores *to* the pointer are ok.
390 if (StoreInst *SI = dyn_cast<StoreInst>(I))
391 return SI->getOperand(0) != V;
393 // Otherwise, it must be a GEP.
394 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I);
395 if (GEPI == 0) return false;
397 if (GEPI->getNumOperands() < 3 || !isa<Constant>(GEPI->getOperand(1)) ||
398 !cast<Constant>(GEPI->getOperand(1))->isNullValue())
401 for (Value::use_iterator I = GEPI->use_begin(), E = GEPI->use_end();
403 if (!isSafeSROAElementUse(*I))
409 /// IsUserOfGlobalSafeForSRA - U is a direct user of the specified global value.
410 /// Look at it and its uses and decide whether it is safe to SROA this global.
412 static bool IsUserOfGlobalSafeForSRA(User *U, GlobalValue *GV) {
413 // The user of the global must be a GEP Inst or a ConstantExpr GEP.
414 if (!isa<GetElementPtrInst>(U) &&
415 (!isa<ConstantExpr>(U) ||
416 cast<ConstantExpr>(U)->getOpcode() != Instruction::GetElementPtr))
419 // Check to see if this ConstantExpr GEP is SRA'able. In particular, we
420 // don't like < 3 operand CE's, and we don't like non-constant integer
421 // indices. This enforces that all uses are 'gep GV, 0, C, ...' for some
423 if (U->getNumOperands() < 3 || !isa<Constant>(U->getOperand(1)) ||
424 !cast<Constant>(U->getOperand(1))->isNullValue() ||
425 !isa<ConstantInt>(U->getOperand(2)))
428 gep_type_iterator GEPI = gep_type_begin(U), E = gep_type_end(U);
429 ++GEPI; // Skip over the pointer index.
431 // If this is a use of an array allocation, do a bit more checking for sanity.
432 if (const ArrayType *AT = dyn_cast<ArrayType>(*GEPI)) {
433 uint64_t NumElements = AT->getNumElements();
434 ConstantInt *Idx = cast<ConstantInt>(U->getOperand(2));
436 // Check to make sure that index falls within the array. If not,
437 // something funny is going on, so we won't do the optimization.
439 if (Idx->getZExtValue() >= NumElements)
442 // We cannot scalar repl this level of the array unless any array
443 // sub-indices are in-range constants. In particular, consider:
444 // A[0][i]. We cannot know that the user isn't doing invalid things like
445 // allowing i to index an out-of-range subscript that accesses A[1].
447 // Scalar replacing *just* the outer index of the array is probably not
448 // going to be a win anyway, so just give up.
449 for (++GEPI; // Skip array index.
452 uint64_t NumElements;
453 if (const ArrayType *SubArrayTy = dyn_cast<ArrayType>(*GEPI))
454 NumElements = SubArrayTy->getNumElements();
455 else if (const VectorType *SubVectorTy = dyn_cast<VectorType>(*GEPI))
456 NumElements = SubVectorTy->getNumElements();
458 assert((*GEPI)->isStructTy() &&
459 "Indexed GEP type is not array, vector, or struct!");
463 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPI.getOperand());
464 if (!IdxVal || IdxVal->getZExtValue() >= NumElements)
469 for (Value::use_iterator I = U->use_begin(), E = U->use_end(); I != E; ++I)
470 if (!isSafeSROAElementUse(*I))
475 /// GlobalUsersSafeToSRA - Look at all uses of the global and decide whether it
476 /// is safe for us to perform this transformation.
478 static bool GlobalUsersSafeToSRA(GlobalValue *GV) {
479 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end();
481 if (!IsUserOfGlobalSafeForSRA(*UI, GV))
488 /// SRAGlobal - Perform scalar replacement of aggregates on the specified global
489 /// variable. This opens the door for other optimizations by exposing the
490 /// behavior of the program in a more fine-grained way. We have determined that
491 /// this transformation is safe already. We return the first global variable we
492 /// insert so that the caller can reprocess it.
493 static GlobalVariable *SRAGlobal(GlobalVariable *GV, const TargetData &TD) {
494 // Make sure this global only has simple uses that we can SRA.
495 if (!GlobalUsersSafeToSRA(GV))
498 assert(GV->hasLocalLinkage() && !GV->isConstant());
499 Constant *Init = GV->getInitializer();
500 const Type *Ty = Init->getType();
502 std::vector<GlobalVariable*> NewGlobals;
503 Module::GlobalListType &Globals = GV->getParent()->getGlobalList();
505 // Get the alignment of the global, either explicit or target-specific.
506 unsigned StartAlignment = GV->getAlignment();
507 if (StartAlignment == 0)
508 StartAlignment = TD.getABITypeAlignment(GV->getType());
510 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
511 NewGlobals.reserve(STy->getNumElements());
512 const StructLayout &Layout = *TD.getStructLayout(STy);
513 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
514 Constant *In = getAggregateConstantElement(Init,
515 ConstantInt::get(Type::getInt32Ty(STy->getContext()), i));
516 assert(In && "Couldn't get element of initializer?");
517 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(i), false,
518 GlobalVariable::InternalLinkage,
519 In, GV->getName()+"."+Twine(i),
521 GV->getType()->getAddressSpace());
522 Globals.insert(GV, NGV);
523 NewGlobals.push_back(NGV);
525 // Calculate the known alignment of the field. If the original aggregate
526 // had 256 byte alignment for example, something might depend on that:
527 // propagate info to each field.
528 uint64_t FieldOffset = Layout.getElementOffset(i);
529 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, FieldOffset);
530 if (NewAlign > TD.getABITypeAlignment(STy->getElementType(i)))
531 NGV->setAlignment(NewAlign);
533 } else if (const SequentialType *STy = dyn_cast<SequentialType>(Ty)) {
534 unsigned NumElements = 0;
535 if (const ArrayType *ATy = dyn_cast<ArrayType>(STy))
536 NumElements = ATy->getNumElements();
538 NumElements = cast<VectorType>(STy)->getNumElements();
540 if (NumElements > 16 && GV->hasNUsesOrMore(16))
541 return 0; // It's not worth it.
542 NewGlobals.reserve(NumElements);
544 uint64_t EltSize = TD.getTypeAllocSize(STy->getElementType());
545 unsigned EltAlign = TD.getABITypeAlignment(STy->getElementType());
546 for (unsigned i = 0, e = NumElements; i != e; ++i) {
547 Constant *In = getAggregateConstantElement(Init,
548 ConstantInt::get(Type::getInt32Ty(Init->getContext()), i));
549 assert(In && "Couldn't get element of initializer?");
551 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false,
552 GlobalVariable::InternalLinkage,
553 In, GV->getName()+"."+Twine(i),
555 GV->getType()->getAddressSpace());
556 Globals.insert(GV, NGV);
557 NewGlobals.push_back(NGV);
559 // Calculate the known alignment of the field. If the original aggregate
560 // had 256 byte alignment for example, something might depend on that:
561 // propagate info to each field.
562 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, EltSize*i);
563 if (NewAlign > EltAlign)
564 NGV->setAlignment(NewAlign);
568 if (NewGlobals.empty())
571 DEBUG(dbgs() << "PERFORMING GLOBAL SRA ON: " << *GV);
573 Constant *NullInt =Constant::getNullValue(Type::getInt32Ty(GV->getContext()));
575 // Loop over all of the uses of the global, replacing the constantexpr geps,
576 // with smaller constantexpr geps or direct references.
577 while (!GV->use_empty()) {
578 User *GEP = GV->use_back();
579 assert(((isa<ConstantExpr>(GEP) &&
580 cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)||
581 isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!");
583 // Ignore the 1th operand, which has to be zero or else the program is quite
584 // broken (undefined). Get the 2nd operand, which is the structure or array
586 unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
587 if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access.
589 Value *NewPtr = NewGlobals[Val];
591 // Form a shorter GEP if needed.
592 if (GEP->getNumOperands() > 3) {
593 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) {
594 SmallVector<Constant*, 8> Idxs;
595 Idxs.push_back(NullInt);
596 for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i)
597 Idxs.push_back(CE->getOperand(i));
598 NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr),
599 &Idxs[0], Idxs.size());
601 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP);
602 SmallVector<Value*, 8> Idxs;
603 Idxs.push_back(NullInt);
604 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i)
605 Idxs.push_back(GEPI->getOperand(i));
606 NewPtr = GetElementPtrInst::Create(NewPtr, Idxs.begin(), Idxs.end(),
607 GEPI->getName()+"."+Twine(Val),GEPI);
610 GEP->replaceAllUsesWith(NewPtr);
612 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP))
613 GEPI->eraseFromParent();
615 cast<ConstantExpr>(GEP)->destroyConstant();
618 // Delete the old global, now that it is dead.
622 // Loop over the new globals array deleting any globals that are obviously
623 // dead. This can arise due to scalarization of a structure or an array that
624 // has elements that are dead.
625 unsigned FirstGlobal = 0;
626 for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i)
627 if (NewGlobals[i]->use_empty()) {
628 Globals.erase(NewGlobals[i]);
629 if (FirstGlobal == i) ++FirstGlobal;
632 return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : 0;
635 /// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified
636 /// value will trap if the value is dynamically null. PHIs keeps track of any
637 /// phi nodes we've seen to avoid reprocessing them.
638 static bool AllUsesOfValueWillTrapIfNull(const Value *V,
639 SmallPtrSet<const PHINode*, 8> &PHIs) {
640 for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;
644 if (isa<LoadInst>(U)) {
646 } else if (const StoreInst *SI = dyn_cast<StoreInst>(U)) {
647 if (SI->getOperand(0) == V) {
648 //cerr << "NONTRAPPING USE: " << *U;
649 return false; // Storing the value.
651 } else if (const CallInst *CI = dyn_cast<CallInst>(U)) {
652 if (CI->getCalledValue() != V) {
653 //cerr << "NONTRAPPING USE: " << *U;
654 return false; // Not calling the ptr
656 } else if (const InvokeInst *II = dyn_cast<InvokeInst>(U)) {
657 if (II->getCalledValue() != V) {
658 //cerr << "NONTRAPPING USE: " << *U;
659 return false; // Not calling the ptr
661 } else if (const BitCastInst *CI = dyn_cast<BitCastInst>(U)) {
662 if (!AllUsesOfValueWillTrapIfNull(CI, PHIs)) return false;
663 } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
664 if (!AllUsesOfValueWillTrapIfNull(GEPI, PHIs)) return false;
665 } else if (const PHINode *PN = dyn_cast<PHINode>(U)) {
666 // If we've already seen this phi node, ignore it, it has already been
668 if (PHIs.insert(PN) && !AllUsesOfValueWillTrapIfNull(PN, PHIs))
670 } else if (isa<ICmpInst>(U) &&
671 isa<ConstantPointerNull>(UI->getOperand(1))) {
672 // Ignore icmp X, null
674 //cerr << "NONTRAPPING USE: " << *U;
681 /// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads
682 /// from GV will trap if the loaded value is null. Note that this also permits
683 /// comparisons of the loaded value against null, as a special case.
684 static bool AllUsesOfLoadedValueWillTrapIfNull(const GlobalVariable *GV) {
685 for (Value::const_use_iterator UI = GV->use_begin(), E = GV->use_end();
689 if (const LoadInst *LI = dyn_cast<LoadInst>(U)) {
690 SmallPtrSet<const PHINode*, 8> PHIs;
691 if (!AllUsesOfValueWillTrapIfNull(LI, PHIs))
693 } else if (isa<StoreInst>(U)) {
694 // Ignore stores to the global.
696 // We don't know or understand this user, bail out.
697 //cerr << "UNKNOWN USER OF GLOBAL!: " << *U;
704 static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) {
705 bool Changed = false;
706 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) {
707 Instruction *I = cast<Instruction>(*UI++);
708 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
709 LI->setOperand(0, NewV);
711 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
712 if (SI->getOperand(1) == V) {
713 SI->setOperand(1, NewV);
716 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
718 if (CS.getCalledValue() == V) {
719 // Calling through the pointer! Turn into a direct call, but be careful
720 // that the pointer is not also being passed as an argument.
721 CS.setCalledFunction(NewV);
723 bool PassedAsArg = false;
724 for (unsigned i = 0, e = CS.arg_size(); i != e; ++i)
725 if (CS.getArgument(i) == V) {
727 CS.setArgument(i, NewV);
731 // Being passed as an argument also. Be careful to not invalidate UI!
735 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
736 Changed |= OptimizeAwayTrappingUsesOfValue(CI,
737 ConstantExpr::getCast(CI->getOpcode(),
738 NewV, CI->getType()));
739 if (CI->use_empty()) {
741 CI->eraseFromParent();
743 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
744 // Should handle GEP here.
745 SmallVector<Constant*, 8> Idxs;
746 Idxs.reserve(GEPI->getNumOperands()-1);
747 for (User::op_iterator i = GEPI->op_begin() + 1, e = GEPI->op_end();
749 if (Constant *C = dyn_cast<Constant>(*i))
753 if (Idxs.size() == GEPI->getNumOperands()-1)
754 Changed |= OptimizeAwayTrappingUsesOfValue(GEPI,
755 ConstantExpr::getGetElementPtr(NewV, &Idxs[0],
757 if (GEPI->use_empty()) {
759 GEPI->eraseFromParent();
768 /// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null
769 /// value stored into it. If there are uses of the loaded value that would trap
770 /// if the loaded value is dynamically null, then we know that they cannot be
771 /// reachable with a null optimize away the load.
772 static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV) {
773 bool Changed = false;
775 // Keep track of whether we are able to remove all the uses of the global
776 // other than the store that defines it.
777 bool AllNonStoreUsesGone = true;
779 // Replace all uses of loads with uses of uses of the stored value.
780 for (Value::use_iterator GUI = GV->use_begin(), E = GV->use_end(); GUI != E;){
781 User *GlobalUser = *GUI++;
782 if (LoadInst *LI = dyn_cast<LoadInst>(GlobalUser)) {
783 Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV);
784 // If we were able to delete all uses of the loads
785 if (LI->use_empty()) {
786 LI->eraseFromParent();
789 AllNonStoreUsesGone = false;
791 } else if (isa<StoreInst>(GlobalUser)) {
792 // Ignore the store that stores "LV" to the global.
793 assert(GlobalUser->getOperand(1) == GV &&
794 "Must be storing *to* the global");
796 AllNonStoreUsesGone = false;
798 // If we get here we could have other crazy uses that are transitively
800 assert((isa<PHINode>(GlobalUser) || isa<SelectInst>(GlobalUser) ||
801 isa<ConstantExpr>(GlobalUser)) && "Only expect load and stores!");
806 DEBUG(dbgs() << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV);
810 // If we nuked all of the loads, then none of the stores are needed either,
811 // nor is the global.
812 if (AllNonStoreUsesGone) {
813 DEBUG(dbgs() << " *** GLOBAL NOW DEAD!\n");
814 CleanupConstantGlobalUsers(GV, 0);
815 if (GV->use_empty()) {
816 GV->eraseFromParent();
824 /// ConstantPropUsersOf - Walk the use list of V, constant folding all of the
825 /// instructions that are foldable.
826 static void ConstantPropUsersOf(Value *V) {
827 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; )
828 if (Instruction *I = dyn_cast<Instruction>(*UI++))
829 if (Constant *NewC = ConstantFoldInstruction(I)) {
830 I->replaceAllUsesWith(NewC);
832 // Advance UI to the next non-I use to avoid invalidating it!
833 // Instructions could multiply use V.
834 while (UI != E && *UI == I)
836 I->eraseFromParent();
840 /// OptimizeGlobalAddressOfMalloc - This function takes the specified global
841 /// variable, and transforms the program as if it always contained the result of
842 /// the specified malloc. Because it is always the result of the specified
843 /// malloc, there is no reason to actually DO the malloc. Instead, turn the
844 /// malloc into a global, and any loads of GV as uses of the new global.
845 static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV,
848 ConstantInt *NElements,
850 DEBUG(errs() << "PROMOTING GLOBAL: " << *GV << " CALL = " << *CI << '\n');
852 const Type *GlobalType;
853 if (NElements->getZExtValue() == 1)
854 GlobalType = AllocTy;
856 // If we have an array allocation, the global variable is of an array.
857 GlobalType = ArrayType::get(AllocTy, NElements->getZExtValue());
859 // Create the new global variable. The contents of the malloc'd memory is
860 // undefined, so initialize with an undef value.
861 GlobalVariable *NewGV = new GlobalVariable(*GV->getParent(),
863 GlobalValue::InternalLinkage,
864 UndefValue::get(GlobalType),
865 GV->getName()+".body",
867 GV->isThreadLocal());
869 // If there are bitcast users of the malloc (which is typical, usually we have
870 // a malloc + bitcast) then replace them with uses of the new global. Update
871 // other users to use the global as well.
872 BitCastInst *TheBC = 0;
873 while (!CI->use_empty()) {
874 Instruction *User = cast<Instruction>(CI->use_back());
875 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
876 if (BCI->getType() == NewGV->getType()) {
877 BCI->replaceAllUsesWith(NewGV);
878 BCI->eraseFromParent();
880 BCI->setOperand(0, NewGV);
884 TheBC = new BitCastInst(NewGV, CI->getType(), "newgv", CI);
885 User->replaceUsesOfWith(CI, TheBC);
889 Constant *RepValue = NewGV;
890 if (NewGV->getType() != GV->getType()->getElementType())
891 RepValue = ConstantExpr::getBitCast(RepValue,
892 GV->getType()->getElementType());
894 // If there is a comparison against null, we will insert a global bool to
895 // keep track of whether the global was initialized yet or not.
896 GlobalVariable *InitBool =
897 new GlobalVariable(Type::getInt1Ty(GV->getContext()), false,
898 GlobalValue::InternalLinkage,
899 ConstantInt::getFalse(GV->getContext()),
900 GV->getName()+".init", GV->isThreadLocal());
901 bool InitBoolUsed = false;
903 // Loop over all uses of GV, processing them in turn.
904 while (!GV->use_empty()) {
905 if (StoreInst *SI = dyn_cast<StoreInst>(GV->use_back())) {
906 // The global is initialized when the store to it occurs.
907 new StoreInst(ConstantInt::getTrue(GV->getContext()), InitBool, SI);
908 SI->eraseFromParent();
912 LoadInst *LI = cast<LoadInst>(GV->use_back());
913 while (!LI->use_empty()) {
914 Use &LoadUse = LI->use_begin().getUse();
915 if (!isa<ICmpInst>(LoadUse.getUser())) {
920 ICmpInst *ICI = cast<ICmpInst>(LoadUse.getUser());
921 // Replace the cmp X, 0 with a use of the bool value.
922 Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", ICI);
924 switch (ICI->getPredicate()) {
925 default: llvm_unreachable("Unknown ICmp Predicate!");
926 case ICmpInst::ICMP_ULT:
927 case ICmpInst::ICMP_SLT: // X < null -> always false
928 LV = ConstantInt::getFalse(GV->getContext());
930 case ICmpInst::ICMP_ULE:
931 case ICmpInst::ICMP_SLE:
932 case ICmpInst::ICMP_EQ:
933 LV = BinaryOperator::CreateNot(LV, "notinit", ICI);
935 case ICmpInst::ICMP_NE:
936 case ICmpInst::ICMP_UGE:
937 case ICmpInst::ICMP_SGE:
938 case ICmpInst::ICMP_UGT:
939 case ICmpInst::ICMP_SGT:
942 ICI->replaceAllUsesWith(LV);
943 ICI->eraseFromParent();
945 LI->eraseFromParent();
948 // If the initialization boolean was used, insert it, otherwise delete it.
950 while (!InitBool->use_empty()) // Delete initializations
951 cast<StoreInst>(InitBool->use_back())->eraseFromParent();
954 GV->getParent()->getGlobalList().insert(GV, InitBool);
956 // Now the GV is dead, nuke it and the malloc..
957 GV->eraseFromParent();
958 CI->eraseFromParent();
960 // To further other optimizations, loop over all users of NewGV and try to
961 // constant prop them. This will promote GEP instructions with constant
962 // indices into GEP constant-exprs, which will allow global-opt to hack on it.
963 ConstantPropUsersOf(NewGV);
964 if (RepValue != NewGV)
965 ConstantPropUsersOf(RepValue);
970 /// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking
971 /// to make sure that there are no complex uses of V. We permit simple things
972 /// like dereferencing the pointer, but not storing through the address, unless
973 /// it is to the specified global.
974 static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(const Instruction *V,
975 const GlobalVariable *GV,
976 SmallPtrSet<const PHINode*, 8> &PHIs) {
977 for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end();
979 const Instruction *Inst = cast<Instruction>(*UI);
981 if (isa<LoadInst>(Inst) || isa<CmpInst>(Inst)) {
982 continue; // Fine, ignore.
985 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
986 if (SI->getOperand(0) == V && SI->getOperand(1) != GV)
987 return false; // Storing the pointer itself... bad.
988 continue; // Otherwise, storing through it, or storing into GV... fine.
991 // Must index into the array and into the struct.
992 if (isa<GetElementPtrInst>(Inst) && Inst->getNumOperands() >= 3) {
993 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Inst, GV, PHIs))
998 if (const PHINode *PN = dyn_cast<PHINode>(Inst)) {
999 // PHIs are ok if all uses are ok. Don't infinitely recurse through PHI
1001 if (PHIs.insert(PN))
1002 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(PN, GV, PHIs))
1007 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Inst)) {
1008 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(BCI, GV, PHIs))
1018 /// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV
1019 /// somewhere. Transform all uses of the allocation into loads from the
1020 /// global and uses of the resultant pointer. Further, delete the store into
1021 /// GV. This assumes that these value pass the
1022 /// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate.
1023 static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc,
1024 GlobalVariable *GV) {
1025 while (!Alloc->use_empty()) {
1026 Instruction *U = cast<Instruction>(*Alloc->use_begin());
1027 Instruction *InsertPt = U;
1028 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
1029 // If this is the store of the allocation into the global, remove it.
1030 if (SI->getOperand(1) == GV) {
1031 SI->eraseFromParent();
1034 } else if (PHINode *PN = dyn_cast<PHINode>(U)) {
1035 // Insert the load in the corresponding predecessor, not right before the
1037 InsertPt = PN->getIncomingBlock(Alloc->use_begin())->getTerminator();
1038 } else if (isa<BitCastInst>(U)) {
1039 // Must be bitcast between the malloc and store to initialize the global.
1040 ReplaceUsesOfMallocWithGlobal(U, GV);
1041 U->eraseFromParent();
1043 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
1044 // If this is a "GEP bitcast" and the user is a store to the global, then
1045 // just process it as a bitcast.
1046 if (GEPI->hasAllZeroIndices() && GEPI->hasOneUse())
1047 if (StoreInst *SI = dyn_cast<StoreInst>(GEPI->use_back()))
1048 if (SI->getOperand(1) == GV) {
1049 // Must be bitcast GEP between the malloc and store to initialize
1051 ReplaceUsesOfMallocWithGlobal(GEPI, GV);
1052 GEPI->eraseFromParent();
1057 // Insert a load from the global, and use it instead of the malloc.
1058 Value *NL = new LoadInst(GV, GV->getName()+".val", InsertPt);
1059 U->replaceUsesOfWith(Alloc, NL);
1063 /// LoadUsesSimpleEnoughForHeapSRA - Verify that all uses of V (a load, or a phi
1064 /// of a load) are simple enough to perform heap SRA on. This permits GEP's
1065 /// that index through the array and struct field, icmps of null, and PHIs.
1066 static bool LoadUsesSimpleEnoughForHeapSRA(const Value *V,
1067 SmallPtrSet<const PHINode*, 32> &LoadUsingPHIs,
1068 SmallPtrSet<const PHINode*, 32> &LoadUsingPHIsPerLoad) {
1069 // We permit two users of the load: setcc comparing against the null
1070 // pointer, and a getelementptr of a specific form.
1071 for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;
1073 const Instruction *User = cast<Instruction>(*UI);
1075 // Comparison against null is ok.
1076 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(User)) {
1077 if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
1082 // getelementptr is also ok, but only a simple form.
1083 if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1084 // Must index into the array and into the struct.
1085 if (GEPI->getNumOperands() < 3)
1088 // Otherwise the GEP is ok.
1092 if (const PHINode *PN = dyn_cast<PHINode>(User)) {
1093 if (!LoadUsingPHIsPerLoad.insert(PN))
1094 // This means some phi nodes are dependent on each other.
1095 // Avoid infinite looping!
1097 if (!LoadUsingPHIs.insert(PN))
1098 // If we have already analyzed this PHI, then it is safe.
1101 // Make sure all uses of the PHI are simple enough to transform.
1102 if (!LoadUsesSimpleEnoughForHeapSRA(PN,
1103 LoadUsingPHIs, LoadUsingPHIsPerLoad))
1109 // Otherwise we don't know what this is, not ok.
1117 /// AllGlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from
1118 /// GV are simple enough to perform HeapSRA, return true.
1119 static bool AllGlobalLoadUsesSimpleEnoughForHeapSRA(const GlobalVariable *GV,
1120 Instruction *StoredVal) {
1121 SmallPtrSet<const PHINode*, 32> LoadUsingPHIs;
1122 SmallPtrSet<const PHINode*, 32> LoadUsingPHIsPerLoad;
1123 for (Value::const_use_iterator UI = GV->use_begin(), E = GV->use_end();
1125 if (const LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
1126 if (!LoadUsesSimpleEnoughForHeapSRA(LI, LoadUsingPHIs,
1127 LoadUsingPHIsPerLoad))
1129 LoadUsingPHIsPerLoad.clear();
1132 // If we reach here, we know that all uses of the loads and transitive uses
1133 // (through PHI nodes) are simple enough to transform. However, we don't know
1134 // that all inputs the to the PHI nodes are in the same equivalence sets.
1135 // Check to verify that all operands of the PHIs are either PHIS that can be
1136 // transformed, loads from GV, or MI itself.
1137 for (SmallPtrSet<const PHINode*, 32>::const_iterator I = LoadUsingPHIs.begin()
1138 , E = LoadUsingPHIs.end(); I != E; ++I) {
1139 const PHINode *PN = *I;
1140 for (unsigned op = 0, e = PN->getNumIncomingValues(); op != e; ++op) {
1141 Value *InVal = PN->getIncomingValue(op);
1143 // PHI of the stored value itself is ok.
1144 if (InVal == StoredVal) continue;
1146 if (const PHINode *InPN = dyn_cast<PHINode>(InVal)) {
1147 // One of the PHIs in our set is (optimistically) ok.
1148 if (LoadUsingPHIs.count(InPN))
1153 // Load from GV is ok.
1154 if (const LoadInst *LI = dyn_cast<LoadInst>(InVal))
1155 if (LI->getOperand(0) == GV)
1160 // Anything else is rejected.
1168 static Value *GetHeapSROAValue(Value *V, unsigned FieldNo,
1169 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1170 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1171 std::vector<Value*> &FieldVals = InsertedScalarizedValues[V];
1173 if (FieldNo >= FieldVals.size())
1174 FieldVals.resize(FieldNo+1);
1176 // If we already have this value, just reuse the previously scalarized
1178 if (Value *FieldVal = FieldVals[FieldNo])
1181 // Depending on what instruction this is, we have several cases.
1183 if (LoadInst *LI = dyn_cast<LoadInst>(V)) {
1184 // This is a scalarized version of the load from the global. Just create
1185 // a new Load of the scalarized global.
1186 Result = new LoadInst(GetHeapSROAValue(LI->getOperand(0), FieldNo,
1187 InsertedScalarizedValues,
1189 LI->getName()+".f"+Twine(FieldNo), LI);
1190 } else if (PHINode *PN = dyn_cast<PHINode>(V)) {
1191 // PN's type is pointer to struct. Make a new PHI of pointer to struct
1193 const StructType *ST =
1194 cast<StructType>(cast<PointerType>(PN->getType())->getElementType());
1197 PHINode::Create(PointerType::getUnqual(ST->getElementType(FieldNo)),
1198 PN->getNumIncomingValues(),
1199 PN->getName()+".f"+Twine(FieldNo), PN);
1201 PHIsToRewrite.push_back(std::make_pair(PN, FieldNo));
1203 llvm_unreachable("Unknown usable value");
1207 return FieldVals[FieldNo] = Result;
1210 /// RewriteHeapSROALoadUser - Given a load instruction and a value derived from
1211 /// the load, rewrite the derived value to use the HeapSRoA'd load.
1212 static void RewriteHeapSROALoadUser(Instruction *LoadUser,
1213 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1214 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1215 // If this is a comparison against null, handle it.
1216 if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) {
1217 assert(isa<ConstantPointerNull>(SCI->getOperand(1)));
1218 // If we have a setcc of the loaded pointer, we can use a setcc of any
1220 Value *NPtr = GetHeapSROAValue(SCI->getOperand(0), 0,
1221 InsertedScalarizedValues, PHIsToRewrite);
1223 Value *New = new ICmpInst(SCI, SCI->getPredicate(), NPtr,
1224 Constant::getNullValue(NPtr->getType()),
1226 SCI->replaceAllUsesWith(New);
1227 SCI->eraseFromParent();
1231 // Handle 'getelementptr Ptr, Idx, i32 FieldNo ...'
1232 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) {
1233 assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2))
1234 && "Unexpected GEPI!");
1236 // Load the pointer for this field.
1237 unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
1238 Value *NewPtr = GetHeapSROAValue(GEPI->getOperand(0), FieldNo,
1239 InsertedScalarizedValues, PHIsToRewrite);
1241 // Create the new GEP idx vector.
1242 SmallVector<Value*, 8> GEPIdx;
1243 GEPIdx.push_back(GEPI->getOperand(1));
1244 GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end());
1246 Value *NGEPI = GetElementPtrInst::Create(NewPtr,
1247 GEPIdx.begin(), GEPIdx.end(),
1248 GEPI->getName(), GEPI);
1249 GEPI->replaceAllUsesWith(NGEPI);
1250 GEPI->eraseFromParent();
1254 // Recursively transform the users of PHI nodes. This will lazily create the
1255 // PHIs that are needed for individual elements. Keep track of what PHIs we
1256 // see in InsertedScalarizedValues so that we don't get infinite loops (very
1257 // antisocial). If the PHI is already in InsertedScalarizedValues, it has
1258 // already been seen first by another load, so its uses have already been
1260 PHINode *PN = cast<PHINode>(LoadUser);
1262 DenseMap<Value*, std::vector<Value*> >::iterator InsertPos;
1263 tie(InsertPos, Inserted) =
1264 InsertedScalarizedValues.insert(std::make_pair(PN, std::vector<Value*>()));
1265 if (!Inserted) return;
1267 // If this is the first time we've seen this PHI, recursively process all
1269 for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end(); UI != E; ) {
1270 Instruction *User = cast<Instruction>(*UI++);
1271 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite);
1275 /// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global. Ptr
1276 /// is a value loaded from the global. Eliminate all uses of Ptr, making them
1277 /// use FieldGlobals instead. All uses of loaded values satisfy
1278 /// AllGlobalLoadUsesSimpleEnoughForHeapSRA.
1279 static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load,
1280 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1281 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1282 for (Value::use_iterator UI = Load->use_begin(), E = Load->use_end();
1284 Instruction *User = cast<Instruction>(*UI++);
1285 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite);
1288 if (Load->use_empty()) {
1289 Load->eraseFromParent();
1290 InsertedScalarizedValues.erase(Load);
1294 /// PerformHeapAllocSRoA - CI is an allocation of an array of structures. Break
1295 /// it up into multiple allocations of arrays of the fields.
1296 static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, CallInst *CI,
1297 Value* NElems, TargetData *TD) {
1298 DEBUG(dbgs() << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *CI << '\n');
1299 const Type* MAT = getMallocAllocatedType(CI);
1300 const StructType *STy = cast<StructType>(MAT);
1302 // There is guaranteed to be at least one use of the malloc (storing
1303 // it into GV). If there are other uses, change them to be uses of
1304 // the global to simplify later code. This also deletes the store
1306 ReplaceUsesOfMallocWithGlobal(CI, GV);
1308 // Okay, at this point, there are no users of the malloc. Insert N
1309 // new mallocs at the same place as CI, and N globals.
1310 std::vector<Value*> FieldGlobals;
1311 std::vector<Value*> FieldMallocs;
1313 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){
1314 const Type *FieldTy = STy->getElementType(FieldNo);
1315 const PointerType *PFieldTy = PointerType::getUnqual(FieldTy);
1317 GlobalVariable *NGV =
1318 new GlobalVariable(*GV->getParent(),
1319 PFieldTy, false, GlobalValue::InternalLinkage,
1320 Constant::getNullValue(PFieldTy),
1321 GV->getName() + ".f" + Twine(FieldNo), GV,
1322 GV->isThreadLocal());
1323 FieldGlobals.push_back(NGV);
1325 unsigned TypeSize = TD->getTypeAllocSize(FieldTy);
1326 if (const StructType *ST = dyn_cast<StructType>(FieldTy))
1327 TypeSize = TD->getStructLayout(ST)->getSizeInBytes();
1328 const Type *IntPtrTy = TD->getIntPtrType(CI->getContext());
1329 Value *NMI = CallInst::CreateMalloc(CI, IntPtrTy, FieldTy,
1330 ConstantInt::get(IntPtrTy, TypeSize),
1332 CI->getName() + ".f" + Twine(FieldNo));
1333 FieldMallocs.push_back(NMI);
1334 new StoreInst(NMI, NGV, CI);
1337 // The tricky aspect of this transformation is handling the case when malloc
1338 // fails. In the original code, malloc failing would set the result pointer
1339 // of malloc to null. In this case, some mallocs could succeed and others
1340 // could fail. As such, we emit code that looks like this:
1341 // F0 = malloc(field0)
1342 // F1 = malloc(field1)
1343 // F2 = malloc(field2)
1344 // if (F0 == 0 || F1 == 0 || F2 == 0) {
1345 // if (F0) { free(F0); F0 = 0; }
1346 // if (F1) { free(F1); F1 = 0; }
1347 // if (F2) { free(F2); F2 = 0; }
1349 // The malloc can also fail if its argument is too large.
1350 Constant *ConstantZero = ConstantInt::get(CI->getArgOperand(0)->getType(), 0);
1351 Value *RunningOr = new ICmpInst(CI, ICmpInst::ICMP_SLT, CI->getArgOperand(0),
1352 ConstantZero, "isneg");
1353 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) {
1354 Value *Cond = new ICmpInst(CI, ICmpInst::ICMP_EQ, FieldMallocs[i],
1355 Constant::getNullValue(FieldMallocs[i]->getType()),
1357 RunningOr = BinaryOperator::CreateOr(RunningOr, Cond, "tmp", CI);
1360 // Split the basic block at the old malloc.
1361 BasicBlock *OrigBB = CI->getParent();
1362 BasicBlock *ContBB = OrigBB->splitBasicBlock(CI, "malloc_cont");
1364 // Create the block to check the first condition. Put all these blocks at the
1365 // end of the function as they are unlikely to be executed.
1366 BasicBlock *NullPtrBlock = BasicBlock::Create(OrigBB->getContext(),
1368 OrigBB->getParent());
1370 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond
1371 // branch on RunningOr.
1372 OrigBB->getTerminator()->eraseFromParent();
1373 BranchInst::Create(NullPtrBlock, ContBB, RunningOr, OrigBB);
1375 // Within the NullPtrBlock, we need to emit a comparison and branch for each
1376 // pointer, because some may be null while others are not.
1377 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1378 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock);
1379 Value *Cmp = new ICmpInst(*NullPtrBlock, ICmpInst::ICMP_NE, GVVal,
1380 Constant::getNullValue(GVVal->getType()),
1382 BasicBlock *FreeBlock = BasicBlock::Create(Cmp->getContext(), "free_it",
1383 OrigBB->getParent());
1384 BasicBlock *NextBlock = BasicBlock::Create(Cmp->getContext(), "next",
1385 OrigBB->getParent());
1386 Instruction *BI = BranchInst::Create(FreeBlock, NextBlock,
1389 // Fill in FreeBlock.
1390 CallInst::CreateFree(GVVal, BI);
1391 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i],
1393 BranchInst::Create(NextBlock, FreeBlock);
1395 NullPtrBlock = NextBlock;
1398 BranchInst::Create(ContBB, NullPtrBlock);
1400 // CI is no longer needed, remove it.
1401 CI->eraseFromParent();
1403 /// InsertedScalarizedLoads - As we process loads, if we can't immediately
1404 /// update all uses of the load, keep track of what scalarized loads are
1405 /// inserted for a given load.
1406 DenseMap<Value*, std::vector<Value*> > InsertedScalarizedValues;
1407 InsertedScalarizedValues[GV] = FieldGlobals;
1409 std::vector<std::pair<PHINode*, unsigned> > PHIsToRewrite;
1411 // Okay, the malloc site is completely handled. All of the uses of GV are now
1412 // loads, and all uses of those loads are simple. Rewrite them to use loads
1413 // of the per-field globals instead.
1414 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;) {
1415 Instruction *User = cast<Instruction>(*UI++);
1417 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1418 RewriteUsesOfLoadForHeapSRoA(LI, InsertedScalarizedValues, PHIsToRewrite);
1422 // Must be a store of null.
1423 StoreInst *SI = cast<StoreInst>(User);
1424 assert(isa<ConstantPointerNull>(SI->getOperand(0)) &&
1425 "Unexpected heap-sra user!");
1427 // Insert a store of null into each global.
1428 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1429 const PointerType *PT = cast<PointerType>(FieldGlobals[i]->getType());
1430 Constant *Null = Constant::getNullValue(PT->getElementType());
1431 new StoreInst(Null, FieldGlobals[i], SI);
1433 // Erase the original store.
1434 SI->eraseFromParent();
1437 // While we have PHIs that are interesting to rewrite, do it.
1438 while (!PHIsToRewrite.empty()) {
1439 PHINode *PN = PHIsToRewrite.back().first;
1440 unsigned FieldNo = PHIsToRewrite.back().second;
1441 PHIsToRewrite.pop_back();
1442 PHINode *FieldPN = cast<PHINode>(InsertedScalarizedValues[PN][FieldNo]);
1443 assert(FieldPN->getNumIncomingValues() == 0 &&"Already processed this phi");
1445 // Add all the incoming values. This can materialize more phis.
1446 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1447 Value *InVal = PN->getIncomingValue(i);
1448 InVal = GetHeapSROAValue(InVal, FieldNo, InsertedScalarizedValues,
1450 FieldPN->addIncoming(InVal, PN->getIncomingBlock(i));
1454 // Drop all inter-phi links and any loads that made it this far.
1455 for (DenseMap<Value*, std::vector<Value*> >::iterator
1456 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1458 if (PHINode *PN = dyn_cast<PHINode>(I->first))
1459 PN->dropAllReferences();
1460 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1461 LI->dropAllReferences();
1464 // Delete all the phis and loads now that inter-references are dead.
1465 for (DenseMap<Value*, std::vector<Value*> >::iterator
1466 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1468 if (PHINode *PN = dyn_cast<PHINode>(I->first))
1469 PN->eraseFromParent();
1470 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1471 LI->eraseFromParent();
1474 // The old global is now dead, remove it.
1475 GV->eraseFromParent();
1478 return cast<GlobalVariable>(FieldGlobals[0]);
1481 /// TryToOptimizeStoreOfMallocToGlobal - This function is called when we see a
1482 /// pointer global variable with a single value stored it that is a malloc or
1484 static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV,
1486 const Type *AllocTy,
1487 Module::global_iterator &GVI,
1492 // If this is a malloc of an abstract type, don't touch it.
1493 if (!AllocTy->isSized())
1496 // We can't optimize this global unless all uses of it are *known* to be
1497 // of the malloc value, not of the null initializer value (consider a use
1498 // that compares the global's value against zero to see if the malloc has
1499 // been reached). To do this, we check to see if all uses of the global
1500 // would trap if the global were null: this proves that they must all
1501 // happen after the malloc.
1502 if (!AllUsesOfLoadedValueWillTrapIfNull(GV))
1505 // We can't optimize this if the malloc itself is used in a complex way,
1506 // for example, being stored into multiple globals. This allows the
1507 // malloc to be stored into the specified global, loaded setcc'd, and
1508 // GEP'd. These are all things we could transform to using the global
1510 SmallPtrSet<const PHINode*, 8> PHIs;
1511 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(CI, GV, PHIs))
1514 // If we have a global that is only initialized with a fixed size malloc,
1515 // transform the program to use global memory instead of malloc'd memory.
1516 // This eliminates dynamic allocation, avoids an indirection accessing the
1517 // data, and exposes the resultant global to further GlobalOpt.
1518 // We cannot optimize the malloc if we cannot determine malloc array size.
1519 Value *NElems = getMallocArraySize(CI, TD, true);
1523 if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems))
1524 // Restrict this transformation to only working on small allocations
1525 // (2048 bytes currently), as we don't want to introduce a 16M global or
1527 if (NElements->getZExtValue() * TD->getTypeAllocSize(AllocTy) < 2048) {
1528 GVI = OptimizeGlobalAddressOfMalloc(GV, CI, AllocTy, NElements, TD);
1532 // If the allocation is an array of structures, consider transforming this
1533 // into multiple malloc'd arrays, one for each field. This is basically
1534 // SRoA for malloc'd memory.
1536 // If this is an allocation of a fixed size array of structs, analyze as a
1537 // variable size array. malloc [100 x struct],1 -> malloc struct, 100
1538 if (NElems == ConstantInt::get(CI->getArgOperand(0)->getType(), 1))
1539 if (const ArrayType *AT = dyn_cast<ArrayType>(AllocTy))
1540 AllocTy = AT->getElementType();
1542 const StructType *AllocSTy = dyn_cast<StructType>(AllocTy);
1546 // This the structure has an unreasonable number of fields, leave it
1548 if (AllocSTy->getNumElements() <= 16 && AllocSTy->getNumElements() != 0 &&
1549 AllGlobalLoadUsesSimpleEnoughForHeapSRA(GV, CI)) {
1551 // If this is a fixed size array, transform the Malloc to be an alloc of
1552 // structs. malloc [100 x struct],1 -> malloc struct, 100
1553 if (const ArrayType *AT = dyn_cast<ArrayType>(getMallocAllocatedType(CI))) {
1554 const Type *IntPtrTy = TD->getIntPtrType(CI->getContext());
1555 unsigned TypeSize = TD->getStructLayout(AllocSTy)->getSizeInBytes();
1556 Value *AllocSize = ConstantInt::get(IntPtrTy, TypeSize);
1557 Value *NumElements = ConstantInt::get(IntPtrTy, AT->getNumElements());
1558 Instruction *Malloc = CallInst::CreateMalloc(CI, IntPtrTy, AllocSTy,
1559 AllocSize, NumElements,
1561 Instruction *Cast = new BitCastInst(Malloc, CI->getType(), "tmp", CI);
1562 CI->replaceAllUsesWith(Cast);
1563 CI->eraseFromParent();
1564 CI = dyn_cast<BitCastInst>(Malloc) ?
1565 extractMallocCallFromBitCast(Malloc) : cast<CallInst>(Malloc);
1568 GVI = PerformHeapAllocSRoA(GV, CI, getMallocArraySize(CI, TD, true),TD);
1575 // OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge
1576 // that only one value (besides its initializer) is ever stored to the global.
1577 static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal,
1578 Module::global_iterator &GVI,
1580 // Ignore no-op GEPs and bitcasts.
1581 StoredOnceVal = StoredOnceVal->stripPointerCasts();
1583 // If we are dealing with a pointer global that is initialized to null and
1584 // only has one (non-null) value stored into it, then we can optimize any
1585 // users of the loaded value (often calls and loads) that would trap if the
1587 if (GV->getInitializer()->getType()->isPointerTy() &&
1588 GV->getInitializer()->isNullValue()) {
1589 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) {
1590 if (GV->getInitializer()->getType() != SOVC->getType())
1592 ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType());
1594 // Optimize away any trapping uses of the loaded value.
1595 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC))
1597 } else if (CallInst *CI = extractMallocCall(StoredOnceVal)) {
1598 const Type* MallocType = getMallocAllocatedType(CI);
1599 if (MallocType && TryToOptimizeStoreOfMallocToGlobal(GV, CI, MallocType,
1608 /// TryToShrinkGlobalToBoolean - At this point, we have learned that the only
1609 /// two values ever stored into GV are its initializer and OtherVal. See if we
1610 /// can shrink the global into a boolean and select between the two values
1611 /// whenever it is used. This exposes the values to other scalar optimizations.
1612 static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) {
1613 const Type *GVElType = GV->getType()->getElementType();
1615 // If GVElType is already i1, it is already shrunk. If the type of the GV is
1616 // an FP value, pointer or vector, don't do this optimization because a select
1617 // between them is very expensive and unlikely to lead to later
1618 // simplification. In these cases, we typically end up with "cond ? v1 : v2"
1619 // where v1 and v2 both require constant pool loads, a big loss.
1620 if (GVElType == Type::getInt1Ty(GV->getContext()) ||
1621 GVElType->isFloatingPointTy() ||
1622 GVElType->isPointerTy() || GVElType->isVectorTy())
1625 // Walk the use list of the global seeing if all the uses are load or store.
1626 // If there is anything else, bail out.
1627 for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I){
1629 if (!isa<LoadInst>(U) && !isa<StoreInst>(U))
1633 DEBUG(dbgs() << " *** SHRINKING TO BOOL: " << *GV);
1635 // Create the new global, initializing it to false.
1636 GlobalVariable *NewGV = new GlobalVariable(Type::getInt1Ty(GV->getContext()),
1638 GlobalValue::InternalLinkage,
1639 ConstantInt::getFalse(GV->getContext()),
1641 GV->isThreadLocal());
1642 GV->getParent()->getGlobalList().insert(GV, NewGV);
1644 Constant *InitVal = GV->getInitializer();
1645 assert(InitVal->getType() != Type::getInt1Ty(GV->getContext()) &&
1646 "No reason to shrink to bool!");
1648 // If initialized to zero and storing one into the global, we can use a cast
1649 // instead of a select to synthesize the desired value.
1650 bool IsOneZero = false;
1651 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal))
1652 IsOneZero = InitVal->isNullValue() && CI->isOne();
1654 while (!GV->use_empty()) {
1655 Instruction *UI = cast<Instruction>(GV->use_back());
1656 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
1657 // Change the store into a boolean store.
1658 bool StoringOther = SI->getOperand(0) == OtherVal;
1659 // Only do this if we weren't storing a loaded value.
1661 if (StoringOther || SI->getOperand(0) == InitVal)
1662 StoreVal = ConstantInt::get(Type::getInt1Ty(GV->getContext()),
1665 // Otherwise, we are storing a previously loaded copy. To do this,
1666 // change the copy from copying the original value to just copying the
1668 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0));
1670 // If we've already replaced the input, StoredVal will be a cast or
1671 // select instruction. If not, it will be a load of the original
1673 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
1674 assert(LI->getOperand(0) == GV && "Not a copy!");
1675 // Insert a new load, to preserve the saved value.
1676 StoreVal = new LoadInst(NewGV, LI->getName()+".b", LI);
1678 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) &&
1679 "This is not a form that we understand!");
1680 StoreVal = StoredVal->getOperand(0);
1681 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!");
1684 new StoreInst(StoreVal, NewGV, SI);
1686 // Change the load into a load of bool then a select.
1687 LoadInst *LI = cast<LoadInst>(UI);
1688 LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", LI);
1691 NSI = new ZExtInst(NLI, LI->getType(), "", LI);
1693 NSI = SelectInst::Create(NLI, OtherVal, InitVal, "", LI);
1695 LI->replaceAllUsesWith(NSI);
1697 UI->eraseFromParent();
1700 GV->eraseFromParent();
1705 /// ProcessInternalGlobal - Analyze the specified global variable and optimize
1706 /// it if possible. If we make a change, return true.
1707 bool GlobalOpt::ProcessGlobal(GlobalVariable *GV,
1708 Module::global_iterator &GVI) {
1709 if (!GV->hasLocalLinkage())
1712 // Do more involved optimizations if the global is internal.
1713 GV->removeDeadConstantUsers();
1715 if (GV->use_empty()) {
1716 DEBUG(dbgs() << "GLOBAL DEAD: " << *GV);
1717 GV->eraseFromParent();
1722 SmallPtrSet<const PHINode*, 16> PHIUsers;
1725 if (AnalyzeGlobal(GV, GS, PHIUsers))
1728 if (!GS.isCompared && !GV->hasUnnamedAddr()) {
1729 GV->setUnnamedAddr(true);
1733 if (GV->isConstant() || !GV->hasInitializer())
1736 return ProcessInternalGlobal(GV, GVI, PHIUsers, GS);
1739 /// ProcessInternalGlobal - Analyze the specified global variable and optimize
1740 /// it if possible. If we make a change, return true.
1741 bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV,
1742 Module::global_iterator &GVI,
1743 const SmallPtrSet<const PHINode*, 16> &PHIUsers,
1744 const GlobalStatus &GS) {
1745 // If this is a first class global and has only one accessing function
1746 // and this function is main (which we know is not recursive we can make
1747 // this global a local variable) we replace the global with a local alloca
1748 // in this function.
1750 // NOTE: It doesn't make sense to promote non single-value types since we
1751 // are just replacing static memory to stack memory.
1753 // If the global is in different address space, don't bring it to stack.
1754 if (!GS.HasMultipleAccessingFunctions &&
1755 GS.AccessingFunction && !GS.HasNonInstructionUser &&
1756 GV->getType()->getElementType()->isSingleValueType() &&
1757 GS.AccessingFunction->getName() == "main" &&
1758 GS.AccessingFunction->hasExternalLinkage() &&
1759 GV->getType()->getAddressSpace() == 0) {
1760 DEBUG(dbgs() << "LOCALIZING GLOBAL: " << *GV);
1761 Instruction& FirstI = const_cast<Instruction&>(*GS.AccessingFunction
1762 ->getEntryBlock().begin());
1763 const Type* ElemTy = GV->getType()->getElementType();
1764 // FIXME: Pass Global's alignment when globals have alignment
1765 AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), &FirstI);
1766 if (!isa<UndefValue>(GV->getInitializer()))
1767 new StoreInst(GV->getInitializer(), Alloca, &FirstI);
1769 GV->replaceAllUsesWith(Alloca);
1770 GV->eraseFromParent();
1775 // If the global is never loaded (but may be stored to), it is dead.
1778 DEBUG(dbgs() << "GLOBAL NEVER LOADED: " << *GV);
1780 // Delete any stores we can find to the global. We may not be able to
1781 // make it completely dead though.
1782 bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer());
1784 // If the global is dead now, delete it.
1785 if (GV->use_empty()) {
1786 GV->eraseFromParent();
1792 } else if (GS.StoredType <= GlobalStatus::isInitializerStored) {
1793 DEBUG(dbgs() << "MARKING CONSTANT: " << *GV);
1794 GV->setConstant(true);
1796 // Clean up any obviously simplifiable users now.
1797 CleanupConstantGlobalUsers(GV, GV->getInitializer());
1799 // If the global is dead now, just nuke it.
1800 if (GV->use_empty()) {
1801 DEBUG(dbgs() << " *** Marking constant allowed us to simplify "
1802 << "all users and delete global!\n");
1803 GV->eraseFromParent();
1809 } else if (!GV->getInitializer()->getType()->isSingleValueType()) {
1810 if (TargetData *TD = getAnalysisIfAvailable<TargetData>())
1811 if (GlobalVariable *FirstNewGV = SRAGlobal(GV, *TD)) {
1812 GVI = FirstNewGV; // Don't skip the newly produced globals!
1815 } else if (GS.StoredType == GlobalStatus::isStoredOnce) {
1816 // If the initial value for the global was an undef value, and if only
1817 // one other value was stored into it, we can just change the
1818 // initializer to be the stored value, then delete all stores to the
1819 // global. This allows us to mark it constant.
1820 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1821 if (isa<UndefValue>(GV->getInitializer())) {
1822 // Change the initial value here.
1823 GV->setInitializer(SOVConstant);
1825 // Clean up any obviously simplifiable users now.
1826 CleanupConstantGlobalUsers(GV, GV->getInitializer());
1828 if (GV->use_empty()) {
1829 DEBUG(dbgs() << " *** Substituting initializer allowed us to "
1830 << "simplify all users and delete global!\n");
1831 GV->eraseFromParent();
1840 // Try to optimize globals based on the knowledge that only one value
1841 // (besides its initializer) is ever stored to the global.
1842 if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI,
1843 getAnalysisIfAvailable<TargetData>()))
1846 // Otherwise, if the global was not a boolean, we can shrink it to be a
1848 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1849 if (TryToShrinkGlobalToBoolean(GV, SOVConstant)) {
1858 /// ChangeCalleesToFastCall - Walk all of the direct calls of the specified
1859 /// function, changing them to FastCC.
1860 static void ChangeCalleesToFastCall(Function *F) {
1861 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1862 CallSite User(cast<Instruction>(*UI));
1863 User.setCallingConv(CallingConv::Fast);
1867 static AttrListPtr StripNest(const AttrListPtr &Attrs) {
1868 for (unsigned i = 0, e = Attrs.getNumSlots(); i != e; ++i) {
1869 if ((Attrs.getSlot(i).Attrs & Attribute::Nest) == 0)
1872 // There can be only one.
1873 return Attrs.removeAttr(Attrs.getSlot(i).Index, Attribute::Nest);
1879 static void RemoveNestAttribute(Function *F) {
1880 F->setAttributes(StripNest(F->getAttributes()));
1881 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1882 CallSite User(cast<Instruction>(*UI));
1883 User.setAttributes(StripNest(User.getAttributes()));
1887 bool GlobalOpt::OptimizeFunctions(Module &M) {
1888 bool Changed = false;
1889 // Optimize functions.
1890 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) {
1892 // Functions without names cannot be referenced outside this module.
1893 if (!F->hasName() && !F->isDeclaration())
1894 F->setLinkage(GlobalValue::InternalLinkage);
1895 F->removeDeadConstantUsers();
1896 if (F->use_empty() && (F->hasLocalLinkage() || F->hasLinkOnceLinkage())) {
1897 F->eraseFromParent();
1900 } else if (F->hasLocalLinkage()) {
1901 if (F->getCallingConv() == CallingConv::C && !F->isVarArg() &&
1902 !F->hasAddressTaken()) {
1903 // If this function has C calling conventions, is not a varargs
1904 // function, and is only called directly, promote it to use the Fast
1905 // calling convention.
1906 F->setCallingConv(CallingConv::Fast);
1907 ChangeCalleesToFastCall(F);
1912 if (F->getAttributes().hasAttrSomewhere(Attribute::Nest) &&
1913 !F->hasAddressTaken()) {
1914 // The function is not used by a trampoline intrinsic, so it is safe
1915 // to remove the 'nest' attribute.
1916 RemoveNestAttribute(F);
1925 bool GlobalOpt::OptimizeGlobalVars(Module &M) {
1926 bool Changed = false;
1927 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
1929 GlobalVariable *GV = GVI++;
1930 // Global variables without names cannot be referenced outside this module.
1931 if (!GV->hasName() && !GV->isDeclaration())
1932 GV->setLinkage(GlobalValue::InternalLinkage);
1933 // Simplify the initializer.
1934 if (GV->hasInitializer())
1935 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GV->getInitializer())) {
1936 TargetData *TD = getAnalysisIfAvailable<TargetData>();
1937 Constant *New = ConstantFoldConstantExpression(CE, TD);
1938 if (New && New != CE)
1939 GV->setInitializer(New);
1942 Changed |= ProcessGlobal(GV, GVI);
1947 /// FindGlobalCtors - Find the llvm.globalctors list, verifying that all
1948 /// initializers have an init priority of 65535.
1949 GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) {
1950 GlobalVariable *GV = M.getGlobalVariable("llvm.global_ctors");
1951 if (GV == 0) return 0;
1953 // Found it, verify it's an array of { int, void()* }.
1954 const ArrayType *ATy =dyn_cast<ArrayType>(GV->getType()->getElementType());
1956 const StructType *STy = dyn_cast<StructType>(ATy->getElementType());
1957 if (!STy || STy->getNumElements() != 2 ||
1958 !STy->getElementType(0)->isIntegerTy(32)) return 0;
1959 const PointerType *PFTy = dyn_cast<PointerType>(STy->getElementType(1));
1960 if (!PFTy) return 0;
1961 const FunctionType *FTy = dyn_cast<FunctionType>(PFTy->getElementType());
1962 if (!FTy || !FTy->getReturnType()->isVoidTy() ||
1963 FTy->isVarArg() || FTy->getNumParams() != 0)
1966 // Verify that the initializer is simple enough for us to handle. We are
1967 // only allowed to optimize the initializer if it is unique.
1968 if (!GV->hasUniqueInitializer()) return 0;
1970 ConstantArray *CA = dyn_cast<ConstantArray>(GV->getInitializer());
1973 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) {
1974 ConstantStruct *CS = dyn_cast<ConstantStruct>(*i);
1975 if (CS == 0) return 0;
1977 if (isa<ConstantPointerNull>(CS->getOperand(1)))
1980 // Must have a function or null ptr.
1981 if (!isa<Function>(CS->getOperand(1)))
1984 // Init priority must be standard.
1985 ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0));
1986 if (!CI || CI->getZExtValue() != 65535)
1993 /// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand,
1994 /// return a list of the functions and null terminator as a vector.
1995 static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) {
1996 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer());
1997 std::vector<Function*> Result;
1998 Result.reserve(CA->getNumOperands());
1999 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) {
2000 ConstantStruct *CS = cast<ConstantStruct>(*i);
2001 Result.push_back(dyn_cast<Function>(CS->getOperand(1)));
2006 /// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the
2007 /// specified array, returning the new global to use.
2008 static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL,
2009 const std::vector<Function*> &Ctors) {
2010 // If we made a change, reassemble the initializer list.
2011 std::vector<Constant*> CSVals;
2012 CSVals.push_back(ConstantInt::get(Type::getInt32Ty(GCL->getContext()),65535));
2013 CSVals.push_back(0);
2015 // Create the new init list.
2016 std::vector<Constant*> CAList;
2017 for (unsigned i = 0, e = Ctors.size(); i != e; ++i) {
2019 CSVals[1] = Ctors[i];
2021 const Type *FTy = FunctionType::get(Type::getVoidTy(GCL->getContext()),
2023 const PointerType *PFTy = PointerType::getUnqual(FTy);
2024 CSVals[1] = Constant::getNullValue(PFTy);
2025 CSVals[0] = ConstantInt::get(Type::getInt32Ty(GCL->getContext()),
2028 CAList.push_back(ConstantStruct::get(GCL->getContext(), CSVals, false));
2031 // Create the array initializer.
2032 const Type *StructTy =
2033 cast<ArrayType>(GCL->getType()->getElementType())->getElementType();
2034 Constant *CA = ConstantArray::get(ArrayType::get(StructTy,
2035 CAList.size()), CAList);
2037 // If we didn't change the number of elements, don't create a new GV.
2038 if (CA->getType() == GCL->getInitializer()->getType()) {
2039 GCL->setInitializer(CA);
2043 // Create the new global and insert it next to the existing list.
2044 GlobalVariable *NGV = new GlobalVariable(CA->getType(), GCL->isConstant(),
2045 GCL->getLinkage(), CA, "",
2046 GCL->isThreadLocal());
2047 GCL->getParent()->getGlobalList().insert(GCL, NGV);
2050 // Nuke the old list, replacing any uses with the new one.
2051 if (!GCL->use_empty()) {
2053 if (V->getType() != GCL->getType())
2054 V = ConstantExpr::getBitCast(V, GCL->getType());
2055 GCL->replaceAllUsesWith(V);
2057 GCL->eraseFromParent();
2066 static Constant *getVal(DenseMap<Value*, Constant*> &ComputedValues, Value *V) {
2067 if (Constant *CV = dyn_cast<Constant>(V)) return CV;
2068 Constant *R = ComputedValues[V];
2069 assert(R && "Reference to an uncomputed value!");
2074 isSimpleEnoughValueToCommit(Constant *C,
2075 SmallPtrSet<Constant*, 8> &SimpleConstants);
2078 /// isSimpleEnoughValueToCommit - Return true if the specified constant can be
2079 /// handled by the code generator. We don't want to generate something like:
2080 /// void *X = &X/42;
2081 /// because the code generator doesn't have a relocation that can handle that.
2083 /// This function should be called if C was not found (but just got inserted)
2084 /// in SimpleConstants to avoid having to rescan the same constants all the
2086 static bool isSimpleEnoughValueToCommitHelper(Constant *C,
2087 SmallPtrSet<Constant*, 8> &SimpleConstants) {
2088 // Simple integer, undef, constant aggregate zero, global addresses, etc are
2090 if (C->getNumOperands() == 0 || isa<BlockAddress>(C) ||
2091 isa<GlobalValue>(C))
2094 // Aggregate values are safe if all their elements are.
2095 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C) ||
2096 isa<ConstantVector>(C)) {
2097 for (unsigned i = 0, e = C->getNumOperands(); i != e; ++i) {
2098 Constant *Op = cast<Constant>(C->getOperand(i));
2099 if (!isSimpleEnoughValueToCommit(Op, SimpleConstants))
2105 // We don't know exactly what relocations are allowed in constant expressions,
2106 // so we allow &global+constantoffset, which is safe and uniformly supported
2108 ConstantExpr *CE = cast<ConstantExpr>(C);
2109 switch (CE->getOpcode()) {
2110 case Instruction::BitCast:
2111 case Instruction::IntToPtr:
2112 case Instruction::PtrToInt:
2113 // These casts are always fine if the casted value is.
2114 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants);
2116 // GEP is fine if it is simple + constant offset.
2117 case Instruction::GetElementPtr:
2118 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
2119 if (!isa<ConstantInt>(CE->getOperand(i)))
2121 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants);
2123 case Instruction::Add:
2124 // We allow simple+cst.
2125 if (!isa<ConstantInt>(CE->getOperand(1)))
2127 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants);
2133 isSimpleEnoughValueToCommit(Constant *C,
2134 SmallPtrSet<Constant*, 8> &SimpleConstants) {
2135 // If we already checked this constant, we win.
2136 if (!SimpleConstants.insert(C)) return true;
2137 // Check the constant.
2138 return isSimpleEnoughValueToCommitHelper(C, SimpleConstants);
2142 /// isSimpleEnoughPointerToCommit - Return true if this constant is simple
2143 /// enough for us to understand. In particular, if it is a cast to anything
2144 /// other than from one pointer type to another pointer type, we punt.
2145 /// We basically just support direct accesses to globals and GEP's of
2146 /// globals. This should be kept up to date with CommitValueTo.
2147 static bool isSimpleEnoughPointerToCommit(Constant *C) {
2148 // Conservatively, avoid aggregate types. This is because we don't
2149 // want to worry about them partially overlapping other stores.
2150 if (!cast<PointerType>(C->getType())->getElementType()->isSingleValueType())
2153 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
2154 // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or
2155 // external globals.
2156 return GV->hasUniqueInitializer();
2158 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
2159 // Handle a constantexpr gep.
2160 if (CE->getOpcode() == Instruction::GetElementPtr &&
2161 isa<GlobalVariable>(CE->getOperand(0)) &&
2162 cast<GEPOperator>(CE)->isInBounds()) {
2163 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2164 // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or
2165 // external globals.
2166 if (!GV->hasUniqueInitializer())
2169 // The first index must be zero.
2170 ConstantInt *CI = dyn_cast<ConstantInt>(*llvm::next(CE->op_begin()));
2171 if (!CI || !CI->isZero()) return false;
2173 // The remaining indices must be compile-time known integers within the
2174 // notional bounds of the corresponding static array types.
2175 if (!CE->isGEPWithNoNotionalOverIndexing())
2178 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
2180 // A constantexpr bitcast from a pointer to another pointer is a no-op,
2181 // and we know how to evaluate it by moving the bitcast from the pointer
2182 // operand to the value operand.
2183 } else if (CE->getOpcode() == Instruction::BitCast &&
2184 isa<GlobalVariable>(CE->getOperand(0))) {
2185 // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or
2186 // external globals.
2187 return cast<GlobalVariable>(CE->getOperand(0))->hasUniqueInitializer();
2194 /// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global
2195 /// initializer. This returns 'Init' modified to reflect 'Val' stored into it.
2196 /// At this point, the GEP operands of Addr [0, OpNo) have been stepped into.
2197 static Constant *EvaluateStoreInto(Constant *Init, Constant *Val,
2198 ConstantExpr *Addr, unsigned OpNo) {
2199 // Base case of the recursion.
2200 if (OpNo == Addr->getNumOperands()) {
2201 assert(Val->getType() == Init->getType() && "Type mismatch!");
2205 std::vector<Constant*> Elts;
2206 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
2208 // Break up the constant into its elements.
2209 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
2210 for (User::op_iterator i = CS->op_begin(), e = CS->op_end(); i != e; ++i)
2211 Elts.push_back(cast<Constant>(*i));
2212 } else if (isa<ConstantAggregateZero>(Init)) {
2213 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
2214 Elts.push_back(Constant::getNullValue(STy->getElementType(i)));
2215 } else if (isa<UndefValue>(Init)) {
2216 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
2217 Elts.push_back(UndefValue::get(STy->getElementType(i)));
2219 llvm_unreachable("This code is out of sync with "
2220 " ConstantFoldLoadThroughGEPConstantExpr");
2223 // Replace the element that we are supposed to.
2224 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo));
2225 unsigned Idx = CU->getZExtValue();
2226 assert(Idx < STy->getNumElements() && "Struct index out of range!");
2227 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1);
2229 // Return the modified struct.
2230 return ConstantStruct::get(Init->getContext(), &Elts[0], Elts.size(),
2233 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo));
2234 const SequentialType *InitTy = cast<SequentialType>(Init->getType());
2237 if (const ArrayType *ATy = dyn_cast<ArrayType>(InitTy))
2238 NumElts = ATy->getNumElements();
2240 NumElts = cast<VectorType>(InitTy)->getNumElements();
2243 // Break up the array into elements.
2244 if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
2245 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i)
2246 Elts.push_back(cast<Constant>(*i));
2247 } else if (ConstantVector *CV = dyn_cast<ConstantVector>(Init)) {
2248 for (User::op_iterator i = CV->op_begin(), e = CV->op_end(); i != e; ++i)
2249 Elts.push_back(cast<Constant>(*i));
2250 } else if (isa<ConstantAggregateZero>(Init)) {
2251 Elts.assign(NumElts, Constant::getNullValue(InitTy->getElementType()));
2253 assert(isa<UndefValue>(Init) && "This code is out of sync with "
2254 " ConstantFoldLoadThroughGEPConstantExpr");
2255 Elts.assign(NumElts, UndefValue::get(InitTy->getElementType()));
2258 assert(CI->getZExtValue() < NumElts);
2259 Elts[CI->getZExtValue()] =
2260 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1);
2262 if (Init->getType()->isArrayTy())
2263 return ConstantArray::get(cast<ArrayType>(InitTy), Elts);
2264 return ConstantVector::get(Elts);
2268 /// CommitValueTo - We have decided that Addr (which satisfies the predicate
2269 /// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen.
2270 static void CommitValueTo(Constant *Val, Constant *Addr) {
2271 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
2272 assert(GV->hasInitializer());
2273 GV->setInitializer(Val);
2277 ConstantExpr *CE = cast<ConstantExpr>(Addr);
2278 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2279 GV->setInitializer(EvaluateStoreInto(GV->getInitializer(), Val, CE, 2));
2282 /// ComputeLoadResult - Return the value that would be computed by a load from
2283 /// P after the stores reflected by 'memory' have been performed. If we can't
2284 /// decide, return null.
2285 static Constant *ComputeLoadResult(Constant *P,
2286 const DenseMap<Constant*, Constant*> &Memory) {
2287 // If this memory location has been recently stored, use the stored value: it
2288 // is the most up-to-date.
2289 DenseMap<Constant*, Constant*>::const_iterator I = Memory.find(P);
2290 if (I != Memory.end()) return I->second;
2293 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
2294 if (GV->hasDefinitiveInitializer())
2295 return GV->getInitializer();
2299 // Handle a constantexpr getelementptr.
2300 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P))
2301 if (CE->getOpcode() == Instruction::GetElementPtr &&
2302 isa<GlobalVariable>(CE->getOperand(0))) {
2303 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2304 if (GV->hasDefinitiveInitializer())
2305 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
2308 return 0; // don't know how to evaluate.
2311 /// EvaluateFunction - Evaluate a call to function F, returning true if
2312 /// successful, false if we can't evaluate it. ActualArgs contains the formal
2313 /// arguments for the function.
2314 static bool EvaluateFunction(Function *F, Constant *&RetVal,
2315 const SmallVectorImpl<Constant*> &ActualArgs,
2316 std::vector<Function*> &CallStack,
2317 DenseMap<Constant*, Constant*> &MutatedMemory,
2318 std::vector<GlobalVariable*> &AllocaTmps,
2319 SmallPtrSet<Constant*, 8> &SimpleConstants,
2320 const TargetData *TD) {
2321 // Check to see if this function is already executing (recursion). If so,
2322 // bail out. TODO: we might want to accept limited recursion.
2323 if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end())
2326 CallStack.push_back(F);
2328 /// Values - As we compute SSA register values, we store their contents here.
2329 DenseMap<Value*, Constant*> Values;
2331 // Initialize arguments to the incoming values specified.
2333 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E;
2335 Values[AI] = ActualArgs[ArgNo];
2337 /// ExecutedBlocks - We only handle non-looping, non-recursive code. As such,
2338 /// we can only evaluate any one basic block at most once. This set keeps
2339 /// track of what we have executed so we can detect recursive cases etc.
2340 SmallPtrSet<BasicBlock*, 32> ExecutedBlocks;
2342 // CurInst - The current instruction we're evaluating.
2343 BasicBlock::iterator CurInst = F->begin()->begin();
2345 // This is the main evaluation loop.
2347 Constant *InstResult = 0;
2349 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) {
2350 if (SI->isVolatile()) return false; // no volatile accesses.
2351 Constant *Ptr = getVal(Values, SI->getOperand(1));
2352 if (!isSimpleEnoughPointerToCommit(Ptr))
2353 // If this is too complex for us to commit, reject it.
2356 Constant *Val = getVal(Values, SI->getOperand(0));
2358 // If this might be too difficult for the backend to handle (e.g. the addr
2359 // of one global variable divided by another) then we can't commit it.
2360 if (!isSimpleEnoughValueToCommit(Val, SimpleConstants))
2363 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
2364 if (CE->getOpcode() == Instruction::BitCast) {
2365 // If we're evaluating a store through a bitcast, then we need
2366 // to pull the bitcast off the pointer type and push it onto the
2368 Ptr = CE->getOperand(0);
2370 const Type *NewTy=cast<PointerType>(Ptr->getType())->getElementType();
2372 // In order to push the bitcast onto the stored value, a bitcast
2373 // from NewTy to Val's type must be legal. If it's not, we can try
2374 // introspecting NewTy to find a legal conversion.
2375 while (!Val->getType()->canLosslesslyBitCastTo(NewTy)) {
2376 // If NewTy is a struct, we can convert the pointer to the struct
2377 // into a pointer to its first member.
2378 // FIXME: This could be extended to support arrays as well.
2379 if (const StructType *STy = dyn_cast<StructType>(NewTy)) {
2380 NewTy = STy->getTypeAtIndex(0U);
2382 const IntegerType *IdxTy =IntegerType::get(NewTy->getContext(), 32);
2383 Constant *IdxZero = ConstantInt::get(IdxTy, 0, false);
2384 Constant * const IdxList[] = {IdxZero, IdxZero};
2386 Ptr = ConstantExpr::getGetElementPtr(Ptr, IdxList, 2);
2388 // If we can't improve the situation by introspecting NewTy,
2389 // we have to give up.
2395 // If we found compatible types, go ahead and push the bitcast
2396 // onto the stored value.
2397 Val = ConstantExpr::getBitCast(Val, NewTy);
2400 MutatedMemory[Ptr] = Val;
2401 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) {
2402 InstResult = ConstantExpr::get(BO->getOpcode(),
2403 getVal(Values, BO->getOperand(0)),
2404 getVal(Values, BO->getOperand(1)));
2405 } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) {
2406 InstResult = ConstantExpr::getCompare(CI->getPredicate(),
2407 getVal(Values, CI->getOperand(0)),
2408 getVal(Values, CI->getOperand(1)));
2409 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) {
2410 InstResult = ConstantExpr::getCast(CI->getOpcode(),
2411 getVal(Values, CI->getOperand(0)),
2413 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) {
2414 InstResult = ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)),
2415 getVal(Values, SI->getOperand(1)),
2416 getVal(Values, SI->getOperand(2)));
2417 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) {
2418 Constant *P = getVal(Values, GEP->getOperand(0));
2419 SmallVector<Constant*, 8> GEPOps;
2420 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end();
2422 GEPOps.push_back(getVal(Values, *i));
2423 InstResult = cast<GEPOperator>(GEP)->isInBounds() ?
2424 ConstantExpr::getInBoundsGetElementPtr(P, &GEPOps[0], GEPOps.size()) :
2425 ConstantExpr::getGetElementPtr(P, &GEPOps[0], GEPOps.size());
2426 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) {
2427 if (LI->isVolatile()) return false; // no volatile accesses.
2428 InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)),
2430 if (InstResult == 0) return false; // Could not evaluate load.
2431 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) {
2432 if (AI->isArrayAllocation()) return false; // Cannot handle array allocs.
2433 const Type *Ty = AI->getType()->getElementType();
2434 AllocaTmps.push_back(new GlobalVariable(Ty, false,
2435 GlobalValue::InternalLinkage,
2436 UndefValue::get(Ty),
2438 InstResult = AllocaTmps.back();
2439 } else if (CallInst *CI = dyn_cast<CallInst>(CurInst)) {
2441 // Debug info can safely be ignored here.
2442 if (isa<DbgInfoIntrinsic>(CI)) {
2447 // Cannot handle inline asm.
2448 if (isa<InlineAsm>(CI->getCalledValue())) return false;
2450 // Resolve function pointers.
2451 Function *Callee = dyn_cast<Function>(getVal(Values,
2452 CI->getCalledValue()));
2453 if (!Callee) return false; // Cannot resolve.
2455 SmallVector<Constant*, 8> Formals;
2457 for (User::op_iterator i = CS.arg_begin(), e = CS.arg_end();
2459 Formals.push_back(getVal(Values, *i));
2461 if (Callee->isDeclaration()) {
2462 // If this is a function we can constant fold, do it.
2463 if (Constant *C = ConstantFoldCall(Callee, Formals.data(),
2470 if (Callee->getFunctionType()->isVarArg())
2474 // Execute the call, if successful, use the return value.
2475 if (!EvaluateFunction(Callee, RetVal, Formals, CallStack,
2476 MutatedMemory, AllocaTmps, SimpleConstants, TD))
2478 InstResult = RetVal;
2480 } else if (isa<TerminatorInst>(CurInst)) {
2481 BasicBlock *NewBB = 0;
2482 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) {
2483 if (BI->isUnconditional()) {
2484 NewBB = BI->getSuccessor(0);
2487 dyn_cast<ConstantInt>(getVal(Values, BI->getCondition()));
2488 if (!Cond) return false; // Cannot determine.
2490 NewBB = BI->getSuccessor(!Cond->getZExtValue());
2492 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) {
2494 dyn_cast<ConstantInt>(getVal(Values, SI->getCondition()));
2495 if (!Val) return false; // Cannot determine.
2496 NewBB = SI->getSuccessor(SI->findCaseValue(Val));
2497 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(CurInst)) {
2498 Value *Val = getVal(Values, IBI->getAddress())->stripPointerCasts();
2499 if (BlockAddress *BA = dyn_cast<BlockAddress>(Val))
2500 NewBB = BA->getBasicBlock();
2502 return false; // Cannot determine.
2503 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) {
2504 if (RI->getNumOperands())
2505 RetVal = getVal(Values, RI->getOperand(0));
2507 CallStack.pop_back(); // return from fn.
2508 return true; // We succeeded at evaluating this ctor!
2510 // invoke, unwind, unreachable.
2511 return false; // Cannot handle this terminator.
2514 // Okay, we succeeded in evaluating this control flow. See if we have
2515 // executed the new block before. If so, we have a looping function,
2516 // which we cannot evaluate in reasonable time.
2517 if (!ExecutedBlocks.insert(NewBB))
2518 return false; // looped!
2520 // Okay, we have never been in this block before. Check to see if there
2521 // are any PHI nodes. If so, evaluate them with information about where
2523 BasicBlock *OldBB = CurInst->getParent();
2524 CurInst = NewBB->begin();
2526 for (; (PN = dyn_cast<PHINode>(CurInst)); ++CurInst)
2527 Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB));
2529 // Do NOT increment CurInst. We know that the terminator had no value.
2532 // Did not know how to evaluate this!
2536 if (!CurInst->use_empty()) {
2537 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(InstResult))
2538 InstResult = ConstantFoldConstantExpression(CE, TD);
2540 Values[CurInst] = InstResult;
2543 // Advance program counter.
2548 /// EvaluateStaticConstructor - Evaluate static constructors in the function, if
2549 /// we can. Return true if we can, false otherwise.
2550 static bool EvaluateStaticConstructor(Function *F, const TargetData *TD) {
2551 /// MutatedMemory - For each store we execute, we update this map. Loads
2552 /// check this to get the most up-to-date value. If evaluation is successful,
2553 /// this state is committed to the process.
2554 DenseMap<Constant*, Constant*> MutatedMemory;
2556 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable
2557 /// to represent its body. This vector is needed so we can delete the
2558 /// temporary globals when we are done.
2559 std::vector<GlobalVariable*> AllocaTmps;
2561 /// CallStack - This is used to detect recursion. In pathological situations
2562 /// we could hit exponential behavior, but at least there is nothing
2564 std::vector<Function*> CallStack;
2566 /// SimpleConstants - These are constants we have checked and know to be
2567 /// simple enough to live in a static initializer of a global.
2568 SmallPtrSet<Constant*, 8> SimpleConstants;
2570 // Call the function.
2571 Constant *RetValDummy;
2572 bool EvalSuccess = EvaluateFunction(F, RetValDummy,
2573 SmallVector<Constant*, 0>(), CallStack,
2574 MutatedMemory, AllocaTmps,
2575 SimpleConstants, TD);
2578 // We succeeded at evaluation: commit the result.
2579 DEBUG(dbgs() << "FULLY EVALUATED GLOBAL CTOR FUNCTION '"
2580 << F->getName() << "' to " << MutatedMemory.size()
2582 for (DenseMap<Constant*, Constant*>::iterator I = MutatedMemory.begin(),
2583 E = MutatedMemory.end(); I != E; ++I)
2584 CommitValueTo(I->second, I->first);
2587 // At this point, we are done interpreting. If we created any 'alloca'
2588 // temporaries, release them now.
2589 while (!AllocaTmps.empty()) {
2590 GlobalVariable *Tmp = AllocaTmps.back();
2591 AllocaTmps.pop_back();
2593 // If there are still users of the alloca, the program is doing something
2594 // silly, e.g. storing the address of the alloca somewhere and using it
2595 // later. Since this is undefined, we'll just make it be null.
2596 if (!Tmp->use_empty())
2597 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType()));
2606 /// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible.
2607 /// Return true if anything changed.
2608 bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) {
2609 std::vector<Function*> Ctors = ParseGlobalCtors(GCL);
2610 bool MadeChange = false;
2611 if (Ctors.empty()) return false;
2613 const TargetData *TD = getAnalysisIfAvailable<TargetData>();
2614 // Loop over global ctors, optimizing them when we can.
2615 for (unsigned i = 0; i != Ctors.size(); ++i) {
2616 Function *F = Ctors[i];
2617 // Found a null terminator in the middle of the list, prune off the rest of
2620 if (i != Ctors.size()-1) {
2627 // We cannot simplify external ctor functions.
2628 if (F->empty()) continue;
2630 // If we can evaluate the ctor at compile time, do.
2631 if (EvaluateStaticConstructor(F, TD)) {
2632 Ctors.erase(Ctors.begin()+i);
2635 ++NumCtorsEvaluated;
2640 if (!MadeChange) return false;
2642 GCL = InstallGlobalCtors(GCL, Ctors);
2646 bool GlobalOpt::OptimizeGlobalAliases(Module &M) {
2647 bool Changed = false;
2649 for (Module::alias_iterator I = M.alias_begin(), E = M.alias_end();
2651 Module::alias_iterator J = I++;
2652 // Aliases without names cannot be referenced outside this module.
2653 if (!J->hasName() && !J->isDeclaration())
2654 J->setLinkage(GlobalValue::InternalLinkage);
2655 // If the aliasee may change at link time, nothing can be done - bail out.
2656 if (J->mayBeOverridden())
2659 Constant *Aliasee = J->getAliasee();
2660 GlobalValue *Target = cast<GlobalValue>(Aliasee->stripPointerCasts());
2661 Target->removeDeadConstantUsers();
2662 bool hasOneUse = Target->hasOneUse() && Aliasee->hasOneUse();
2664 // Make all users of the alias use the aliasee instead.
2665 if (!J->use_empty()) {
2666 J->replaceAllUsesWith(Aliasee);
2667 ++NumAliasesResolved;
2671 // If the alias is externally visible, we may still be able to simplify it.
2672 if (!J->hasLocalLinkage()) {
2673 // If the aliasee has internal linkage, give it the name and linkage
2674 // of the alias, and delete the alias. This turns:
2675 // define internal ... @f(...)
2676 // @a = alias ... @f
2678 // define ... @a(...)
2679 if (!Target->hasLocalLinkage())
2682 // Do not perform the transform if multiple aliases potentially target the
2683 // aliasee. This check also ensures that it is safe to replace the section
2684 // and other attributes of the aliasee with those of the alias.
2688 // Give the aliasee the name, linkage and other attributes of the alias.
2689 Target->takeName(J);
2690 Target->setLinkage(J->getLinkage());
2691 Target->GlobalValue::copyAttributesFrom(J);
2694 // Delete the alias.
2695 M.getAliasList().erase(J);
2696 ++NumAliasesRemoved;
2703 static Function *FindCXAAtExit(Module &M) {
2704 Function *Fn = M.getFunction("__cxa_atexit");
2709 const FunctionType *FTy = Fn->getFunctionType();
2711 // Checking that the function has the right return type, the right number of
2712 // parameters and that they all have pointer types should be enough.
2713 if (!FTy->getReturnType()->isIntegerTy() ||
2714 FTy->getNumParams() != 3 ||
2715 !FTy->getParamType(0)->isPointerTy() ||
2716 !FTy->getParamType(1)->isPointerTy() ||
2717 !FTy->getParamType(2)->isPointerTy())
2723 /// cxxDtorIsEmpty - Returns whether the given function is an empty C++
2724 /// destructor and can therefore be eliminated.
2725 /// Note that we assume that other optimization passes have already simplified
2726 /// the code so we only look for a function with a single basic block, where
2727 /// the only allowed instructions are 'ret' or 'call' to empty C++ dtor.
2728 static bool cxxDtorIsEmpty(const Function &Fn,
2729 SmallPtrSet<const Function *, 8> &CalledFunctions) {
2730 // FIXME: We could eliminate C++ destructors if they're readonly/readnone and
2731 // nounwind, but that doesn't seem worth doing.
2732 if (Fn.isDeclaration())
2735 if (++Fn.begin() != Fn.end())
2738 const BasicBlock &EntryBlock = Fn.getEntryBlock();
2739 for (BasicBlock::const_iterator I = EntryBlock.begin(), E = EntryBlock.end();
2741 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
2742 // Ignore debug intrinsics.
2743 if (isa<DbgInfoIntrinsic>(CI))
2746 const Function *CalledFn = CI->getCalledFunction();
2751 SmallPtrSet<const Function *, 8> NewCalledFunctions(CalledFunctions);
2753 // Don't treat recursive functions as empty.
2754 if (!NewCalledFunctions.insert(CalledFn))
2757 if (!cxxDtorIsEmpty(*CalledFn, NewCalledFunctions))
2759 } else if (isa<ReturnInst>(*I))
2768 bool GlobalOpt::OptimizeEmptyGlobalCXXDtors(Function *CXAAtExitFn) {
2769 /// Itanium C++ ABI p3.3.5:
2771 /// After constructing a global (or local static) object, that will require
2772 /// destruction on exit, a termination function is registered as follows:
2774 /// extern "C" int __cxa_atexit ( void (*f)(void *), void *p, void *d );
2776 /// This registration, e.g. __cxa_atexit(f,p,d), is intended to cause the
2777 /// call f(p) when DSO d is unloaded, before all such termination calls
2778 /// registered before this one. It returns zero if registration is
2779 /// successful, nonzero on failure.
2781 // This pass will look for calls to __cxa_atexit where the function is trivial
2783 bool Changed = false;
2785 for (Function::use_iterator I = CXAAtExitFn->use_begin(),
2786 E = CXAAtExitFn->use_end(); I != E;) {
2787 // We're only interested in calls. Theoretically, we could handle invoke
2788 // instructions as well, but neither llvm-gcc nor clang generate invokes
2790 CallInst *CI = dyn_cast<CallInst>(*I++);
2795 dyn_cast<Function>(CI->getArgOperand(0)->stripPointerCasts());
2799 SmallPtrSet<const Function *, 8> CalledFunctions;
2800 if (!cxxDtorIsEmpty(*DtorFn, CalledFunctions))
2803 // Just remove the call.
2804 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
2805 CI->eraseFromParent();
2807 ++NumCXXDtorsRemoved;
2815 bool GlobalOpt::runOnModule(Module &M) {
2816 bool Changed = false;
2818 // Try to find the llvm.globalctors list.
2819 GlobalVariable *GlobalCtors = FindGlobalCtors(M);
2821 Function *CXAAtExitFn = FindCXAAtExit(M);
2823 bool LocalChange = true;
2824 while (LocalChange) {
2825 LocalChange = false;
2827 // Delete functions that are trivially dead, ccc -> fastcc
2828 LocalChange |= OptimizeFunctions(M);
2830 // Optimize global_ctors list.
2832 LocalChange |= OptimizeGlobalCtorsList(GlobalCtors);
2834 // Optimize non-address-taken globals.
2835 LocalChange |= OptimizeGlobalVars(M);
2837 // Resolve aliases, when possible.
2838 LocalChange |= OptimizeGlobalAliases(M);
2840 // Try to remove trivial global destructors.
2842 LocalChange |= OptimizeEmptyGlobalCXXDtors(CXAAtExitFn);
2844 Changed |= LocalChange;
2847 // TODO: Move all global ctors functions to the end of the module for code