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->getName()+".f"+Twine(FieldNo), PN);
1199 PHIsToRewrite.push_back(std::make_pair(PN, FieldNo));
1201 llvm_unreachable("Unknown usable value");
1205 return FieldVals[FieldNo] = Result;
1208 /// RewriteHeapSROALoadUser - Given a load instruction and a value derived from
1209 /// the load, rewrite the derived value to use the HeapSRoA'd load.
1210 static void RewriteHeapSROALoadUser(Instruction *LoadUser,
1211 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1212 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1213 // If this is a comparison against null, handle it.
1214 if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) {
1215 assert(isa<ConstantPointerNull>(SCI->getOperand(1)));
1216 // If we have a setcc of the loaded pointer, we can use a setcc of any
1218 Value *NPtr = GetHeapSROAValue(SCI->getOperand(0), 0,
1219 InsertedScalarizedValues, PHIsToRewrite);
1221 Value *New = new ICmpInst(SCI, SCI->getPredicate(), NPtr,
1222 Constant::getNullValue(NPtr->getType()),
1224 SCI->replaceAllUsesWith(New);
1225 SCI->eraseFromParent();
1229 // Handle 'getelementptr Ptr, Idx, i32 FieldNo ...'
1230 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) {
1231 assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2))
1232 && "Unexpected GEPI!");
1234 // Load the pointer for this field.
1235 unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
1236 Value *NewPtr = GetHeapSROAValue(GEPI->getOperand(0), FieldNo,
1237 InsertedScalarizedValues, PHIsToRewrite);
1239 // Create the new GEP idx vector.
1240 SmallVector<Value*, 8> GEPIdx;
1241 GEPIdx.push_back(GEPI->getOperand(1));
1242 GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end());
1244 Value *NGEPI = GetElementPtrInst::Create(NewPtr,
1245 GEPIdx.begin(), GEPIdx.end(),
1246 GEPI->getName(), GEPI);
1247 GEPI->replaceAllUsesWith(NGEPI);
1248 GEPI->eraseFromParent();
1252 // Recursively transform the users of PHI nodes. This will lazily create the
1253 // PHIs that are needed for individual elements. Keep track of what PHIs we
1254 // see in InsertedScalarizedValues so that we don't get infinite loops (very
1255 // antisocial). If the PHI is already in InsertedScalarizedValues, it has
1256 // already been seen first by another load, so its uses have already been
1258 PHINode *PN = cast<PHINode>(LoadUser);
1260 DenseMap<Value*, std::vector<Value*> >::iterator InsertPos;
1261 tie(InsertPos, Inserted) =
1262 InsertedScalarizedValues.insert(std::make_pair(PN, std::vector<Value*>()));
1263 if (!Inserted) return;
1265 // If this is the first time we've seen this PHI, recursively process all
1267 for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end(); UI != E; ) {
1268 Instruction *User = cast<Instruction>(*UI++);
1269 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite);
1273 /// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global. Ptr
1274 /// is a value loaded from the global. Eliminate all uses of Ptr, making them
1275 /// use FieldGlobals instead. All uses of loaded values satisfy
1276 /// AllGlobalLoadUsesSimpleEnoughForHeapSRA.
1277 static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load,
1278 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1279 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1280 for (Value::use_iterator UI = Load->use_begin(), E = Load->use_end();
1282 Instruction *User = cast<Instruction>(*UI++);
1283 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite);
1286 if (Load->use_empty()) {
1287 Load->eraseFromParent();
1288 InsertedScalarizedValues.erase(Load);
1292 /// PerformHeapAllocSRoA - CI is an allocation of an array of structures. Break
1293 /// it up into multiple allocations of arrays of the fields.
1294 static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, CallInst *CI,
1295 Value* NElems, TargetData *TD) {
1296 DEBUG(dbgs() << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *CI << '\n');
1297 const Type* MAT = getMallocAllocatedType(CI);
1298 const StructType *STy = cast<StructType>(MAT);
1300 // There is guaranteed to be at least one use of the malloc (storing
1301 // it into GV). If there are other uses, change them to be uses of
1302 // the global to simplify later code. This also deletes the store
1304 ReplaceUsesOfMallocWithGlobal(CI, GV);
1306 // Okay, at this point, there are no users of the malloc. Insert N
1307 // new mallocs at the same place as CI, and N globals.
1308 std::vector<Value*> FieldGlobals;
1309 std::vector<Value*> FieldMallocs;
1311 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){
1312 const Type *FieldTy = STy->getElementType(FieldNo);
1313 const PointerType *PFieldTy = PointerType::getUnqual(FieldTy);
1315 GlobalVariable *NGV =
1316 new GlobalVariable(*GV->getParent(),
1317 PFieldTy, false, GlobalValue::InternalLinkage,
1318 Constant::getNullValue(PFieldTy),
1319 GV->getName() + ".f" + Twine(FieldNo), GV,
1320 GV->isThreadLocal());
1321 FieldGlobals.push_back(NGV);
1323 unsigned TypeSize = TD->getTypeAllocSize(FieldTy);
1324 if (const StructType *ST = dyn_cast<StructType>(FieldTy))
1325 TypeSize = TD->getStructLayout(ST)->getSizeInBytes();
1326 const Type *IntPtrTy = TD->getIntPtrType(CI->getContext());
1327 Value *NMI = CallInst::CreateMalloc(CI, IntPtrTy, FieldTy,
1328 ConstantInt::get(IntPtrTy, TypeSize),
1330 CI->getName() + ".f" + Twine(FieldNo));
1331 FieldMallocs.push_back(NMI);
1332 new StoreInst(NMI, NGV, CI);
1335 // The tricky aspect of this transformation is handling the case when malloc
1336 // fails. In the original code, malloc failing would set the result pointer
1337 // of malloc to null. In this case, some mallocs could succeed and others
1338 // could fail. As such, we emit code that looks like this:
1339 // F0 = malloc(field0)
1340 // F1 = malloc(field1)
1341 // F2 = malloc(field2)
1342 // if (F0 == 0 || F1 == 0 || F2 == 0) {
1343 // if (F0) { free(F0); F0 = 0; }
1344 // if (F1) { free(F1); F1 = 0; }
1345 // if (F2) { free(F2); F2 = 0; }
1347 // The malloc can also fail if its argument is too large.
1348 Constant *ConstantZero = ConstantInt::get(CI->getArgOperand(0)->getType(), 0);
1349 Value *RunningOr = new ICmpInst(CI, ICmpInst::ICMP_SLT, CI->getArgOperand(0),
1350 ConstantZero, "isneg");
1351 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) {
1352 Value *Cond = new ICmpInst(CI, ICmpInst::ICMP_EQ, FieldMallocs[i],
1353 Constant::getNullValue(FieldMallocs[i]->getType()),
1355 RunningOr = BinaryOperator::CreateOr(RunningOr, Cond, "tmp", CI);
1358 // Split the basic block at the old malloc.
1359 BasicBlock *OrigBB = CI->getParent();
1360 BasicBlock *ContBB = OrigBB->splitBasicBlock(CI, "malloc_cont");
1362 // Create the block to check the first condition. Put all these blocks at the
1363 // end of the function as they are unlikely to be executed.
1364 BasicBlock *NullPtrBlock = BasicBlock::Create(OrigBB->getContext(),
1366 OrigBB->getParent());
1368 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond
1369 // branch on RunningOr.
1370 OrigBB->getTerminator()->eraseFromParent();
1371 BranchInst::Create(NullPtrBlock, ContBB, RunningOr, OrigBB);
1373 // Within the NullPtrBlock, we need to emit a comparison and branch for each
1374 // pointer, because some may be null while others are not.
1375 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1376 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock);
1377 Value *Cmp = new ICmpInst(*NullPtrBlock, ICmpInst::ICMP_NE, GVVal,
1378 Constant::getNullValue(GVVal->getType()),
1380 BasicBlock *FreeBlock = BasicBlock::Create(Cmp->getContext(), "free_it",
1381 OrigBB->getParent());
1382 BasicBlock *NextBlock = BasicBlock::Create(Cmp->getContext(), "next",
1383 OrigBB->getParent());
1384 Instruction *BI = BranchInst::Create(FreeBlock, NextBlock,
1387 // Fill in FreeBlock.
1388 CallInst::CreateFree(GVVal, BI);
1389 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i],
1391 BranchInst::Create(NextBlock, FreeBlock);
1393 NullPtrBlock = NextBlock;
1396 BranchInst::Create(ContBB, NullPtrBlock);
1398 // CI is no longer needed, remove it.
1399 CI->eraseFromParent();
1401 /// InsertedScalarizedLoads - As we process loads, if we can't immediately
1402 /// update all uses of the load, keep track of what scalarized loads are
1403 /// inserted for a given load.
1404 DenseMap<Value*, std::vector<Value*> > InsertedScalarizedValues;
1405 InsertedScalarizedValues[GV] = FieldGlobals;
1407 std::vector<std::pair<PHINode*, unsigned> > PHIsToRewrite;
1409 // Okay, the malloc site is completely handled. All of the uses of GV are now
1410 // loads, and all uses of those loads are simple. Rewrite them to use loads
1411 // of the per-field globals instead.
1412 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;) {
1413 Instruction *User = cast<Instruction>(*UI++);
1415 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1416 RewriteUsesOfLoadForHeapSRoA(LI, InsertedScalarizedValues, PHIsToRewrite);
1420 // Must be a store of null.
1421 StoreInst *SI = cast<StoreInst>(User);
1422 assert(isa<ConstantPointerNull>(SI->getOperand(0)) &&
1423 "Unexpected heap-sra user!");
1425 // Insert a store of null into each global.
1426 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1427 const PointerType *PT = cast<PointerType>(FieldGlobals[i]->getType());
1428 Constant *Null = Constant::getNullValue(PT->getElementType());
1429 new StoreInst(Null, FieldGlobals[i], SI);
1431 // Erase the original store.
1432 SI->eraseFromParent();
1435 // While we have PHIs that are interesting to rewrite, do it.
1436 while (!PHIsToRewrite.empty()) {
1437 PHINode *PN = PHIsToRewrite.back().first;
1438 unsigned FieldNo = PHIsToRewrite.back().second;
1439 PHIsToRewrite.pop_back();
1440 PHINode *FieldPN = cast<PHINode>(InsertedScalarizedValues[PN][FieldNo]);
1441 assert(FieldPN->getNumIncomingValues() == 0 &&"Already processed this phi");
1443 // Add all the incoming values. This can materialize more phis.
1444 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1445 Value *InVal = PN->getIncomingValue(i);
1446 InVal = GetHeapSROAValue(InVal, FieldNo, InsertedScalarizedValues,
1448 FieldPN->addIncoming(InVal, PN->getIncomingBlock(i));
1452 // Drop all inter-phi links and any loads that made it this far.
1453 for (DenseMap<Value*, std::vector<Value*> >::iterator
1454 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1456 if (PHINode *PN = dyn_cast<PHINode>(I->first))
1457 PN->dropAllReferences();
1458 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1459 LI->dropAllReferences();
1462 // Delete all the phis and loads now that inter-references are dead.
1463 for (DenseMap<Value*, std::vector<Value*> >::iterator
1464 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1466 if (PHINode *PN = dyn_cast<PHINode>(I->first))
1467 PN->eraseFromParent();
1468 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1469 LI->eraseFromParent();
1472 // The old global is now dead, remove it.
1473 GV->eraseFromParent();
1476 return cast<GlobalVariable>(FieldGlobals[0]);
1479 /// TryToOptimizeStoreOfMallocToGlobal - This function is called when we see a
1480 /// pointer global variable with a single value stored it that is a malloc or
1482 static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV,
1484 const Type *AllocTy,
1485 Module::global_iterator &GVI,
1490 // If this is a malloc of an abstract type, don't touch it.
1491 if (!AllocTy->isSized())
1494 // We can't optimize this global unless all uses of it are *known* to be
1495 // of the malloc value, not of the null initializer value (consider a use
1496 // that compares the global's value against zero to see if the malloc has
1497 // been reached). To do this, we check to see if all uses of the global
1498 // would trap if the global were null: this proves that they must all
1499 // happen after the malloc.
1500 if (!AllUsesOfLoadedValueWillTrapIfNull(GV))
1503 // We can't optimize this if the malloc itself is used in a complex way,
1504 // for example, being stored into multiple globals. This allows the
1505 // malloc to be stored into the specified global, loaded setcc'd, and
1506 // GEP'd. These are all things we could transform to using the global
1508 SmallPtrSet<const PHINode*, 8> PHIs;
1509 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(CI, GV, PHIs))
1512 // If we have a global that is only initialized with a fixed size malloc,
1513 // transform the program to use global memory instead of malloc'd memory.
1514 // This eliminates dynamic allocation, avoids an indirection accessing the
1515 // data, and exposes the resultant global to further GlobalOpt.
1516 // We cannot optimize the malloc if we cannot determine malloc array size.
1517 Value *NElems = getMallocArraySize(CI, TD, true);
1521 if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems))
1522 // Restrict this transformation to only working on small allocations
1523 // (2048 bytes currently), as we don't want to introduce a 16M global or
1525 if (NElements->getZExtValue() * TD->getTypeAllocSize(AllocTy) < 2048) {
1526 GVI = OptimizeGlobalAddressOfMalloc(GV, CI, AllocTy, NElements, TD);
1530 // If the allocation is an array of structures, consider transforming this
1531 // into multiple malloc'd arrays, one for each field. This is basically
1532 // SRoA for malloc'd memory.
1534 // If this is an allocation of a fixed size array of structs, analyze as a
1535 // variable size array. malloc [100 x struct],1 -> malloc struct, 100
1536 if (NElems == ConstantInt::get(CI->getArgOperand(0)->getType(), 1))
1537 if (const ArrayType *AT = dyn_cast<ArrayType>(AllocTy))
1538 AllocTy = AT->getElementType();
1540 const StructType *AllocSTy = dyn_cast<StructType>(AllocTy);
1544 // This the structure has an unreasonable number of fields, leave it
1546 if (AllocSTy->getNumElements() <= 16 && AllocSTy->getNumElements() != 0 &&
1547 AllGlobalLoadUsesSimpleEnoughForHeapSRA(GV, CI)) {
1549 // If this is a fixed size array, transform the Malloc to be an alloc of
1550 // structs. malloc [100 x struct],1 -> malloc struct, 100
1551 if (const ArrayType *AT = dyn_cast<ArrayType>(getMallocAllocatedType(CI))) {
1552 const Type *IntPtrTy = TD->getIntPtrType(CI->getContext());
1553 unsigned TypeSize = TD->getStructLayout(AllocSTy)->getSizeInBytes();
1554 Value *AllocSize = ConstantInt::get(IntPtrTy, TypeSize);
1555 Value *NumElements = ConstantInt::get(IntPtrTy, AT->getNumElements());
1556 Instruction *Malloc = CallInst::CreateMalloc(CI, IntPtrTy, AllocSTy,
1557 AllocSize, NumElements,
1559 Instruction *Cast = new BitCastInst(Malloc, CI->getType(), "tmp", CI);
1560 CI->replaceAllUsesWith(Cast);
1561 CI->eraseFromParent();
1562 CI = dyn_cast<BitCastInst>(Malloc) ?
1563 extractMallocCallFromBitCast(Malloc) : cast<CallInst>(Malloc);
1566 GVI = PerformHeapAllocSRoA(GV, CI, getMallocArraySize(CI, TD, true),TD);
1573 // OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge
1574 // that only one value (besides its initializer) is ever stored to the global.
1575 static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal,
1576 Module::global_iterator &GVI,
1578 // Ignore no-op GEPs and bitcasts.
1579 StoredOnceVal = StoredOnceVal->stripPointerCasts();
1581 // If we are dealing with a pointer global that is initialized to null and
1582 // only has one (non-null) value stored into it, then we can optimize any
1583 // users of the loaded value (often calls and loads) that would trap if the
1585 if (GV->getInitializer()->getType()->isPointerTy() &&
1586 GV->getInitializer()->isNullValue()) {
1587 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) {
1588 if (GV->getInitializer()->getType() != SOVC->getType())
1590 ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType());
1592 // Optimize away any trapping uses of the loaded value.
1593 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC))
1595 } else if (CallInst *CI = extractMallocCall(StoredOnceVal)) {
1596 const Type* MallocType = getMallocAllocatedType(CI);
1597 if (MallocType && TryToOptimizeStoreOfMallocToGlobal(GV, CI, MallocType,
1606 /// TryToShrinkGlobalToBoolean - At this point, we have learned that the only
1607 /// two values ever stored into GV are its initializer and OtherVal. See if we
1608 /// can shrink the global into a boolean and select between the two values
1609 /// whenever it is used. This exposes the values to other scalar optimizations.
1610 static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) {
1611 const Type *GVElType = GV->getType()->getElementType();
1613 // If GVElType is already i1, it is already shrunk. If the type of the GV is
1614 // an FP value, pointer or vector, don't do this optimization because a select
1615 // between them is very expensive and unlikely to lead to later
1616 // simplification. In these cases, we typically end up with "cond ? v1 : v2"
1617 // where v1 and v2 both require constant pool loads, a big loss.
1618 if (GVElType == Type::getInt1Ty(GV->getContext()) ||
1619 GVElType->isFloatingPointTy() ||
1620 GVElType->isPointerTy() || GVElType->isVectorTy())
1623 // Walk the use list of the global seeing if all the uses are load or store.
1624 // If there is anything else, bail out.
1625 for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I){
1627 if (!isa<LoadInst>(U) && !isa<StoreInst>(U))
1631 DEBUG(dbgs() << " *** SHRINKING TO BOOL: " << *GV);
1633 // Create the new global, initializing it to false.
1634 GlobalVariable *NewGV = new GlobalVariable(Type::getInt1Ty(GV->getContext()),
1636 GlobalValue::InternalLinkage,
1637 ConstantInt::getFalse(GV->getContext()),
1639 GV->isThreadLocal());
1640 GV->getParent()->getGlobalList().insert(GV, NewGV);
1642 Constant *InitVal = GV->getInitializer();
1643 assert(InitVal->getType() != Type::getInt1Ty(GV->getContext()) &&
1644 "No reason to shrink to bool!");
1646 // If initialized to zero and storing one into the global, we can use a cast
1647 // instead of a select to synthesize the desired value.
1648 bool IsOneZero = false;
1649 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal))
1650 IsOneZero = InitVal->isNullValue() && CI->isOne();
1652 while (!GV->use_empty()) {
1653 Instruction *UI = cast<Instruction>(GV->use_back());
1654 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
1655 // Change the store into a boolean store.
1656 bool StoringOther = SI->getOperand(0) == OtherVal;
1657 // Only do this if we weren't storing a loaded value.
1659 if (StoringOther || SI->getOperand(0) == InitVal)
1660 StoreVal = ConstantInt::get(Type::getInt1Ty(GV->getContext()),
1663 // Otherwise, we are storing a previously loaded copy. To do this,
1664 // change the copy from copying the original value to just copying the
1666 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0));
1668 // If we've already replaced the input, StoredVal will be a cast or
1669 // select instruction. If not, it will be a load of the original
1671 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
1672 assert(LI->getOperand(0) == GV && "Not a copy!");
1673 // Insert a new load, to preserve the saved value.
1674 StoreVal = new LoadInst(NewGV, LI->getName()+".b", LI);
1676 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) &&
1677 "This is not a form that we understand!");
1678 StoreVal = StoredVal->getOperand(0);
1679 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!");
1682 new StoreInst(StoreVal, NewGV, SI);
1684 // Change the load into a load of bool then a select.
1685 LoadInst *LI = cast<LoadInst>(UI);
1686 LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", LI);
1689 NSI = new ZExtInst(NLI, LI->getType(), "", LI);
1691 NSI = SelectInst::Create(NLI, OtherVal, InitVal, "", LI);
1693 LI->replaceAllUsesWith(NSI);
1695 UI->eraseFromParent();
1698 GV->eraseFromParent();
1703 /// ProcessInternalGlobal - Analyze the specified global variable and optimize
1704 /// it if possible. If we make a change, return true.
1705 bool GlobalOpt::ProcessGlobal(GlobalVariable *GV,
1706 Module::global_iterator &GVI) {
1707 if (!GV->hasLocalLinkage())
1710 // Do more involved optimizations if the global is internal.
1711 GV->removeDeadConstantUsers();
1713 if (GV->use_empty()) {
1714 DEBUG(dbgs() << "GLOBAL DEAD: " << *GV);
1715 GV->eraseFromParent();
1720 SmallPtrSet<const PHINode*, 16> PHIUsers;
1723 if (AnalyzeGlobal(GV, GS, PHIUsers))
1726 if (!GS.isCompared && !GV->hasUnnamedAddr()) {
1727 GV->setUnnamedAddr(true);
1731 if (GV->isConstant() || !GV->hasInitializer())
1734 return ProcessInternalGlobal(GV, GVI, PHIUsers, GS);
1737 /// ProcessInternalGlobal - Analyze the specified global variable and optimize
1738 /// it if possible. If we make a change, return true.
1739 bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV,
1740 Module::global_iterator &GVI,
1741 const SmallPtrSet<const PHINode*, 16> &PHIUsers,
1742 const GlobalStatus &GS) {
1743 // If this is a first class global and has only one accessing function
1744 // and this function is main (which we know is not recursive we can make
1745 // this global a local variable) we replace the global with a local alloca
1746 // in this function.
1748 // NOTE: It doesn't make sense to promote non single-value types since we
1749 // are just replacing static memory to stack memory.
1751 // If the global is in different address space, don't bring it to stack.
1752 if (!GS.HasMultipleAccessingFunctions &&
1753 GS.AccessingFunction && !GS.HasNonInstructionUser &&
1754 GV->getType()->getElementType()->isSingleValueType() &&
1755 GS.AccessingFunction->getName() == "main" &&
1756 GS.AccessingFunction->hasExternalLinkage() &&
1757 GV->getType()->getAddressSpace() == 0) {
1758 DEBUG(dbgs() << "LOCALIZING GLOBAL: " << *GV);
1759 Instruction& FirstI = const_cast<Instruction&>(*GS.AccessingFunction
1760 ->getEntryBlock().begin());
1761 const Type* ElemTy = GV->getType()->getElementType();
1762 // FIXME: Pass Global's alignment when globals have alignment
1763 AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), &FirstI);
1764 if (!isa<UndefValue>(GV->getInitializer()))
1765 new StoreInst(GV->getInitializer(), Alloca, &FirstI);
1767 GV->replaceAllUsesWith(Alloca);
1768 GV->eraseFromParent();
1773 // If the global is never loaded (but may be stored to), it is dead.
1776 DEBUG(dbgs() << "GLOBAL NEVER LOADED: " << *GV);
1778 // Delete any stores we can find to the global. We may not be able to
1779 // make it completely dead though.
1780 bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer());
1782 // If the global is dead now, delete it.
1783 if (GV->use_empty()) {
1784 GV->eraseFromParent();
1790 } else if (GS.StoredType <= GlobalStatus::isInitializerStored) {
1791 DEBUG(dbgs() << "MARKING CONSTANT: " << *GV);
1792 GV->setConstant(true);
1794 // Clean up any obviously simplifiable users now.
1795 CleanupConstantGlobalUsers(GV, GV->getInitializer());
1797 // If the global is dead now, just nuke it.
1798 if (GV->use_empty()) {
1799 DEBUG(dbgs() << " *** Marking constant allowed us to simplify "
1800 << "all users and delete global!\n");
1801 GV->eraseFromParent();
1807 } else if (!GV->getInitializer()->getType()->isSingleValueType()) {
1808 if (TargetData *TD = getAnalysisIfAvailable<TargetData>())
1809 if (GlobalVariable *FirstNewGV = SRAGlobal(GV, *TD)) {
1810 GVI = FirstNewGV; // Don't skip the newly produced globals!
1813 } else if (GS.StoredType == GlobalStatus::isStoredOnce) {
1814 // If the initial value for the global was an undef value, and if only
1815 // one other value was stored into it, we can just change the
1816 // initializer to be the stored value, then delete all stores to the
1817 // global. This allows us to mark it constant.
1818 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1819 if (isa<UndefValue>(GV->getInitializer())) {
1820 // Change the initial value here.
1821 GV->setInitializer(SOVConstant);
1823 // Clean up any obviously simplifiable users now.
1824 CleanupConstantGlobalUsers(GV, GV->getInitializer());
1826 if (GV->use_empty()) {
1827 DEBUG(dbgs() << " *** Substituting initializer allowed us to "
1828 << "simplify all users and delete global!\n");
1829 GV->eraseFromParent();
1838 // Try to optimize globals based on the knowledge that only one value
1839 // (besides its initializer) is ever stored to the global.
1840 if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI,
1841 getAnalysisIfAvailable<TargetData>()))
1844 // Otherwise, if the global was not a boolean, we can shrink it to be a
1846 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1847 if (TryToShrinkGlobalToBoolean(GV, SOVConstant)) {
1856 /// ChangeCalleesToFastCall - Walk all of the direct calls of the specified
1857 /// function, changing them to FastCC.
1858 static void ChangeCalleesToFastCall(Function *F) {
1859 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1860 CallSite User(cast<Instruction>(*UI));
1861 User.setCallingConv(CallingConv::Fast);
1865 static AttrListPtr StripNest(const AttrListPtr &Attrs) {
1866 for (unsigned i = 0, e = Attrs.getNumSlots(); i != e; ++i) {
1867 if ((Attrs.getSlot(i).Attrs & Attribute::Nest) == 0)
1870 // There can be only one.
1871 return Attrs.removeAttr(Attrs.getSlot(i).Index, Attribute::Nest);
1877 static void RemoveNestAttribute(Function *F) {
1878 F->setAttributes(StripNest(F->getAttributes()));
1879 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1880 CallSite User(cast<Instruction>(*UI));
1881 User.setAttributes(StripNest(User.getAttributes()));
1885 bool GlobalOpt::OptimizeFunctions(Module &M) {
1886 bool Changed = false;
1887 // Optimize functions.
1888 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) {
1890 // Functions without names cannot be referenced outside this module.
1891 if (!F->hasName() && !F->isDeclaration())
1892 F->setLinkage(GlobalValue::InternalLinkage);
1893 F->removeDeadConstantUsers();
1894 if (F->use_empty() && (F->hasLocalLinkage() || F->hasLinkOnceLinkage())) {
1895 F->eraseFromParent();
1898 } else if (F->hasLocalLinkage()) {
1899 if (F->getCallingConv() == CallingConv::C && !F->isVarArg() &&
1900 !F->hasAddressTaken()) {
1901 // If this function has C calling conventions, is not a varargs
1902 // function, and is only called directly, promote it to use the Fast
1903 // calling convention.
1904 F->setCallingConv(CallingConv::Fast);
1905 ChangeCalleesToFastCall(F);
1910 if (F->getAttributes().hasAttrSomewhere(Attribute::Nest) &&
1911 !F->hasAddressTaken()) {
1912 // The function is not used by a trampoline intrinsic, so it is safe
1913 // to remove the 'nest' attribute.
1914 RemoveNestAttribute(F);
1923 bool GlobalOpt::OptimizeGlobalVars(Module &M) {
1924 bool Changed = false;
1925 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
1927 GlobalVariable *GV = GVI++;
1928 // Global variables without names cannot be referenced outside this module.
1929 if (!GV->hasName() && !GV->isDeclaration())
1930 GV->setLinkage(GlobalValue::InternalLinkage);
1931 // Simplify the initializer.
1932 if (GV->hasInitializer())
1933 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GV->getInitializer())) {
1934 TargetData *TD = getAnalysisIfAvailable<TargetData>();
1935 Constant *New = ConstantFoldConstantExpression(CE, TD);
1936 if (New && New != CE)
1937 GV->setInitializer(New);
1940 Changed |= ProcessGlobal(GV, GVI);
1945 /// FindGlobalCtors - Find the llvm.globalctors list, verifying that all
1946 /// initializers have an init priority of 65535.
1947 GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) {
1948 GlobalVariable *GV = M.getGlobalVariable("llvm.global_ctors");
1949 if (GV == 0) return 0;
1951 // Found it, verify it's an array of { int, void()* }.
1952 const ArrayType *ATy =dyn_cast<ArrayType>(GV->getType()->getElementType());
1954 const StructType *STy = dyn_cast<StructType>(ATy->getElementType());
1955 if (!STy || STy->getNumElements() != 2 ||
1956 !STy->getElementType(0)->isIntegerTy(32)) return 0;
1957 const PointerType *PFTy = dyn_cast<PointerType>(STy->getElementType(1));
1958 if (!PFTy) return 0;
1959 const FunctionType *FTy = dyn_cast<FunctionType>(PFTy->getElementType());
1960 if (!FTy || !FTy->getReturnType()->isVoidTy() ||
1961 FTy->isVarArg() || FTy->getNumParams() != 0)
1964 // Verify that the initializer is simple enough for us to handle. We are
1965 // only allowed to optimize the initializer if it is unique.
1966 if (!GV->hasUniqueInitializer()) return 0;
1968 ConstantArray *CA = dyn_cast<ConstantArray>(GV->getInitializer());
1971 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) {
1972 ConstantStruct *CS = dyn_cast<ConstantStruct>(*i);
1973 if (CS == 0) return 0;
1975 if (isa<ConstantPointerNull>(CS->getOperand(1)))
1978 // Must have a function or null ptr.
1979 if (!isa<Function>(CS->getOperand(1)))
1982 // Init priority must be standard.
1983 ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0));
1984 if (!CI || CI->getZExtValue() != 65535)
1991 /// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand,
1992 /// return a list of the functions and null terminator as a vector.
1993 static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) {
1994 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer());
1995 std::vector<Function*> Result;
1996 Result.reserve(CA->getNumOperands());
1997 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) {
1998 ConstantStruct *CS = cast<ConstantStruct>(*i);
1999 Result.push_back(dyn_cast<Function>(CS->getOperand(1)));
2004 /// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the
2005 /// specified array, returning the new global to use.
2006 static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL,
2007 const std::vector<Function*> &Ctors) {
2008 // If we made a change, reassemble the initializer list.
2009 std::vector<Constant*> CSVals;
2010 CSVals.push_back(ConstantInt::get(Type::getInt32Ty(GCL->getContext()),65535));
2011 CSVals.push_back(0);
2013 // Create the new init list.
2014 std::vector<Constant*> CAList;
2015 for (unsigned i = 0, e = Ctors.size(); i != e; ++i) {
2017 CSVals[1] = Ctors[i];
2019 const Type *FTy = FunctionType::get(Type::getVoidTy(GCL->getContext()),
2021 const PointerType *PFTy = PointerType::getUnqual(FTy);
2022 CSVals[1] = Constant::getNullValue(PFTy);
2023 CSVals[0] = ConstantInt::get(Type::getInt32Ty(GCL->getContext()),
2026 CAList.push_back(ConstantStruct::get(GCL->getContext(), CSVals, false));
2029 // Create the array initializer.
2030 const Type *StructTy =
2031 cast<ArrayType>(GCL->getType()->getElementType())->getElementType();
2032 Constant *CA = ConstantArray::get(ArrayType::get(StructTy,
2033 CAList.size()), CAList);
2035 // If we didn't change the number of elements, don't create a new GV.
2036 if (CA->getType() == GCL->getInitializer()->getType()) {
2037 GCL->setInitializer(CA);
2041 // Create the new global and insert it next to the existing list.
2042 GlobalVariable *NGV = new GlobalVariable(CA->getType(), GCL->isConstant(),
2043 GCL->getLinkage(), CA, "",
2044 GCL->isThreadLocal());
2045 GCL->getParent()->getGlobalList().insert(GCL, NGV);
2048 // Nuke the old list, replacing any uses with the new one.
2049 if (!GCL->use_empty()) {
2051 if (V->getType() != GCL->getType())
2052 V = ConstantExpr::getBitCast(V, GCL->getType());
2053 GCL->replaceAllUsesWith(V);
2055 GCL->eraseFromParent();
2064 static Constant *getVal(DenseMap<Value*, Constant*> &ComputedValues, Value *V) {
2065 if (Constant *CV = dyn_cast<Constant>(V)) return CV;
2066 Constant *R = ComputedValues[V];
2067 assert(R && "Reference to an uncomputed value!");
2072 isSimpleEnoughValueToCommit(Constant *C,
2073 SmallPtrSet<Constant*, 8> &SimpleConstants);
2076 /// isSimpleEnoughValueToCommit - Return true if the specified constant can be
2077 /// handled by the code generator. We don't want to generate something like:
2078 /// void *X = &X/42;
2079 /// because the code generator doesn't have a relocation that can handle that.
2081 /// This function should be called if C was not found (but just got inserted)
2082 /// in SimpleConstants to avoid having to rescan the same constants all the
2084 static bool isSimpleEnoughValueToCommitHelper(Constant *C,
2085 SmallPtrSet<Constant*, 8> &SimpleConstants) {
2086 // Simple integer, undef, constant aggregate zero, global addresses, etc are
2088 if (C->getNumOperands() == 0 || isa<BlockAddress>(C) ||
2089 isa<GlobalValue>(C))
2092 // Aggregate values are safe if all their elements are.
2093 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C) ||
2094 isa<ConstantVector>(C)) {
2095 for (unsigned i = 0, e = C->getNumOperands(); i != e; ++i) {
2096 Constant *Op = cast<Constant>(C->getOperand(i));
2097 if (!isSimpleEnoughValueToCommit(Op, SimpleConstants))
2103 // We don't know exactly what relocations are allowed in constant expressions,
2104 // so we allow &global+constantoffset, which is safe and uniformly supported
2106 ConstantExpr *CE = cast<ConstantExpr>(C);
2107 switch (CE->getOpcode()) {
2108 case Instruction::BitCast:
2109 case Instruction::IntToPtr:
2110 case Instruction::PtrToInt:
2111 // These casts are always fine if the casted value is.
2112 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants);
2114 // GEP is fine if it is simple + constant offset.
2115 case Instruction::GetElementPtr:
2116 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
2117 if (!isa<ConstantInt>(CE->getOperand(i)))
2119 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants);
2121 case Instruction::Add:
2122 // We allow simple+cst.
2123 if (!isa<ConstantInt>(CE->getOperand(1)))
2125 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants);
2131 isSimpleEnoughValueToCommit(Constant *C,
2132 SmallPtrSet<Constant*, 8> &SimpleConstants) {
2133 // If we already checked this constant, we win.
2134 if (!SimpleConstants.insert(C)) return true;
2135 // Check the constant.
2136 return isSimpleEnoughValueToCommitHelper(C, SimpleConstants);
2140 /// isSimpleEnoughPointerToCommit - Return true if this constant is simple
2141 /// enough for us to understand. In particular, if it is a cast to anything
2142 /// other than from one pointer type to another pointer type, we punt.
2143 /// We basically just support direct accesses to globals and GEP's of
2144 /// globals. This should be kept up to date with CommitValueTo.
2145 static bool isSimpleEnoughPointerToCommit(Constant *C) {
2146 // Conservatively, avoid aggregate types. This is because we don't
2147 // want to worry about them partially overlapping other stores.
2148 if (!cast<PointerType>(C->getType())->getElementType()->isSingleValueType())
2151 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
2152 // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or
2153 // external globals.
2154 return GV->hasUniqueInitializer();
2156 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
2157 // Handle a constantexpr gep.
2158 if (CE->getOpcode() == Instruction::GetElementPtr &&
2159 isa<GlobalVariable>(CE->getOperand(0)) &&
2160 cast<GEPOperator>(CE)->isInBounds()) {
2161 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2162 // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or
2163 // external globals.
2164 if (!GV->hasUniqueInitializer())
2167 // The first index must be zero.
2168 ConstantInt *CI = dyn_cast<ConstantInt>(*llvm::next(CE->op_begin()));
2169 if (!CI || !CI->isZero()) return false;
2171 // The remaining indices must be compile-time known integers within the
2172 // notional bounds of the corresponding static array types.
2173 if (!CE->isGEPWithNoNotionalOverIndexing())
2176 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
2178 // A constantexpr bitcast from a pointer to another pointer is a no-op,
2179 // and we know how to evaluate it by moving the bitcast from the pointer
2180 // operand to the value operand.
2181 } else if (CE->getOpcode() == Instruction::BitCast &&
2182 isa<GlobalVariable>(CE->getOperand(0))) {
2183 // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or
2184 // external globals.
2185 return cast<GlobalVariable>(CE->getOperand(0))->hasUniqueInitializer();
2192 /// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global
2193 /// initializer. This returns 'Init' modified to reflect 'Val' stored into it.
2194 /// At this point, the GEP operands of Addr [0, OpNo) have been stepped into.
2195 static Constant *EvaluateStoreInto(Constant *Init, Constant *Val,
2196 ConstantExpr *Addr, unsigned OpNo) {
2197 // Base case of the recursion.
2198 if (OpNo == Addr->getNumOperands()) {
2199 assert(Val->getType() == Init->getType() && "Type mismatch!");
2203 std::vector<Constant*> Elts;
2204 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
2206 // Break up the constant into its elements.
2207 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
2208 for (User::op_iterator i = CS->op_begin(), e = CS->op_end(); i != e; ++i)
2209 Elts.push_back(cast<Constant>(*i));
2210 } else if (isa<ConstantAggregateZero>(Init)) {
2211 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
2212 Elts.push_back(Constant::getNullValue(STy->getElementType(i)));
2213 } else if (isa<UndefValue>(Init)) {
2214 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
2215 Elts.push_back(UndefValue::get(STy->getElementType(i)));
2217 llvm_unreachable("This code is out of sync with "
2218 " ConstantFoldLoadThroughGEPConstantExpr");
2221 // Replace the element that we are supposed to.
2222 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo));
2223 unsigned Idx = CU->getZExtValue();
2224 assert(Idx < STy->getNumElements() && "Struct index out of range!");
2225 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1);
2227 // Return the modified struct.
2228 return ConstantStruct::get(Init->getContext(), &Elts[0], Elts.size(),
2231 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo));
2232 const SequentialType *InitTy = cast<SequentialType>(Init->getType());
2235 if (const ArrayType *ATy = dyn_cast<ArrayType>(InitTy))
2236 NumElts = ATy->getNumElements();
2238 NumElts = cast<VectorType>(InitTy)->getNumElements();
2241 // Break up the array into elements.
2242 if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
2243 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i)
2244 Elts.push_back(cast<Constant>(*i));
2245 } else if (ConstantVector *CV = dyn_cast<ConstantVector>(Init)) {
2246 for (User::op_iterator i = CV->op_begin(), e = CV->op_end(); i != e; ++i)
2247 Elts.push_back(cast<Constant>(*i));
2248 } else if (isa<ConstantAggregateZero>(Init)) {
2249 Elts.assign(NumElts, Constant::getNullValue(InitTy->getElementType()));
2251 assert(isa<UndefValue>(Init) && "This code is out of sync with "
2252 " ConstantFoldLoadThroughGEPConstantExpr");
2253 Elts.assign(NumElts, UndefValue::get(InitTy->getElementType()));
2256 assert(CI->getZExtValue() < NumElts);
2257 Elts[CI->getZExtValue()] =
2258 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1);
2260 if (Init->getType()->isArrayTy())
2261 return ConstantArray::get(cast<ArrayType>(InitTy), Elts);
2262 return ConstantVector::get(Elts);
2266 /// CommitValueTo - We have decided that Addr (which satisfies the predicate
2267 /// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen.
2268 static void CommitValueTo(Constant *Val, Constant *Addr) {
2269 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
2270 assert(GV->hasInitializer());
2271 GV->setInitializer(Val);
2275 ConstantExpr *CE = cast<ConstantExpr>(Addr);
2276 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2277 GV->setInitializer(EvaluateStoreInto(GV->getInitializer(), Val, CE, 2));
2280 /// ComputeLoadResult - Return the value that would be computed by a load from
2281 /// P after the stores reflected by 'memory' have been performed. If we can't
2282 /// decide, return null.
2283 static Constant *ComputeLoadResult(Constant *P,
2284 const DenseMap<Constant*, Constant*> &Memory) {
2285 // If this memory location has been recently stored, use the stored value: it
2286 // is the most up-to-date.
2287 DenseMap<Constant*, Constant*>::const_iterator I = Memory.find(P);
2288 if (I != Memory.end()) return I->second;
2291 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
2292 if (GV->hasDefinitiveInitializer())
2293 return GV->getInitializer();
2297 // Handle a constantexpr getelementptr.
2298 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P))
2299 if (CE->getOpcode() == Instruction::GetElementPtr &&
2300 isa<GlobalVariable>(CE->getOperand(0))) {
2301 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2302 if (GV->hasDefinitiveInitializer())
2303 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
2306 return 0; // don't know how to evaluate.
2309 /// EvaluateFunction - Evaluate a call to function F, returning true if
2310 /// successful, false if we can't evaluate it. ActualArgs contains the formal
2311 /// arguments for the function.
2312 static bool EvaluateFunction(Function *F, Constant *&RetVal,
2313 const SmallVectorImpl<Constant*> &ActualArgs,
2314 std::vector<Function*> &CallStack,
2315 DenseMap<Constant*, Constant*> &MutatedMemory,
2316 std::vector<GlobalVariable*> &AllocaTmps,
2317 SmallPtrSet<Constant*, 8> &SimpleConstants,
2318 const TargetData *TD) {
2319 // Check to see if this function is already executing (recursion). If so,
2320 // bail out. TODO: we might want to accept limited recursion.
2321 if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end())
2324 CallStack.push_back(F);
2326 /// Values - As we compute SSA register values, we store their contents here.
2327 DenseMap<Value*, Constant*> Values;
2329 // Initialize arguments to the incoming values specified.
2331 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E;
2333 Values[AI] = ActualArgs[ArgNo];
2335 /// ExecutedBlocks - We only handle non-looping, non-recursive code. As such,
2336 /// we can only evaluate any one basic block at most once. This set keeps
2337 /// track of what we have executed so we can detect recursive cases etc.
2338 SmallPtrSet<BasicBlock*, 32> ExecutedBlocks;
2340 // CurInst - The current instruction we're evaluating.
2341 BasicBlock::iterator CurInst = F->begin()->begin();
2343 // This is the main evaluation loop.
2345 Constant *InstResult = 0;
2347 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) {
2348 if (SI->isVolatile()) return false; // no volatile accesses.
2349 Constant *Ptr = getVal(Values, SI->getOperand(1));
2350 if (!isSimpleEnoughPointerToCommit(Ptr))
2351 // If this is too complex for us to commit, reject it.
2354 Constant *Val = getVal(Values, SI->getOperand(0));
2356 // If this might be too difficult for the backend to handle (e.g. the addr
2357 // of one global variable divided by another) then we can't commit it.
2358 if (!isSimpleEnoughValueToCommit(Val, SimpleConstants))
2361 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
2362 if (CE->getOpcode() == Instruction::BitCast) {
2363 // If we're evaluating a store through a bitcast, then we need
2364 // to pull the bitcast off the pointer type and push it onto the
2366 Ptr = CE->getOperand(0);
2368 const Type *NewTy=cast<PointerType>(Ptr->getType())->getElementType();
2370 // In order to push the bitcast onto the stored value, a bitcast
2371 // from NewTy to Val's type must be legal. If it's not, we can try
2372 // introspecting NewTy to find a legal conversion.
2373 while (!Val->getType()->canLosslesslyBitCastTo(NewTy)) {
2374 // If NewTy is a struct, we can convert the pointer to the struct
2375 // into a pointer to its first member.
2376 // FIXME: This could be extended to support arrays as well.
2377 if (const StructType *STy = dyn_cast<StructType>(NewTy)) {
2378 NewTy = STy->getTypeAtIndex(0U);
2380 const IntegerType *IdxTy =IntegerType::get(NewTy->getContext(), 32);
2381 Constant *IdxZero = ConstantInt::get(IdxTy, 0, false);
2382 Constant * const IdxList[] = {IdxZero, IdxZero};
2384 Ptr = ConstantExpr::getGetElementPtr(Ptr, IdxList, 2);
2386 // If we can't improve the situation by introspecting NewTy,
2387 // we have to give up.
2393 // If we found compatible types, go ahead and push the bitcast
2394 // onto the stored value.
2395 Val = ConstantExpr::getBitCast(Val, NewTy);
2398 MutatedMemory[Ptr] = Val;
2399 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) {
2400 InstResult = ConstantExpr::get(BO->getOpcode(),
2401 getVal(Values, BO->getOperand(0)),
2402 getVal(Values, BO->getOperand(1)));
2403 } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) {
2404 InstResult = ConstantExpr::getCompare(CI->getPredicate(),
2405 getVal(Values, CI->getOperand(0)),
2406 getVal(Values, CI->getOperand(1)));
2407 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) {
2408 InstResult = ConstantExpr::getCast(CI->getOpcode(),
2409 getVal(Values, CI->getOperand(0)),
2411 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) {
2412 InstResult = ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)),
2413 getVal(Values, SI->getOperand(1)),
2414 getVal(Values, SI->getOperand(2)));
2415 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) {
2416 Constant *P = getVal(Values, GEP->getOperand(0));
2417 SmallVector<Constant*, 8> GEPOps;
2418 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end();
2420 GEPOps.push_back(getVal(Values, *i));
2421 InstResult = cast<GEPOperator>(GEP)->isInBounds() ?
2422 ConstantExpr::getInBoundsGetElementPtr(P, &GEPOps[0], GEPOps.size()) :
2423 ConstantExpr::getGetElementPtr(P, &GEPOps[0], GEPOps.size());
2424 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) {
2425 if (LI->isVolatile()) return false; // no volatile accesses.
2426 InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)),
2428 if (InstResult == 0) return false; // Could not evaluate load.
2429 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) {
2430 if (AI->isArrayAllocation()) return false; // Cannot handle array allocs.
2431 const Type *Ty = AI->getType()->getElementType();
2432 AllocaTmps.push_back(new GlobalVariable(Ty, false,
2433 GlobalValue::InternalLinkage,
2434 UndefValue::get(Ty),
2436 InstResult = AllocaTmps.back();
2437 } else if (CallInst *CI = dyn_cast<CallInst>(CurInst)) {
2439 // Debug info can safely be ignored here.
2440 if (isa<DbgInfoIntrinsic>(CI)) {
2445 // Cannot handle inline asm.
2446 if (isa<InlineAsm>(CI->getCalledValue())) return false;
2448 // Resolve function pointers.
2449 Function *Callee = dyn_cast<Function>(getVal(Values,
2450 CI->getCalledValue()));
2451 if (!Callee) return false; // Cannot resolve.
2453 SmallVector<Constant*, 8> Formals;
2455 for (User::op_iterator i = CS.arg_begin(), e = CS.arg_end();
2457 Formals.push_back(getVal(Values, *i));
2459 if (Callee->isDeclaration()) {
2460 // If this is a function we can constant fold, do it.
2461 if (Constant *C = ConstantFoldCall(Callee, Formals.data(),
2468 if (Callee->getFunctionType()->isVarArg())
2472 // Execute the call, if successful, use the return value.
2473 if (!EvaluateFunction(Callee, RetVal, Formals, CallStack,
2474 MutatedMemory, AllocaTmps, SimpleConstants, TD))
2476 InstResult = RetVal;
2478 } else if (isa<TerminatorInst>(CurInst)) {
2479 BasicBlock *NewBB = 0;
2480 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) {
2481 if (BI->isUnconditional()) {
2482 NewBB = BI->getSuccessor(0);
2485 dyn_cast<ConstantInt>(getVal(Values, BI->getCondition()));
2486 if (!Cond) return false; // Cannot determine.
2488 NewBB = BI->getSuccessor(!Cond->getZExtValue());
2490 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) {
2492 dyn_cast<ConstantInt>(getVal(Values, SI->getCondition()));
2493 if (!Val) return false; // Cannot determine.
2494 NewBB = SI->getSuccessor(SI->findCaseValue(Val));
2495 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(CurInst)) {
2496 Value *Val = getVal(Values, IBI->getAddress())->stripPointerCasts();
2497 if (BlockAddress *BA = dyn_cast<BlockAddress>(Val))
2498 NewBB = BA->getBasicBlock();
2500 return false; // Cannot determine.
2501 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) {
2502 if (RI->getNumOperands())
2503 RetVal = getVal(Values, RI->getOperand(0));
2505 CallStack.pop_back(); // return from fn.
2506 return true; // We succeeded at evaluating this ctor!
2508 // invoke, unwind, unreachable.
2509 return false; // Cannot handle this terminator.
2512 // Okay, we succeeded in evaluating this control flow. See if we have
2513 // executed the new block before. If so, we have a looping function,
2514 // which we cannot evaluate in reasonable time.
2515 if (!ExecutedBlocks.insert(NewBB))
2516 return false; // looped!
2518 // Okay, we have never been in this block before. Check to see if there
2519 // are any PHI nodes. If so, evaluate them with information about where
2521 BasicBlock *OldBB = CurInst->getParent();
2522 CurInst = NewBB->begin();
2524 for (; (PN = dyn_cast<PHINode>(CurInst)); ++CurInst)
2525 Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB));
2527 // Do NOT increment CurInst. We know that the terminator had no value.
2530 // Did not know how to evaluate this!
2534 if (!CurInst->use_empty()) {
2535 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(InstResult))
2536 InstResult = ConstantFoldConstantExpression(CE, TD);
2538 Values[CurInst] = InstResult;
2541 // Advance program counter.
2546 /// EvaluateStaticConstructor - Evaluate static constructors in the function, if
2547 /// we can. Return true if we can, false otherwise.
2548 static bool EvaluateStaticConstructor(Function *F, const TargetData *TD) {
2549 /// MutatedMemory - For each store we execute, we update this map. Loads
2550 /// check this to get the most up-to-date value. If evaluation is successful,
2551 /// this state is committed to the process.
2552 DenseMap<Constant*, Constant*> MutatedMemory;
2554 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable
2555 /// to represent its body. This vector is needed so we can delete the
2556 /// temporary globals when we are done.
2557 std::vector<GlobalVariable*> AllocaTmps;
2559 /// CallStack - This is used to detect recursion. In pathological situations
2560 /// we could hit exponential behavior, but at least there is nothing
2562 std::vector<Function*> CallStack;
2564 /// SimpleConstants - These are constants we have checked and know to be
2565 /// simple enough to live in a static initializer of a global.
2566 SmallPtrSet<Constant*, 8> SimpleConstants;
2568 // Call the function.
2569 Constant *RetValDummy;
2570 bool EvalSuccess = EvaluateFunction(F, RetValDummy,
2571 SmallVector<Constant*, 0>(), CallStack,
2572 MutatedMemory, AllocaTmps,
2573 SimpleConstants, TD);
2576 // We succeeded at evaluation: commit the result.
2577 DEBUG(dbgs() << "FULLY EVALUATED GLOBAL CTOR FUNCTION '"
2578 << F->getName() << "' to " << MutatedMemory.size()
2580 for (DenseMap<Constant*, Constant*>::iterator I = MutatedMemory.begin(),
2581 E = MutatedMemory.end(); I != E; ++I)
2582 CommitValueTo(I->second, I->first);
2585 // At this point, we are done interpreting. If we created any 'alloca'
2586 // temporaries, release them now.
2587 while (!AllocaTmps.empty()) {
2588 GlobalVariable *Tmp = AllocaTmps.back();
2589 AllocaTmps.pop_back();
2591 // If there are still users of the alloca, the program is doing something
2592 // silly, e.g. storing the address of the alloca somewhere and using it
2593 // later. Since this is undefined, we'll just make it be null.
2594 if (!Tmp->use_empty())
2595 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType()));
2604 /// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible.
2605 /// Return true if anything changed.
2606 bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) {
2607 std::vector<Function*> Ctors = ParseGlobalCtors(GCL);
2608 bool MadeChange = false;
2609 if (Ctors.empty()) return false;
2611 const TargetData *TD = getAnalysisIfAvailable<TargetData>();
2612 // Loop over global ctors, optimizing them when we can.
2613 for (unsigned i = 0; i != Ctors.size(); ++i) {
2614 Function *F = Ctors[i];
2615 // Found a null terminator in the middle of the list, prune off the rest of
2618 if (i != Ctors.size()-1) {
2625 // We cannot simplify external ctor functions.
2626 if (F->empty()) continue;
2628 // If we can evaluate the ctor at compile time, do.
2629 if (EvaluateStaticConstructor(F, TD)) {
2630 Ctors.erase(Ctors.begin()+i);
2633 ++NumCtorsEvaluated;
2638 if (!MadeChange) return false;
2640 GCL = InstallGlobalCtors(GCL, Ctors);
2644 bool GlobalOpt::OptimizeGlobalAliases(Module &M) {
2645 bool Changed = false;
2647 for (Module::alias_iterator I = M.alias_begin(), E = M.alias_end();
2649 Module::alias_iterator J = I++;
2650 // Aliases without names cannot be referenced outside this module.
2651 if (!J->hasName() && !J->isDeclaration())
2652 J->setLinkage(GlobalValue::InternalLinkage);
2653 // If the aliasee may change at link time, nothing can be done - bail out.
2654 if (J->mayBeOverridden())
2657 Constant *Aliasee = J->getAliasee();
2658 GlobalValue *Target = cast<GlobalValue>(Aliasee->stripPointerCasts());
2659 Target->removeDeadConstantUsers();
2660 bool hasOneUse = Target->hasOneUse() && Aliasee->hasOneUse();
2662 // Make all users of the alias use the aliasee instead.
2663 if (!J->use_empty()) {
2664 J->replaceAllUsesWith(Aliasee);
2665 ++NumAliasesResolved;
2669 // If the alias is externally visible, we may still be able to simplify it.
2670 if (!J->hasLocalLinkage()) {
2671 // If the aliasee has internal linkage, give it the name and linkage
2672 // of the alias, and delete the alias. This turns:
2673 // define internal ... @f(...)
2674 // @a = alias ... @f
2676 // define ... @a(...)
2677 if (!Target->hasLocalLinkage())
2680 // Do not perform the transform if multiple aliases potentially target the
2681 // aliasee. This check also ensures that it is safe to replace the section
2682 // and other attributes of the aliasee with those of the alias.
2686 // Give the aliasee the name, linkage and other attributes of the alias.
2687 Target->takeName(J);
2688 Target->setLinkage(J->getLinkage());
2689 Target->GlobalValue::copyAttributesFrom(J);
2692 // Delete the alias.
2693 M.getAliasList().erase(J);
2694 ++NumAliasesRemoved;
2701 static Function *FindCXAAtExit(Module &M) {
2702 Function *Fn = M.getFunction("__cxa_atexit");
2707 const FunctionType *FTy = Fn->getFunctionType();
2709 // Checking that the function has the right return type, the right number of
2710 // parameters and that they all have pointer types should be enough.
2711 if (!FTy->getReturnType()->isIntegerTy() ||
2712 FTy->getNumParams() != 3 ||
2713 !FTy->getParamType(0)->isPointerTy() ||
2714 !FTy->getParamType(1)->isPointerTy() ||
2715 !FTy->getParamType(2)->isPointerTy())
2721 /// cxxDtorIsEmpty - Returns whether the given function is an empty C++
2722 /// destructor and can therefore be eliminated.
2723 /// Note that we assume that other optimization passes have already simplified
2724 /// the code so we only look for a function with a single basic block, where
2725 /// the only allowed instructions are 'ret' or 'call' to empty C++ dtor.
2726 static bool cxxDtorIsEmpty(const Function &Fn,
2727 SmallPtrSet<const Function *, 8> &CalledFunctions) {
2728 // FIXME: We could eliminate C++ destructors if they're readonly/readnone and
2729 // unwind, but that doesn't seem worth doing.
2730 if (Fn.isDeclaration())
2733 if (++Fn.begin() != Fn.end())
2736 const BasicBlock &EntryBlock = Fn.getEntryBlock();
2737 for (BasicBlock::const_iterator I = EntryBlock.begin(), E = EntryBlock.end();
2739 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
2740 const Function *CalledFn = CI->getCalledFunction();
2745 // Don't treat recursive functions as empty.
2746 if (!CalledFunctions.insert(CalledFn))
2749 if (!cxxDtorIsEmpty(*CalledFn, CalledFunctions))
2751 } else if (isa<ReturnInst>(*I))
2760 bool GlobalOpt::OptimizeEmptyGlobalCXXDtors(Function *CXAAtExitFn) {
2761 /// Itanium C++ ABI p3.3.5:
2763 /// After constructing a global (or local static) object, that will require
2764 /// destruction on exit, a termination function is registered as follows:
2766 /// extern "C" int __cxa_atexit ( void (*f)(void *), void *p, void *d );
2768 /// This registration, e.g. __cxa_atexit(f,p,d), is intended to cause the
2769 /// call f(p) when DSO d is unloaded, before all such termination calls
2770 /// registered before this one. It returns zero if registration is
2771 /// successful, nonzero on failure.
2773 // This pass will look for calls to __cxa_atexit where the function is trivial
2775 bool Changed = false;
2777 for (Function::use_iterator I = CXAAtExitFn->use_begin(),
2778 E = CXAAtExitFn->use_end(); I != E;) {
2783 // We're only interested in calls. Theoretically, we could handle invoke
2784 // instructions as well, but neither llvm-gcc nor clang generate invokes
2790 dyn_cast<Function>(CS.getArgument(0)->stripPointerCasts());
2794 SmallPtrSet<const Function *, 8> CalledFunctions;
2795 if (!cxxDtorIsEmpty(*DtorFn, CalledFunctions))
2798 // Just remove the call.
2799 CS->replaceAllUsesWith(Constant::getNullValue(CS.getType()));
2800 CS->eraseFromParent();
2802 ++NumCXXDtorsRemoved;
2810 bool GlobalOpt::runOnModule(Module &M) {
2811 bool Changed = false;
2813 // Try to find the llvm.globalctors list.
2814 GlobalVariable *GlobalCtors = FindGlobalCtors(M);
2816 Function *CXAAtExitFn = FindCXAAtExit(M);
2818 bool LocalChange = true;
2819 while (LocalChange) {
2820 LocalChange = false;
2822 // Delete functions that are trivially dead, ccc -> fastcc
2823 LocalChange |= OptimizeFunctions(M);
2825 // Optimize global_ctors list.
2827 LocalChange |= OptimizeGlobalCtorsList(GlobalCtors);
2829 // Optimize non-address-taken globals.
2830 LocalChange |= OptimizeGlobalVars(M);
2832 // Resolve aliases, when possible.
2833 LocalChange |= OptimizeGlobalAliases(M);
2835 // Try to remove trivial global destructors.
2837 LocalChange |= OptimizeEmptyGlobalCXXDtors(CXAAtExitFn);
2839 Changed |= LocalChange;
2842 // TODO: Move all global ctors functions to the end of the module for code