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/LLVMContext.h"
24 #include "llvm/Module.h"
25 #include "llvm/Pass.h"
26 #include "llvm/Analysis/ConstantFolding.h"
27 #include "llvm/Analysis/MallocHelper.h"
28 #include "llvm/Target/TargetData.h"
29 #include "llvm/Support/CallSite.h"
30 #include "llvm/Support/Compiler.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/GetElementPtrTypeIterator.h"
34 #include "llvm/Support/MathExtras.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include "llvm/ADT/DenseMap.h"
37 #include "llvm/ADT/SmallPtrSet.h"
38 #include "llvm/ADT/SmallVector.h"
39 #include "llvm/ADT/Statistic.h"
40 #include "llvm/ADT/STLExtras.h"
44 STATISTIC(NumMarked , "Number of globals marked constant");
45 STATISTIC(NumSRA , "Number of aggregate globals broken into scalars");
46 STATISTIC(NumHeapSRA , "Number of heap objects SRA'd");
47 STATISTIC(NumSubstitute,"Number of globals with initializers stored into them");
48 STATISTIC(NumDeleted , "Number of globals deleted");
49 STATISTIC(NumFnDeleted , "Number of functions deleted");
50 STATISTIC(NumGlobUses , "Number of global uses devirtualized");
51 STATISTIC(NumLocalized , "Number of globals localized");
52 STATISTIC(NumShrunkToBool , "Number of global vars shrunk to booleans");
53 STATISTIC(NumFastCallFns , "Number of functions converted to fastcc");
54 STATISTIC(NumCtorsEvaluated, "Number of static ctors evaluated");
55 STATISTIC(NumNestRemoved , "Number of nest attributes removed");
56 STATISTIC(NumAliasesResolved, "Number of global aliases resolved");
57 STATISTIC(NumAliasesRemoved, "Number of global aliases eliminated");
60 struct VISIBILITY_HIDDEN GlobalOpt : public ModulePass {
61 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
63 static char ID; // Pass identification, replacement for typeid
64 GlobalOpt() : ModulePass(&ID) {}
66 bool runOnModule(Module &M);
69 GlobalVariable *FindGlobalCtors(Module &M);
70 bool OptimizeFunctions(Module &M);
71 bool OptimizeGlobalVars(Module &M);
72 bool OptimizeGlobalAliases(Module &M);
73 bool OptimizeGlobalCtorsList(GlobalVariable *&GCL);
74 bool ProcessInternalGlobal(GlobalVariable *GV,Module::global_iterator &GVI);
78 char GlobalOpt::ID = 0;
79 static RegisterPass<GlobalOpt> X("globalopt", "Global Variable Optimizer");
81 ModulePass *llvm::createGlobalOptimizerPass() { return new GlobalOpt(); }
85 /// GlobalStatus - As we analyze each global, keep track of some information
86 /// about it. If we find out that the address of the global is taken, none of
87 /// this info will be accurate.
88 struct VISIBILITY_HIDDEN GlobalStatus {
89 /// isLoaded - True if the global is ever loaded. If the global isn't ever
90 /// loaded it can be deleted.
93 /// StoredType - Keep track of what stores to the global look like.
96 /// NotStored - There is no store to this global. It can thus be marked
100 /// isInitializerStored - This global is stored to, but the only thing
101 /// stored is the constant it was initialized with. This is only tracked
102 /// for scalar globals.
105 /// isStoredOnce - This global is stored to, but only its initializer and
106 /// one other value is ever stored to it. If this global isStoredOnce, we
107 /// track the value stored to it in StoredOnceValue below. This is only
108 /// tracked for scalar globals.
111 /// isStored - This global is stored to by multiple values or something else
112 /// that we cannot track.
116 /// StoredOnceValue - If only one value (besides the initializer constant) is
117 /// ever stored to this global, keep track of what value it is.
118 Value *StoredOnceValue;
120 /// AccessingFunction/HasMultipleAccessingFunctions - These start out
121 /// null/false. When the first accessing function is noticed, it is recorded.
122 /// When a second different accessing function is noticed,
123 /// HasMultipleAccessingFunctions is set to true.
124 Function *AccessingFunction;
125 bool HasMultipleAccessingFunctions;
127 /// HasNonInstructionUser - Set to true if this global has a user that is not
128 /// an instruction (e.g. a constant expr or GV initializer).
129 bool HasNonInstructionUser;
131 /// HasPHIUser - Set to true if this global has a user that is a PHI node.
134 GlobalStatus() : isLoaded(false), StoredType(NotStored), StoredOnceValue(0),
135 AccessingFunction(0), HasMultipleAccessingFunctions(false),
136 HasNonInstructionUser(false), HasPHIUser(false) {}
141 // SafeToDestroyConstant - It is safe to destroy a constant iff it is only used
142 // by constants itself. Note that constants cannot be cyclic, so this test is
143 // pretty easy to implement recursively.
145 static bool SafeToDestroyConstant(Constant *C) {
146 if (isa<GlobalValue>(C)) return false;
148 for (Value::use_iterator UI = C->use_begin(), E = C->use_end(); UI != E; ++UI)
149 if (Constant *CU = dyn_cast<Constant>(*UI)) {
150 if (!SafeToDestroyConstant(CU)) return false;
157 /// AnalyzeGlobal - Look at all uses of the global and fill in the GlobalStatus
158 /// structure. If the global has its address taken, return true to indicate we
159 /// can't do anything with it.
161 static bool AnalyzeGlobal(Value *V, GlobalStatus &GS,
162 SmallPtrSet<PHINode*, 16> &PHIUsers) {
163 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
164 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(*UI)) {
165 GS.HasNonInstructionUser = true;
167 if (AnalyzeGlobal(CE, GS, PHIUsers)) return true;
169 } else if (Instruction *I = dyn_cast<Instruction>(*UI)) {
170 if (!GS.HasMultipleAccessingFunctions) {
171 Function *F = I->getParent()->getParent();
172 if (GS.AccessingFunction == 0)
173 GS.AccessingFunction = F;
174 else if (GS.AccessingFunction != F)
175 GS.HasMultipleAccessingFunctions = true;
177 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
179 if (LI->isVolatile()) return true; // Don't hack on volatile loads.
180 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
181 // Don't allow a store OF the address, only stores TO the address.
182 if (SI->getOperand(0) == V) return true;
184 if (SI->isVolatile()) return true; // Don't hack on volatile stores.
186 // If this is a direct store to the global (i.e., the global is a scalar
187 // value, not an aggregate), keep more specific information about
189 if (GS.StoredType != GlobalStatus::isStored) {
190 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(SI->getOperand(1))){
191 Value *StoredVal = SI->getOperand(0);
192 if (StoredVal == GV->getInitializer()) {
193 if (GS.StoredType < GlobalStatus::isInitializerStored)
194 GS.StoredType = GlobalStatus::isInitializerStored;
195 } else if (isa<LoadInst>(StoredVal) &&
196 cast<LoadInst>(StoredVal)->getOperand(0) == GV) {
198 if (GS.StoredType < GlobalStatus::isInitializerStored)
199 GS.StoredType = GlobalStatus::isInitializerStored;
200 } else if (GS.StoredType < GlobalStatus::isStoredOnce) {
201 GS.StoredType = GlobalStatus::isStoredOnce;
202 GS.StoredOnceValue = StoredVal;
203 } else if (GS.StoredType == GlobalStatus::isStoredOnce &&
204 GS.StoredOnceValue == StoredVal) {
207 GS.StoredType = GlobalStatus::isStored;
210 GS.StoredType = GlobalStatus::isStored;
213 } else if (isa<GetElementPtrInst>(I)) {
214 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
215 } else if (isa<SelectInst>(I)) {
216 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
217 } else if (PHINode *PN = dyn_cast<PHINode>(I)) {
218 // PHI nodes we can check just like select or GEP instructions, but we
219 // have to be careful about infinite recursion.
220 if (PHIUsers.insert(PN)) // Not already visited.
221 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
222 GS.HasPHIUser = true;
223 } else if (isa<CmpInst>(I)) {
224 } else if (isa<MemTransferInst>(I)) {
225 if (I->getOperand(1) == V)
226 GS.StoredType = GlobalStatus::isStored;
227 if (I->getOperand(2) == V)
229 } else if (isa<MemSetInst>(I)) {
230 assert(I->getOperand(1) == V && "Memset only takes one pointer!");
231 GS.StoredType = GlobalStatus::isStored;
233 return true; // Any other non-load instruction might take address!
235 } else if (Constant *C = dyn_cast<Constant>(*UI)) {
236 GS.HasNonInstructionUser = true;
237 // We might have a dead and dangling constant hanging off of here.
238 if (!SafeToDestroyConstant(C))
241 GS.HasNonInstructionUser = true;
242 // Otherwise must be some other user.
249 static Constant *getAggregateConstantElement(Constant *Agg, Constant *Idx,
250 LLVMContext &Context) {
251 ConstantInt *CI = dyn_cast<ConstantInt>(Idx);
253 unsigned IdxV = CI->getZExtValue();
255 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg)) {
256 if (IdxV < CS->getNumOperands()) return CS->getOperand(IdxV);
257 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg)) {
258 if (IdxV < CA->getNumOperands()) return CA->getOperand(IdxV);
259 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Agg)) {
260 if (IdxV < CP->getNumOperands()) return CP->getOperand(IdxV);
261 } else if (isa<ConstantAggregateZero>(Agg)) {
262 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
263 if (IdxV < STy->getNumElements())
264 return Constant::getNullValue(STy->getElementType(IdxV));
265 } else if (const SequentialType *STy =
266 dyn_cast<SequentialType>(Agg->getType())) {
267 return Constant::getNullValue(STy->getElementType());
269 } else if (isa<UndefValue>(Agg)) {
270 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
271 if (IdxV < STy->getNumElements())
272 return UndefValue::get(STy->getElementType(IdxV));
273 } else if (const SequentialType *STy =
274 dyn_cast<SequentialType>(Agg->getType())) {
275 return UndefValue::get(STy->getElementType());
282 /// CleanupConstantGlobalUsers - We just marked GV constant. Loop over all
283 /// users of the global, cleaning up the obvious ones. This is largely just a
284 /// quick scan over the use list to clean up the easy and obvious cruft. This
285 /// returns true if it made a change.
286 static bool CleanupConstantGlobalUsers(Value *V, Constant *Init,
287 LLVMContext &Context) {
288 bool Changed = false;
289 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;) {
292 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
294 // Replace the load with the initializer.
295 LI->replaceAllUsesWith(Init);
296 LI->eraseFromParent();
299 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
300 // Store must be unreachable or storing Init into the global.
301 SI->eraseFromParent();
303 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
304 if (CE->getOpcode() == Instruction::GetElementPtr) {
305 Constant *SubInit = 0;
307 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
308 Changed |= CleanupConstantGlobalUsers(CE, SubInit, Context);
309 } else if (CE->getOpcode() == Instruction::BitCast &&
310 isa<PointerType>(CE->getType())) {
311 // Pointer cast, delete any stores and memsets to the global.
312 Changed |= CleanupConstantGlobalUsers(CE, 0, Context);
315 if (CE->use_empty()) {
316 CE->destroyConstant();
319 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
320 // Do not transform "gepinst (gep constexpr (GV))" here, because forming
321 // "gepconstexpr (gep constexpr (GV))" will cause the two gep's to fold
322 // and will invalidate our notion of what Init is.
323 Constant *SubInit = 0;
324 if (!isa<ConstantExpr>(GEP->getOperand(0))) {
326 dyn_cast_or_null<ConstantExpr>(ConstantFoldInstruction(GEP, Context));
327 if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr)
328 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
330 Changed |= CleanupConstantGlobalUsers(GEP, SubInit, Context);
332 if (GEP->use_empty()) {
333 GEP->eraseFromParent();
336 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv
337 if (MI->getRawDest() == V) {
338 MI->eraseFromParent();
342 } else if (Constant *C = dyn_cast<Constant>(U)) {
343 // If we have a chain of dead constantexprs or other things dangling from
344 // us, and if they are all dead, nuke them without remorse.
345 if (SafeToDestroyConstant(C)) {
346 C->destroyConstant();
347 // This could have invalidated UI, start over from scratch.
348 CleanupConstantGlobalUsers(V, Init, Context);
356 /// isSafeSROAElementUse - Return true if the specified instruction is a safe
357 /// user of a derived expression from a global that we want to SROA.
358 static bool isSafeSROAElementUse(Value *V) {
359 // We might have a dead and dangling constant hanging off of here.
360 if (Constant *C = dyn_cast<Constant>(V))
361 return SafeToDestroyConstant(C);
363 Instruction *I = dyn_cast<Instruction>(V);
364 if (!I) return false;
367 if (isa<LoadInst>(I)) return true;
369 // Stores *to* the pointer are ok.
370 if (StoreInst *SI = dyn_cast<StoreInst>(I))
371 return SI->getOperand(0) != V;
373 // Otherwise, it must be a GEP.
374 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I);
375 if (GEPI == 0) return false;
377 if (GEPI->getNumOperands() < 3 || !isa<Constant>(GEPI->getOperand(1)) ||
378 !cast<Constant>(GEPI->getOperand(1))->isNullValue())
381 for (Value::use_iterator I = GEPI->use_begin(), E = GEPI->use_end();
383 if (!isSafeSROAElementUse(*I))
389 /// IsUserOfGlobalSafeForSRA - U is a direct user of the specified global value.
390 /// Look at it and its uses and decide whether it is safe to SROA this global.
392 static bool IsUserOfGlobalSafeForSRA(User *U, GlobalValue *GV) {
393 // The user of the global must be a GEP Inst or a ConstantExpr GEP.
394 if (!isa<GetElementPtrInst>(U) &&
395 (!isa<ConstantExpr>(U) ||
396 cast<ConstantExpr>(U)->getOpcode() != Instruction::GetElementPtr))
399 // Check to see if this ConstantExpr GEP is SRA'able. In particular, we
400 // don't like < 3 operand CE's, and we don't like non-constant integer
401 // indices. This enforces that all uses are 'gep GV, 0, C, ...' for some
403 if (U->getNumOperands() < 3 || !isa<Constant>(U->getOperand(1)) ||
404 !cast<Constant>(U->getOperand(1))->isNullValue() ||
405 !isa<ConstantInt>(U->getOperand(2)))
408 gep_type_iterator GEPI = gep_type_begin(U), E = gep_type_end(U);
409 ++GEPI; // Skip over the pointer index.
411 // If this is a use of an array allocation, do a bit more checking for sanity.
412 if (const ArrayType *AT = dyn_cast<ArrayType>(*GEPI)) {
413 uint64_t NumElements = AT->getNumElements();
414 ConstantInt *Idx = cast<ConstantInt>(U->getOperand(2));
416 // Check to make sure that index falls within the array. If not,
417 // something funny is going on, so we won't do the optimization.
419 if (Idx->getZExtValue() >= NumElements)
422 // We cannot scalar repl this level of the array unless any array
423 // sub-indices are in-range constants. In particular, consider:
424 // A[0][i]. We cannot know that the user isn't doing invalid things like
425 // allowing i to index an out-of-range subscript that accesses A[1].
427 // Scalar replacing *just* the outer index of the array is probably not
428 // going to be a win anyway, so just give up.
429 for (++GEPI; // Skip array index.
432 uint64_t NumElements;
433 if (const ArrayType *SubArrayTy = dyn_cast<ArrayType>(*GEPI))
434 NumElements = SubArrayTy->getNumElements();
435 else if (const VectorType *SubVectorTy = dyn_cast<VectorType>(*GEPI))
436 NumElements = SubVectorTy->getNumElements();
438 assert(isa<StructType>(*GEPI) &&
439 "Indexed GEP type is not array, vector, or struct!");
443 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPI.getOperand());
444 if (!IdxVal || IdxVal->getZExtValue() >= NumElements)
449 for (Value::use_iterator I = U->use_begin(), E = U->use_end(); I != E; ++I)
450 if (!isSafeSROAElementUse(*I))
455 /// GlobalUsersSafeToSRA - Look at all uses of the global and decide whether it
456 /// is safe for us to perform this transformation.
458 static bool GlobalUsersSafeToSRA(GlobalValue *GV) {
459 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end();
461 if (!IsUserOfGlobalSafeForSRA(*UI, GV))
468 /// SRAGlobal - Perform scalar replacement of aggregates on the specified global
469 /// variable. This opens the door for other optimizations by exposing the
470 /// behavior of the program in a more fine-grained way. We have determined that
471 /// this transformation is safe already. We return the first global variable we
472 /// insert so that the caller can reprocess it.
473 static GlobalVariable *SRAGlobal(GlobalVariable *GV, const TargetData &TD,
474 LLVMContext &Context) {
475 // Make sure this global only has simple uses that we can SRA.
476 if (!GlobalUsersSafeToSRA(GV))
479 assert(GV->hasLocalLinkage() && !GV->isConstant());
480 Constant *Init = GV->getInitializer();
481 const Type *Ty = Init->getType();
483 std::vector<GlobalVariable*> NewGlobals;
484 Module::GlobalListType &Globals = GV->getParent()->getGlobalList();
486 // Get the alignment of the global, either explicit or target-specific.
487 unsigned StartAlignment = GV->getAlignment();
488 if (StartAlignment == 0)
489 StartAlignment = TD.getABITypeAlignment(GV->getType());
491 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
492 NewGlobals.reserve(STy->getNumElements());
493 const StructLayout &Layout = *TD.getStructLayout(STy);
494 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
495 Constant *In = getAggregateConstantElement(Init,
496 ConstantInt::get(Type::getInt32Ty(Context), i),
498 assert(In && "Couldn't get element of initializer?");
499 GlobalVariable *NGV = new GlobalVariable(Context,
500 STy->getElementType(i), false,
501 GlobalVariable::InternalLinkage,
502 In, GV->getName()+"."+Twine(i),
504 GV->getType()->getAddressSpace());
505 Globals.insert(GV, NGV);
506 NewGlobals.push_back(NGV);
508 // Calculate the known alignment of the field. If the original aggregate
509 // had 256 byte alignment for example, something might depend on that:
510 // propagate info to each field.
511 uint64_t FieldOffset = Layout.getElementOffset(i);
512 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, FieldOffset);
513 if (NewAlign > TD.getABITypeAlignment(STy->getElementType(i)))
514 NGV->setAlignment(NewAlign);
516 } else if (const SequentialType *STy = dyn_cast<SequentialType>(Ty)) {
517 unsigned NumElements = 0;
518 if (const ArrayType *ATy = dyn_cast<ArrayType>(STy))
519 NumElements = ATy->getNumElements();
521 NumElements = cast<VectorType>(STy)->getNumElements();
523 if (NumElements > 16 && GV->hasNUsesOrMore(16))
524 return 0; // It's not worth it.
525 NewGlobals.reserve(NumElements);
527 uint64_t EltSize = TD.getTypeAllocSize(STy->getElementType());
528 unsigned EltAlign = TD.getABITypeAlignment(STy->getElementType());
529 for (unsigned i = 0, e = NumElements; i != e; ++i) {
530 Constant *In = getAggregateConstantElement(Init,
531 ConstantInt::get(Type::getInt32Ty(Context), i),
533 assert(In && "Couldn't get element of initializer?");
535 GlobalVariable *NGV = new GlobalVariable(Context,
536 STy->getElementType(), false,
537 GlobalVariable::InternalLinkage,
538 In, GV->getName()+"."+Twine(i),
540 GV->getType()->getAddressSpace());
541 Globals.insert(GV, NGV);
542 NewGlobals.push_back(NGV);
544 // Calculate the known alignment of the field. If the original aggregate
545 // had 256 byte alignment for example, something might depend on that:
546 // propagate info to each field.
547 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, EltSize*i);
548 if (NewAlign > EltAlign)
549 NGV->setAlignment(NewAlign);
553 if (NewGlobals.empty())
556 DEBUG(errs() << "PERFORMING GLOBAL SRA ON: " << *GV);
558 Constant *NullInt = Constant::getNullValue(Type::getInt32Ty(Context));
560 // Loop over all of the uses of the global, replacing the constantexpr geps,
561 // with smaller constantexpr geps or direct references.
562 while (!GV->use_empty()) {
563 User *GEP = GV->use_back();
564 assert(((isa<ConstantExpr>(GEP) &&
565 cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)||
566 isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!");
568 // Ignore the 1th operand, which has to be zero or else the program is quite
569 // broken (undefined). Get the 2nd operand, which is the structure or array
571 unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
572 if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access.
574 Value *NewPtr = NewGlobals[Val];
576 // Form a shorter GEP if needed.
577 if (GEP->getNumOperands() > 3) {
578 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) {
579 SmallVector<Constant*, 8> Idxs;
580 Idxs.push_back(NullInt);
581 for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i)
582 Idxs.push_back(CE->getOperand(i));
583 NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr),
584 &Idxs[0], Idxs.size());
586 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP);
587 SmallVector<Value*, 8> Idxs;
588 Idxs.push_back(NullInt);
589 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i)
590 Idxs.push_back(GEPI->getOperand(i));
591 NewPtr = GetElementPtrInst::Create(NewPtr, Idxs.begin(), Idxs.end(),
592 GEPI->getName()+"."+Twine(Val),GEPI);
595 GEP->replaceAllUsesWith(NewPtr);
597 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP))
598 GEPI->eraseFromParent();
600 cast<ConstantExpr>(GEP)->destroyConstant();
603 // Delete the old global, now that it is dead.
607 // Loop over the new globals array deleting any globals that are obviously
608 // dead. This can arise due to scalarization of a structure or an array that
609 // has elements that are dead.
610 unsigned FirstGlobal = 0;
611 for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i)
612 if (NewGlobals[i]->use_empty()) {
613 Globals.erase(NewGlobals[i]);
614 if (FirstGlobal == i) ++FirstGlobal;
617 return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : 0;
620 /// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified
621 /// value will trap if the value is dynamically null. PHIs keeps track of any
622 /// phi nodes we've seen to avoid reprocessing them.
623 static bool AllUsesOfValueWillTrapIfNull(Value *V,
624 SmallPtrSet<PHINode*, 8> &PHIs) {
625 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
626 if (isa<LoadInst>(*UI)) {
628 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
629 if (SI->getOperand(0) == V) {
630 //cerr << "NONTRAPPING USE: " << **UI;
631 return false; // Storing the value.
633 } else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
634 if (CI->getOperand(0) != V) {
635 //cerr << "NONTRAPPING USE: " << **UI;
636 return false; // Not calling the ptr
638 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
639 if (II->getOperand(0) != V) {
640 //cerr << "NONTRAPPING USE: " << **UI;
641 return false; // Not calling the ptr
643 } else if (BitCastInst *CI = dyn_cast<BitCastInst>(*UI)) {
644 if (!AllUsesOfValueWillTrapIfNull(CI, PHIs)) return false;
645 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) {
646 if (!AllUsesOfValueWillTrapIfNull(GEPI, PHIs)) return false;
647 } else if (PHINode *PN = dyn_cast<PHINode>(*UI)) {
648 // If we've already seen this phi node, ignore it, it has already been
651 return AllUsesOfValueWillTrapIfNull(PN, PHIs);
652 } else if (isa<ICmpInst>(*UI) &&
653 isa<ConstantPointerNull>(UI->getOperand(1))) {
654 // Ignore setcc X, null
656 //cerr << "NONTRAPPING USE: " << **UI;
662 /// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads
663 /// from GV will trap if the loaded value is null. Note that this also permits
664 /// comparisons of the loaded value against null, as a special case.
665 static bool AllUsesOfLoadedValueWillTrapIfNull(GlobalVariable *GV) {
666 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI!=E; ++UI)
667 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
668 SmallPtrSet<PHINode*, 8> PHIs;
669 if (!AllUsesOfValueWillTrapIfNull(LI, PHIs))
671 } else if (isa<StoreInst>(*UI)) {
672 // Ignore stores to the global.
674 // We don't know or understand this user, bail out.
675 //cerr << "UNKNOWN USER OF GLOBAL!: " << **UI;
682 static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV,
683 LLVMContext &Context) {
684 bool Changed = false;
685 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) {
686 Instruction *I = cast<Instruction>(*UI++);
687 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
688 LI->setOperand(0, NewV);
690 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
691 if (SI->getOperand(1) == V) {
692 SI->setOperand(1, NewV);
695 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
696 if (I->getOperand(0) == V) {
697 // Calling through the pointer! Turn into a direct call, but be careful
698 // that the pointer is not also being passed as an argument.
699 I->setOperand(0, NewV);
701 bool PassedAsArg = false;
702 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i)
703 if (I->getOperand(i) == V) {
705 I->setOperand(i, NewV);
709 // Being passed as an argument also. Be careful to not invalidate UI!
713 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
714 Changed |= OptimizeAwayTrappingUsesOfValue(CI,
715 ConstantExpr::getCast(CI->getOpcode(),
716 NewV, CI->getType()), Context);
717 if (CI->use_empty()) {
719 CI->eraseFromParent();
721 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
722 // Should handle GEP here.
723 SmallVector<Constant*, 8> Idxs;
724 Idxs.reserve(GEPI->getNumOperands()-1);
725 for (User::op_iterator i = GEPI->op_begin() + 1, e = GEPI->op_end();
727 if (Constant *C = dyn_cast<Constant>(*i))
731 if (Idxs.size() == GEPI->getNumOperands()-1)
732 Changed |= OptimizeAwayTrappingUsesOfValue(GEPI,
733 ConstantExpr::getGetElementPtr(NewV, &Idxs[0],
734 Idxs.size()), Context);
735 if (GEPI->use_empty()) {
737 GEPI->eraseFromParent();
746 /// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null
747 /// value stored into it. If there are uses of the loaded value that would trap
748 /// if the loaded value is dynamically null, then we know that they cannot be
749 /// reachable with a null optimize away the load.
750 static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV,
751 LLVMContext &Context) {
752 bool Changed = false;
754 // Keep track of whether we are able to remove all the uses of the global
755 // other than the store that defines it.
756 bool AllNonStoreUsesGone = true;
758 // Replace all uses of loads with uses of uses of the stored value.
759 for (Value::use_iterator GUI = GV->use_begin(), E = GV->use_end(); GUI != E;){
760 User *GlobalUser = *GUI++;
761 if (LoadInst *LI = dyn_cast<LoadInst>(GlobalUser)) {
762 Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV, Context);
763 // If we were able to delete all uses of the loads
764 if (LI->use_empty()) {
765 LI->eraseFromParent();
768 AllNonStoreUsesGone = false;
770 } else if (isa<StoreInst>(GlobalUser)) {
771 // Ignore the store that stores "LV" to the global.
772 assert(GlobalUser->getOperand(1) == GV &&
773 "Must be storing *to* the global");
775 AllNonStoreUsesGone = false;
777 // If we get here we could have other crazy uses that are transitively
779 assert((isa<PHINode>(GlobalUser) || isa<SelectInst>(GlobalUser) ||
780 isa<ConstantExpr>(GlobalUser)) && "Only expect load and stores!");
785 DEBUG(errs() << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV);
789 // If we nuked all of the loads, then none of the stores are needed either,
790 // nor is the global.
791 if (AllNonStoreUsesGone) {
792 DEBUG(errs() << " *** GLOBAL NOW DEAD!\n");
793 CleanupConstantGlobalUsers(GV, 0, Context);
794 if (GV->use_empty()) {
795 GV->eraseFromParent();
803 /// ConstantPropUsersOf - Walk the use list of V, constant folding all of the
804 /// instructions that are foldable.
805 static void ConstantPropUsersOf(Value *V, LLVMContext &Context) {
806 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; )
807 if (Instruction *I = dyn_cast<Instruction>(*UI++))
808 if (Constant *NewC = ConstantFoldInstruction(I, Context)) {
809 I->replaceAllUsesWith(NewC);
811 // Advance UI to the next non-I use to avoid invalidating it!
812 // Instructions could multiply use V.
813 while (UI != E && *UI == I)
815 I->eraseFromParent();
819 /// OptimizeGlobalAddressOfMalloc - This function takes the specified global
820 /// variable, and transforms the program as if it always contained the result of
821 /// the specified malloc. Because it is always the result of the specified
822 /// malloc, there is no reason to actually DO the malloc. Instead, turn the
823 /// malloc into a global, and any loads of GV as uses of the new global.
824 static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV,
827 LLVMContext &Context,
829 DEBUG(errs() << "PROMOTING MALLOC GLOBAL: " << *GV
830 << " CALL = " << *CI << " BCI = " << *BCI << '\n');
832 const Type *IntPtrTy = TD->getIntPtrType(Context);
834 Value* ArraySize = getMallocArraySize(CI, Context, TD);
835 assert(ArraySize && "not a malloc whose array size can be determined");
836 ConstantInt *NElements = cast<ConstantInt>(ArraySize);
837 if (NElements->getZExtValue() != 1) {
838 // If we have an array allocation, transform it to a single element
839 // allocation to make the code below simpler.
840 Type *NewTy = ArrayType::get(getMallocAllocatedType(CI),
841 NElements->getZExtValue());
842 Value* NewM = CallInst::CreateMalloc(CI, IntPtrTy, NewTy);
843 Instruction* NewMI = cast<Instruction>(NewM);
845 Indices[0] = Indices[1] = Constant::getNullValue(IntPtrTy);
846 Value *NewGEP = GetElementPtrInst::Create(NewMI, Indices, Indices + 2,
847 NewMI->getName()+".el0", CI);
848 BCI->replaceAllUsesWith(NewGEP);
849 BCI->eraseFromParent();
850 CI->eraseFromParent();
851 BCI = cast<BitCastInst>(NewMI);
852 CI = extractMallocCallFromBitCast(NewMI);
855 // Create the new global variable. The contents of the malloc'd memory is
856 // undefined, so initialize with an undef value.
857 const Type *MAT = getMallocAllocatedType(CI);
858 Constant *Init = UndefValue::get(MAT);
859 GlobalVariable *NewGV = new GlobalVariable(*GV->getParent(),
861 GlobalValue::InternalLinkage, Init,
862 GV->getName()+".body",
864 GV->isThreadLocal());
866 // Anything that used the malloc now uses the global directly.
867 BCI->replaceAllUsesWith(NewGV);
869 Constant *RepValue = NewGV;
870 if (NewGV->getType() != GV->getType()->getElementType())
871 RepValue = ConstantExpr::getBitCast(RepValue,
872 GV->getType()->getElementType());
874 // If there is a comparison against null, we will insert a global bool to
875 // keep track of whether the global was initialized yet or not.
876 GlobalVariable *InitBool =
877 new GlobalVariable(Context, Type::getInt1Ty(Context), false,
878 GlobalValue::InternalLinkage,
879 ConstantInt::getFalse(Context), GV->getName()+".init",
880 GV->isThreadLocal());
881 bool InitBoolUsed = false;
883 // Loop over all uses of GV, processing them in turn.
884 std::vector<StoreInst*> Stores;
885 while (!GV->use_empty())
886 if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) {
887 while (!LI->use_empty()) {
888 Use &LoadUse = LI->use_begin().getUse();
889 if (!isa<ICmpInst>(LoadUse.getUser()))
892 ICmpInst *ICI = cast<ICmpInst>(LoadUse.getUser());
893 // Replace the cmp X, 0 with a use of the bool value.
894 Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", ICI);
896 switch (ICI->getPredicate()) {
897 default: llvm_unreachable("Unknown ICmp Predicate!");
898 case ICmpInst::ICMP_ULT:
899 case ICmpInst::ICMP_SLT:
900 LV = ConstantInt::getFalse(Context); // X < null -> always false
902 case ICmpInst::ICMP_ULE:
903 case ICmpInst::ICMP_SLE:
904 case ICmpInst::ICMP_EQ:
905 LV = BinaryOperator::CreateNot(LV, "notinit", ICI);
907 case ICmpInst::ICMP_NE:
908 case ICmpInst::ICMP_UGE:
909 case ICmpInst::ICMP_SGE:
910 case ICmpInst::ICMP_UGT:
911 case ICmpInst::ICMP_SGT:
914 ICI->replaceAllUsesWith(LV);
915 ICI->eraseFromParent();
918 LI->eraseFromParent();
920 StoreInst *SI = cast<StoreInst>(GV->use_back());
921 // The global is initialized when the store to it occurs.
922 new StoreInst(ConstantInt::getTrue(Context), InitBool, SI);
923 SI->eraseFromParent();
926 // If the initialization boolean was used, insert it, otherwise delete it.
928 while (!InitBool->use_empty()) // Delete initializations
929 cast<Instruction>(InitBool->use_back())->eraseFromParent();
932 GV->getParent()->getGlobalList().insert(GV, InitBool);
935 // Now the GV is dead, nuke it and the malloc.
936 GV->eraseFromParent();
937 BCI->eraseFromParent();
938 CI->eraseFromParent();
940 // To further other optimizations, loop over all users of NewGV and try to
941 // constant prop them. This will promote GEP instructions with constant
942 // indices into GEP constant-exprs, which will allow global-opt to hack on it.
943 ConstantPropUsersOf(NewGV, Context);
944 if (RepValue != NewGV)
945 ConstantPropUsersOf(RepValue, Context);
950 /// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking
951 /// to make sure that there are no complex uses of V. We permit simple things
952 /// like dereferencing the pointer, but not storing through the address, unless
953 /// it is to the specified global.
954 static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Instruction *V,
956 SmallPtrSet<PHINode*, 8> &PHIs) {
957 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
958 Instruction *Inst = cast<Instruction>(*UI);
960 if (isa<LoadInst>(Inst) || isa<CmpInst>(Inst)) {
961 continue; // Fine, ignore.
964 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
965 if (SI->getOperand(0) == V && SI->getOperand(1) != GV)
966 return false; // Storing the pointer itself... bad.
967 continue; // Otherwise, storing through it, or storing into GV... fine.
970 if (isa<GetElementPtrInst>(Inst)) {
971 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Inst, GV, PHIs))
976 if (PHINode *PN = dyn_cast<PHINode>(Inst)) {
977 // PHIs are ok if all uses are ok. Don't infinitely recurse through PHI
980 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(PN, GV, PHIs))
985 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Inst)) {
986 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(BCI, GV, PHIs))
996 /// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV
997 /// somewhere. Transform all uses of the allocation into loads from the
998 /// global and uses of the resultant pointer. Further, delete the store into
999 /// GV. This assumes that these value pass the
1000 /// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate.
1001 static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc,
1002 GlobalVariable *GV) {
1003 while (!Alloc->use_empty()) {
1004 Instruction *U = cast<Instruction>(*Alloc->use_begin());
1005 Instruction *InsertPt = U;
1006 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
1007 // If this is the store of the allocation into the global, remove it.
1008 if (SI->getOperand(1) == GV) {
1009 SI->eraseFromParent();
1012 } else if (PHINode *PN = dyn_cast<PHINode>(U)) {
1013 // Insert the load in the corresponding predecessor, not right before the
1015 InsertPt = PN->getIncomingBlock(Alloc->use_begin())->getTerminator();
1016 } else if (isa<BitCastInst>(U)) {
1017 // Must be bitcast between the malloc and store to initialize the global.
1018 ReplaceUsesOfMallocWithGlobal(U, GV);
1019 U->eraseFromParent();
1021 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
1022 // If this is a "GEP bitcast" and the user is a store to the global, then
1023 // just process it as a bitcast.
1024 if (GEPI->hasAllZeroIndices() && GEPI->hasOneUse())
1025 if (StoreInst *SI = dyn_cast<StoreInst>(GEPI->use_back()))
1026 if (SI->getOperand(1) == GV) {
1027 // Must be bitcast GEP between the malloc and store to initialize
1029 ReplaceUsesOfMallocWithGlobal(GEPI, GV);
1030 GEPI->eraseFromParent();
1035 // Insert a load from the global, and use it instead of the malloc.
1036 Value *NL = new LoadInst(GV, GV->getName()+".val", InsertPt);
1037 U->replaceUsesOfWith(Alloc, NL);
1041 /// LoadUsesSimpleEnoughForHeapSRA - Verify that all uses of V (a load, or a phi
1042 /// of a load) are simple enough to perform heap SRA on. This permits GEP's
1043 /// that index through the array and struct field, icmps of null, and PHIs.
1044 static bool LoadUsesSimpleEnoughForHeapSRA(Value *V,
1045 SmallPtrSet<PHINode*, 32> &LoadUsingPHIs,
1046 SmallPtrSet<PHINode*, 32> &LoadUsingPHIsPerLoad) {
1047 // We permit two users of the load: setcc comparing against the null
1048 // pointer, and a getelementptr of a specific form.
1049 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
1050 Instruction *User = cast<Instruction>(*UI);
1052 // Comparison against null is ok.
1053 if (ICmpInst *ICI = dyn_cast<ICmpInst>(User)) {
1054 if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
1059 // getelementptr is also ok, but only a simple form.
1060 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1061 // Must index into the array and into the struct.
1062 if (GEPI->getNumOperands() < 3)
1065 // Otherwise the GEP is ok.
1069 if (PHINode *PN = dyn_cast<PHINode>(User)) {
1070 if (!LoadUsingPHIsPerLoad.insert(PN))
1071 // This means some phi nodes are dependent on each other.
1072 // Avoid infinite looping!
1074 if (!LoadUsingPHIs.insert(PN))
1075 // If we have already analyzed this PHI, then it is safe.
1078 // Make sure all uses of the PHI are simple enough to transform.
1079 if (!LoadUsesSimpleEnoughForHeapSRA(PN,
1080 LoadUsingPHIs, LoadUsingPHIsPerLoad))
1086 // Otherwise we don't know what this is, not ok.
1094 /// AllGlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from
1095 /// GV are simple enough to perform HeapSRA, return true.
1096 static bool AllGlobalLoadUsesSimpleEnoughForHeapSRA(GlobalVariable *GV,
1097 Instruction *StoredVal) {
1098 SmallPtrSet<PHINode*, 32> LoadUsingPHIs;
1099 SmallPtrSet<PHINode*, 32> LoadUsingPHIsPerLoad;
1100 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;
1102 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
1103 if (!LoadUsesSimpleEnoughForHeapSRA(LI, LoadUsingPHIs,
1104 LoadUsingPHIsPerLoad))
1106 LoadUsingPHIsPerLoad.clear();
1109 // If we reach here, we know that all uses of the loads and transitive uses
1110 // (through PHI nodes) are simple enough to transform. However, we don't know
1111 // that all inputs the to the PHI nodes are in the same equivalence sets.
1112 // Check to verify that all operands of the PHIs are either PHIS that can be
1113 // transformed, loads from GV, or MI itself.
1114 for (SmallPtrSet<PHINode*, 32>::iterator I = LoadUsingPHIs.begin(),
1115 E = LoadUsingPHIs.end(); I != E; ++I) {
1117 for (unsigned op = 0, e = PN->getNumIncomingValues(); op != e; ++op) {
1118 Value *InVal = PN->getIncomingValue(op);
1120 // PHI of the stored value itself is ok.
1121 if (InVal == StoredVal) continue;
1123 if (PHINode *InPN = dyn_cast<PHINode>(InVal)) {
1124 // One of the PHIs in our set is (optimistically) ok.
1125 if (LoadUsingPHIs.count(InPN))
1130 // Load from GV is ok.
1131 if (LoadInst *LI = dyn_cast<LoadInst>(InVal))
1132 if (LI->getOperand(0) == GV)
1137 // Anything else is rejected.
1145 static Value *GetHeapSROAValue(Value *V, unsigned FieldNo,
1146 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1147 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite,
1148 LLVMContext &Context) {
1149 std::vector<Value*> &FieldVals = InsertedScalarizedValues[V];
1151 if (FieldNo >= FieldVals.size())
1152 FieldVals.resize(FieldNo+1);
1154 // If we already have this value, just reuse the previously scalarized
1156 if (Value *FieldVal = FieldVals[FieldNo])
1159 // Depending on what instruction this is, we have several cases.
1161 if (LoadInst *LI = dyn_cast<LoadInst>(V)) {
1162 // This is a scalarized version of the load from the global. Just create
1163 // a new Load of the scalarized global.
1164 Result = new LoadInst(GetHeapSROAValue(LI->getOperand(0), FieldNo,
1165 InsertedScalarizedValues,
1166 PHIsToRewrite, Context),
1167 LI->getName()+".f"+Twine(FieldNo), LI);
1168 } else if (PHINode *PN = dyn_cast<PHINode>(V)) {
1169 // PN's type is pointer to struct. Make a new PHI of pointer to struct
1171 const StructType *ST =
1172 cast<StructType>(cast<PointerType>(PN->getType())->getElementType());
1175 PHINode::Create(PointerType::getUnqual(ST->getElementType(FieldNo)),
1176 PN->getName()+".f"+Twine(FieldNo), PN);
1177 PHIsToRewrite.push_back(std::make_pair(PN, FieldNo));
1179 llvm_unreachable("Unknown usable value");
1183 return FieldVals[FieldNo] = Result;
1186 /// RewriteHeapSROALoadUser - Given a load instruction and a value derived from
1187 /// the load, rewrite the derived value to use the HeapSRoA'd load.
1188 static void RewriteHeapSROALoadUser(Instruction *LoadUser,
1189 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1190 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite,
1191 LLVMContext &Context) {
1192 // If this is a comparison against null, handle it.
1193 if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) {
1194 assert(isa<ConstantPointerNull>(SCI->getOperand(1)));
1195 // If we have a setcc of the loaded pointer, we can use a setcc of any
1197 Value *NPtr = GetHeapSROAValue(SCI->getOperand(0), 0,
1198 InsertedScalarizedValues, PHIsToRewrite,
1201 Value *New = new ICmpInst(SCI, SCI->getPredicate(), NPtr,
1202 Constant::getNullValue(NPtr->getType()),
1204 SCI->replaceAllUsesWith(New);
1205 SCI->eraseFromParent();
1209 // Handle 'getelementptr Ptr, Idx, i32 FieldNo ...'
1210 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) {
1211 assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2))
1212 && "Unexpected GEPI!");
1214 // Load the pointer for this field.
1215 unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
1216 Value *NewPtr = GetHeapSROAValue(GEPI->getOperand(0), FieldNo,
1217 InsertedScalarizedValues, PHIsToRewrite,
1220 // Create the new GEP idx vector.
1221 SmallVector<Value*, 8> GEPIdx;
1222 GEPIdx.push_back(GEPI->getOperand(1));
1223 GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end());
1225 Value *NGEPI = GetElementPtrInst::Create(NewPtr,
1226 GEPIdx.begin(), GEPIdx.end(),
1227 GEPI->getName(), GEPI);
1228 GEPI->replaceAllUsesWith(NGEPI);
1229 GEPI->eraseFromParent();
1233 // Recursively transform the users of PHI nodes. This will lazily create the
1234 // PHIs that are needed for individual elements. Keep track of what PHIs we
1235 // see in InsertedScalarizedValues so that we don't get infinite loops (very
1236 // antisocial). If the PHI is already in InsertedScalarizedValues, it has
1237 // already been seen first by another load, so its uses have already been
1239 PHINode *PN = cast<PHINode>(LoadUser);
1241 DenseMap<Value*, std::vector<Value*> >::iterator InsertPos;
1242 tie(InsertPos, Inserted) =
1243 InsertedScalarizedValues.insert(std::make_pair(PN, std::vector<Value*>()));
1244 if (!Inserted) return;
1246 // If this is the first time we've seen this PHI, recursively process all
1248 for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end(); UI != E; ) {
1249 Instruction *User = cast<Instruction>(*UI++);
1250 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite,
1255 /// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global. Ptr
1256 /// is a value loaded from the global. Eliminate all uses of Ptr, making them
1257 /// use FieldGlobals instead. All uses of loaded values satisfy
1258 /// AllGlobalLoadUsesSimpleEnoughForHeapSRA.
1259 static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load,
1260 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1261 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite,
1262 LLVMContext &Context) {
1263 for (Value::use_iterator UI = Load->use_begin(), E = Load->use_end();
1265 Instruction *User = cast<Instruction>(*UI++);
1266 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite,
1270 if (Load->use_empty()) {
1271 Load->eraseFromParent();
1272 InsertedScalarizedValues.erase(Load);
1276 /// PerformHeapAllocSRoA - CI is an allocation of an array of structures. Break
1277 /// it up into multiple allocations of arrays of the fields.
1278 static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV,
1279 CallInst *CI, BitCastInst* BCI,
1280 LLVMContext &Context,
1282 DEBUG(errs() << "SROA HEAP ALLOC: " << *GV << " MALLOC CALL = " << *CI
1283 << " BITCAST = " << *BCI << '\n');
1284 const Type* MAT = getMallocAllocatedType(CI);
1285 const StructType *STy = cast<StructType>(MAT);
1286 Value* ArraySize = getMallocArraySize(CI, Context, TD);
1287 assert(ArraySize && "not a malloc whose array size can be determined");
1289 // There is guaranteed to be at least one use of the malloc (storing
1290 // it into GV). If there are other uses, change them to be uses of
1291 // the global to simplify later code. This also deletes the store
1293 ReplaceUsesOfMallocWithGlobal(BCI, GV);
1295 // Okay, at this point, there are no users of the malloc. Insert N
1296 // new mallocs at the same place as CI, and N globals.
1297 std::vector<Value*> FieldGlobals;
1298 std::vector<Value*> FieldMallocs;
1300 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){
1301 const Type *FieldTy = STy->getElementType(FieldNo);
1302 const PointerType *PFieldTy = PointerType::getUnqual(FieldTy);
1304 GlobalVariable *NGV =
1305 new GlobalVariable(*GV->getParent(),
1306 PFieldTy, false, GlobalValue::InternalLinkage,
1307 Constant::getNullValue(PFieldTy),
1308 GV->getName() + ".f" + Twine(FieldNo), GV,
1309 GV->isThreadLocal());
1310 FieldGlobals.push_back(NGV);
1312 Value *NMI = CallInst::CreateMalloc(CI, TD->getIntPtrType(Context),
1314 BCI->getName() + ".f" + Twine(FieldNo));
1315 FieldMallocs.push_back(NMI);
1316 new StoreInst(NMI, NGV, BCI);
1319 // The tricky aspect of this transformation is handling the case when malloc
1320 // fails. In the original code, malloc failing would set the result pointer
1321 // of malloc to null. In this case, some mallocs could succeed and others
1322 // could fail. As such, we emit code that looks like this:
1323 // F0 = malloc(field0)
1324 // F1 = malloc(field1)
1325 // F2 = malloc(field2)
1326 // if (F0 == 0 || F1 == 0 || F2 == 0) {
1327 // if (F0) { free(F0); F0 = 0; }
1328 // if (F1) { free(F1); F1 = 0; }
1329 // if (F2) { free(F2); F2 = 0; }
1331 Value *RunningOr = 0;
1332 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) {
1333 Value *Cond = new ICmpInst(BCI, ICmpInst::ICMP_EQ, FieldMallocs[i],
1334 Constant::getNullValue(FieldMallocs[i]->getType()),
1337 RunningOr = Cond; // First seteq
1339 RunningOr = BinaryOperator::CreateOr(RunningOr, Cond, "tmp", BCI);
1342 // Split the basic block at the old malloc.
1343 BasicBlock *OrigBB = BCI->getParent();
1344 BasicBlock *ContBB = OrigBB->splitBasicBlock(BCI, "malloc_cont");
1346 // Create the block to check the first condition. Put all these blocks at the
1347 // end of the function as they are unlikely to be executed.
1348 BasicBlock *NullPtrBlock = BasicBlock::Create(Context, "malloc_ret_null",
1349 OrigBB->getParent());
1351 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond
1352 // branch on RunningOr.
1353 OrigBB->getTerminator()->eraseFromParent();
1354 BranchInst::Create(NullPtrBlock, ContBB, RunningOr, OrigBB);
1356 // Within the NullPtrBlock, we need to emit a comparison and branch for each
1357 // pointer, because some may be null while others are not.
1358 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1359 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock);
1360 Value *Cmp = new ICmpInst(*NullPtrBlock, ICmpInst::ICMP_NE, GVVal,
1361 Constant::getNullValue(GVVal->getType()),
1363 BasicBlock *FreeBlock = BasicBlock::Create(Context, "free_it",
1364 OrigBB->getParent());
1365 BasicBlock *NextBlock = BasicBlock::Create(Context, "next",
1366 OrigBB->getParent());
1367 BranchInst::Create(FreeBlock, NextBlock, Cmp, NullPtrBlock);
1369 // Fill in FreeBlock.
1370 new FreeInst(GVVal, FreeBlock);
1371 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i],
1373 BranchInst::Create(NextBlock, FreeBlock);
1375 NullPtrBlock = NextBlock;
1378 BranchInst::Create(ContBB, NullPtrBlock);
1380 // CI and BCI are no longer needed, remove them.
1381 BCI->eraseFromParent();
1382 CI->eraseFromParent();
1384 /// InsertedScalarizedLoads - As we process loads, if we can't immediately
1385 /// update all uses of the load, keep track of what scalarized loads are
1386 /// inserted for a given load.
1387 DenseMap<Value*, std::vector<Value*> > InsertedScalarizedValues;
1388 InsertedScalarizedValues[GV] = FieldGlobals;
1390 std::vector<std::pair<PHINode*, unsigned> > PHIsToRewrite;
1392 // Okay, the malloc site is completely handled. All of the uses of GV are now
1393 // loads, and all uses of those loads are simple. Rewrite them to use loads
1394 // of the per-field globals instead.
1395 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;) {
1396 Instruction *User = cast<Instruction>(*UI++);
1398 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1399 RewriteUsesOfLoadForHeapSRoA(LI, InsertedScalarizedValues, PHIsToRewrite,
1404 // Must be a store of null.
1405 StoreInst *SI = cast<StoreInst>(User);
1406 assert(isa<ConstantPointerNull>(SI->getOperand(0)) &&
1407 "Unexpected heap-sra user!");
1409 // Insert a store of null into each global.
1410 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1411 const PointerType *PT = cast<PointerType>(FieldGlobals[i]->getType());
1412 Constant *Null = Constant::getNullValue(PT->getElementType());
1413 new StoreInst(Null, FieldGlobals[i], SI);
1415 // Erase the original store.
1416 SI->eraseFromParent();
1419 // While we have PHIs that are interesting to rewrite, do it.
1420 while (!PHIsToRewrite.empty()) {
1421 PHINode *PN = PHIsToRewrite.back().first;
1422 unsigned FieldNo = PHIsToRewrite.back().second;
1423 PHIsToRewrite.pop_back();
1424 PHINode *FieldPN = cast<PHINode>(InsertedScalarizedValues[PN][FieldNo]);
1425 assert(FieldPN->getNumIncomingValues() == 0 &&"Already processed this phi");
1427 // Add all the incoming values. This can materialize more phis.
1428 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1429 Value *InVal = PN->getIncomingValue(i);
1430 InVal = GetHeapSROAValue(InVal, FieldNo, InsertedScalarizedValues,
1431 PHIsToRewrite, Context);
1432 FieldPN->addIncoming(InVal, PN->getIncomingBlock(i));
1436 // Drop all inter-phi links and any loads that made it this far.
1437 for (DenseMap<Value*, std::vector<Value*> >::iterator
1438 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1440 if (PHINode *PN = dyn_cast<PHINode>(I->first))
1441 PN->dropAllReferences();
1442 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1443 LI->dropAllReferences();
1446 // Delete all the phis and loads now that inter-references are dead.
1447 for (DenseMap<Value*, std::vector<Value*> >::iterator
1448 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1450 if (PHINode *PN = dyn_cast<PHINode>(I->first))
1451 PN->eraseFromParent();
1452 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1453 LI->eraseFromParent();
1456 // The old global is now dead, remove it.
1457 GV->eraseFromParent();
1460 return cast<GlobalVariable>(FieldGlobals[0]);
1463 /// TryToOptimizeStoreOfMallocToGlobal - This function is called when we see a
1464 /// pointer global variable with a single value stored it that is a malloc or
1466 static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV,
1469 Module::global_iterator &GVI,
1471 LLVMContext &Context) {
1472 // If we can't figure out the type being malloced, then we can't optimize.
1473 const Type *AllocTy = getMallocAllocatedType(CI);
1476 // If this is a malloc of an abstract type, don't touch it.
1477 if (!AllocTy->isSized())
1480 // We can't optimize this global unless all uses of it are *known* to be
1481 // of the malloc value, not of the null initializer value (consider a use
1482 // that compares the global's value against zero to see if the malloc has
1483 // been reached). To do this, we check to see if all uses of the global
1484 // would trap if the global were null: this proves that they must all
1485 // happen after the malloc.
1486 if (!AllUsesOfLoadedValueWillTrapIfNull(GV))
1489 // We can't optimize this if the malloc itself is used in a complex way,
1490 // for example, being stored into multiple globals. This allows the
1491 // malloc to be stored into the specified global, loaded setcc'd, and
1492 // GEP'd. These are all things we could transform to using the global
1495 SmallPtrSet<PHINode*, 8> PHIs;
1496 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(BCI, GV, PHIs))
1500 // If we have a global that is only initialized with a fixed size malloc,
1501 // transform the program to use global memory instead of malloc'd memory.
1502 // This eliminates dynamic allocation, avoids an indirection accessing the
1503 // data, and exposes the resultant global to further GlobalOpt.
1504 Value *NElems = getMallocArraySize(CI, Context, TD);
1505 // We cannot optimize the malloc if we cannot determine malloc array size.
1507 if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems))
1508 // Restrict this transformation to only working on small allocations
1509 // (2048 bytes currently), as we don't want to introduce a 16M global or
1512 NElements->getZExtValue() * TD->getTypeAllocSize(AllocTy) < 2048) {
1513 GVI = OptimizeGlobalAddressOfMalloc(GV, CI, BCI, Context, TD);
1517 // If the allocation is an array of structures, consider transforming this
1518 // into multiple malloc'd arrays, one for each field. This is basically
1519 // SRoA for malloc'd memory.
1521 // If this is an allocation of a fixed size array of structs, analyze as a
1522 // variable size array. malloc [100 x struct],1 -> malloc struct, 100
1523 if (!isArrayMalloc(CI, Context, TD))
1524 if (const ArrayType *AT = dyn_cast<ArrayType>(AllocTy))
1525 AllocTy = AT->getElementType();
1527 if (const StructType *AllocSTy = dyn_cast<StructType>(AllocTy)) {
1528 // This the structure has an unreasonable number of fields, leave it
1530 if (AllocSTy->getNumElements() <= 16 && AllocSTy->getNumElements() != 0 &&
1531 AllGlobalLoadUsesSimpleEnoughForHeapSRA(GV, BCI)) {
1533 // If this is a fixed size array, transform the Malloc to be an alloc of
1534 // structs. malloc [100 x struct],1 -> malloc struct, 100
1535 if (const ArrayType *AT =
1536 dyn_cast<ArrayType>(getMallocAllocatedType(CI))) {
1537 Value* NumElements = ConstantInt::get(Type::getInt32Ty(Context),
1538 AT->getNumElements());
1539 Value* NewMI = CallInst::CreateMalloc(CI, TD->getIntPtrType(Context),
1540 AllocSTy, NumElements,
1542 Value *Cast = new BitCastInst(NewMI, getMallocType(CI), "tmp", CI);
1543 BCI->replaceAllUsesWith(Cast);
1544 BCI->eraseFromParent();
1545 CI->eraseFromParent();
1546 BCI = cast<BitCastInst>(NewMI);
1547 CI = extractMallocCallFromBitCast(NewMI);
1550 GVI = PerformHeapAllocSRoA(GV, CI, BCI, Context, TD);
1559 // OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge
1560 // that only one value (besides its initializer) is ever stored to the global.
1561 static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal,
1562 Module::global_iterator &GVI,
1563 TargetData *TD, LLVMContext &Context) {
1564 // Ignore no-op GEPs and bitcasts.
1565 StoredOnceVal = StoredOnceVal->stripPointerCasts();
1567 // If we are dealing with a pointer global that is initialized to null and
1568 // only has one (non-null) value stored into it, then we can optimize any
1569 // users of the loaded value (often calls and loads) that would trap if the
1571 if (isa<PointerType>(GV->getInitializer()->getType()) &&
1572 GV->getInitializer()->isNullValue()) {
1573 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) {
1574 if (GV->getInitializer()->getType() != SOVC->getType())
1576 ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType());
1578 // Optimize away any trapping uses of the loaded value.
1579 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC, Context))
1581 } else if (CallInst *CI = extractMallocCall(StoredOnceVal)) {
1582 if (getMallocAllocatedType(CI)) {
1583 BitCastInst* BCI = NULL;
1584 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
1586 BCI = dyn_cast<BitCastInst>(cast<Instruction>(*UI++));
1588 TryToOptimizeStoreOfMallocToGlobal(GV, CI, BCI, GVI, TD, Context))
1597 /// TryToShrinkGlobalToBoolean - At this point, we have learned that the only
1598 /// two values ever stored into GV are its initializer and OtherVal. See if we
1599 /// can shrink the global into a boolean and select between the two values
1600 /// whenever it is used. This exposes the values to other scalar optimizations.
1601 static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal,
1602 LLVMContext &Context) {
1603 const Type *GVElType = GV->getType()->getElementType();
1605 // If GVElType is already i1, it is already shrunk. If the type of the GV is
1606 // an FP value, pointer or vector, don't do this optimization because a select
1607 // between them is very expensive and unlikely to lead to later
1608 // simplification. In these cases, we typically end up with "cond ? v1 : v2"
1609 // where v1 and v2 both require constant pool loads, a big loss.
1610 if (GVElType == Type::getInt1Ty(Context) || GVElType->isFloatingPoint() ||
1611 isa<PointerType>(GVElType) || isa<VectorType>(GVElType))
1614 // Walk the use list of the global seeing if all the uses are load or store.
1615 // If there is anything else, bail out.
1616 for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I)
1617 if (!isa<LoadInst>(I) && !isa<StoreInst>(I))
1620 DEBUG(errs() << " *** SHRINKING TO BOOL: " << *GV);
1622 // Create the new global, initializing it to false.
1623 GlobalVariable *NewGV = new GlobalVariable(Context,
1624 Type::getInt1Ty(Context), false,
1625 GlobalValue::InternalLinkage, ConstantInt::getFalse(Context),
1627 GV->isThreadLocal());
1628 GV->getParent()->getGlobalList().insert(GV, NewGV);
1630 Constant *InitVal = GV->getInitializer();
1631 assert(InitVal->getType() != Type::getInt1Ty(Context) &&
1632 "No reason to shrink to bool!");
1634 // If initialized to zero and storing one into the global, we can use a cast
1635 // instead of a select to synthesize the desired value.
1636 bool IsOneZero = false;
1637 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal))
1638 IsOneZero = InitVal->isNullValue() && CI->isOne();
1640 while (!GV->use_empty()) {
1641 Instruction *UI = cast<Instruction>(GV->use_back());
1642 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
1643 // Change the store into a boolean store.
1644 bool StoringOther = SI->getOperand(0) == OtherVal;
1645 // Only do this if we weren't storing a loaded value.
1647 if (StoringOther || SI->getOperand(0) == InitVal)
1648 StoreVal = ConstantInt::get(Type::getInt1Ty(Context), StoringOther);
1650 // Otherwise, we are storing a previously loaded copy. To do this,
1651 // change the copy from copying the original value to just copying the
1653 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0));
1655 // If we're already replaced the input, StoredVal will be a cast or
1656 // select instruction. If not, it will be a load of the original
1658 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
1659 assert(LI->getOperand(0) == GV && "Not a copy!");
1660 // Insert a new load, to preserve the saved value.
1661 StoreVal = new LoadInst(NewGV, LI->getName()+".b", LI);
1663 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) &&
1664 "This is not a form that we understand!");
1665 StoreVal = StoredVal->getOperand(0);
1666 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!");
1669 new StoreInst(StoreVal, NewGV, SI);
1671 // Change the load into a load of bool then a select.
1672 LoadInst *LI = cast<LoadInst>(UI);
1673 LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", LI);
1676 NSI = new ZExtInst(NLI, LI->getType(), "", LI);
1678 NSI = SelectInst::Create(NLI, OtherVal, InitVal, "", LI);
1680 LI->replaceAllUsesWith(NSI);
1682 UI->eraseFromParent();
1685 GV->eraseFromParent();
1690 /// ProcessInternalGlobal - Analyze the specified global variable and optimize
1691 /// it if possible. If we make a change, return true.
1692 bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV,
1693 Module::global_iterator &GVI) {
1694 SmallPtrSet<PHINode*, 16> PHIUsers;
1696 GV->removeDeadConstantUsers();
1698 if (GV->use_empty()) {
1699 DEBUG(errs() << "GLOBAL DEAD: " << *GV);
1700 GV->eraseFromParent();
1705 if (!AnalyzeGlobal(GV, GS, PHIUsers)) {
1707 cerr << "Global: " << *GV;
1708 cerr << " isLoaded = " << GS.isLoaded << "\n";
1709 cerr << " StoredType = ";
1710 switch (GS.StoredType) {
1711 case GlobalStatus::NotStored: cerr << "NEVER STORED\n"; break;
1712 case GlobalStatus::isInitializerStored: cerr << "INIT STORED\n"; break;
1713 case GlobalStatus::isStoredOnce: cerr << "STORED ONCE\n"; break;
1714 case GlobalStatus::isStored: cerr << "stored\n"; break;
1716 if (GS.StoredType == GlobalStatus::isStoredOnce && GS.StoredOnceValue)
1717 cerr << " StoredOnceValue = " << *GS.StoredOnceValue << "\n";
1718 if (GS.AccessingFunction && !GS.HasMultipleAccessingFunctions)
1719 cerr << " AccessingFunction = " << GS.AccessingFunction->getName()
1721 cerr << " HasMultipleAccessingFunctions = "
1722 << GS.HasMultipleAccessingFunctions << "\n";
1723 cerr << " HasNonInstructionUser = " << GS.HasNonInstructionUser<<"\n";
1727 // If this is a first class global and has only one accessing function
1728 // and this function is main (which we know is not recursive we can make
1729 // this global a local variable) we replace the global with a local alloca
1730 // in this function.
1732 // NOTE: It doesn't make sense to promote non single-value types since we
1733 // are just replacing static memory to stack memory.
1735 // If the global is in different address space, don't bring it to stack.
1736 if (!GS.HasMultipleAccessingFunctions &&
1737 GS.AccessingFunction && !GS.HasNonInstructionUser &&
1738 GV->getType()->getElementType()->isSingleValueType() &&
1739 GS.AccessingFunction->getName() == "main" &&
1740 GS.AccessingFunction->hasExternalLinkage() &&
1741 GV->getType()->getAddressSpace() == 0) {
1742 DEBUG(errs() << "LOCALIZING GLOBAL: " << *GV);
1743 Instruction* FirstI = GS.AccessingFunction->getEntryBlock().begin();
1744 const Type* ElemTy = GV->getType()->getElementType();
1745 // FIXME: Pass Global's alignment when globals have alignment
1746 AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), FirstI);
1747 if (!isa<UndefValue>(GV->getInitializer()))
1748 new StoreInst(GV->getInitializer(), Alloca, FirstI);
1750 GV->replaceAllUsesWith(Alloca);
1751 GV->eraseFromParent();
1756 // If the global is never loaded (but may be stored to), it is dead.
1759 DEBUG(errs() << "GLOBAL NEVER LOADED: " << *GV);
1761 // Delete any stores we can find to the global. We may not be able to
1762 // make it completely dead though.
1763 bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer(),
1766 // If the global is dead now, delete it.
1767 if (GV->use_empty()) {
1768 GV->eraseFromParent();
1774 } else if (GS.StoredType <= GlobalStatus::isInitializerStored) {
1775 DEBUG(errs() << "MARKING CONSTANT: " << *GV);
1776 GV->setConstant(true);
1778 // Clean up any obviously simplifiable users now.
1779 CleanupConstantGlobalUsers(GV, GV->getInitializer(), GV->getContext());
1781 // If the global is dead now, just nuke it.
1782 if (GV->use_empty()) {
1783 DEBUG(errs() << " *** Marking constant allowed us to simplify "
1784 << "all users and delete global!\n");
1785 GV->eraseFromParent();
1791 } else if (!GV->getInitializer()->getType()->isSingleValueType()) {
1792 if (TargetData *TD = getAnalysisIfAvailable<TargetData>())
1793 if (GlobalVariable *FirstNewGV = SRAGlobal(GV, *TD,
1794 GV->getContext())) {
1795 GVI = FirstNewGV; // Don't skip the newly produced globals!
1798 } else if (GS.StoredType == GlobalStatus::isStoredOnce) {
1799 // If the initial value for the global was an undef value, and if only
1800 // one other value was stored into it, we can just change the
1801 // initializer to be the stored value, then delete all stores to the
1802 // global. This allows us to mark it constant.
1803 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1804 if (isa<UndefValue>(GV->getInitializer())) {
1805 // Change the initial value here.
1806 GV->setInitializer(SOVConstant);
1808 // Clean up any obviously simplifiable users now.
1809 CleanupConstantGlobalUsers(GV, GV->getInitializer(),
1812 if (GV->use_empty()) {
1813 DEBUG(errs() << " *** Substituting initializer allowed us to "
1814 << "simplify all users and delete global!\n");
1815 GV->eraseFromParent();
1824 // Try to optimize globals based on the knowledge that only one value
1825 // (besides its initializer) is ever stored to the global.
1826 if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI,
1827 getAnalysisIfAvailable<TargetData>(),
1831 // Otherwise, if the global was not a boolean, we can shrink it to be a
1833 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1834 if (TryToShrinkGlobalToBoolean(GV, SOVConstant, GV->getContext())) {
1843 /// ChangeCalleesToFastCall - Walk all of the direct calls of the specified
1844 /// function, changing them to FastCC.
1845 static void ChangeCalleesToFastCall(Function *F) {
1846 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1847 CallSite User(cast<Instruction>(*UI));
1848 User.setCallingConv(CallingConv::Fast);
1852 static AttrListPtr StripNest(const AttrListPtr &Attrs) {
1853 for (unsigned i = 0, e = Attrs.getNumSlots(); i != e; ++i) {
1854 if ((Attrs.getSlot(i).Attrs & Attribute::Nest) == 0)
1857 // There can be only one.
1858 return Attrs.removeAttr(Attrs.getSlot(i).Index, Attribute::Nest);
1864 static void RemoveNestAttribute(Function *F) {
1865 F->setAttributes(StripNest(F->getAttributes()));
1866 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1867 CallSite User(cast<Instruction>(*UI));
1868 User.setAttributes(StripNest(User.getAttributes()));
1872 bool GlobalOpt::OptimizeFunctions(Module &M) {
1873 bool Changed = false;
1874 // Optimize functions.
1875 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) {
1877 // Functions without names cannot be referenced outside this module.
1878 if (!F->hasName() && !F->isDeclaration())
1879 F->setLinkage(GlobalValue::InternalLinkage);
1880 F->removeDeadConstantUsers();
1881 if (F->use_empty() && (F->hasLocalLinkage() ||
1882 F->hasLinkOnceLinkage())) {
1883 M.getFunctionList().erase(F);
1886 } else if (F->hasLocalLinkage()) {
1887 if (F->getCallingConv() == CallingConv::C && !F->isVarArg() &&
1888 !F->hasAddressTaken()) {
1889 // If this function has C calling conventions, is not a varargs
1890 // function, and is only called directly, promote it to use the Fast
1891 // calling convention.
1892 F->setCallingConv(CallingConv::Fast);
1893 ChangeCalleesToFastCall(F);
1898 if (F->getAttributes().hasAttrSomewhere(Attribute::Nest) &&
1899 !F->hasAddressTaken()) {
1900 // The function is not used by a trampoline intrinsic, so it is safe
1901 // to remove the 'nest' attribute.
1902 RemoveNestAttribute(F);
1911 bool GlobalOpt::OptimizeGlobalVars(Module &M) {
1912 bool Changed = false;
1913 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
1915 GlobalVariable *GV = GVI++;
1916 // Global variables without names cannot be referenced outside this module.
1917 if (!GV->hasName() && !GV->isDeclaration())
1918 GV->setLinkage(GlobalValue::InternalLinkage);
1919 if (!GV->isConstant() && GV->hasLocalLinkage() &&
1920 GV->hasInitializer())
1921 Changed |= ProcessInternalGlobal(GV, GVI);
1926 /// FindGlobalCtors - Find the llvm.globalctors list, verifying that all
1927 /// initializers have an init priority of 65535.
1928 GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) {
1929 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
1931 if (I->getName() == "llvm.global_ctors") {
1932 // Found it, verify it's an array of { int, void()* }.
1933 const ArrayType *ATy =dyn_cast<ArrayType>(I->getType()->getElementType());
1935 const StructType *STy = dyn_cast<StructType>(ATy->getElementType());
1936 if (!STy || STy->getNumElements() != 2 ||
1937 STy->getElementType(0) != Type::getInt32Ty(M.getContext())) return 0;
1938 const PointerType *PFTy = dyn_cast<PointerType>(STy->getElementType(1));
1939 if (!PFTy) return 0;
1940 const FunctionType *FTy = dyn_cast<FunctionType>(PFTy->getElementType());
1941 if (!FTy || FTy->getReturnType() != Type::getVoidTy(M.getContext()) ||
1942 FTy->isVarArg() || FTy->getNumParams() != 0)
1945 // Verify that the initializer is simple enough for us to handle.
1946 if (!I->hasDefinitiveInitializer()) return 0;
1947 ConstantArray *CA = dyn_cast<ConstantArray>(I->getInitializer());
1949 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i)
1950 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(*i)) {
1951 if (isa<ConstantPointerNull>(CS->getOperand(1)))
1954 // Must have a function or null ptr.
1955 if (!isa<Function>(CS->getOperand(1)))
1958 // Init priority must be standard.
1959 ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0));
1960 if (!CI || CI->getZExtValue() != 65535)
1971 /// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand,
1972 /// return a list of the functions and null terminator as a vector.
1973 static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) {
1974 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer());
1975 std::vector<Function*> Result;
1976 Result.reserve(CA->getNumOperands());
1977 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) {
1978 ConstantStruct *CS = cast<ConstantStruct>(*i);
1979 Result.push_back(dyn_cast<Function>(CS->getOperand(1)));
1984 /// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the
1985 /// specified array, returning the new global to use.
1986 static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL,
1987 const std::vector<Function*> &Ctors,
1988 LLVMContext &Context) {
1989 // If we made a change, reassemble the initializer list.
1990 std::vector<Constant*> CSVals;
1991 CSVals.push_back(ConstantInt::get(Type::getInt32Ty(Context), 65535));
1992 CSVals.push_back(0);
1994 // Create the new init list.
1995 std::vector<Constant*> CAList;
1996 for (unsigned i = 0, e = Ctors.size(); i != e; ++i) {
1998 CSVals[1] = Ctors[i];
2000 const Type *FTy = FunctionType::get(Type::getVoidTy(Context), false);
2001 const PointerType *PFTy = PointerType::getUnqual(FTy);
2002 CSVals[1] = Constant::getNullValue(PFTy);
2003 CSVals[0] = ConstantInt::get(Type::getInt32Ty(Context), 2147483647);
2005 CAList.push_back(ConstantStruct::get(Context, CSVals, false));
2008 // Create the array initializer.
2009 const Type *StructTy =
2010 cast<ArrayType>(GCL->getType()->getElementType())->getElementType();
2011 Constant *CA = ConstantArray::get(ArrayType::get(StructTy,
2012 CAList.size()), CAList);
2014 // If we didn't change the number of elements, don't create a new GV.
2015 if (CA->getType() == GCL->getInitializer()->getType()) {
2016 GCL->setInitializer(CA);
2020 // Create the new global and insert it next to the existing list.
2021 GlobalVariable *NGV = new GlobalVariable(Context, CA->getType(),
2023 GCL->getLinkage(), CA, "",
2024 GCL->isThreadLocal());
2025 GCL->getParent()->getGlobalList().insert(GCL, NGV);
2028 // Nuke the old list, replacing any uses with the new one.
2029 if (!GCL->use_empty()) {
2031 if (V->getType() != GCL->getType())
2032 V = ConstantExpr::getBitCast(V, GCL->getType());
2033 GCL->replaceAllUsesWith(V);
2035 GCL->eraseFromParent();
2044 static Constant *getVal(DenseMap<Value*, Constant*> &ComputedValues,
2046 if (Constant *CV = dyn_cast<Constant>(V)) return CV;
2047 Constant *R = ComputedValues[V];
2048 assert(R && "Reference to an uncomputed value!");
2052 /// isSimpleEnoughPointerToCommit - Return true if this constant is simple
2053 /// enough for us to understand. In particular, if it is a cast of something,
2054 /// we punt. We basically just support direct accesses to globals and GEP's of
2055 /// globals. This should be kept up to date with CommitValueTo.
2056 static bool isSimpleEnoughPointerToCommit(Constant *C, LLVMContext &Context) {
2057 // Conservatively, avoid aggregate types. This is because we don't
2058 // want to worry about them partially overlapping other stores.
2059 if (!cast<PointerType>(C->getType())->getElementType()->isSingleValueType())
2062 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
2063 // Do not allow weak/linkonce/dllimport/dllexport linkage or
2064 // external globals.
2065 return GV->hasDefinitiveInitializer();
2067 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2068 // Handle a constantexpr gep.
2069 if (CE->getOpcode() == Instruction::GetElementPtr &&
2070 isa<GlobalVariable>(CE->getOperand(0)) &&
2071 cast<GEPOperator>(CE)->isInBounds()) {
2072 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2073 // Do not allow weak/linkonce/dllimport/dllexport linkage or
2074 // external globals.
2075 if (!GV->hasDefinitiveInitializer())
2078 // The first index must be zero.
2079 ConstantInt *CI = dyn_cast<ConstantInt>(*next(CE->op_begin()));
2080 if (!CI || !CI->isZero()) return false;
2082 // The remaining indices must be compile-time known integers within the
2083 // notional bounds of the corresponding static array types.
2084 if (!CE->isGEPWithNoNotionalOverIndexing())
2087 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
2092 /// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global
2093 /// initializer. This returns 'Init' modified to reflect 'Val' stored into it.
2094 /// At this point, the GEP operands of Addr [0, OpNo) have been stepped into.
2095 static Constant *EvaluateStoreInto(Constant *Init, Constant *Val,
2096 ConstantExpr *Addr, unsigned OpNo,
2097 LLVMContext &Context) {
2098 // Base case of the recursion.
2099 if (OpNo == Addr->getNumOperands()) {
2100 assert(Val->getType() == Init->getType() && "Type mismatch!");
2104 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
2105 std::vector<Constant*> Elts;
2107 // Break up the constant into its elements.
2108 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
2109 for (User::op_iterator i = CS->op_begin(), e = CS->op_end(); i != e; ++i)
2110 Elts.push_back(cast<Constant>(*i));
2111 } else if (isa<ConstantAggregateZero>(Init)) {
2112 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
2113 Elts.push_back(Constant::getNullValue(STy->getElementType(i)));
2114 } else if (isa<UndefValue>(Init)) {
2115 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
2116 Elts.push_back(UndefValue::get(STy->getElementType(i)));
2118 llvm_unreachable("This code is out of sync with "
2119 " ConstantFoldLoadThroughGEPConstantExpr");
2122 // Replace the element that we are supposed to.
2123 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo));
2124 unsigned Idx = CU->getZExtValue();
2125 assert(Idx < STy->getNumElements() && "Struct index out of range!");
2126 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1, Context);
2128 // Return the modified struct.
2129 return ConstantStruct::get(Context, &Elts[0], Elts.size(), STy->isPacked());
2131 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo));
2132 const ArrayType *ATy = cast<ArrayType>(Init->getType());
2134 // Break up the array into elements.
2135 std::vector<Constant*> Elts;
2136 if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
2137 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i)
2138 Elts.push_back(cast<Constant>(*i));
2139 } else if (isa<ConstantAggregateZero>(Init)) {
2140 Constant *Elt = Constant::getNullValue(ATy->getElementType());
2141 Elts.assign(ATy->getNumElements(), Elt);
2142 } else if (isa<UndefValue>(Init)) {
2143 Constant *Elt = UndefValue::get(ATy->getElementType());
2144 Elts.assign(ATy->getNumElements(), Elt);
2146 llvm_unreachable("This code is out of sync with "
2147 " ConstantFoldLoadThroughGEPConstantExpr");
2150 assert(CI->getZExtValue() < ATy->getNumElements());
2151 Elts[CI->getZExtValue()] =
2152 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1, Context);
2153 return ConstantArray::get(ATy, Elts);
2157 /// CommitValueTo - We have decided that Addr (which satisfies the predicate
2158 /// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen.
2159 static void CommitValueTo(Constant *Val, Constant *Addr,
2160 LLVMContext &Context) {
2161 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
2162 assert(GV->hasInitializer());
2163 GV->setInitializer(Val);
2167 ConstantExpr *CE = cast<ConstantExpr>(Addr);
2168 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2170 Constant *Init = GV->getInitializer();
2171 Init = EvaluateStoreInto(Init, Val, CE, 2, Context);
2172 GV->setInitializer(Init);
2175 /// ComputeLoadResult - Return the value that would be computed by a load from
2176 /// P after the stores reflected by 'memory' have been performed. If we can't
2177 /// decide, return null.
2178 static Constant *ComputeLoadResult(Constant *P,
2179 const DenseMap<Constant*, Constant*> &Memory,
2180 LLVMContext &Context) {
2181 // If this memory location has been recently stored, use the stored value: it
2182 // is the most up-to-date.
2183 DenseMap<Constant*, Constant*>::const_iterator I = Memory.find(P);
2184 if (I != Memory.end()) return I->second;
2187 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
2188 if (GV->hasDefinitiveInitializer())
2189 return GV->getInitializer();
2193 // Handle a constantexpr getelementptr.
2194 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P))
2195 if (CE->getOpcode() == Instruction::GetElementPtr &&
2196 isa<GlobalVariable>(CE->getOperand(0))) {
2197 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2198 if (GV->hasDefinitiveInitializer())
2199 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
2202 return 0; // don't know how to evaluate.
2205 /// EvaluateFunction - Evaluate a call to function F, returning true if
2206 /// successful, false if we can't evaluate it. ActualArgs contains the formal
2207 /// arguments for the function.
2208 static bool EvaluateFunction(Function *F, Constant *&RetVal,
2209 const SmallVectorImpl<Constant*> &ActualArgs,
2210 std::vector<Function*> &CallStack,
2211 DenseMap<Constant*, Constant*> &MutatedMemory,
2212 std::vector<GlobalVariable*> &AllocaTmps) {
2213 // Check to see if this function is already executing (recursion). If so,
2214 // bail out. TODO: we might want to accept limited recursion.
2215 if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end())
2218 LLVMContext &Context = F->getContext();
2220 CallStack.push_back(F);
2222 /// Values - As we compute SSA register values, we store their contents here.
2223 DenseMap<Value*, Constant*> Values;
2225 // Initialize arguments to the incoming values specified.
2227 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E;
2229 Values[AI] = ActualArgs[ArgNo];
2231 /// ExecutedBlocks - We only handle non-looping, non-recursive code. As such,
2232 /// we can only evaluate any one basic block at most once. This set keeps
2233 /// track of what we have executed so we can detect recursive cases etc.
2234 SmallPtrSet<BasicBlock*, 32> ExecutedBlocks;
2236 // CurInst - The current instruction we're evaluating.
2237 BasicBlock::iterator CurInst = F->begin()->begin();
2239 // This is the main evaluation loop.
2241 Constant *InstResult = 0;
2243 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) {
2244 if (SI->isVolatile()) return false; // no volatile accesses.
2245 Constant *Ptr = getVal(Values, SI->getOperand(1));
2246 if (!isSimpleEnoughPointerToCommit(Ptr, Context))
2247 // If this is too complex for us to commit, reject it.
2249 Constant *Val = getVal(Values, SI->getOperand(0));
2250 MutatedMemory[Ptr] = Val;
2251 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) {
2252 InstResult = ConstantExpr::get(BO->getOpcode(),
2253 getVal(Values, BO->getOperand(0)),
2254 getVal(Values, BO->getOperand(1)));
2255 } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) {
2256 InstResult = ConstantExpr::getCompare(CI->getPredicate(),
2257 getVal(Values, CI->getOperand(0)),
2258 getVal(Values, CI->getOperand(1)));
2259 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) {
2260 InstResult = ConstantExpr::getCast(CI->getOpcode(),
2261 getVal(Values, CI->getOperand(0)),
2263 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) {
2265 ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)),
2266 getVal(Values, SI->getOperand(1)),
2267 getVal(Values, SI->getOperand(2)));
2268 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) {
2269 Constant *P = getVal(Values, GEP->getOperand(0));
2270 SmallVector<Constant*, 8> GEPOps;
2271 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end();
2273 GEPOps.push_back(getVal(Values, *i));
2274 InstResult = cast<GEPOperator>(GEP)->isInBounds() ?
2275 ConstantExpr::getInBoundsGetElementPtr(P, &GEPOps[0], GEPOps.size()) :
2276 ConstantExpr::getGetElementPtr(P, &GEPOps[0], GEPOps.size());
2277 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) {
2278 if (LI->isVolatile()) return false; // no volatile accesses.
2279 InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)),
2280 MutatedMemory, Context);
2281 if (InstResult == 0) return false; // Could not evaluate load.
2282 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) {
2283 if (AI->isArrayAllocation()) return false; // Cannot handle array allocs.
2284 const Type *Ty = AI->getType()->getElementType();
2285 AllocaTmps.push_back(new GlobalVariable(Context, Ty, false,
2286 GlobalValue::InternalLinkage,
2287 UndefValue::get(Ty),
2289 InstResult = AllocaTmps.back();
2290 } else if (CallInst *CI = dyn_cast<CallInst>(CurInst)) {
2292 // Debug info can safely be ignored here.
2293 if (isa<DbgInfoIntrinsic>(CI)) {
2298 // Cannot handle inline asm.
2299 if (isa<InlineAsm>(CI->getOperand(0))) return false;
2301 // Resolve function pointers.
2302 Function *Callee = dyn_cast<Function>(getVal(Values, CI->getOperand(0)));
2303 if (!Callee) return false; // Cannot resolve.
2305 SmallVector<Constant*, 8> Formals;
2306 for (User::op_iterator i = CI->op_begin() + 1, e = CI->op_end();
2308 Formals.push_back(getVal(Values, *i));
2310 if (Callee->isDeclaration()) {
2311 // If this is a function we can constant fold, do it.
2312 if (Constant *C = ConstantFoldCall(Callee, Formals.data(),
2319 if (Callee->getFunctionType()->isVarArg())
2323 // Execute the call, if successful, use the return value.
2324 if (!EvaluateFunction(Callee, RetVal, Formals, CallStack,
2325 MutatedMemory, AllocaTmps))
2327 InstResult = RetVal;
2329 } else if (isa<TerminatorInst>(CurInst)) {
2330 BasicBlock *NewBB = 0;
2331 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) {
2332 if (BI->isUnconditional()) {
2333 NewBB = BI->getSuccessor(0);
2336 dyn_cast<ConstantInt>(getVal(Values, BI->getCondition()));
2337 if (!Cond) return false; // Cannot determine.
2339 NewBB = BI->getSuccessor(!Cond->getZExtValue());
2341 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) {
2343 dyn_cast<ConstantInt>(getVal(Values, SI->getCondition()));
2344 if (!Val) return false; // Cannot determine.
2345 NewBB = SI->getSuccessor(SI->findCaseValue(Val));
2346 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) {
2347 if (RI->getNumOperands())
2348 RetVal = getVal(Values, RI->getOperand(0));
2350 CallStack.pop_back(); // return from fn.
2351 return true; // We succeeded at evaluating this ctor!
2353 // invoke, unwind, unreachable.
2354 return false; // Cannot handle this terminator.
2357 // Okay, we succeeded in evaluating this control flow. See if we have
2358 // executed the new block before. If so, we have a looping function,
2359 // which we cannot evaluate in reasonable time.
2360 if (!ExecutedBlocks.insert(NewBB))
2361 return false; // looped!
2363 // Okay, we have never been in this block before. Check to see if there
2364 // are any PHI nodes. If so, evaluate them with information about where
2366 BasicBlock *OldBB = CurInst->getParent();
2367 CurInst = NewBB->begin();
2369 for (; (PN = dyn_cast<PHINode>(CurInst)); ++CurInst)
2370 Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB));
2372 // Do NOT increment CurInst. We know that the terminator had no value.
2375 // Did not know how to evaluate this!
2379 if (!CurInst->use_empty())
2380 Values[CurInst] = InstResult;
2382 // Advance program counter.
2387 /// EvaluateStaticConstructor - Evaluate static constructors in the function, if
2388 /// we can. Return true if we can, false otherwise.
2389 static bool EvaluateStaticConstructor(Function *F) {
2390 /// MutatedMemory - For each store we execute, we update this map. Loads
2391 /// check this to get the most up-to-date value. If evaluation is successful,
2392 /// this state is committed to the process.
2393 DenseMap<Constant*, Constant*> MutatedMemory;
2395 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable
2396 /// to represent its body. This vector is needed so we can delete the
2397 /// temporary globals when we are done.
2398 std::vector<GlobalVariable*> AllocaTmps;
2400 /// CallStack - This is used to detect recursion. In pathological situations
2401 /// we could hit exponential behavior, but at least there is nothing
2403 std::vector<Function*> CallStack;
2405 // Call the function.
2406 Constant *RetValDummy;
2407 bool EvalSuccess = EvaluateFunction(F, RetValDummy,
2408 SmallVector<Constant*, 0>(), CallStack,
2409 MutatedMemory, AllocaTmps);
2411 // We succeeded at evaluation: commit the result.
2412 DEBUG(errs() << "FULLY EVALUATED GLOBAL CTOR FUNCTION '"
2413 << F->getName() << "' to " << MutatedMemory.size()
2415 for (DenseMap<Constant*, Constant*>::iterator I = MutatedMemory.begin(),
2416 E = MutatedMemory.end(); I != E; ++I)
2417 CommitValueTo(I->second, I->first, F->getContext());
2420 // At this point, we are done interpreting. If we created any 'alloca'
2421 // temporaries, release them now.
2422 while (!AllocaTmps.empty()) {
2423 GlobalVariable *Tmp = AllocaTmps.back();
2424 AllocaTmps.pop_back();
2426 // If there are still users of the alloca, the program is doing something
2427 // silly, e.g. storing the address of the alloca somewhere and using it
2428 // later. Since this is undefined, we'll just make it be null.
2429 if (!Tmp->use_empty())
2430 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType()));
2439 /// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible.
2440 /// Return true if anything changed.
2441 bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) {
2442 std::vector<Function*> Ctors = ParseGlobalCtors(GCL);
2443 bool MadeChange = false;
2444 if (Ctors.empty()) return false;
2446 // Loop over global ctors, optimizing them when we can.
2447 for (unsigned i = 0; i != Ctors.size(); ++i) {
2448 Function *F = Ctors[i];
2449 // Found a null terminator in the middle of the list, prune off the rest of
2452 if (i != Ctors.size()-1) {
2459 // We cannot simplify external ctor functions.
2460 if (F->empty()) continue;
2462 // If we can evaluate the ctor at compile time, do.
2463 if (EvaluateStaticConstructor(F)) {
2464 Ctors.erase(Ctors.begin()+i);
2467 ++NumCtorsEvaluated;
2472 if (!MadeChange) return false;
2474 GCL = InstallGlobalCtors(GCL, Ctors, GCL->getContext());
2478 bool GlobalOpt::OptimizeGlobalAliases(Module &M) {
2479 bool Changed = false;
2481 for (Module::alias_iterator I = M.alias_begin(), E = M.alias_end();
2483 Module::alias_iterator J = I++;
2484 // Aliases without names cannot be referenced outside this module.
2485 if (!J->hasName() && !J->isDeclaration())
2486 J->setLinkage(GlobalValue::InternalLinkage);
2487 // If the aliasee may change at link time, nothing can be done - bail out.
2488 if (J->mayBeOverridden())
2491 Constant *Aliasee = J->getAliasee();
2492 GlobalValue *Target = cast<GlobalValue>(Aliasee->stripPointerCasts());
2493 Target->removeDeadConstantUsers();
2494 bool hasOneUse = Target->hasOneUse() && Aliasee->hasOneUse();
2496 // Make all users of the alias use the aliasee instead.
2497 if (!J->use_empty()) {
2498 J->replaceAllUsesWith(Aliasee);
2499 ++NumAliasesResolved;
2503 // If the aliasee has internal linkage, give it the name and linkage
2504 // of the alias, and delete the alias. This turns:
2505 // define internal ... @f(...)
2506 // @a = alias ... @f
2508 // define ... @a(...)
2509 if (!Target->hasLocalLinkage())
2512 // The transform is only useful if the alias does not have internal linkage.
2513 if (J->hasLocalLinkage())
2516 // Do not perform the transform if multiple aliases potentially target the
2517 // aliasee. This check also ensures that it is safe to replace the section
2518 // and other attributes of the aliasee with those of the alias.
2522 // Give the aliasee the name, linkage and other attributes of the alias.
2523 Target->takeName(J);
2524 Target->setLinkage(J->getLinkage());
2525 Target->GlobalValue::copyAttributesFrom(J);
2527 // Delete the alias.
2528 M.getAliasList().erase(J);
2529 ++NumAliasesRemoved;
2536 bool GlobalOpt::runOnModule(Module &M) {
2537 bool Changed = false;
2539 // Try to find the llvm.globalctors list.
2540 GlobalVariable *GlobalCtors = FindGlobalCtors(M);
2542 bool LocalChange = true;
2543 while (LocalChange) {
2544 LocalChange = false;
2546 // Delete functions that are trivially dead, ccc -> fastcc
2547 LocalChange |= OptimizeFunctions(M);
2549 // Optimize global_ctors list.
2551 LocalChange |= OptimizeGlobalCtorsList(GlobalCtors);
2553 // Optimize non-address-taken globals.
2554 LocalChange |= OptimizeGlobalVars(M);
2556 // Resolve aliases, when possible.
2557 LocalChange |= OptimizeGlobalAliases(M);
2558 Changed |= LocalChange;
2561 // TODO: Move all global ctors functions to the end of the module for code