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/Target/TargetData.h"
28 #include "llvm/Support/CallSite.h"
29 #include "llvm/Support/Compiler.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Support/ErrorHandling.h"
32 #include "llvm/Support/GetElementPtrTypeIterator.h"
33 #include "llvm/Support/MathExtras.h"
34 #include "llvm/Support/raw_ostream.h"
35 #include "llvm/ADT/DenseMap.h"
36 #include "llvm/ADT/SmallPtrSet.h"
37 #include "llvm/ADT/SmallVector.h"
38 #include "llvm/ADT/Statistic.h"
39 #include "llvm/ADT/STLExtras.h"
43 STATISTIC(NumMarked , "Number of globals marked constant");
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");
59 struct VISIBILITY_HIDDEN GlobalOpt : public ModulePass {
60 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
61 AU.addRequired<TargetData>();
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, Context);
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, Context);
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.
430 GEPI != E && (isa<ArrayType>(*GEPI) || isa<VectorType>(*GEPI));
432 uint64_t NumElements;
433 if (const ArrayType *SubArrayTy = dyn_cast<ArrayType>(*GEPI))
434 NumElements = SubArrayTy->getNumElements();
436 NumElements = cast<VectorType>(*GEPI)->getNumElements();
438 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPI.getOperand());
439 if (!IdxVal || IdxVal->getZExtValue() >= NumElements)
444 for (Value::use_iterator I = U->use_begin(), E = U->use_end(); I != E; ++I)
445 if (!isSafeSROAElementUse(*I))
450 /// GlobalUsersSafeToSRA - Look at all uses of the global and decide whether it
451 /// is safe for us to perform this transformation.
453 static bool GlobalUsersSafeToSRA(GlobalValue *GV) {
454 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end();
456 if (!IsUserOfGlobalSafeForSRA(*UI, GV))
463 /// SRAGlobal - Perform scalar replacement of aggregates on the specified global
464 /// variable. This opens the door for other optimizations by exposing the
465 /// behavior of the program in a more fine-grained way. We have determined that
466 /// this transformation is safe already. We return the first global variable we
467 /// insert so that the caller can reprocess it.
468 static GlobalVariable *SRAGlobal(GlobalVariable *GV, const TargetData &TD,
469 LLVMContext &Context) {
470 // Make sure this global only has simple uses that we can SRA.
471 if (!GlobalUsersSafeToSRA(GV))
474 assert(GV->hasLocalLinkage() && !GV->isConstant());
475 Constant *Init = GV->getInitializer();
476 const Type *Ty = Init->getType();
478 std::vector<GlobalVariable*> NewGlobals;
479 Module::GlobalListType &Globals = GV->getParent()->getGlobalList();
481 // Get the alignment of the global, either explicit or target-specific.
482 unsigned StartAlignment = GV->getAlignment();
483 if (StartAlignment == 0)
484 StartAlignment = TD.getABITypeAlignment(GV->getType());
486 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
487 NewGlobals.reserve(STy->getNumElements());
488 const StructLayout &Layout = *TD.getStructLayout(STy);
489 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
490 Constant *In = getAggregateConstantElement(Init,
491 ConstantInt::get(Type::Int32Ty, i),
493 assert(In && "Couldn't get element of initializer?");
494 GlobalVariable *NGV = new GlobalVariable(Context,
495 STy->getElementType(i), false,
496 GlobalVariable::InternalLinkage,
497 In, GV->getName()+"."+Twine(i),
499 GV->getType()->getAddressSpace());
500 Globals.insert(GV, NGV);
501 NewGlobals.push_back(NGV);
503 // Calculate the known alignment of the field. If the original aggregate
504 // had 256 byte alignment for example, something might depend on that:
505 // propagate info to each field.
506 uint64_t FieldOffset = Layout.getElementOffset(i);
507 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, FieldOffset);
508 if (NewAlign > TD.getABITypeAlignment(STy->getElementType(i)))
509 NGV->setAlignment(NewAlign);
511 } else if (const SequentialType *STy = dyn_cast<SequentialType>(Ty)) {
512 unsigned NumElements = 0;
513 if (const ArrayType *ATy = dyn_cast<ArrayType>(STy))
514 NumElements = ATy->getNumElements();
516 NumElements = cast<VectorType>(STy)->getNumElements();
518 if (NumElements > 16 && GV->hasNUsesOrMore(16))
519 return 0; // It's not worth it.
520 NewGlobals.reserve(NumElements);
522 uint64_t EltSize = TD.getTypeAllocSize(STy->getElementType());
523 unsigned EltAlign = TD.getABITypeAlignment(STy->getElementType());
524 for (unsigned i = 0, e = NumElements; i != e; ++i) {
525 Constant *In = getAggregateConstantElement(Init,
526 ConstantInt::get(Type::Int32Ty, i),
528 assert(In && "Couldn't get element of initializer?");
530 GlobalVariable *NGV = new GlobalVariable(Context,
531 STy->getElementType(), false,
532 GlobalVariable::InternalLinkage,
533 In, GV->getName()+"."+Twine(i),
535 GV->getType()->getAddressSpace());
536 Globals.insert(GV, NGV);
537 NewGlobals.push_back(NGV);
539 // Calculate the known alignment of the field. If the original aggregate
540 // had 256 byte alignment for example, something might depend on that:
541 // propagate info to each field.
542 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, EltSize*i);
543 if (NewAlign > EltAlign)
544 NGV->setAlignment(NewAlign);
548 if (NewGlobals.empty())
551 DOUT << "PERFORMING GLOBAL SRA ON: " << *GV;
553 Constant *NullInt = Constant::getNullValue(Type::Int32Ty);
555 // Loop over all of the uses of the global, replacing the constantexpr geps,
556 // with smaller constantexpr geps or direct references.
557 while (!GV->use_empty()) {
558 User *GEP = GV->use_back();
559 assert(((isa<ConstantExpr>(GEP) &&
560 cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)||
561 isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!");
563 // Ignore the 1th operand, which has to be zero or else the program is quite
564 // broken (undefined). Get the 2nd operand, which is the structure or array
566 unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
567 if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access.
569 Value *NewPtr = NewGlobals[Val];
571 // Form a shorter GEP if needed.
572 if (GEP->getNumOperands() > 3) {
573 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) {
574 SmallVector<Constant*, 8> Idxs;
575 Idxs.push_back(NullInt);
576 for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i)
577 Idxs.push_back(CE->getOperand(i));
578 NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr),
579 &Idxs[0], Idxs.size());
581 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP);
582 SmallVector<Value*, 8> Idxs;
583 Idxs.push_back(NullInt);
584 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i)
585 Idxs.push_back(GEPI->getOperand(i));
586 NewPtr = GetElementPtrInst::Create(NewPtr, Idxs.begin(), Idxs.end(),
587 GEPI->getName()+"."+Twine(Val),GEPI);
590 GEP->replaceAllUsesWith(NewPtr);
592 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP))
593 GEPI->eraseFromParent();
595 cast<ConstantExpr>(GEP)->destroyConstant();
598 // Delete the old global, now that it is dead.
602 // Loop over the new globals array deleting any globals that are obviously
603 // dead. This can arise due to scalarization of a structure or an array that
604 // has elements that are dead.
605 unsigned FirstGlobal = 0;
606 for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i)
607 if (NewGlobals[i]->use_empty()) {
608 Globals.erase(NewGlobals[i]);
609 if (FirstGlobal == i) ++FirstGlobal;
612 return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : 0;
615 /// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified
616 /// value will trap if the value is dynamically null. PHIs keeps track of any
617 /// phi nodes we've seen to avoid reprocessing them.
618 static bool AllUsesOfValueWillTrapIfNull(Value *V,
619 SmallPtrSet<PHINode*, 8> &PHIs) {
620 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
621 if (isa<LoadInst>(*UI)) {
623 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
624 if (SI->getOperand(0) == V) {
625 //cerr << "NONTRAPPING USE: " << **UI;
626 return false; // Storing the value.
628 } else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
629 if (CI->getOperand(0) != V) {
630 //cerr << "NONTRAPPING USE: " << **UI;
631 return false; // Not calling the ptr
633 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
634 if (II->getOperand(0) != V) {
635 //cerr << "NONTRAPPING USE: " << **UI;
636 return false; // Not calling the ptr
638 } else if (BitCastInst *CI = dyn_cast<BitCastInst>(*UI)) {
639 if (!AllUsesOfValueWillTrapIfNull(CI, PHIs)) return false;
640 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) {
641 if (!AllUsesOfValueWillTrapIfNull(GEPI, PHIs)) return false;
642 } else if (PHINode *PN = dyn_cast<PHINode>(*UI)) {
643 // If we've already seen this phi node, ignore it, it has already been
646 return AllUsesOfValueWillTrapIfNull(PN, PHIs);
647 } else if (isa<ICmpInst>(*UI) &&
648 isa<ConstantPointerNull>(UI->getOperand(1))) {
649 // Ignore setcc X, null
651 //cerr << "NONTRAPPING USE: " << **UI;
657 /// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads
658 /// from GV will trap if the loaded value is null. Note that this also permits
659 /// comparisons of the loaded value against null, as a special case.
660 static bool AllUsesOfLoadedValueWillTrapIfNull(GlobalVariable *GV) {
661 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI!=E; ++UI)
662 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
663 SmallPtrSet<PHINode*, 8> PHIs;
664 if (!AllUsesOfValueWillTrapIfNull(LI, PHIs))
666 } else if (isa<StoreInst>(*UI)) {
667 // Ignore stores to the global.
669 // We don't know or understand this user, bail out.
670 //cerr << "UNKNOWN USER OF GLOBAL!: " << **UI;
677 static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV,
678 LLVMContext &Context) {
679 bool Changed = false;
680 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) {
681 Instruction *I = cast<Instruction>(*UI++);
682 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
683 LI->setOperand(0, NewV);
685 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
686 if (SI->getOperand(1) == V) {
687 SI->setOperand(1, NewV);
690 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
691 if (I->getOperand(0) == V) {
692 // Calling through the pointer! Turn into a direct call, but be careful
693 // that the pointer is not also being passed as an argument.
694 I->setOperand(0, NewV);
696 bool PassedAsArg = false;
697 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i)
698 if (I->getOperand(i) == V) {
700 I->setOperand(i, NewV);
704 // Being passed as an argument also. Be careful to not invalidate UI!
708 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
709 Changed |= OptimizeAwayTrappingUsesOfValue(CI,
710 ConstantExpr::getCast(CI->getOpcode(),
711 NewV, CI->getType()), Context);
712 if (CI->use_empty()) {
714 CI->eraseFromParent();
716 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
717 // Should handle GEP here.
718 SmallVector<Constant*, 8> Idxs;
719 Idxs.reserve(GEPI->getNumOperands()-1);
720 for (User::op_iterator i = GEPI->op_begin() + 1, e = GEPI->op_end();
722 if (Constant *C = dyn_cast<Constant>(*i))
726 if (Idxs.size() == GEPI->getNumOperands()-1)
727 Changed |= OptimizeAwayTrappingUsesOfValue(GEPI,
728 ConstantExpr::getGetElementPtr(NewV, &Idxs[0],
729 Idxs.size()), Context);
730 if (GEPI->use_empty()) {
732 GEPI->eraseFromParent();
741 /// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null
742 /// value stored into it. If there are uses of the loaded value that would trap
743 /// if the loaded value is dynamically null, then we know that they cannot be
744 /// reachable with a null optimize away the load.
745 static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV,
746 LLVMContext &Context) {
747 bool Changed = false;
749 // Keep track of whether we are able to remove all the uses of the global
750 // other than the store that defines it.
751 bool AllNonStoreUsesGone = true;
753 // Replace all uses of loads with uses of uses of the stored value.
754 for (Value::use_iterator GUI = GV->use_begin(), E = GV->use_end(); GUI != E;){
755 User *GlobalUser = *GUI++;
756 if (LoadInst *LI = dyn_cast<LoadInst>(GlobalUser)) {
757 Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV, Context);
758 // If we were able to delete all uses of the loads
759 if (LI->use_empty()) {
760 LI->eraseFromParent();
763 AllNonStoreUsesGone = false;
765 } else if (isa<StoreInst>(GlobalUser)) {
766 // Ignore the store that stores "LV" to the global.
767 assert(GlobalUser->getOperand(1) == GV &&
768 "Must be storing *to* the global");
770 AllNonStoreUsesGone = false;
772 // If we get here we could have other crazy uses that are transitively
774 assert((isa<PHINode>(GlobalUser) || isa<SelectInst>(GlobalUser) ||
775 isa<ConstantExpr>(GlobalUser)) && "Only expect load and stores!");
780 DOUT << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV;
784 // If we nuked all of the loads, then none of the stores are needed either,
785 // nor is the global.
786 if (AllNonStoreUsesGone) {
787 DOUT << " *** GLOBAL NOW DEAD!\n";
788 CleanupConstantGlobalUsers(GV, 0, Context);
789 if (GV->use_empty()) {
790 GV->eraseFromParent();
798 /// ConstantPropUsersOf - Walk the use list of V, constant folding all of the
799 /// instructions that are foldable.
800 static void ConstantPropUsersOf(Value *V, LLVMContext &Context) {
801 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; )
802 if (Instruction *I = dyn_cast<Instruction>(*UI++))
803 if (Constant *NewC = ConstantFoldInstruction(I, Context)) {
804 I->replaceAllUsesWith(NewC);
806 // Advance UI to the next non-I use to avoid invalidating it!
807 // Instructions could multiply use V.
808 while (UI != E && *UI == I)
810 I->eraseFromParent();
814 /// OptimizeGlobalAddressOfMalloc - This function takes the specified global
815 /// variable, and transforms the program as if it always contained the result of
816 /// the specified malloc. Because it is always the result of the specified
817 /// malloc, there is no reason to actually DO the malloc. Instead, turn the
818 /// malloc into a global, and any loads of GV as uses of the new global.
819 static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV,
821 LLVMContext &Context) {
822 DOUT << "PROMOTING MALLOC GLOBAL: " << *GV << " MALLOC = " << *MI;
823 ConstantInt *NElements = cast<ConstantInt>(MI->getArraySize());
825 if (NElements->getZExtValue() != 1) {
826 // If we have an array allocation, transform it to a single element
827 // allocation to make the code below simpler.
828 Type *NewTy = ArrayType::get(MI->getAllocatedType(),
829 NElements->getZExtValue());
831 new MallocInst(NewTy, Constant::getNullValue(Type::Int32Ty),
832 MI->getAlignment(), MI->getName(), MI);
834 Indices[0] = Indices[1] = Constant::getNullValue(Type::Int32Ty);
835 Value *NewGEP = GetElementPtrInst::Create(NewMI, Indices, Indices + 2,
836 NewMI->getName()+".el0", MI);
837 MI->replaceAllUsesWith(NewGEP);
838 MI->eraseFromParent();
842 // Create the new global variable. The contents of the malloc'd memory is
843 // undefined, so initialize with an undef value.
844 // FIXME: This new global should have the alignment returned by malloc. Code
845 // could depend on malloc returning large alignment (on the mac, 16 bytes) but
846 // this would only guarantee some lower alignment.
847 Constant *Init = UndefValue::get(MI->getAllocatedType());
848 GlobalVariable *NewGV = new GlobalVariable(*GV->getParent(),
849 MI->getAllocatedType(), false,
850 GlobalValue::InternalLinkage, Init,
851 GV->getName()+".body",
853 GV->isThreadLocal());
855 // Anything that used the malloc now uses the global directly.
856 MI->replaceAllUsesWith(NewGV);
858 Constant *RepValue = NewGV;
859 if (NewGV->getType() != GV->getType()->getElementType())
860 RepValue = ConstantExpr::getBitCast(RepValue,
861 GV->getType()->getElementType());
863 // If there is a comparison against null, we will insert a global bool to
864 // keep track of whether the global was initialized yet or not.
865 GlobalVariable *InitBool =
866 new GlobalVariable(Context, Type::Int1Ty, false,
867 GlobalValue::InternalLinkage,
868 ConstantInt::getFalse(Context), GV->getName()+".init",
869 GV->isThreadLocal());
870 bool InitBoolUsed = false;
872 // Loop over all uses of GV, processing them in turn.
873 std::vector<StoreInst*> Stores;
874 while (!GV->use_empty())
875 if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) {
876 while (!LI->use_empty()) {
877 Use &LoadUse = LI->use_begin().getUse();
878 if (!isa<ICmpInst>(LoadUse.getUser()))
881 ICmpInst *CI = cast<ICmpInst>(LoadUse.getUser());
882 // Replace the cmp X, 0 with a use of the bool value.
883 Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", CI);
885 switch (CI->getPredicate()) {
886 default: llvm_unreachable("Unknown ICmp Predicate!");
887 case ICmpInst::ICMP_ULT:
888 case ICmpInst::ICMP_SLT:
889 LV = ConstantInt::getFalse(Context); // X < null -> always false
891 case ICmpInst::ICMP_ULE:
892 case ICmpInst::ICMP_SLE:
893 case ICmpInst::ICMP_EQ:
894 LV = BinaryOperator::CreateNot(LV, "notinit", CI);
896 case ICmpInst::ICMP_NE:
897 case ICmpInst::ICMP_UGE:
898 case ICmpInst::ICMP_SGE:
899 case ICmpInst::ICMP_UGT:
900 case ICmpInst::ICMP_SGT:
903 CI->replaceAllUsesWith(LV);
904 CI->eraseFromParent();
907 LI->eraseFromParent();
909 StoreInst *SI = cast<StoreInst>(GV->use_back());
910 // The global is initialized when the store to it occurs.
911 new StoreInst(ConstantInt::getTrue(Context), InitBool, SI);
912 SI->eraseFromParent();
915 // If the initialization boolean was used, insert it, otherwise delete it.
917 while (!InitBool->use_empty()) // Delete initializations
918 cast<Instruction>(InitBool->use_back())->eraseFromParent();
921 GV->getParent()->getGlobalList().insert(GV, InitBool);
924 // Now the GV is dead, nuke it and the malloc.
925 GV->eraseFromParent();
926 MI->eraseFromParent();
928 // To further other optimizations, loop over all users of NewGV and try to
929 // constant prop them. This will promote GEP instructions with constant
930 // indices into GEP constant-exprs, which will allow global-opt to hack on it.
931 ConstantPropUsersOf(NewGV, Context);
932 if (RepValue != NewGV)
933 ConstantPropUsersOf(RepValue, Context);
938 /// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking
939 /// to make sure that there are no complex uses of V. We permit simple things
940 /// like dereferencing the pointer, but not storing through the address, unless
941 /// it is to the specified global.
942 static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Instruction *V,
944 SmallPtrSet<PHINode*, 8> &PHIs) {
945 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
946 Instruction *Inst = cast<Instruction>(*UI);
948 if (isa<LoadInst>(Inst) || isa<CmpInst>(Inst)) {
949 continue; // Fine, ignore.
952 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
953 if (SI->getOperand(0) == V && SI->getOperand(1) != GV)
954 return false; // Storing the pointer itself... bad.
955 continue; // Otherwise, storing through it, or storing into GV... fine.
958 if (isa<GetElementPtrInst>(Inst)) {
959 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Inst, GV, PHIs))
964 if (PHINode *PN = dyn_cast<PHINode>(Inst)) {
965 // PHIs are ok if all uses are ok. Don't infinitely recurse through PHI
968 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(PN, GV, PHIs))
973 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Inst)) {
974 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(BCI, GV, PHIs))
984 /// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV
985 /// somewhere. Transform all uses of the allocation into loads from the
986 /// global and uses of the resultant pointer. Further, delete the store into
987 /// GV. This assumes that these value pass the
988 /// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate.
989 static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc,
990 GlobalVariable *GV) {
991 while (!Alloc->use_empty()) {
992 Instruction *U = cast<Instruction>(*Alloc->use_begin());
993 Instruction *InsertPt = U;
994 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
995 // If this is the store of the allocation into the global, remove it.
996 if (SI->getOperand(1) == GV) {
997 SI->eraseFromParent();
1000 } else if (PHINode *PN = dyn_cast<PHINode>(U)) {
1001 // Insert the load in the corresponding predecessor, not right before the
1003 InsertPt = PN->getIncomingBlock(Alloc->use_begin())->getTerminator();
1004 } else if (isa<BitCastInst>(U)) {
1005 // Must be bitcast between the malloc and store to initialize the global.
1006 ReplaceUsesOfMallocWithGlobal(U, GV);
1007 U->eraseFromParent();
1009 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
1010 // If this is a "GEP bitcast" and the user is a store to the global, then
1011 // just process it as a bitcast.
1012 if (GEPI->hasAllZeroIndices() && GEPI->hasOneUse())
1013 if (StoreInst *SI = dyn_cast<StoreInst>(GEPI->use_back()))
1014 if (SI->getOperand(1) == GV) {
1015 // Must be bitcast GEP between the malloc and store to initialize
1017 ReplaceUsesOfMallocWithGlobal(GEPI, GV);
1018 GEPI->eraseFromParent();
1023 // Insert a load from the global, and use it instead of the malloc.
1024 Value *NL = new LoadInst(GV, GV->getName()+".val", InsertPt);
1025 U->replaceUsesOfWith(Alloc, NL);
1029 /// LoadUsesSimpleEnoughForHeapSRA - Verify that all uses of V (a load, or a phi
1030 /// of a load) are simple enough to perform heap SRA on. This permits GEP's
1031 /// that index through the array and struct field, icmps of null, and PHIs.
1032 static bool LoadUsesSimpleEnoughForHeapSRA(Value *V,
1033 SmallPtrSet<PHINode*, 32> &LoadUsingPHIs,
1034 SmallPtrSet<PHINode*, 32> &LoadUsingPHIsPerLoad) {
1035 // We permit two users of the load: setcc comparing against the null
1036 // pointer, and a getelementptr of a specific form.
1037 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
1038 Instruction *User = cast<Instruction>(*UI);
1040 // Comparison against null is ok.
1041 if (ICmpInst *ICI = dyn_cast<ICmpInst>(User)) {
1042 if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
1047 // getelementptr is also ok, but only a simple form.
1048 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1049 // Must index into the array and into the struct.
1050 if (GEPI->getNumOperands() < 3)
1053 // Otherwise the GEP is ok.
1057 if (PHINode *PN = dyn_cast<PHINode>(User)) {
1058 if (!LoadUsingPHIsPerLoad.insert(PN))
1059 // This means some phi nodes are dependent on each other.
1060 // Avoid infinite looping!
1062 if (!LoadUsingPHIs.insert(PN))
1063 // If we have already analyzed this PHI, then it is safe.
1066 // Make sure all uses of the PHI are simple enough to transform.
1067 if (!LoadUsesSimpleEnoughForHeapSRA(PN,
1068 LoadUsingPHIs, LoadUsingPHIsPerLoad))
1074 // Otherwise we don't know what this is, not ok.
1082 /// AllGlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from
1083 /// GV are simple enough to perform HeapSRA, return true.
1084 static bool AllGlobalLoadUsesSimpleEnoughForHeapSRA(GlobalVariable *GV,
1086 SmallPtrSet<PHINode*, 32> LoadUsingPHIs;
1087 SmallPtrSet<PHINode*, 32> LoadUsingPHIsPerLoad;
1088 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;
1090 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
1091 if (!LoadUsesSimpleEnoughForHeapSRA(LI, LoadUsingPHIs,
1092 LoadUsingPHIsPerLoad))
1094 LoadUsingPHIsPerLoad.clear();
1097 // If we reach here, we know that all uses of the loads and transitive uses
1098 // (through PHI nodes) are simple enough to transform. However, we don't know
1099 // that all inputs the to the PHI nodes are in the same equivalence sets.
1100 // Check to verify that all operands of the PHIs are either PHIS that can be
1101 // transformed, loads from GV, or MI itself.
1102 for (SmallPtrSet<PHINode*, 32>::iterator I = LoadUsingPHIs.begin(),
1103 E = LoadUsingPHIs.end(); I != E; ++I) {
1105 for (unsigned op = 0, e = PN->getNumIncomingValues(); op != e; ++op) {
1106 Value *InVal = PN->getIncomingValue(op);
1108 // PHI of the stored value itself is ok.
1109 if (InVal == MI) continue;
1111 if (PHINode *InPN = dyn_cast<PHINode>(InVal)) {
1112 // One of the PHIs in our set is (optimistically) ok.
1113 if (LoadUsingPHIs.count(InPN))
1118 // Load from GV is ok.
1119 if (LoadInst *LI = dyn_cast<LoadInst>(InVal))
1120 if (LI->getOperand(0) == GV)
1125 // Anything else is rejected.
1133 static Value *GetHeapSROAValue(Value *V, unsigned FieldNo,
1134 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1135 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite,
1136 LLVMContext &Context) {
1137 std::vector<Value*> &FieldVals = InsertedScalarizedValues[V];
1139 if (FieldNo >= FieldVals.size())
1140 FieldVals.resize(FieldNo+1);
1142 // If we already have this value, just reuse the previously scalarized
1144 if (Value *FieldVal = FieldVals[FieldNo])
1147 // Depending on what instruction this is, we have several cases.
1149 if (LoadInst *LI = dyn_cast<LoadInst>(V)) {
1150 // This is a scalarized version of the load from the global. Just create
1151 // a new Load of the scalarized global.
1152 Result = new LoadInst(GetHeapSROAValue(LI->getOperand(0), FieldNo,
1153 InsertedScalarizedValues,
1154 PHIsToRewrite, Context),
1155 LI->getName()+".f"+Twine(FieldNo), LI);
1156 } else if (PHINode *PN = dyn_cast<PHINode>(V)) {
1157 // PN's type is pointer to struct. Make a new PHI of pointer to struct
1159 const StructType *ST =
1160 cast<StructType>(cast<PointerType>(PN->getType())->getElementType());
1163 PHINode::Create(PointerType::getUnqual(ST->getElementType(FieldNo)),
1164 PN->getName()+".f"+Twine(FieldNo), PN);
1165 PHIsToRewrite.push_back(std::make_pair(PN, FieldNo));
1167 llvm_unreachable("Unknown usable value");
1171 return FieldVals[FieldNo] = Result;
1174 /// RewriteHeapSROALoadUser - Given a load instruction and a value derived from
1175 /// the load, rewrite the derived value to use the HeapSRoA'd load.
1176 static void RewriteHeapSROALoadUser(Instruction *LoadUser,
1177 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1178 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite,
1179 LLVMContext &Context) {
1180 // If this is a comparison against null, handle it.
1181 if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) {
1182 assert(isa<ConstantPointerNull>(SCI->getOperand(1)));
1183 // If we have a setcc of the loaded pointer, we can use a setcc of any
1185 Value *NPtr = GetHeapSROAValue(SCI->getOperand(0), 0,
1186 InsertedScalarizedValues, PHIsToRewrite,
1189 Value *New = new ICmpInst(SCI, SCI->getPredicate(), NPtr,
1190 Constant::getNullValue(NPtr->getType()),
1192 SCI->replaceAllUsesWith(New);
1193 SCI->eraseFromParent();
1197 // Handle 'getelementptr Ptr, Idx, i32 FieldNo ...'
1198 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) {
1199 assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2))
1200 && "Unexpected GEPI!");
1202 // Load the pointer for this field.
1203 unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
1204 Value *NewPtr = GetHeapSROAValue(GEPI->getOperand(0), FieldNo,
1205 InsertedScalarizedValues, PHIsToRewrite,
1208 // Create the new GEP idx vector.
1209 SmallVector<Value*, 8> GEPIdx;
1210 GEPIdx.push_back(GEPI->getOperand(1));
1211 GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end());
1213 Value *NGEPI = GetElementPtrInst::Create(NewPtr,
1214 GEPIdx.begin(), GEPIdx.end(),
1215 GEPI->getName(), GEPI);
1216 GEPI->replaceAllUsesWith(NGEPI);
1217 GEPI->eraseFromParent();
1221 // Recursively transform the users of PHI nodes. This will lazily create the
1222 // PHIs that are needed for individual elements. Keep track of what PHIs we
1223 // see in InsertedScalarizedValues so that we don't get infinite loops (very
1224 // antisocial). If the PHI is already in InsertedScalarizedValues, it has
1225 // already been seen first by another load, so its uses have already been
1227 PHINode *PN = cast<PHINode>(LoadUser);
1229 DenseMap<Value*, std::vector<Value*> >::iterator InsertPos;
1230 tie(InsertPos, Inserted) =
1231 InsertedScalarizedValues.insert(std::make_pair(PN, std::vector<Value*>()));
1232 if (!Inserted) return;
1234 // If this is the first time we've seen this PHI, recursively process all
1236 for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end(); UI != E; ) {
1237 Instruction *User = cast<Instruction>(*UI++);
1238 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite,
1243 /// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global. Ptr
1244 /// is a value loaded from the global. Eliminate all uses of Ptr, making them
1245 /// use FieldGlobals instead. All uses of loaded values satisfy
1246 /// AllGlobalLoadUsesSimpleEnoughForHeapSRA.
1247 static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load,
1248 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1249 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite,
1250 LLVMContext &Context) {
1251 for (Value::use_iterator UI = Load->use_begin(), E = Load->use_end();
1253 Instruction *User = cast<Instruction>(*UI++);
1254 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite,
1258 if (Load->use_empty()) {
1259 Load->eraseFromParent();
1260 InsertedScalarizedValues.erase(Load);
1264 /// PerformHeapAllocSRoA - MI is an allocation of an array of structures. Break
1265 /// it up into multiple allocations of arrays of the fields.
1266 static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, MallocInst *MI,
1267 LLVMContext &Context){
1268 DOUT << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *MI;
1269 const StructType *STy = cast<StructType>(MI->getAllocatedType());
1271 // There is guaranteed to be at least one use of the malloc (storing
1272 // it into GV). If there are other uses, change them to be uses of
1273 // the global to simplify later code. This also deletes the store
1275 ReplaceUsesOfMallocWithGlobal(MI, GV);
1277 // Okay, at this point, there are no users of the malloc. Insert N
1278 // new mallocs at the same place as MI, and N globals.
1279 std::vector<Value*> FieldGlobals;
1280 std::vector<MallocInst*> FieldMallocs;
1282 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){
1283 const Type *FieldTy = STy->getElementType(FieldNo);
1284 const Type *PFieldTy = PointerType::getUnqual(FieldTy);
1286 GlobalVariable *NGV =
1287 new GlobalVariable(*GV->getParent(),
1288 PFieldTy, false, GlobalValue::InternalLinkage,
1289 Constant::getNullValue(PFieldTy),
1290 GV->getName() + ".f" + Twine(FieldNo), GV,
1291 GV->isThreadLocal());
1292 FieldGlobals.push_back(NGV);
1294 MallocInst *NMI = new MallocInst(FieldTy, MI->getArraySize(),
1295 MI->getName() + ".f" + Twine(FieldNo), MI);
1296 FieldMallocs.push_back(NMI);
1297 new StoreInst(NMI, NGV, MI);
1300 // The tricky aspect of this transformation is handling the case when malloc
1301 // fails. In the original code, malloc failing would set the result pointer
1302 // of malloc to null. In this case, some mallocs could succeed and others
1303 // could fail. As such, we emit code that looks like this:
1304 // F0 = malloc(field0)
1305 // F1 = malloc(field1)
1306 // F2 = malloc(field2)
1307 // if (F0 == 0 || F1 == 0 || F2 == 0) {
1308 // if (F0) { free(F0); F0 = 0; }
1309 // if (F1) { free(F1); F1 = 0; }
1310 // if (F2) { free(F2); F2 = 0; }
1312 Value *RunningOr = 0;
1313 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) {
1314 Value *Cond = new ICmpInst(MI, ICmpInst::ICMP_EQ, FieldMallocs[i],
1315 Constant::getNullValue(FieldMallocs[i]->getType()),
1318 RunningOr = Cond; // First seteq
1320 RunningOr = BinaryOperator::CreateOr(RunningOr, Cond, "tmp", MI);
1323 // Split the basic block at the old malloc.
1324 BasicBlock *OrigBB = MI->getParent();
1325 BasicBlock *ContBB = OrigBB->splitBasicBlock(MI, "malloc_cont");
1327 // Create the block to check the first condition. Put all these blocks at the
1328 // end of the function as they are unlikely to be executed.
1329 BasicBlock *NullPtrBlock = BasicBlock::Create("malloc_ret_null",
1330 OrigBB->getParent());
1332 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond
1333 // branch on RunningOr.
1334 OrigBB->getTerminator()->eraseFromParent();
1335 BranchInst::Create(NullPtrBlock, ContBB, RunningOr, OrigBB);
1337 // Within the NullPtrBlock, we need to emit a comparison and branch for each
1338 // pointer, because some may be null while others are not.
1339 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1340 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock);
1341 Value *Cmp = new ICmpInst(*NullPtrBlock, ICmpInst::ICMP_NE, GVVal,
1342 Constant::getNullValue(GVVal->getType()),
1344 BasicBlock *FreeBlock = BasicBlock::Create("free_it", OrigBB->getParent());
1345 BasicBlock *NextBlock = BasicBlock::Create("next", OrigBB->getParent());
1346 BranchInst::Create(FreeBlock, NextBlock, Cmp, NullPtrBlock);
1348 // Fill in FreeBlock.
1349 new FreeInst(GVVal, FreeBlock);
1350 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i],
1352 BranchInst::Create(NextBlock, FreeBlock);
1354 NullPtrBlock = NextBlock;
1357 BranchInst::Create(ContBB, NullPtrBlock);
1359 // MI is no longer needed, remove it.
1360 MI->eraseFromParent();
1362 /// InsertedScalarizedLoads - As we process loads, if we can't immediately
1363 /// update all uses of the load, keep track of what scalarized loads are
1364 /// inserted for a given load.
1365 DenseMap<Value*, std::vector<Value*> > InsertedScalarizedValues;
1366 InsertedScalarizedValues[GV] = FieldGlobals;
1368 std::vector<std::pair<PHINode*, unsigned> > PHIsToRewrite;
1370 // Okay, the malloc site is completely handled. All of the uses of GV are now
1371 // loads, and all uses of those loads are simple. Rewrite them to use loads
1372 // of the per-field globals instead.
1373 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;) {
1374 Instruction *User = cast<Instruction>(*UI++);
1376 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1377 RewriteUsesOfLoadForHeapSRoA(LI, InsertedScalarizedValues, PHIsToRewrite,
1382 // Must be a store of null.
1383 StoreInst *SI = cast<StoreInst>(User);
1384 assert(isa<ConstantPointerNull>(SI->getOperand(0)) &&
1385 "Unexpected heap-sra user!");
1387 // Insert a store of null into each global.
1388 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1389 const PointerType *PT = cast<PointerType>(FieldGlobals[i]->getType());
1390 Constant *Null = Constant::getNullValue(PT->getElementType());
1391 new StoreInst(Null, FieldGlobals[i], SI);
1393 // Erase the original store.
1394 SI->eraseFromParent();
1397 // While we have PHIs that are interesting to rewrite, do it.
1398 while (!PHIsToRewrite.empty()) {
1399 PHINode *PN = PHIsToRewrite.back().first;
1400 unsigned FieldNo = PHIsToRewrite.back().second;
1401 PHIsToRewrite.pop_back();
1402 PHINode *FieldPN = cast<PHINode>(InsertedScalarizedValues[PN][FieldNo]);
1403 assert(FieldPN->getNumIncomingValues() == 0 &&"Already processed this phi");
1405 // Add all the incoming values. This can materialize more phis.
1406 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1407 Value *InVal = PN->getIncomingValue(i);
1408 InVal = GetHeapSROAValue(InVal, FieldNo, InsertedScalarizedValues,
1409 PHIsToRewrite, Context);
1410 FieldPN->addIncoming(InVal, PN->getIncomingBlock(i));
1414 // Drop all inter-phi links and any loads that made it this far.
1415 for (DenseMap<Value*, std::vector<Value*> >::iterator
1416 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1418 if (PHINode *PN = dyn_cast<PHINode>(I->first))
1419 PN->dropAllReferences();
1420 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1421 LI->dropAllReferences();
1424 // Delete all the phis and loads now that inter-references are dead.
1425 for (DenseMap<Value*, std::vector<Value*> >::iterator
1426 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1428 if (PHINode *PN = dyn_cast<PHINode>(I->first))
1429 PN->eraseFromParent();
1430 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1431 LI->eraseFromParent();
1434 // The old global is now dead, remove it.
1435 GV->eraseFromParent();
1438 return cast<GlobalVariable>(FieldGlobals[0]);
1441 /// TryToOptimizeStoreOfMallocToGlobal - This function is called when we see a
1442 /// pointer global variable with a single value stored it that is a malloc or
1444 static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV,
1446 Module::global_iterator &GVI,
1448 LLVMContext &Context) {
1449 // If this is a malloc of an abstract type, don't touch it.
1450 if (!MI->getAllocatedType()->isSized())
1453 // We can't optimize this global unless all uses of it are *known* to be
1454 // of the malloc value, not of the null initializer value (consider a use
1455 // that compares the global's value against zero to see if the malloc has
1456 // been reached). To do this, we check to see if all uses of the global
1457 // would trap if the global were null: this proves that they must all
1458 // happen after the malloc.
1459 if (!AllUsesOfLoadedValueWillTrapIfNull(GV))
1462 // We can't optimize this if the malloc itself is used in a complex way,
1463 // for example, being stored into multiple globals. This allows the
1464 // malloc to be stored into the specified global, loaded setcc'd, and
1465 // GEP'd. These are all things we could transform to using the global
1468 SmallPtrSet<PHINode*, 8> PHIs;
1469 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(MI, GV, PHIs))
1474 // If we have a global that is only initialized with a fixed size malloc,
1475 // transform the program to use global memory instead of malloc'd memory.
1476 // This eliminates dynamic allocation, avoids an indirection accessing the
1477 // data, and exposes the resultant global to further GlobalOpt.
1478 if (ConstantInt *NElements = dyn_cast<ConstantInt>(MI->getArraySize())) {
1479 // Restrict this transformation to only working on small allocations
1480 // (2048 bytes currently), as we don't want to introduce a 16M global or
1482 if (NElements->getZExtValue()*
1483 TD.getTypeAllocSize(MI->getAllocatedType()) < 2048) {
1484 GVI = OptimizeGlobalAddressOfMalloc(GV, MI, Context);
1489 // If the allocation is an array of structures, consider transforming this
1490 // into multiple malloc'd arrays, one for each field. This is basically
1491 // SRoA for malloc'd memory.
1492 const Type *AllocTy = MI->getAllocatedType();
1494 // If this is an allocation of a fixed size array of structs, analyze as a
1495 // variable size array. malloc [100 x struct],1 -> malloc struct, 100
1496 if (!MI->isArrayAllocation())
1497 if (const ArrayType *AT = dyn_cast<ArrayType>(AllocTy))
1498 AllocTy = AT->getElementType();
1500 if (const StructType *AllocSTy = dyn_cast<StructType>(AllocTy)) {
1501 // This the structure has an unreasonable number of fields, leave it
1503 if (AllocSTy->getNumElements() <= 16 && AllocSTy->getNumElements() != 0 &&
1504 AllGlobalLoadUsesSimpleEnoughForHeapSRA(GV, MI)) {
1506 // If this is a fixed size array, transform the Malloc to be an alloc of
1507 // structs. malloc [100 x struct],1 -> malloc struct, 100
1508 if (const ArrayType *AT = dyn_cast<ArrayType>(MI->getAllocatedType())) {
1510 new MallocInst(AllocSTy,
1511 ConstantInt::get(Type::Int32Ty, AT->getNumElements()),
1513 NewMI->takeName(MI);
1514 Value *Cast = new BitCastInst(NewMI, MI->getType(), "tmp", MI);
1515 MI->replaceAllUsesWith(Cast);
1516 MI->eraseFromParent();
1520 GVI = PerformHeapAllocSRoA(GV, MI, Context);
1528 // OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge
1529 // that only one value (besides its initializer) is ever stored to the global.
1530 static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal,
1531 Module::global_iterator &GVI,
1532 TargetData &TD, LLVMContext &Context) {
1533 // Ignore no-op GEPs and bitcasts.
1534 StoredOnceVal = StoredOnceVal->stripPointerCasts();
1536 // If we are dealing with a pointer global that is initialized to null and
1537 // only has one (non-null) value stored into it, then we can optimize any
1538 // users of the loaded value (often calls and loads) that would trap if the
1540 if (isa<PointerType>(GV->getInitializer()->getType()) &&
1541 GV->getInitializer()->isNullValue()) {
1542 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) {
1543 if (GV->getInitializer()->getType() != SOVC->getType())
1545 ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType());
1547 // Optimize away any trapping uses of the loaded value.
1548 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC, Context))
1550 } else if (MallocInst *MI = dyn_cast<MallocInst>(StoredOnceVal)) {
1551 if (TryToOptimizeStoreOfMallocToGlobal(GV, MI, GVI, TD, Context))
1559 /// TryToShrinkGlobalToBoolean - At this point, we have learned that the only
1560 /// two values ever stored into GV are its initializer and OtherVal. See if we
1561 /// can shrink the global into a boolean and select between the two values
1562 /// whenever it is used. This exposes the values to other scalar optimizations.
1563 static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal,
1564 LLVMContext &Context) {
1565 const Type *GVElType = GV->getType()->getElementType();
1567 // If GVElType is already i1, it is already shrunk. If the type of the GV is
1568 // an FP value, pointer or vector, don't do this optimization because a select
1569 // between them is very expensive and unlikely to lead to later
1570 // simplification. In these cases, we typically end up with "cond ? v1 : v2"
1571 // where v1 and v2 both require constant pool loads, a big loss.
1572 if (GVElType == Type::Int1Ty || GVElType->isFloatingPoint() ||
1573 isa<PointerType>(GVElType) || isa<VectorType>(GVElType))
1576 // Walk the use list of the global seeing if all the uses are load or store.
1577 // If there is anything else, bail out.
1578 for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I)
1579 if (!isa<LoadInst>(I) && !isa<StoreInst>(I))
1582 DOUT << " *** SHRINKING TO BOOL: " << *GV;
1584 // Create the new global, initializing it to false.
1585 GlobalVariable *NewGV = new GlobalVariable(Context, Type::Int1Ty, false,
1586 GlobalValue::InternalLinkage, ConstantInt::getFalse(Context),
1588 GV->isThreadLocal());
1589 GV->getParent()->getGlobalList().insert(GV, NewGV);
1591 Constant *InitVal = GV->getInitializer();
1592 assert(InitVal->getType() != Type::Int1Ty && "No reason to shrink to bool!");
1594 // If initialized to zero and storing one into the global, we can use a cast
1595 // instead of a select to synthesize the desired value.
1596 bool IsOneZero = false;
1597 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal))
1598 IsOneZero = InitVal->isNullValue() && CI->isOne();
1600 while (!GV->use_empty()) {
1601 Instruction *UI = cast<Instruction>(GV->use_back());
1602 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
1603 // Change the store into a boolean store.
1604 bool StoringOther = SI->getOperand(0) == OtherVal;
1605 // Only do this if we weren't storing a loaded value.
1607 if (StoringOther || SI->getOperand(0) == InitVal)
1608 StoreVal = ConstantInt::get(Type::Int1Ty, StoringOther);
1610 // Otherwise, we are storing a previously loaded copy. To do this,
1611 // change the copy from copying the original value to just copying the
1613 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0));
1615 // If we're already replaced the input, StoredVal will be a cast or
1616 // select instruction. If not, it will be a load of the original
1618 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
1619 assert(LI->getOperand(0) == GV && "Not a copy!");
1620 // Insert a new load, to preserve the saved value.
1621 StoreVal = new LoadInst(NewGV, LI->getName()+".b", LI);
1623 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) &&
1624 "This is not a form that we understand!");
1625 StoreVal = StoredVal->getOperand(0);
1626 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!");
1629 new StoreInst(StoreVal, NewGV, SI);
1631 // Change the load into a load of bool then a select.
1632 LoadInst *LI = cast<LoadInst>(UI);
1633 LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", LI);
1636 NSI = new ZExtInst(NLI, LI->getType(), "", LI);
1638 NSI = SelectInst::Create(NLI, OtherVal, InitVal, "", LI);
1640 LI->replaceAllUsesWith(NSI);
1642 UI->eraseFromParent();
1645 GV->eraseFromParent();
1650 /// ProcessInternalGlobal - Analyze the specified global variable and optimize
1651 /// it if possible. If we make a change, return true.
1652 bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV,
1653 Module::global_iterator &GVI) {
1654 SmallPtrSet<PHINode*, 16> PHIUsers;
1656 GV->removeDeadConstantUsers();
1658 if (GV->use_empty()) {
1659 DOUT << "GLOBAL DEAD: " << *GV;
1660 GV->eraseFromParent();
1665 if (!AnalyzeGlobal(GV, GS, PHIUsers)) {
1667 cerr << "Global: " << *GV;
1668 cerr << " isLoaded = " << GS.isLoaded << "\n";
1669 cerr << " StoredType = ";
1670 switch (GS.StoredType) {
1671 case GlobalStatus::NotStored: cerr << "NEVER STORED\n"; break;
1672 case GlobalStatus::isInitializerStored: cerr << "INIT STORED\n"; break;
1673 case GlobalStatus::isStoredOnce: cerr << "STORED ONCE\n"; break;
1674 case GlobalStatus::isStored: cerr << "stored\n"; break;
1676 if (GS.StoredType == GlobalStatus::isStoredOnce && GS.StoredOnceValue)
1677 cerr << " StoredOnceValue = " << *GS.StoredOnceValue << "\n";
1678 if (GS.AccessingFunction && !GS.HasMultipleAccessingFunctions)
1679 cerr << " AccessingFunction = " << GS.AccessingFunction->getName()
1681 cerr << " HasMultipleAccessingFunctions = "
1682 << GS.HasMultipleAccessingFunctions << "\n";
1683 cerr << " HasNonInstructionUser = " << GS.HasNonInstructionUser<<"\n";
1687 // If this is a first class global and has only one accessing function
1688 // and this function is main (which we know is not recursive we can make
1689 // this global a local variable) we replace the global with a local alloca
1690 // in this function.
1692 // NOTE: It doesn't make sense to promote non single-value types since we
1693 // are just replacing static memory to stack memory.
1695 // If the global is in different address space, don't bring it to stack.
1696 if (!GS.HasMultipleAccessingFunctions &&
1697 GS.AccessingFunction && !GS.HasNonInstructionUser &&
1698 GV->getType()->getElementType()->isSingleValueType() &&
1699 GS.AccessingFunction->getName() == "main" &&
1700 GS.AccessingFunction->hasExternalLinkage() &&
1701 GV->getType()->getAddressSpace() == 0) {
1702 DOUT << "LOCALIZING GLOBAL: " << *GV;
1703 Instruction* FirstI = GS.AccessingFunction->getEntryBlock().begin();
1704 const Type* ElemTy = GV->getType()->getElementType();
1705 // FIXME: Pass Global's alignment when globals have alignment
1706 AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), FirstI);
1707 if (!isa<UndefValue>(GV->getInitializer()))
1708 new StoreInst(GV->getInitializer(), Alloca, FirstI);
1710 GV->replaceAllUsesWith(Alloca);
1711 GV->eraseFromParent();
1716 // If the global is never loaded (but may be stored to), it is dead.
1719 DOUT << "GLOBAL NEVER LOADED: " << *GV;
1721 // Delete any stores we can find to the global. We may not be able to
1722 // make it completely dead though.
1723 bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer(),
1726 // If the global is dead now, delete it.
1727 if (GV->use_empty()) {
1728 GV->eraseFromParent();
1734 } else if (GS.StoredType <= GlobalStatus::isInitializerStored) {
1735 DOUT << "MARKING CONSTANT: " << *GV;
1736 GV->setConstant(true);
1738 // Clean up any obviously simplifiable users now.
1739 CleanupConstantGlobalUsers(GV, GV->getInitializer(), GV->getContext());
1741 // If the global is dead now, just nuke it.
1742 if (GV->use_empty()) {
1743 DOUT << " *** Marking constant allowed us to simplify "
1744 << "all users and delete global!\n";
1745 GV->eraseFromParent();
1751 } else if (!GV->getInitializer()->getType()->isSingleValueType()) {
1752 if (GlobalVariable *FirstNewGV = SRAGlobal(GV,
1753 getAnalysis<TargetData>(),
1754 GV->getContext())) {
1755 GVI = FirstNewGV; // Don't skip the newly produced globals!
1758 } else if (GS.StoredType == GlobalStatus::isStoredOnce) {
1759 // If the initial value for the global was an undef value, and if only
1760 // one other value was stored into it, we can just change the
1761 // initializer to be the stored value, then delete all stores to the
1762 // global. This allows us to mark it constant.
1763 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1764 if (isa<UndefValue>(GV->getInitializer())) {
1765 // Change the initial value here.
1766 GV->setInitializer(SOVConstant);
1768 // Clean up any obviously simplifiable users now.
1769 CleanupConstantGlobalUsers(GV, GV->getInitializer(),
1772 if (GV->use_empty()) {
1773 DOUT << " *** Substituting initializer allowed us to "
1774 << "simplify all users and delete global!\n";
1775 GV->eraseFromParent();
1784 // Try to optimize globals based on the knowledge that only one value
1785 // (besides its initializer) is ever stored to the global.
1786 if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI,
1787 getAnalysis<TargetData>(), GV->getContext()))
1790 // Otherwise, if the global was not a boolean, we can shrink it to be a
1792 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1793 if (TryToShrinkGlobalToBoolean(GV, SOVConstant, GV->getContext())) {
1802 /// ChangeCalleesToFastCall - Walk all of the direct calls of the specified
1803 /// function, changing them to FastCC.
1804 static void ChangeCalleesToFastCall(Function *F) {
1805 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1806 CallSite User(cast<Instruction>(*UI));
1807 User.setCallingConv(CallingConv::Fast);
1811 static AttrListPtr StripNest(const AttrListPtr &Attrs) {
1812 for (unsigned i = 0, e = Attrs.getNumSlots(); i != e; ++i) {
1813 if ((Attrs.getSlot(i).Attrs & Attribute::Nest) == 0)
1816 // There can be only one.
1817 return Attrs.removeAttr(Attrs.getSlot(i).Index, Attribute::Nest);
1823 static void RemoveNestAttribute(Function *F) {
1824 F->setAttributes(StripNest(F->getAttributes()));
1825 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1826 CallSite User(cast<Instruction>(*UI));
1827 User.setAttributes(StripNest(User.getAttributes()));
1831 bool GlobalOpt::OptimizeFunctions(Module &M) {
1832 bool Changed = false;
1833 // Optimize functions.
1834 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) {
1836 // Functions without names cannot be referenced outside this module.
1837 if (!F->hasName() && !F->isDeclaration())
1838 F->setLinkage(GlobalValue::InternalLinkage);
1839 F->removeDeadConstantUsers();
1840 if (F->use_empty() && (F->hasLocalLinkage() ||
1841 F->hasLinkOnceLinkage())) {
1842 M.getFunctionList().erase(F);
1845 } else if (F->hasLocalLinkage()) {
1846 if (F->getCallingConv() == CallingConv::C && !F->isVarArg() &&
1847 !F->hasAddressTaken()) {
1848 // If this function has C calling conventions, is not a varargs
1849 // function, and is only called directly, promote it to use the Fast
1850 // calling convention.
1851 F->setCallingConv(CallingConv::Fast);
1852 ChangeCalleesToFastCall(F);
1857 if (F->getAttributes().hasAttrSomewhere(Attribute::Nest) &&
1858 !F->hasAddressTaken()) {
1859 // The function is not used by a trampoline intrinsic, so it is safe
1860 // to remove the 'nest' attribute.
1861 RemoveNestAttribute(F);
1870 bool GlobalOpt::OptimizeGlobalVars(Module &M) {
1871 bool Changed = false;
1872 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
1874 GlobalVariable *GV = GVI++;
1875 // Global variables without names cannot be referenced outside this module.
1876 if (!GV->hasName() && !GV->isDeclaration())
1877 GV->setLinkage(GlobalValue::InternalLinkage);
1878 if (!GV->isConstant() && GV->hasLocalLinkage() &&
1879 GV->hasInitializer())
1880 Changed |= ProcessInternalGlobal(GV, GVI);
1885 /// FindGlobalCtors - Find the llvm.globalctors list, verifying that all
1886 /// initializers have an init priority of 65535.
1887 GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) {
1888 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
1890 if (I->getName() == "llvm.global_ctors") {
1891 // Found it, verify it's an array of { int, void()* }.
1892 const ArrayType *ATy =dyn_cast<ArrayType>(I->getType()->getElementType());
1894 const StructType *STy = dyn_cast<StructType>(ATy->getElementType());
1895 if (!STy || STy->getNumElements() != 2 ||
1896 STy->getElementType(0) != Type::Int32Ty) return 0;
1897 const PointerType *PFTy = dyn_cast<PointerType>(STy->getElementType(1));
1898 if (!PFTy) return 0;
1899 const FunctionType *FTy = dyn_cast<FunctionType>(PFTy->getElementType());
1900 if (!FTy || FTy->getReturnType() != Type::VoidTy || FTy->isVarArg() ||
1901 FTy->getNumParams() != 0)
1904 // Verify that the initializer is simple enough for us to handle.
1905 if (!I->hasInitializer()) return 0;
1906 ConstantArray *CA = dyn_cast<ConstantArray>(I->getInitializer());
1908 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i)
1909 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(*i)) {
1910 if (isa<ConstantPointerNull>(CS->getOperand(1)))
1913 // Must have a function or null ptr.
1914 if (!isa<Function>(CS->getOperand(1)))
1917 // Init priority must be standard.
1918 ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0));
1919 if (!CI || CI->getZExtValue() != 65535)
1930 /// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand,
1931 /// return a list of the functions and null terminator as a vector.
1932 static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) {
1933 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer());
1934 std::vector<Function*> Result;
1935 Result.reserve(CA->getNumOperands());
1936 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) {
1937 ConstantStruct *CS = cast<ConstantStruct>(*i);
1938 Result.push_back(dyn_cast<Function>(CS->getOperand(1)));
1943 /// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the
1944 /// specified array, returning the new global to use.
1945 static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL,
1946 const std::vector<Function*> &Ctors,
1947 LLVMContext &Context) {
1948 // If we made a change, reassemble the initializer list.
1949 std::vector<Constant*> CSVals;
1950 CSVals.push_back(ConstantInt::get(Type::Int32Ty, 65535));
1951 CSVals.push_back(0);
1953 // Create the new init list.
1954 std::vector<Constant*> CAList;
1955 for (unsigned i = 0, e = Ctors.size(); i != e; ++i) {
1957 CSVals[1] = Ctors[i];
1959 const Type *FTy = FunctionType::get(Type::VoidTy, false);
1960 const PointerType *PFTy = PointerType::getUnqual(FTy);
1961 CSVals[1] = Constant::getNullValue(PFTy);
1962 CSVals[0] = ConstantInt::get(Type::Int32Ty, 2147483647);
1964 CAList.push_back(ConstantStruct::get(Context, CSVals));
1967 // Create the array initializer.
1968 const Type *StructTy =
1969 cast<ArrayType>(GCL->getType()->getElementType())->getElementType();
1970 Constant *CA = ConstantArray::get(ArrayType::get(StructTy,
1971 CAList.size()), CAList);
1973 // If we didn't change the number of elements, don't create a new GV.
1974 if (CA->getType() == GCL->getInitializer()->getType()) {
1975 GCL->setInitializer(CA);
1979 // Create the new global and insert it next to the existing list.
1980 GlobalVariable *NGV = new GlobalVariable(Context, CA->getType(),
1982 GCL->getLinkage(), CA, "",
1983 GCL->isThreadLocal());
1984 GCL->getParent()->getGlobalList().insert(GCL, NGV);
1987 // Nuke the old list, replacing any uses with the new one.
1988 if (!GCL->use_empty()) {
1990 if (V->getType() != GCL->getType())
1991 V = ConstantExpr::getBitCast(V, GCL->getType());
1992 GCL->replaceAllUsesWith(V);
1994 GCL->eraseFromParent();
2003 static Constant *getVal(DenseMap<Value*, Constant*> &ComputedValues,
2005 if (Constant *CV = dyn_cast<Constant>(V)) return CV;
2006 Constant *R = ComputedValues[V];
2007 assert(R && "Reference to an uncomputed value!");
2011 /// isSimpleEnoughPointerToCommit - Return true if this constant is simple
2012 /// enough for us to understand. In particular, if it is a cast of something,
2013 /// we punt. We basically just support direct accesses to globals and GEP's of
2014 /// globals. This should be kept up to date with CommitValueTo.
2015 static bool isSimpleEnoughPointerToCommit(Constant *C, LLVMContext &Context) {
2016 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) {
2017 if (!GV->hasExternalLinkage() && !GV->hasLocalLinkage())
2018 return false; // do not allow weak/linkonce/dllimport/dllexport linkage.
2019 return !GV->isDeclaration(); // reject external globals.
2021 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2022 // Handle a constantexpr gep.
2023 if (CE->getOpcode() == Instruction::GetElementPtr &&
2024 isa<GlobalVariable>(CE->getOperand(0))) {
2025 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2026 if (!GV->hasExternalLinkage() && !GV->hasLocalLinkage())
2027 return false; // do not allow weak/linkonce/dllimport/dllexport linkage.
2028 return GV->hasInitializer() &&
2029 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE,
2035 /// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global
2036 /// initializer. This returns 'Init' modified to reflect 'Val' stored into it.
2037 /// At this point, the GEP operands of Addr [0, OpNo) have been stepped into.
2038 static Constant *EvaluateStoreInto(Constant *Init, Constant *Val,
2039 ConstantExpr *Addr, unsigned OpNo,
2040 LLVMContext &Context) {
2041 // Base case of the recursion.
2042 if (OpNo == Addr->getNumOperands()) {
2043 assert(Val->getType() == Init->getType() && "Type mismatch!");
2047 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
2048 std::vector<Constant*> Elts;
2050 // Break up the constant into its elements.
2051 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
2052 for (User::op_iterator i = CS->op_begin(), e = CS->op_end(); i != e; ++i)
2053 Elts.push_back(cast<Constant>(*i));
2054 } else if (isa<ConstantAggregateZero>(Init)) {
2055 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
2056 Elts.push_back(Constant::getNullValue(STy->getElementType(i)));
2057 } else if (isa<UndefValue>(Init)) {
2058 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
2059 Elts.push_back(UndefValue::get(STy->getElementType(i)));
2061 llvm_unreachable("This code is out of sync with "
2062 " ConstantFoldLoadThroughGEPConstantExpr");
2065 // Replace the element that we are supposed to.
2066 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo));
2067 unsigned Idx = CU->getZExtValue();
2068 assert(Idx < STy->getNumElements() && "Struct index out of range!");
2069 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1, Context);
2071 // Return the modified struct.
2072 return ConstantStruct::get(Context, &Elts[0], Elts.size(), STy->isPacked());
2074 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo));
2075 const ArrayType *ATy = cast<ArrayType>(Init->getType());
2077 // Break up the array into elements.
2078 std::vector<Constant*> Elts;
2079 if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
2080 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i)
2081 Elts.push_back(cast<Constant>(*i));
2082 } else if (isa<ConstantAggregateZero>(Init)) {
2083 Constant *Elt = Constant::getNullValue(ATy->getElementType());
2084 Elts.assign(ATy->getNumElements(), Elt);
2085 } else if (isa<UndefValue>(Init)) {
2086 Constant *Elt = UndefValue::get(ATy->getElementType());
2087 Elts.assign(ATy->getNumElements(), Elt);
2089 llvm_unreachable("This code is out of sync with "
2090 " ConstantFoldLoadThroughGEPConstantExpr");
2093 assert(CI->getZExtValue() < ATy->getNumElements());
2094 Elts[CI->getZExtValue()] =
2095 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1, Context);
2096 return ConstantArray::get(ATy, Elts);
2100 /// CommitValueTo - We have decided that Addr (which satisfies the predicate
2101 /// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen.
2102 static void CommitValueTo(Constant *Val, Constant *Addr,
2103 LLVMContext &Context) {
2104 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
2105 assert(GV->hasInitializer());
2106 GV->setInitializer(Val);
2110 ConstantExpr *CE = cast<ConstantExpr>(Addr);
2111 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2113 Constant *Init = GV->getInitializer();
2114 Init = EvaluateStoreInto(Init, Val, CE, 2, Context);
2115 GV->setInitializer(Init);
2118 /// ComputeLoadResult - Return the value that would be computed by a load from
2119 /// P after the stores reflected by 'memory' have been performed. If we can't
2120 /// decide, return null.
2121 static Constant *ComputeLoadResult(Constant *P,
2122 const DenseMap<Constant*, Constant*> &Memory,
2123 LLVMContext &Context) {
2124 // If this memory location has been recently stored, use the stored value: it
2125 // is the most up-to-date.
2126 DenseMap<Constant*, Constant*>::const_iterator I = Memory.find(P);
2127 if (I != Memory.end()) return I->second;
2130 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
2131 if (GV->hasInitializer())
2132 return GV->getInitializer();
2136 // Handle a constantexpr getelementptr.
2137 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P))
2138 if (CE->getOpcode() == Instruction::GetElementPtr &&
2139 isa<GlobalVariable>(CE->getOperand(0))) {
2140 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2141 if (GV->hasInitializer())
2142 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE,
2146 return 0; // don't know how to evaluate.
2149 /// EvaluateFunction - Evaluate a call to function F, returning true if
2150 /// successful, false if we can't evaluate it. ActualArgs contains the formal
2151 /// arguments for the function.
2152 static bool EvaluateFunction(Function *F, Constant *&RetVal,
2153 const std::vector<Constant*> &ActualArgs,
2154 std::vector<Function*> &CallStack,
2155 DenseMap<Constant*, Constant*> &MutatedMemory,
2156 std::vector<GlobalVariable*> &AllocaTmps) {
2157 // Check to see if this function is already executing (recursion). If so,
2158 // bail out. TODO: we might want to accept limited recursion.
2159 if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end())
2162 LLVMContext &Context = F->getContext();
2164 CallStack.push_back(F);
2166 /// Values - As we compute SSA register values, we store their contents here.
2167 DenseMap<Value*, Constant*> Values;
2169 // Initialize arguments to the incoming values specified.
2171 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E;
2173 Values[AI] = ActualArgs[ArgNo];
2175 /// ExecutedBlocks - We only handle non-looping, non-recursive code. As such,
2176 /// we can only evaluate any one basic block at most once. This set keeps
2177 /// track of what we have executed so we can detect recursive cases etc.
2178 SmallPtrSet<BasicBlock*, 32> ExecutedBlocks;
2180 // CurInst - The current instruction we're evaluating.
2181 BasicBlock::iterator CurInst = F->begin()->begin();
2183 // This is the main evaluation loop.
2185 Constant *InstResult = 0;
2187 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) {
2188 if (SI->isVolatile()) return false; // no volatile accesses.
2189 Constant *Ptr = getVal(Values, SI->getOperand(1));
2190 if (!isSimpleEnoughPointerToCommit(Ptr, Context))
2191 // If this is too complex for us to commit, reject it.
2193 Constant *Val = getVal(Values, SI->getOperand(0));
2194 MutatedMemory[Ptr] = Val;
2195 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) {
2196 InstResult = ConstantExpr::get(BO->getOpcode(),
2197 getVal(Values, BO->getOperand(0)),
2198 getVal(Values, BO->getOperand(1)));
2199 } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) {
2200 InstResult = ConstantExpr::getCompare(CI->getPredicate(),
2201 getVal(Values, CI->getOperand(0)),
2202 getVal(Values, CI->getOperand(1)));
2203 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) {
2204 InstResult = ConstantExpr::getCast(CI->getOpcode(),
2205 getVal(Values, CI->getOperand(0)),
2207 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) {
2209 ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)),
2210 getVal(Values, SI->getOperand(1)),
2211 getVal(Values, SI->getOperand(2)));
2212 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) {
2213 Constant *P = getVal(Values, GEP->getOperand(0));
2214 SmallVector<Constant*, 8> GEPOps;
2215 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end();
2217 GEPOps.push_back(getVal(Values, *i));
2219 ConstantExpr::getGetElementPtr(P, &GEPOps[0], GEPOps.size());
2220 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) {
2221 if (LI->isVolatile()) return false; // no volatile accesses.
2222 InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)),
2223 MutatedMemory, Context);
2224 if (InstResult == 0) return false; // Could not evaluate load.
2225 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) {
2226 if (AI->isArrayAllocation()) return false; // Cannot handle array allocs.
2227 const Type *Ty = AI->getType()->getElementType();
2228 AllocaTmps.push_back(new GlobalVariable(Context, Ty, false,
2229 GlobalValue::InternalLinkage,
2230 UndefValue::get(Ty),
2232 InstResult = AllocaTmps.back();
2233 } else if (CallInst *CI = dyn_cast<CallInst>(CurInst)) {
2235 // Debug info can safely be ignored here.
2236 if (isa<DbgInfoIntrinsic>(CI)) {
2241 // Cannot handle inline asm.
2242 if (isa<InlineAsm>(CI->getOperand(0))) return false;
2244 // Resolve function pointers.
2245 Function *Callee = dyn_cast<Function>(getVal(Values, CI->getOperand(0)));
2246 if (!Callee) return false; // Cannot resolve.
2248 std::vector<Constant*> Formals;
2249 for (User::op_iterator i = CI->op_begin() + 1, e = CI->op_end();
2251 Formals.push_back(getVal(Values, *i));
2253 if (Callee->isDeclaration()) {
2254 // If this is a function we can constant fold, do it.
2255 if (Constant *C = ConstantFoldCall(Callee, &Formals[0],
2262 if (Callee->getFunctionType()->isVarArg())
2266 // Execute the call, if successful, use the return value.
2267 if (!EvaluateFunction(Callee, RetVal, Formals, CallStack,
2268 MutatedMemory, AllocaTmps))
2270 InstResult = RetVal;
2272 } else if (isa<TerminatorInst>(CurInst)) {
2273 BasicBlock *NewBB = 0;
2274 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) {
2275 if (BI->isUnconditional()) {
2276 NewBB = BI->getSuccessor(0);
2279 dyn_cast<ConstantInt>(getVal(Values, BI->getCondition()));
2280 if (!Cond) return false; // Cannot determine.
2282 NewBB = BI->getSuccessor(!Cond->getZExtValue());
2284 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) {
2286 dyn_cast<ConstantInt>(getVal(Values, SI->getCondition()));
2287 if (!Val) return false; // Cannot determine.
2288 NewBB = SI->getSuccessor(SI->findCaseValue(Val));
2289 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) {
2290 if (RI->getNumOperands())
2291 RetVal = getVal(Values, RI->getOperand(0));
2293 CallStack.pop_back(); // return from fn.
2294 return true; // We succeeded at evaluating this ctor!
2296 // invoke, unwind, unreachable.
2297 return false; // Cannot handle this terminator.
2300 // Okay, we succeeded in evaluating this control flow. See if we have
2301 // executed the new block before. If so, we have a looping function,
2302 // which we cannot evaluate in reasonable time.
2303 if (!ExecutedBlocks.insert(NewBB))
2304 return false; // looped!
2306 // Okay, we have never been in this block before. Check to see if there
2307 // are any PHI nodes. If so, evaluate them with information about where
2309 BasicBlock *OldBB = CurInst->getParent();
2310 CurInst = NewBB->begin();
2312 for (; (PN = dyn_cast<PHINode>(CurInst)); ++CurInst)
2313 Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB));
2315 // Do NOT increment CurInst. We know that the terminator had no value.
2318 // Did not know how to evaluate this!
2322 if (!CurInst->use_empty())
2323 Values[CurInst] = InstResult;
2325 // Advance program counter.
2330 /// EvaluateStaticConstructor - Evaluate static constructors in the function, if
2331 /// we can. Return true if we can, false otherwise.
2332 static bool EvaluateStaticConstructor(Function *F) {
2333 /// MutatedMemory - For each store we execute, we update this map. Loads
2334 /// check this to get the most up-to-date value. If evaluation is successful,
2335 /// this state is committed to the process.
2336 DenseMap<Constant*, Constant*> MutatedMemory;
2338 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable
2339 /// to represent its body. This vector is needed so we can delete the
2340 /// temporary globals when we are done.
2341 std::vector<GlobalVariable*> AllocaTmps;
2343 /// CallStack - This is used to detect recursion. In pathological situations
2344 /// we could hit exponential behavior, but at least there is nothing
2346 std::vector<Function*> CallStack;
2348 // Call the function.
2349 Constant *RetValDummy;
2350 bool EvalSuccess = EvaluateFunction(F, RetValDummy, std::vector<Constant*>(),
2351 CallStack, MutatedMemory, AllocaTmps);
2353 // We succeeded at evaluation: commit the result.
2354 DEBUG(errs() << "FULLY EVALUATED GLOBAL CTOR FUNCTION '"
2355 << F->getName() << "' to " << MutatedMemory.size()
2357 for (DenseMap<Constant*, Constant*>::iterator I = MutatedMemory.begin(),
2358 E = MutatedMemory.end(); I != E; ++I)
2359 CommitValueTo(I->second, I->first, F->getContext());
2362 // At this point, we are done interpreting. If we created any 'alloca'
2363 // temporaries, release them now.
2364 while (!AllocaTmps.empty()) {
2365 GlobalVariable *Tmp = AllocaTmps.back();
2366 AllocaTmps.pop_back();
2368 // If there are still users of the alloca, the program is doing something
2369 // silly, e.g. storing the address of the alloca somewhere and using it
2370 // later. Since this is undefined, we'll just make it be null.
2371 if (!Tmp->use_empty())
2372 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType()));
2381 /// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible.
2382 /// Return true if anything changed.
2383 bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) {
2384 std::vector<Function*> Ctors = ParseGlobalCtors(GCL);
2385 bool MadeChange = false;
2386 if (Ctors.empty()) return false;
2388 // Loop over global ctors, optimizing them when we can.
2389 for (unsigned i = 0; i != Ctors.size(); ++i) {
2390 Function *F = Ctors[i];
2391 // Found a null terminator in the middle of the list, prune off the rest of
2394 if (i != Ctors.size()-1) {
2401 // We cannot simplify external ctor functions.
2402 if (F->empty()) continue;
2404 // If we can evaluate the ctor at compile time, do.
2405 if (EvaluateStaticConstructor(F)) {
2406 Ctors.erase(Ctors.begin()+i);
2409 ++NumCtorsEvaluated;
2414 if (!MadeChange) return false;
2416 GCL = InstallGlobalCtors(GCL, Ctors, GCL->getContext());
2420 bool GlobalOpt::OptimizeGlobalAliases(Module &M) {
2421 bool Changed = false;
2423 for (Module::alias_iterator I = M.alias_begin(), E = M.alias_end();
2425 Module::alias_iterator J = I++;
2426 // Aliases without names cannot be referenced outside this module.
2427 if (!J->hasName() && !J->isDeclaration())
2428 J->setLinkage(GlobalValue::InternalLinkage);
2429 // If the aliasee may change at link time, nothing can be done - bail out.
2430 if (J->mayBeOverridden())
2433 Constant *Aliasee = J->getAliasee();
2434 GlobalValue *Target = cast<GlobalValue>(Aliasee->stripPointerCasts());
2435 Target->removeDeadConstantUsers();
2436 bool hasOneUse = Target->hasOneUse() && Aliasee->hasOneUse();
2438 // Make all users of the alias use the aliasee instead.
2439 if (!J->use_empty()) {
2440 J->replaceAllUsesWith(Aliasee);
2441 ++NumAliasesResolved;
2445 // If the aliasee has internal linkage, give it the name and linkage
2446 // of the alias, and delete the alias. This turns:
2447 // define internal ... @f(...)
2448 // @a = alias ... @f
2450 // define ... @a(...)
2451 if (!Target->hasLocalLinkage())
2454 // The transform is only useful if the alias does not have internal linkage.
2455 if (J->hasLocalLinkage())
2458 // Do not perform the transform if multiple aliases potentially target the
2459 // aliasee. This check also ensures that it is safe to replace the section
2460 // and other attributes of the aliasee with those of the alias.
2464 // Give the aliasee the name, linkage and other attributes of the alias.
2465 Target->takeName(J);
2466 Target->setLinkage(J->getLinkage());
2467 Target->GlobalValue::copyAttributesFrom(J);
2469 // Delete the alias.
2470 M.getAliasList().erase(J);
2471 ++NumAliasesRemoved;
2478 bool GlobalOpt::runOnModule(Module &M) {
2479 bool Changed = false;
2481 // Try to find the llvm.globalctors list.
2482 GlobalVariable *GlobalCtors = FindGlobalCtors(M);
2484 bool LocalChange = true;
2485 while (LocalChange) {
2486 LocalChange = false;
2488 // Delete functions that are trivially dead, ccc -> fastcc
2489 LocalChange |= OptimizeFunctions(M);
2491 // Optimize global_ctors list.
2493 LocalChange |= OptimizeGlobalCtorsList(GlobalCtors);
2495 // Optimize non-address-taken globals.
2496 LocalChange |= OptimizeGlobalVars(M);
2498 // Resolve aliases, when possible.
2499 LocalChange |= OptimizeGlobalAliases(M);
2500 Changed |= LocalChange;
2503 // TODO: Move all global ctors functions to the end of the module for code